annotate x86/dsputil_mmx.c @ 11118:001eb7e3e2d3 libavcodec

Enable SSE2 (put|avg)_pixels_16_sse2 SVQ1 chroma has been special-cased aligned to 16-bytes since at least r15466 Other architectures also assume 16-byte alignment here too but set STRIDE_ALIGN to 16.
author conrad
date Wed, 10 Feb 2010 02:02:06 +0000
parents abb3b23bda35
children 98970e51365a
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
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
25 #include "libavutil/x86_cpu.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
26 #include "libavcodec/dsputil.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
27 #include "libavcodec/mpegvideo.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
28 #include "libavcodec/simple_idct.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
29 #include "dsputil_mmx.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
30 #include "vp3dsp_mmx.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
31 #include "vp3dsp_sse2.h"
8817
492f8911992c add MMX version of vp6_filter_diag
aurel
parents: 8816
diff changeset
32 #include "vp6dsp_mmx.h"
8818
17cc6df384a6 add SSE2 version of vp6_filter_diag
aurel
parents: 8817
diff changeset
33 #include "vp6dsp_sse2.h"
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
34 #include "idct_xvid.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
35
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
36 //#undef NDEBUG
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
37 //#include <assert.h>
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
38
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
39 int mm_flags; /* multimedia extension flags */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
40
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
41 /* pixel operations */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
42 DECLARE_ALIGNED_8 (const uint64_t, ff_bone) = 0x0101010101010101ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
43 DECLARE_ALIGNED_8 (const uint64_t, ff_wtwo) = 0x0002000200020002ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
44
10961
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10766
diff changeset
45 DECLARE_ALIGNED_16(const uint64_t, ff_pdw_80000000)[2] =
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
46 {0x8000000080000000ULL, 0x8000000080000000ULL};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
47
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
48 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_3 ) = 0x0003000300030003ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
49 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_4 ) = 0x0004000400040004ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
50 DECLARE_ALIGNED_16(const xmm_reg, ff_pw_5 ) = {0x0005000500050005ULL, 0x0005000500050005ULL};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
51 DECLARE_ALIGNED_16(const xmm_reg, ff_pw_8 ) = {0x0008000800080008ULL, 0x0008000800080008ULL};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
52 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_15 ) = 0x000F000F000F000FULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
53 DECLARE_ALIGNED_16(const xmm_reg, ff_pw_16 ) = {0x0010001000100010ULL, 0x0010001000100010ULL};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
54 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_20 ) = 0x0014001400140014ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
55 DECLARE_ALIGNED_16(const xmm_reg, ff_pw_28 ) = {0x001C001C001C001CULL, 0x001C001C001C001CULL};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
56 DECLARE_ALIGNED_16(const xmm_reg, ff_pw_32 ) = {0x0020002000200020ULL, 0x0020002000200020ULL};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
57 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_42 ) = 0x002A002A002A002AULL;
8816
53f9f3994ec8 convert ff_pw_64 into an xmm_reg for future use in vp6 sse code
aurel
parents: 8798
diff changeset
58 DECLARE_ALIGNED_16(const xmm_reg, ff_pw_64 ) = {0x0040004000400040ULL, 0x0040004000400040ULL};
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
59 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_96 ) = 0x0060006000600060ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
60 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_128) = 0x0080008000800080ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
61 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_255) = 0x00ff00ff00ff00ffULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
62
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
63 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_1 ) = 0x0101010101010101ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
64 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_3 ) = 0x0303030303030303ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
65 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_7 ) = 0x0707070707070707ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
66 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_1F ) = 0x1F1F1F1F1F1F1F1FULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
67 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_3F ) = 0x3F3F3F3F3F3F3F3FULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
68 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_81 ) = 0x8181818181818181ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
69 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_A1 ) = 0xA1A1A1A1A1A1A1A1ULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
70 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_FC ) = 0xFCFCFCFCFCFCFCFCULL;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
71
10961
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10766
diff changeset
72 DECLARE_ALIGNED_16(const double, ff_pd_1)[2] = { 1.0, 1.0 };
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10766
diff changeset
73 DECLARE_ALIGNED_16(const double, ff_pd_2)[2] = { 2.0, 2.0 };
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
74
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
75 #define JUMPALIGN() __asm__ volatile (ASMALIGN(3)::)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
76 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%" #regd ", %%" #regd ::)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
77
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
78 #define MOVQ_BFE(regd) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
79 __asm__ volatile ( \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
80 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
81 "paddb %%" #regd ", %%" #regd " \n\t" ::)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
82
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
83 #ifndef PIC
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
84 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_bone))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
85 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_wtwo))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
86 #else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
87 // for shared library it's better to use this way for accessing constants
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
88 // pcmpeqd -> -1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
89 #define MOVQ_BONE(regd) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
90 __asm__ volatile ( \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
91 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
92 "psrlw $15, %%" #regd " \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
93 "packuswb %%" #regd ", %%" #regd " \n\t" ::)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
94
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
95 #define MOVQ_WTWO(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 "psllw $1, %%" #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 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
102
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
103 // using regr as temporary and for the output result
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
104 // first argument is unmodifed and second is trashed
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
105 // regfe is supposed to contain 0xfefefefefefefefe
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
106 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
107 "movq " #rega ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
108 "pand " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
109 "pxor " #rega ", " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
110 "pand " #regfe "," #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
111 "psrlq $1, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
112 "paddb " #regb ", " #regr " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
113
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
114 #define PAVGB_MMX(rega, regb, regr, regfe) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
115 "movq " #rega ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
116 "por " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
117 "pxor " #rega ", " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
118 "pand " #regfe "," #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
119 "psrlq $1, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
120 "psubb " #regb ", " #regr " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
121
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
122 // mm6 is supposed to contain 0xfefefefefefefefe
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
123 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
124 "movq " #rega ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
125 "movq " #regc ", " #regp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
126 "pand " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
127 "pand " #regd ", " #regp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
128 "pxor " #rega ", " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
129 "pxor " #regc ", " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
130 "pand %%mm6, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
131 "pand %%mm6, " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
132 "psrlq $1, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
133 "psrlq $1, " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
134 "paddb " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
135 "paddb " #regd ", " #regp " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
136
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
137 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
138 "movq " #rega ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
139 "movq " #regc ", " #regp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
140 "por " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
141 "por " #regd ", " #regp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
142 "pxor " #rega ", " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
143 "pxor " #regc ", " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
144 "pand %%mm6, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
145 "pand %%mm6, " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
146 "psrlq $1, " #regd " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
147 "psrlq $1, " #regb " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
148 "psubb " #regb ", " #regr " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
149 "psubb " #regd ", " #regp " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
150
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
151 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
152 /* MMX no rounding */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
153 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
154 #define SET_RND MOVQ_WONE
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
155 #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
156 #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
157 #define OP_AVG(a, b, c, e) PAVGB_MMX(a, b, c, e)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
158
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
159 #include "dsputil_mmx_rnd_template.c"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
160
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
161 #undef DEF
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
162 #undef SET_RND
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
163 #undef PAVGBP
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
164 #undef PAVGB
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
165 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
166 /* MMX rounding */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
167
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
168 #define DEF(x, y) x ## _ ## y ##_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
169 #define SET_RND MOVQ_WTWO
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
170 #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
171 #define PAVGB(a, b, c, e) PAVGB_MMX(a, b, c, e)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
172
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
173 #include "dsputil_mmx_rnd_template.c"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
174
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
175 #undef DEF
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
176 #undef SET_RND
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
177 #undef PAVGBP
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
178 #undef PAVGB
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
179 #undef OP_AVG
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
180
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
181 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
182 /* 3Dnow specific */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
183
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
184 #define DEF(x) x ## _3dnow
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
185 #define PAVGB "pavgusb"
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
186 #define OP_AVG PAVGB
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
187
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
188 #include "dsputil_mmx_avg_template.c"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
189
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
190 #undef DEF
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
191 #undef PAVGB
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
192 #undef OP_AVG
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
193
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
194 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
195 /* MMX2 specific */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
196
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
197 #define DEF(x) x ## _mmx2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
198
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
199 /* Introduced only in MMX2 set */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
200 #define PAVGB "pavgb"
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
201 #define OP_AVG PAVGB
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
202
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
203 #include "dsputil_mmx_avg_template.c"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
204
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
205 #undef DEF
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
206 #undef PAVGB
9445
41245484dc0b avg_ pixel functions need to use (dst+pix+1)>>1 to average with existing
conrad
parents: 9441
diff changeset
207 #undef OP_AVG
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
208
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
209 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
210 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
211 #define put_pixels16_mmx2 put_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
212 #define put_pixels8_mmx2 put_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
213 #define put_pixels4_mmx2 put_pixels4_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
214 #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
215 #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
216 #define put_pixels16_3dnow put_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
217 #define put_pixels8_3dnow put_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
218 #define put_pixels4_3dnow put_pixels4_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
219 #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
220 #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
221
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
222 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
223 /* standard MMX */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
224
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
225 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
226 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
227 const DCTELEM *p;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
228 uint8_t *pix;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
229
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
230 /* read the pixels */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
231 p = block;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
232 pix = pixels;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
233 /* unrolled loop */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
234 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
235 "movq %3, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
236 "movq 8%3, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
237 "movq 16%3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
238 "movq 24%3, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
239 "movq 32%3, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
240 "movq 40%3, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
241 "movq 48%3, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
242 "movq 56%3, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
243 "packuswb %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
244 "packuswb %%mm3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
245 "packuswb %%mm5, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
246 "packuswb %%mm7, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
247 "movq %%mm0, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
248 "movq %%mm2, (%0, %1) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
249 "movq %%mm4, (%0, %1, 2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
250 "movq %%mm6, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
251 ::"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
252 :"memory");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
253 pix += line_size*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
254 p += 32;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
255
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
256 // if here would be an exact copy of the code above
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
257 // compiler would generate some very strange code
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
258 // thus using "r"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
259 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
260 "movq (%3), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
261 "movq 8(%3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
262 "movq 16(%3), %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
263 "movq 24(%3), %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
264 "movq 32(%3), %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
265 "movq 40(%3), %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
266 "movq 48(%3), %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
267 "movq 56(%3), %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
268 "packuswb %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
269 "packuswb %%mm3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
270 "packuswb %%mm5, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
271 "packuswb %%mm7, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
272 "movq %%mm0, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
273 "movq %%mm2, (%0, %1) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
274 "movq %%mm4, (%0, %1, 2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
275 "movq %%mm6, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
276 ::"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
277 :"memory");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
278 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
279
10961
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10766
diff changeset
280 DECLARE_ASM_CONST(8, uint8_t, ff_vector128)[8] =
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
281 { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 };
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
282
9337
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
283 #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
284 "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
285 "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
286 "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
287 "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
288 "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
289 "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
290 "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
291 "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
292 "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
293 "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
294 "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
295 "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
296 "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
297 "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
298 "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
299 "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
300
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
301 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
302 {
9337
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
303 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
304 x86_reg line_skip3;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
305
9337
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
306 __asm__ volatile (
a0d54042ea37 Rewrite put_signed_pixels_clamped_mmx() to eliminate mmx.h from dsputil_mmx.c.
alexc
parents: 8818
diff changeset
307 "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
308 "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
309 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
310 "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
311 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
312 :"+&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
313 :"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
314 :"memory");
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
315 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
316
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
317 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
318 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
319 const DCTELEM *p;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
320 uint8_t *pix;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
321 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
322
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
323 /* read the pixels */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
324 p = block;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
325 pix = pixels;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
326 MOVQ_ZERO(mm7);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
327 i = 4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
328 do {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
329 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
330 "movq (%2), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
331 "movq 8(%2), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
332 "movq 16(%2), %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
333 "movq 24(%2), %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
334 "movq %0, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
335 "movq %1, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
336 "movq %%mm4, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
337 "punpcklbw %%mm7, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
338 "punpckhbw %%mm7, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
339 "paddsw %%mm4, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
340 "paddsw %%mm5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
341 "movq %%mm6, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
342 "punpcklbw %%mm7, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
343 "punpckhbw %%mm7, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
344 "paddsw %%mm6, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
345 "paddsw %%mm5, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
346 "packuswb %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
347 "packuswb %%mm3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
348 "movq %%mm0, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
349 "movq %%mm2, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
350 :"+m"(*pix), "+m"(*(pix+line_size))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
351 :"r"(p)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
352 :"memory");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
353 pix += line_size*2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
354 p += 16;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
355 } while (--i);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
356 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
357
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
358 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
359 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
360 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
361 "lea (%3, %3), %%"REG_a" \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
362 ASMALIGN(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
363 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
364 "movd (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
365 "movd (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
366 "movd %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
367 "movd %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
368 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
369 "add %%"REG_a", %2 \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 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
377 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
378 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
379 : "r"((x86_reg)line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
380 : "%"REG_a, "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
381 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
382 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
383
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
384 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
385 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
386 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
387 "lea (%3, %3), %%"REG_a" \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
388 ASMALIGN(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
389 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
390 "movq (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
391 "movq (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
392 "movq %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
393 "movq %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
394 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
395 "add %%"REG_a", %2 \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 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
403 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
404 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
405 : "r"((x86_reg)line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
406 : "%"REG_a, "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
407 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
408 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
409
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
410 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
411 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
412 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
413 "lea (%3, %3), %%"REG_a" \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
414 ASMALIGN(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
415 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
416 "movq (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
417 "movq 8(%1), %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
418 "movq (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
419 "movq 8(%1, %3), %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
420 "movq %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
421 "movq %%mm4, 8(%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
422 "movq %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
423 "movq %%mm5, 8(%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
424 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
425 "add %%"REG_a", %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
426 "movq (%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
427 "movq 8(%1), %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
428 "movq (%1, %3), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
429 "movq 8(%1, %3), %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
430 "movq %%mm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
431 "movq %%mm4, 8(%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
432 "movq %%mm1, (%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
433 "movq %%mm5, 8(%2, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
434 "add %%"REG_a", %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
435 "add %%"REG_a", %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
436 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
437 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
438 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
439 : "r"((x86_reg)line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
440 : "%"REG_a, "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
441 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
442 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
443
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
444 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
445 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
446 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
447 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
448 "movdqu (%1), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
449 "movdqu (%1,%3), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
450 "movdqu (%1,%3,2), %%xmm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
451 "movdqu (%1,%4), %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
452 "movdqa %%xmm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
453 "movdqa %%xmm1, (%2,%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
454 "movdqa %%xmm2, (%2,%3,2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
455 "movdqa %%xmm3, (%2,%4) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
456 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
457 "lea (%1,%3,4), %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
458 "lea (%2,%3,4), %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
459 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
460 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
461 : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
462 : "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
463 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
464 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
465
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
466 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
467 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
468 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
469 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
470 "movdqu (%1), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
471 "movdqu (%1,%3), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
472 "movdqu (%1,%3,2), %%xmm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
473 "movdqu (%1,%4), %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
474 "pavgb (%2), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
475 "pavgb (%2,%3), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
476 "pavgb (%2,%3,2), %%xmm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
477 "pavgb (%2,%4), %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
478 "movdqa %%xmm0, (%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
479 "movdqa %%xmm1, (%2,%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
480 "movdqa %%xmm2, (%2,%3,2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
481 "movdqa %%xmm3, (%2,%4) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
482 "subl $4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
483 "lea (%1,%3,4), %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
484 "lea (%2,%3,4), %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
485 "jnz 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
486 : "+g"(h), "+r" (pixels), "+r" (block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
487 : "r"((x86_reg)line_size), "r"((x86_reg)3L*line_size)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
488 : "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
489 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
490 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
491
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
492 #define CLEAR_BLOCKS(name,n) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
493 static void name(DCTELEM *blocks)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
494 {\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
495 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
496 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
497 "mov %1, %%"REG_a" \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
498 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
499 "movq %%mm7, (%0, %%"REG_a") \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
500 "movq %%mm7, 8(%0, %%"REG_a") \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
501 "movq %%mm7, 16(%0, %%"REG_a") \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
502 "movq %%mm7, 24(%0, %%"REG_a") \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
503 "add $32, %%"REG_a" \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
504 " js 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
505 : : "r" (((uint8_t *)blocks)+128*n),\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
506 "i" (-128*n)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
507 : "%"REG_a\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
508 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
509 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
510 CLEAR_BLOCKS(clear_blocks_mmx, 6)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
511 CLEAR_BLOCKS(clear_block_mmx, 1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
512
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
513 static void clear_block_sse(DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
514 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
515 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
516 "xorps %%xmm0, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
517 "movaps %%xmm0, (%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
518 "movaps %%xmm0, 16(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
519 "movaps %%xmm0, 32(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
520 "movaps %%xmm0, 48(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
521 "movaps %%xmm0, 64(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
522 "movaps %%xmm0, 80(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
523 "movaps %%xmm0, 96(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
524 "movaps %%xmm0, 112(%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
525 :: "r"(block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
526 : "memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
527 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
528 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
529
9861
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
530 static void clear_blocks_sse(DCTELEM *blocks)
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
531 {\
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
532 __asm__ volatile(
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
533 "xorps %%xmm0, %%xmm0 \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
534 "mov %1, %%"REG_a" \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
535 "1: \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
536 "movaps %%xmm0, (%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
537 "movaps %%xmm0, 16(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
538 "movaps %%xmm0, 32(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
539 "movaps %%xmm0, 48(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
540 "movaps %%xmm0, 64(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
541 "movaps %%xmm0, 80(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
542 "movaps %%xmm0, 96(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
543 "movaps %%xmm0, 112(%0, %%"REG_a") \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
544 "add $128, %%"REG_a" \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
545 " js 1b \n"
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
546 : : "r" (((uint8_t *)blocks)+128*6),
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
547 "i" (-128*6)
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
548 : "%"REG_a
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
549 );
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
550 }
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
551
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
552 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
553 x86_reg i=0;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
554 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
555 "jmp 2f \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
556 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
557 "movq (%1, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
558 "movq (%2, %0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
559 "paddb %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
560 "movq %%mm1, (%2, %0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
561 "movq 8(%1, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
562 "movq 8(%2, %0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
563 "paddb %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
564 "movq %%mm1, 8(%2, %0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
565 "add $16, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
566 "2: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
567 "cmp %3, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
568 " js 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
569 : "+r" (i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
570 : "r"(src), "r"(dst), "r"((x86_reg)w-15)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
571 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
572 for(; i<w; i++)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
573 dst[i+0] += src[i+0];
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
574 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
575
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
576 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
577 x86_reg i=0;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
578 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
579 "jmp 2f \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
580 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
581 "movq (%2, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
582 "movq 8(%2, %0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
583 "paddb (%3, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
584 "paddb 8(%3, %0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
585 "movq %%mm0, (%1, %0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
586 "movq %%mm1, 8(%1, %0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
587 "add $16, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
588 "2: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
589 "cmp %4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
590 " js 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
591 : "+r" (i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
592 : "r"(dst), "r"(src1), "r"(src2), "r"((x86_reg)w-15)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
593 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
594 for(; i<w; i++)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
595 dst[i] = src1[i] + src2[i];
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
596 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
597
8798
a5c8210814d7 Add check whether the compiler/assembler supports 10 or more operands.
diego
parents: 8760
diff changeset
598 #if HAVE_7REGS && HAVE_TEN_OPERANDS
10431
546b7ebeaf07 huffyuv: add some const qualifiers
lorenm
parents: 10430
diff changeset
599 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
600 x86_reg w2 = -w;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
601 x86_reg x;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
602 int l = *left & 0xff;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
603 int tl = *left_top & 0xff;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
604 int t;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
605 __asm__ volatile(
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
606 "mov %7, %3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
607 "1: \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
608 "movzx (%3,%4), %2 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
609 "mov %2, %k3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
610 "sub %b1, %b3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
611 "add %b0, %b3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
612 "mov %2, %1 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
613 "cmp %0, %2 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
614 "cmovg %0, %2 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
615 "cmovg %1, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
616 "cmp %k3, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
617 "cmovg %k3, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
618 "mov %7, %3 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
619 "cmp %2, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
620 "cmovl %2, %0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
621 "add (%6,%4), %b0 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
622 "mov %b0, (%5,%4) \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
623 "inc %4 \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
624 "jl 1b \n"
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
625 :"+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
626 :"r"(dst+w), "r"(diff+w), "rm"(top+w)
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
627 );
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
628 *left = l;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
629 *left_top = tl;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
630 }
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
631 #endif
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
632
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
633 #define H263_LOOP_FILTER \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
634 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
635 "movq %0, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
636 "movq %0, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
637 "movq %3, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
638 "movq %3, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
639 "punpcklbw %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
640 "punpckhbw %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
641 "punpcklbw %%mm7, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
642 "punpckhbw %%mm7, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
643 "psubw %%mm2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
644 "psubw %%mm3, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
645 "movq %1, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
646 "movq %1, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
647 "movq %2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
648 "movq %2, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
649 "punpcklbw %%mm7, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
650 "punpckhbw %%mm7, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
651 "punpcklbw %%mm7, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
652 "punpckhbw %%mm7, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
653 "psubw %%mm2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
654 "psubw %%mm3, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
655 "psllw $2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
656 "psllw $2, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
657 "paddw %%mm0, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
658 "paddw %%mm1, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
659 "pxor %%mm6, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
660 "pcmpgtw %%mm4, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
661 "pcmpgtw %%mm5, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
662 "pxor %%mm6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
663 "pxor %%mm7, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
664 "psubw %%mm6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
665 "psubw %%mm7, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
666 "psrlw $3, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
667 "psrlw $3, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
668 "packuswb %%mm5, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
669 "packsswb %%mm7, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
670 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
671 "movd %4, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
672 "punpcklbw %%mm2, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
673 "punpcklbw %%mm2, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
674 "punpcklbw %%mm2, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
675 "psubusb %%mm4, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
676 "movq %%mm2, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
677 "psubusb %%mm4, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
678 "psubb %%mm3, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
679 "movq %1, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
680 "movq %2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
681 "pxor %%mm6, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
682 "pxor %%mm6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
683 "paddusb %%mm2, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
684 "psubusb %%mm2, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
685 "pxor %%mm6, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
686 "pxor %%mm6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
687 "paddusb %%mm2, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
688 "packsswb %%mm1, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
689 "pcmpgtb %%mm0, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
690 "pxor %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
691 "psubb %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
692 "movq %%mm0, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
693 "psubusb %%mm2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
694 "psubb %%mm0, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
695 "pand %5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
696 "psrlw $2, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
697 "pxor %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
698 "psubb %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
699 "movq %0, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
700 "movq %3, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
701 "psubb %%mm1, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
702 "paddb %%mm1, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
703
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
704 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
705 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
706 const int strength= ff_h263_loop_filter_strength[qscale];
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
707
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
708 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
709
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
710 H263_LOOP_FILTER
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
711
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
712 "movq %%mm3, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
713 "movq %%mm4, %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
714 "movq %%mm5, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
715 "movq %%mm6, %3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
716 : "+m" (*(uint64_t*)(src - 2*stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
717 "+m" (*(uint64_t*)(src - 1*stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
718 "+m" (*(uint64_t*)(src + 0*stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
719 "+m" (*(uint64_t*)(src + 1*stride))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
720 : "g" (2*strength), "m"(ff_pb_FC)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
721 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
722 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
723 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
724
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
725 static inline void transpose4x4(uint8_t *dst, uint8_t *src, int dst_stride, int src_stride){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
726 __asm__ volatile( //FIXME could save 1 instruction if done as 8x4 ...
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
727 "movd %4, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
728 "movd %5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
729 "movd %6, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
730 "movd %7, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
731 "punpcklbw %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
732 "punpcklbw %%mm3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
733 "movq %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
734 "punpcklwd %%mm2, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
735 "punpckhwd %%mm2, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
736 "movd %%mm0, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
737 "punpckhdq %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
738 "movd %%mm0, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
739 "movd %%mm1, %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
740 "punpckhdq %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
741 "movd %%mm1, %3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
742
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
743 : "=m" (*(uint32_t*)(dst + 0*dst_stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
744 "=m" (*(uint32_t*)(dst + 1*dst_stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
745 "=m" (*(uint32_t*)(dst + 2*dst_stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
746 "=m" (*(uint32_t*)(dst + 3*dst_stride))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
747 : "m" (*(uint32_t*)(src + 0*src_stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
748 "m" (*(uint32_t*)(src + 1*src_stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
749 "m" (*(uint32_t*)(src + 2*src_stride)),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
750 "m" (*(uint32_t*)(src + 3*src_stride))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
751 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
752 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
753
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
754 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
755 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
756 const int strength= ff_h263_loop_filter_strength[qscale];
10961
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10766
diff changeset
757 DECLARE_ALIGNED(8, uint64_t, temp)[4];
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
758 uint8_t *btemp= (uint8_t*)temp;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
759
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
760 src -= 2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
761
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
762 transpose4x4(btemp , src , 8, stride);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
763 transpose4x4(btemp+4, src + 4*stride, 8, stride);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
764 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
765 H263_LOOP_FILTER // 5 3 4 6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
766
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
767 : "+m" (temp[0]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
768 "+m" (temp[1]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
769 "+m" (temp[2]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
770 "+m" (temp[3])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
771 : "g" (2*strength), "m"(ff_pb_FC)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
772 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
773
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
774 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
775 "movq %%mm5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
776 "movq %%mm4, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
777 "punpcklbw %%mm3, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
778 "punpcklbw %%mm6, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
779 "punpckhbw %%mm3, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
780 "punpckhbw %%mm6, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
781 "movq %%mm5, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
782 "movq %%mm1, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
783 "punpcklwd %%mm4, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
784 "punpcklwd %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
785 "punpckhwd %%mm4, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
786 "punpckhwd %%mm0, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
787 "movd %%mm5, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
788 "punpckhdq %%mm5, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
789 "movd %%mm5, (%0,%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
790 "movd %%mm3, (%0,%2,2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
791 "punpckhdq %%mm3, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
792 "movd %%mm3, (%0,%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
793 "movd %%mm1, (%1) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
794 "punpckhdq %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
795 "movd %%mm1, (%1,%2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
796 "movd %%mm6, (%1,%2,2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
797 "punpckhdq %%mm6, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
798 "movd %%mm6, (%1,%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
799 :: "r" (src),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
800 "r" (src + 4*stride),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
801 "r" ((x86_reg) stride ),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
802 "r" ((x86_reg)(3*stride))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
803 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
804 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
805 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
806
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
807 /* draw the edges of width 'w' of an image of size width, height
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
808 this mmx version can only handle w==8 || w==16 */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
809 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
810 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
811 uint8_t *ptr, *last_line;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
812 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
813
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
814 last_line = buf + (height - 1) * wrap;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
815 /* left and right */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
816 ptr = buf;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
817 if(w==8)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
818 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
819 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
820 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
821 "movd (%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
822 "punpcklbw %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
823 "punpcklwd %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
824 "punpckldq %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
825 "movq %%mm0, -8(%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
826 "movq -8(%0, %2), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
827 "punpckhbw %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
828 "punpckhwd %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
829 "punpckhdq %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
830 "movq %%mm1, (%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 else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
839 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
840 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
841 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
842 "movd (%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
843 "punpcklbw %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
844 "punpcklwd %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
845 "punpckldq %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
846 "movq %%mm0, -8(%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
847 "movq %%mm0, -16(%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
848 "movq -8(%0, %2), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
849 "punpckhbw %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
850 "punpckhwd %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
851 "punpckhdq %%mm1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
852 "movq %%mm1, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
853 "movq %%mm1, 8(%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
854 "add %1, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
855 "cmp %3, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
856 " jb 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
857 : "+r" (ptr)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
858 : "r" ((x86_reg)wrap), "r" ((x86_reg)width), "r" (ptr + wrap*height)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
859 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
860 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
861
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
862 for(i=0;i<w;i+=4) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
863 /* top and bottom (and hopefully also the corners) */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
864 ptr= buf - (i + 1) * wrap - w;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
865 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
866 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
867 "movq (%1, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
868 "movq %%mm0, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
869 "movq %%mm0, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
870 "movq %%mm0, (%0, %2, 2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
871 "movq %%mm0, (%0, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
872 "add $8, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
873 "cmp %4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
874 " jb 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
875 : "+r" (ptr)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
876 : "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
877 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
878 ptr= last_line + (i + 1) * wrap - w;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
879 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
880 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
881 "movq (%1, %0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
882 "movq %%mm0, (%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
883 "movq %%mm0, (%0, %2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
884 "movq %%mm0, (%0, %2, 2) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
885 "movq %%mm0, (%0, %3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
886 "add $8, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
887 "cmp %4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
888 " jb 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
889 : "+r" (ptr)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
890 : "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
891 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
892 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
893 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
894
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
895 #define PAETH(cpu, abs3)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
896 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
897 {\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
898 x86_reg i = -bpp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
899 x86_reg end = w-3;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
900 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
901 "pxor %%mm7, %%mm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
902 "movd (%1,%0), %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
903 "movd (%2,%0), %%mm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
904 "punpcklbw %%mm7, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
905 "punpcklbw %%mm7, %%mm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
906 "add %4, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
907 "1: \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
908 "movq %%mm1, %%mm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
909 "movd (%2,%0), %%mm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
910 "movq %%mm2, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
911 "punpcklbw %%mm7, %%mm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
912 "movq %%mm2, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
913 "psubw %%mm1, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
914 "psubw %%mm0, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
915 "movq %%mm3, %%mm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
916 "paddw %%mm4, %%mm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
917 abs3\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
918 "movq %%mm4, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
919 "pminsw %%mm5, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
920 "pcmpgtw %%mm6, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
921 "pcmpgtw %%mm5, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
922 "movq %%mm4, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
923 "pand %%mm3, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
924 "pandn %%mm3, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
925 "pandn %%mm0, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
926 "movd (%3,%0), %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
927 "pand %%mm1, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
928 "pand %%mm4, %%mm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
929 "punpcklbw %%mm7, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
930 "movq %6, %%mm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
931 "paddw %%mm6, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
932 "paddw %%mm2, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
933 "paddw %%mm3, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
934 "pand %%mm5, %%mm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
935 "movq %%mm0, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
936 "packuswb %%mm3, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
937 "movd %%mm3, (%1,%0) \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
938 "add %4, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
939 "cmp %5, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
940 "jle 1b \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
941 :"+r"(i)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
942 :"r"(dst), "r"(top), "r"(src), "r"((x86_reg)bpp), "g"(end),\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
943 "m"(ff_pw_255)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
944 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
945 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
946 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
947
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
948 #define ABS3_MMX2\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
949 "psubw %%mm5, %%mm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
950 "pmaxsw %%mm7, %%mm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
951 "pxor %%mm6, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
952 "pxor %%mm7, %%mm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
953 "psubw %%mm3, %%mm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
954 "psubw %%mm4, %%mm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
955 "pmaxsw %%mm6, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
956 "pmaxsw %%mm7, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
957 "pxor %%mm7, %%mm7 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
958
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
959 #define ABS3_SSSE3\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
960 "pabsw %%mm3, %%mm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
961 "pabsw %%mm4, %%mm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
962 "pabsw %%mm5, %%mm5 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
963
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
964 PAETH(mmx2, ABS3_MMX2)
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
965 #if HAVE_SSSE3
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
966 PAETH(ssse3, ABS3_SSSE3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
967 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
968
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
969 #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
970 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
971 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
972 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
973 "movq "#in7", " #m3 " \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
974 "movq "#in0", %%mm5 \n\t" /* D */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
975 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
976 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
977 "movq "#in1", %%mm5 \n\t" /* C */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
978 "movq "#in2", %%mm6 \n\t" /* B */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
979 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
980 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
981 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
982 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
983 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
984 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
985 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
986 "psraw $5, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
987 "packuswb %%mm5, %%mm5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
988 OP(%%mm5, out, %%mm7, d)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
989
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
990 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
991 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
992 uint64_t temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
993 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
994 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
995 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
996 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
997 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
998 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
999 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1000 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1001 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1002 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1003 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1004 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1005 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1006 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1007 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1008 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1009 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1010 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1011 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1012 "paddw %%mm3, %%mm5 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1013 "paddw %%mm2, %%mm6 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1014 "paddw %%mm5, %%mm5 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1015 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1016 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1017 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1018 "paddw %%mm4, %%mm0 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1019 "paddw %%mm1, %%mm5 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1020 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1021 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1022 "paddw %6, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1023 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1024 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1025 "movq %%mm0, %5 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1026 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1027 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1028 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1029 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1030 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1031 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1032 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1033 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1034 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1035 "paddw %%mm0, %%mm2 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1036 "paddw %%mm5, %%mm3 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1037 "paddw %%mm2, %%mm2 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1038 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1039 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1040 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1041 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1042 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1043 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1044 "paddw %%mm2, %%mm1 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1045 "paddw %%mm6, %%mm4 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1046 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1047 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1048 "paddw %6, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1049 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1050 "psraw $5, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1051 "movq %5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1052 "packuswb %%mm3, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1053 OP_MMX2(%%mm1, (%1),%%mm4, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1054 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1055 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1056 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1057 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1058 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1059 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1060 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1061 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1062 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1063 "paddw %%mm1, %%mm5 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1064 "paddw %%mm4, %%mm0 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1065 "paddw %%mm5, %%mm5 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1066 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1067 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1068 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1069 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1070 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1071 "paddw %%mm3, %%mm2 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1072 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1073 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1074 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1075 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1076 "paddw %%mm2, %%mm6 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1077 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1078 "paddw %6, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1079 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1080 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1081 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1082 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1083 "paddw %%mm5, %%mm3 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1084 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1085 "paddw %%mm4, %%mm6 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1086 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1087 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1088 "paddw %%mm1, %%mm4 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1089 "paddw %%mm2, %%mm5 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1090 "paddw %%mm6, %%mm6 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1091 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1092 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1093 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1094 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1095 "paddw %6, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1096 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1097 "psraw $5, %%mm4 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1098 "packuswb %%mm4, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1099 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1100 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1101 "add %3, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1102 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1103 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1104 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1105 : "+a"(src), "+c"(dst), "+D"(h)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1106 : "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
1107 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1108 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1109 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1110 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1111 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
1112 int i;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1113 int16_t temp[16];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1114 /* quick HACK, XXX FIXME MUST be optimized */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1115 for(i=0; i<h; i++)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1116 {\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1117 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
1118 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
1119 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
1120 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
1121 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
1122 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
1123 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
1124 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
1125 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
1126 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
1127 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
1128 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
1129 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
1130 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
1131 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
1132 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
1133 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1134 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1135 "movq 8(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1136 "paddw %2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1137 "paddw %2, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1138 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1139 "psraw $5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1140 "packuswb %%mm1, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1141 OP_3DNOW(%%mm0, (%1), %%mm1, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1142 "movq 16(%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1143 "movq 24(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1144 "paddw %2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1145 "paddw %2, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1146 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1147 "psraw $5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1148 "packuswb %%mm1, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1149 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1150 :: "r"(temp), "r"(dst), "m"(ROUNDER)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1151 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1152 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1153 dst+=dstStride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1154 src+=srcStride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1155 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1156 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1157 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1158 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
1159 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1160 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1161 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1162 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1163 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1164 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1165 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1166 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1167 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1168 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1169 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1170 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1171 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1172 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1173 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1174 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1175 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1176 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1177 "paddw %%mm3, %%mm5 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1178 "paddw %%mm2, %%mm6 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1179 "paddw %%mm5, %%mm5 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1180 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1181 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1182 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1183 "paddw %%mm4, %%mm0 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1184 "paddw %%mm1, %%mm5 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1185 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1186 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1187 "paddw %5, %%mm6 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1188 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1189 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1190 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1191 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1192 "movd 5(%0), %%mm5 \n\t" /* FGHI */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1193 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1194 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1195 "paddw %%mm5, %%mm1 \n\t" /* a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1196 "paddw %%mm6, %%mm2 \n\t" /* b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1197 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1198 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1199 "paddw %%mm6, %%mm3 \n\t" /* c */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1200 "paddw %%mm5, %%mm4 \n\t" /* d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1201 "paddw %%mm2, %%mm2 \n\t" /* 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1202 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1203 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1204 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1205 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1206 "paddw %5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1207 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1208 "psraw $5, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1209 "packuswb %%mm3, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1210 OP_MMX2(%%mm0, (%1), %%mm4, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1211 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1212 "add %3, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1213 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1214 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1215 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1216 : "+a"(src), "+c"(dst), "+d"(h)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1217 : "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
1218 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1219 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1220 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1221 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1222 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
1223 int i;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1224 int16_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1225 /* quick HACK, XXX FIXME MUST be optimized */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1226 for(i=0; i<h; i++)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1227 {\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1228 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
1229 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
1230 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
1231 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
1232 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
1233 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
1234 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
1235 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
1236 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1237 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1238 "movq 8(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1239 "paddw %2, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1240 "paddw %2, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1241 "psraw $5, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1242 "psraw $5, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1243 "packuswb %%mm1, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1244 OP_3DNOW(%%mm0, (%1), %%mm1, q)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1245 :: "r"(temp), "r"(dst), "m"(ROUNDER)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1246 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1247 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1248 dst+=dstStride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1249 src+=srcStride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1250 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1251 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1252
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1253 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1254 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1255 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
1256 uint64_t temp[17*4];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1257 uint64_t *temp_ptr= temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1258 int count= 17;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1259 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1260 /*FIXME unroll */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1261 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1262 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1263 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1264 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1265 "movq (%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1266 "movq 8(%0), %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1267 "movq 8(%0), %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1268 "punpcklbw %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1269 "punpckhbw %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1270 "punpcklbw %%mm7, %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1271 "punpckhbw %%mm7, %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1272 "movq %%mm0, (%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1273 "movq %%mm1, 17*8(%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1274 "movq %%mm2, 2*17*8(%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1275 "movq %%mm3, 3*17*8(%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1276 "add $8, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1277 "add %3, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1278 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1279 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1280 : "+r" (src), "+r" (temp_ptr), "+r"(count)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1281 : "r" ((x86_reg)srcStride)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1282 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1283 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1284 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1285 temp_ptr= temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1286 count=4;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1287 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1288 /*FIXME reorder for speed */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1289 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1290 /*"pxor %%mm7, %%mm7 \n\t"*/\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1291 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1292 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1293 "movq 8(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1294 "movq 16(%0), %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1295 "movq 24(%0), %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1296 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
1297 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
1298 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1299 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
1300 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1301 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
1302 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1303 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
1304 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
1305 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1306 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
1307 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
1308 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1309 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
1310 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
1311 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1312 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
1313 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
1314 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1315 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
1316 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1317 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
1318 "add %4, %1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1319 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
1320 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
1321 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1322 "add $136, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1323 "add %6, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1324 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1325 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1326 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1327 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1328 : "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
1329 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1330 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1331 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1332 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1333 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
1334 uint64_t temp[9*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1335 uint64_t *temp_ptr= temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1336 int count= 9;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1337 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1338 /*FIXME unroll */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1339 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1340 "pxor %%mm7, %%mm7 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1341 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1342 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1343 "movq (%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1344 "punpcklbw %%mm7, %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1345 "punpckhbw %%mm7, %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1346 "movq %%mm0, (%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1347 "movq %%mm1, 9*8(%1) \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1348 "add $8, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1349 "add %3, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1350 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1351 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1352 : "+r" (src), "+r" (temp_ptr), "+r"(count)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1353 : "r" ((x86_reg)srcStride)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1354 : "memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1355 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1356 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1357 temp_ptr= temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1358 count=2;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1359 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1360 /*FIXME reorder for speed */\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1361 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1362 /*"pxor %%mm7, %%mm7 \n\t"*/\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1363 "1: \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1364 "movq (%0), %%mm0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1365 "movq 8(%0), %%mm1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1366 "movq 16(%0), %%mm2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1367 "movq 24(%0), %%mm3 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1368 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
1369 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
1370 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1371 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
1372 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1373 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
1374 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1375 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
1376 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1377 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
1378 "add %4, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1379 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
1380 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
1381 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1382 "add $72, %0 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1383 "add %6, %1 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1384 "decl %2 \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1385 " jnz 1b \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1386 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1387 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1388 : "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
1389 : "memory"\
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 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1393 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1394 OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1395 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1396 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1397 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1398 uint64_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1399 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1400 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1401 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1402 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1403 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1404 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1405 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1406 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1407 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1408 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1409 uint64_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1410 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1411 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1412 OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1413 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1414 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1415 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1416 uint64_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1417 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1418 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1419 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1420 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1421 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1422 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1423 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1424 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1425 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1426 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1427 uint64_t temp[8];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1428 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1429 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1430 OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1431 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1432 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1433 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1434 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1435 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1436 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1437 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1438 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1439 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1440 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1441 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1442 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1443 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1444 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1445 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1446 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1447 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1448 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1449 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1450 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1451 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1452 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1453 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1454 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1455 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1456 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1457 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1458 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1459 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1460 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1461 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1462 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1463 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1464 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1465 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1466 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 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_mc21_ ## 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) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1471 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1472 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1473 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1474 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1475 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1476 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1477 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1478 uint8_t * const halfH= ((uint8_t*)half) + 64;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1479 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1480 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1481 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1482 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1483 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1484 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1485 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1486 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1487 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1488 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1489 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1490 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1491 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1492 uint64_t half[8 + 9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1493 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1494 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1495 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1496 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1497 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1498 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1499 uint64_t half[9];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1500 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1501 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1502 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1503 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1504 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1505 OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1506 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1507 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1508 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1509 uint64_t temp[32];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1510 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1511 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1512 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1513 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1514 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1515 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1516 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1517 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1518 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1519 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1520 uint64_t temp[32];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1521 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1522 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1523 OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1524 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1525 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1526 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1527 uint64_t temp[32];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1528 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1529 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1530 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1531 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1532 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1533 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1534 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1535 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1536 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1537 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1538 uint64_t temp[32];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1539 uint8_t * const half= (uint8_t*)temp;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1540 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1541 OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1542 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1543 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1544 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1545 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1546 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1547 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1548 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1549 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1550 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1551 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1552 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1553 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1554 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1555 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1556 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1557 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1558 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1559 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1560 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1561 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1562 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1563 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1564 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1565 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1566 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1567 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1568 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1569 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1570 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1571 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1572 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1573 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1574 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1575 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1576 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1577 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 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_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1580 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1581 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1582 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1583 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1584 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1585 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1586 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1587 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1588 uint64_t half[16*2 + 17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1589 uint8_t * const halfH= ((uint8_t*)half) + 256;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1590 uint8_t * const halfHV= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1591 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1592 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1593 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1594 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1595 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1596 uint64_t half[17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1597 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1598 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1599 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1600 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1601 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1602 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1603 uint64_t half[17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1604 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1605 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1606 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1607 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1608 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1609 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1610 uint64_t half[17*2];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1611 uint8_t * const halfH= ((uint8_t*)half);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1612 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1613 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1614 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1615
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1616 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1617 #define AVG_3DNOW_OP(a,b,temp, size) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1618 "mov" #size " " #b ", " #temp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1619 "pavgusb " #temp ", " #a " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1620 "mov" #size " " #a ", " #b " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1621 #define AVG_MMX2_OP(a,b,temp, size) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1622 "mov" #size " " #b ", " #temp " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1623 "pavgb " #temp ", " #a " \n\t"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1624 "mov" #size " " #a ", " #b " \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1625
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1626 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1627 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1628 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1629 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1630 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1631 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1632 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1633 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1634 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1635
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1636 /***********************************/
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1637 /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1638
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1639 #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1640 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
1641 OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1642 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1643 #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1644 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
1645 OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src+S0, stride, SIZE, S1, S2);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1646 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1647
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1648 #define QPEL_2TAP(OPNAME, SIZE, MMX)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1649 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1650 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1651 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1652 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1653 OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1654 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1655 OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1656 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1657 OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1658 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
1659 OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src+1, stride, SIZE);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1660 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1661 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
1662 OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src+stride, stride, SIZE);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1663 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1664 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0, 1, 0)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1665 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1, -1, 0)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1666 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0, stride, 0)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1667 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride, -stride, 0)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1668 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0, stride, 1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1669 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1, stride, -1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1670 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride, -stride, 1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1671 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride+1, -stride, -1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1672
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1673 QPEL_2TAP(put_, 16, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1674 QPEL_2TAP(avg_, 16, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1675 QPEL_2TAP(put_, 8, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1676 QPEL_2TAP(avg_, 8, mmx2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1677 QPEL_2TAP(put_, 16, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1678 QPEL_2TAP(avg_, 16, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1679 QPEL_2TAP(put_, 8, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1680 QPEL_2TAP(avg_, 8, 3dnow)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1681
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1682
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1683 #if 0
8527
f8bf438c6000 Add missing 'void' keyword to parameterless function declarations.
diego
parents: 8519
diff changeset
1684 static void just_return(void) { return; }
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1685 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1686
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1687 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
1688 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
1689 const int w = 8;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1690 const int ix = ox>>(16+shift);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1691 const int iy = oy>>(16+shift);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1692 const int oxs = ox>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1693 const int oys = oy>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1694 const int dxxs = dxx>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1695 const int dxys = dxy>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1696 const int dyxs = dyx>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1697 const int dyys = dyy>>4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1698 const uint16_t r4[4] = {r,r,r,r};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1699 const uint16_t dxy4[4] = {dxys,dxys,dxys,dxys};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1700 const uint16_t dyy4[4] = {dyys,dyys,dyys,dyys};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1701 const uint64_t shift2 = 2*shift;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1702 uint8_t edge_buf[(h+1)*stride];
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1703 int x, y;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1704
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1705 const int dxw = (dxx-(1<<(16+shift)))*(w-1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1706 const int dyh = (dyy-(1<<(16+shift)))*(h-1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1707 const int dxh = dxy*(h-1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1708 const int dyw = dyx*(w-1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1709 if( // non-constant fullpel offset (3% of blocks)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1710 ((ox^(ox+dxw)) | (ox^(ox+dxh)) | (ox^(ox+dxw+dxh)) |
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1711 (oy^(oy+dyw)) | (oy^(oy+dyh)) | (oy^(oy+dyw+dyh))) >> (16+shift)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1712 // uses more than 16 bits of subpel mv (only at huge resolution)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1713 || (dxx|dxy|dyx|dyy)&15 )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1714 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1715 //FIXME could still use mmx for some of the rows
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1716 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
1717 return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1718 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1719
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1720 src += ix + iy*stride;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1721 if( (unsigned)ix >= width-w ||
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1722 (unsigned)iy >= height-h )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1723 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1724 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
1725 src = edge_buf;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1726 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1727
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1728 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1729 "movd %0, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1730 "pxor %%mm7, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1731 "punpcklwd %%mm6, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1732 "punpcklwd %%mm6, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1733 :: "r"(1<<shift)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1734 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1735
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1736 for(x=0; x<w; x+=4){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1737 uint16_t dx4[4] = { oxs - dxys + dxxs*(x+0),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1738 oxs - dxys + dxxs*(x+1),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1739 oxs - dxys + dxxs*(x+2),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1740 oxs - dxys + dxxs*(x+3) };
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1741 uint16_t dy4[4] = { oys - dyys + dyxs*(x+0),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1742 oys - dyys + dyxs*(x+1),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1743 oys - dyys + dyxs*(x+2),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1744 oys - dyys + dyxs*(x+3) };
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1745
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1746 for(y=0; y<h; y++){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1747 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1748 "movq %0, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1749 "movq %1, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1750 "paddw %2, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1751 "paddw %3, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1752 "movq %%mm4, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1753 "movq %%mm5, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1754 "psrlw $12, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1755 "psrlw $12, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1756 : "+m"(*dx4), "+m"(*dy4)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1757 : "m"(*dxy4), "m"(*dyy4)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1758 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1759
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1760 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1761 "movq %%mm6, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1762 "movq %%mm6, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1763 "psubw %%mm4, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1764 "psubw %%mm5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1765 "movq %%mm2, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1766 "movq %%mm4, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1767 "pmullw %%mm1, %%mm0 \n\t" // (s-dx)*(s-dy)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1768 "pmullw %%mm5, %%mm3 \n\t" // dx*dy
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1769 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1770 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1771
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1772 "movd %4, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1773 "movd %3, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1774 "punpcklbw %%mm7, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1775 "punpcklbw %%mm7, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1776 "pmullw %%mm5, %%mm3 \n\t" // src[1,1] * dx*dy
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1777 "pmullw %%mm4, %%mm2 \n\t" // src[0,1] * (s-dx)*dy
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1778
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1779 "movd %2, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1780 "movd %1, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1781 "punpcklbw %%mm7, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1782 "punpcklbw %%mm7, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1783 "pmullw %%mm5, %%mm1 \n\t" // src[1,0] * dx*(s-dy)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1784 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1785 "paddw %5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1786 "paddw %%mm3, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1787 "paddw %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1788 "paddw %%mm2, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1789
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1790 "psrlw %6, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1791 "packuswb %%mm0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1792 "movd %%mm0, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1793
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1794 : "=m"(dst[x+y*stride])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1795 : "m"(src[0]), "m"(src[1]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1796 "m"(src[stride]), "m"(src[stride+1]),
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1797 "m"(*r4), "m"(shift2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1798 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1799 src += stride;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1800 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1801 src += 4-h*stride;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1802 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1803 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1804
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1805 #define PREFETCH(name, op) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1806 static void name(void *mem, int stride, int h){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1807 const uint8_t *p= mem;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1808 do{\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1809 __asm__ volatile(#op" %0" :: "m"(*p));\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1810 p+= stride;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1811 }while(--h);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1812 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1813 PREFETCH(prefetch_mmx2, prefetcht0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1814 PREFETCH(prefetch_3dnow, prefetch)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1815 #undef PREFETCH
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1816
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1817 #include "h264dsp_mmx.c"
8519
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
1818 #include "rv40dsp_mmx.c"
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1819
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1820 /* CAVS specific */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1821 void ff_cavsdsp_init_mmx2(DSPContext* c, AVCodecContext *avctx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1822 void ff_cavsdsp_init_3dnow(DSPContext* c, AVCodecContext *avctx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1823
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1824 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
1825 put_pixels8_mmx(dst, src, stride, 8);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1826 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1827 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
1828 avg_pixels8_mmx(dst, src, stride, 8);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1829 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1830 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
1831 put_pixels16_mmx(dst, src, stride, 16);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1832 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1833 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
1834 avg_pixels16_mmx(dst, src, stride, 16);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1835 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1836
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1837 /* VC1 specific */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1838 void ff_vc1dsp_init_mmx(DSPContext* dsp, AVCodecContext *avctx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1839
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1840 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
1841 put_pixels8_mmx(dst, src, stride, 8);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1842 }
9441
e14cd3ac3806 VC1: extend MMX qpel MC to include MMX2 avg qpel
conrad
parents: 9440
diff changeset
1843 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
1844 avg_pixels8_mmx2(dst, src, stride, 8);
e14cd3ac3806 VC1: extend MMX qpel MC to include MMX2 avg qpel
conrad
parents: 9440
diff changeset
1845 }
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1846
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1847 /* external functions, from idct_mmx.c */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1848 void ff_mmx_idct(DCTELEM *block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1849 void ff_mmxext_idct(DCTELEM *block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1850
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1851 /* XXX: those functions should be suppressed ASAP when all IDCTs are
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1852 converted */
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
1853 #if CONFIG_GPL
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1854 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1855 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1856 ff_mmx_idct (block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1857 put_pixels_clamped_mmx(block, dest, line_size);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1858 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1859 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1860 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1861 ff_mmx_idct (block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1862 add_pixels_clamped_mmx(block, dest, line_size);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1863 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1864 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1865 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1866 ff_mmxext_idct (block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1867 put_pixels_clamped_mmx(block, dest, line_size);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1868 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1869 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1870 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1871 ff_mmxext_idct (block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1872 add_pixels_clamped_mmx(block, dest, line_size);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1873 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1874 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1875 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
1876 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1877 ff_idct_xvid_mmx (block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1878 put_pixels_clamped_mmx(block, dest, line_size);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1879 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1880 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
1881 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1882 ff_idct_xvid_mmx (block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1883 add_pixels_clamped_mmx(block, dest, line_size);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1884 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1885 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
1886 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1887 ff_idct_xvid_mmx2 (block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1888 put_pixels_clamped_mmx(block, dest, line_size);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1889 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1890 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
1891 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1892 ff_idct_xvid_mmx2 (block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1893 add_pixels_clamped_mmx(block, dest, line_size);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1894 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1895
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1896 static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1897 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1898 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1899 __asm__ volatile("pxor %%mm7, %%mm7":);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1900 for(i=0; i<blocksize; i+=2) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1901 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1902 "movq %0, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1903 "movq %1, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1904 "movq %%mm0, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1905 "movq %%mm1, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1906 "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1907 "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1908 "pslld $31, %%mm2 \n\t" // keep only the sign bit
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1909 "pxor %%mm2, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1910 "movq %%mm3, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1911 "pand %%mm1, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1912 "pandn %%mm1, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1913 "pfadd %%mm0, %%mm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1914 "pfsub %%mm4, %%mm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1915 "movq %%mm3, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1916 "movq %%mm0, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1917 :"+m"(mag[i]), "+m"(ang[i])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1918 ::"memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1919 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1920 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1921 __asm__ volatile("femms");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1922 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1923 static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1924 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1925 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1926
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1927 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1928 "movaps %0, %%xmm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1929 ::"m"(ff_pdw_80000000[0])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1930 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1931 for(i=0; i<blocksize; i+=4) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1932 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1933 "movaps %0, %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1934 "movaps %1, %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1935 "xorps %%xmm2, %%xmm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1936 "xorps %%xmm3, %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1937 "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1938 "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1939 "andps %%xmm5, %%xmm2 \n\t" // keep only the sign bit
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1940 "xorps %%xmm2, %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1941 "movaps %%xmm3, %%xmm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1942 "andps %%xmm1, %%xmm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1943 "andnps %%xmm1, %%xmm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1944 "addps %%xmm0, %%xmm3 \n\t" // a = m + ((a<0) & (a ^ sign(m)))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1945 "subps %%xmm4, %%xmm0 \n\t" // m = m + ((a>0) & (a ^ sign(m)))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1946 "movaps %%xmm3, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1947 "movaps %%xmm0, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1948 :"+m"(mag[i]), "+m"(ang[i])
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1949 ::"memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1950 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1951 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1952 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1953
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1954 #define IF1(x) x
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1955 #define IF0(x)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1956
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1957 #define MIX5(mono,stereo)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1958 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1959 "movss 0(%2), %%xmm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1960 "movss 8(%2), %%xmm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1961 "movss 24(%2), %%xmm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1962 "shufps $0, %%xmm5, %%xmm5 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1963 "shufps $0, %%xmm6, %%xmm6 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1964 "shufps $0, %%xmm7, %%xmm7 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1965 "1: \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1966 "movaps (%0,%1), %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1967 "movaps 0x400(%0,%1), %%xmm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1968 "movaps 0x800(%0,%1), %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1969 "movaps 0xc00(%0,%1), %%xmm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1970 "movaps 0x1000(%0,%1), %%xmm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1971 "mulps %%xmm5, %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1972 "mulps %%xmm6, %%xmm1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1973 "mulps %%xmm5, %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1974 "mulps %%xmm7, %%xmm3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1975 "mulps %%xmm7, %%xmm4 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1976 stereo("addps %%xmm1, %%xmm0 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1977 "addps %%xmm1, %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1978 "addps %%xmm3, %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1979 "addps %%xmm4, %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1980 mono("addps %%xmm2, %%xmm0 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1981 "movaps %%xmm0, (%0,%1) \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1982 stereo("movaps %%xmm2, 0x400(%0,%1) \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1983 "add $16, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1984 "jl 1b \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1985 :"+&r"(i)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1986 :"r"(samples[0]+len), "r"(matrix)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1987 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1988 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1989
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1990 #define MIX_MISC(stereo)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1991 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1992 "1: \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1993 "movaps (%3,%0), %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1994 stereo("movaps %%xmm0, %%xmm1 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1995 "mulps %%xmm6, %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1996 stereo("mulps %%xmm7, %%xmm1 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1997 "lea 1024(%3,%0), %1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1998 "mov %5, %2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1999 "2: \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2000 "movaps (%1), %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2001 stereo("movaps %%xmm2, %%xmm3 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2002 "mulps (%4,%2), %%xmm2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2003 stereo("mulps 16(%4,%2), %%xmm3 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2004 "addps %%xmm2, %%xmm0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2005 stereo("addps %%xmm3, %%xmm1 \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2006 "add $1024, %1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2007 "add $32, %2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2008 "jl 2b \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2009 "movaps %%xmm0, (%3,%0) \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2010 stereo("movaps %%xmm1, 1024(%3,%0) \n")\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2011 "add $16, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2012 "jl 1b \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2013 :"+&r"(i), "=&r"(j), "=&r"(k)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2014 :"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
2015 :"memory"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2016 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2017
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2018 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
2019 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2020 int (*matrix_cmp)[2] = (int(*)[2])matrix;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2021 intptr_t i,j,k;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2022
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2023 i = -len*sizeof(float);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2024 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
2025 MIX5(IF0,IF1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2026 } 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
2027 MIX5(IF1,IF0);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2028 } else {
10961
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10766
diff changeset
2029 DECLARE_ALIGNED_16(float, matrix_simd)[in_ch][2][4];
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2030 j = 2*in_ch*sizeof(float);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2031 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2032 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2033 "sub $8, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2034 "movss (%2,%0), %%xmm6 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2035 "movss 4(%2,%0), %%xmm7 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2036 "shufps $0, %%xmm6, %%xmm6 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2037 "shufps $0, %%xmm7, %%xmm7 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2038 "movaps %%xmm6, (%1,%0,4) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2039 "movaps %%xmm7, 16(%1,%0,4) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2040 "jg 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2041 :"+&r"(j)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2042 :"r"(matrix_simd), "r"(matrix)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2043 :"memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2044 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2045 if(out_ch == 2) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2046 MIX_MISC(IF1);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2047 } else {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2048 MIX_MISC(IF0);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2049 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2050 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2051 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2052
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2053 static void vector_fmul_3dnow(float *dst, const float *src, int len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2054 x86_reg i = (len-4)*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2055 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2056 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2057 "movq (%1,%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2058 "movq 8(%1,%0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2059 "pfmul (%2,%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2060 "pfmul 8(%2,%0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2061 "movq %%mm0, (%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2062 "movq %%mm1, 8(%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2063 "sub $16, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2064 "jge 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2065 "femms \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2066 :"+r"(i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2067 :"r"(dst), "r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2068 :"memory"
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 static void vector_fmul_sse(float *dst, const float *src, int len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2072 x86_reg i = (len-8)*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2073 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2074 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2075 "movaps (%1,%0), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2076 "movaps 16(%1,%0), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2077 "mulps (%2,%0), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2078 "mulps 16(%2,%0), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2079 "movaps %%xmm0, (%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2080 "movaps %%xmm1, 16(%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2081 "sub $32, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2082 "jge 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2083 :"+r"(i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2084 :"r"(dst), "r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2085 :"memory"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2086 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2087 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2088
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2089 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
2090 x86_reg i = len*4-16;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2091 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2092 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2093 "pswapd 8(%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2094 "pswapd (%1), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2095 "pfmul (%3,%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2096 "pfmul 8(%3,%0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2097 "movq %%mm0, (%2,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2098 "movq %%mm1, 8(%2,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2099 "add $16, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2100 "sub $16, %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), "+r"(src1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2103 :"r"(dst), "r"(src0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2104 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2105 __asm__ volatile("femms");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2106 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2107 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
2108 x86_reg i = len*4-32;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2109 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2110 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2111 "movaps 16(%1), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2112 "movaps (%1), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2113 "shufps $0x1b, %%xmm0, %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2114 "shufps $0x1b, %%xmm1, %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2115 "mulps (%3,%0), %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2116 "mulps 16(%3,%0), %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2117 "movaps %%xmm0, (%2,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2118 "movaps %%xmm1, 16(%2,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2119 "add $32, %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2120 "sub $32, %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2121 "jge 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2122 :"+r"(i), "+r"(src1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2123 :"r"(dst), "r"(src0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2124 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2125 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2126
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10107
diff changeset
2127 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
2128 const float *src2, int len){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2129 x86_reg i = (len-4)*4;
10301
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2130 __asm__ volatile(
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2131 "1: \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2132 "movq (%2,%0), %%mm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2133 "movq 8(%2,%0), %%mm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2134 "pfmul (%3,%0), %%mm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2135 "pfmul 8(%3,%0), %%mm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2136 "pfadd (%4,%0), %%mm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2137 "pfadd 8(%4,%0), %%mm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2138 "movq %%mm0, (%1,%0) \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2139 "movq %%mm1, 8(%1,%0) \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2140 "sub $16, %0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2141 "jge 1b \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2142 :"+r"(i)
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2143 :"r"(dst), "r"(src0), "r"(src1), "r"(src2)
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2144 :"memory"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2145 );
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2146 __asm__ volatile("femms");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2147 }
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10107
diff changeset
2148 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
2149 const float *src2, int len){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2150 x86_reg i = (len-8)*4;
10301
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2151 __asm__ volatile(
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2152 "1: \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2153 "movaps (%2,%0), %%xmm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2154 "movaps 16(%2,%0), %%xmm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2155 "mulps (%3,%0), %%xmm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2156 "mulps 16(%3,%0), %%xmm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2157 "addps (%4,%0), %%xmm0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2158 "addps 16(%4,%0), %%xmm1 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2159 "movaps %%xmm0, (%1,%0) \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2160 "movaps %%xmm1, 16(%1,%0) \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2161 "sub $32, %0 \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2162 "jge 1b \n\t"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2163 :"+r"(i)
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2164 :"r"(dst), "r"(src0), "r"(src1), "r"(src2)
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2165 :"memory"
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
2166 );
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2167 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2168
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2169 static void vector_fmul_window_3dnow2(float *dst, const float *src0, const float *src1,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2170 const float *win, float add_bias, int len){
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2171 #if HAVE_6REGS
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2172 if(add_bias == 0){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2173 x86_reg i = -len*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2174 x86_reg j = len*4-8;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2175 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2176 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2177 "pswapd (%5,%1), %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2178 "movq (%5,%0), %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2179 "pswapd (%4,%1), %%mm5 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2180 "movq (%3,%0), %%mm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2181 "movq %%mm0, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2182 "movq %%mm1, %%mm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2183 "pfmul %%mm4, %%mm2 \n" // src0[len+i]*win[len+i]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2184 "pfmul %%mm5, %%mm3 \n" // src1[ j]*win[len+j]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2185 "pfmul %%mm4, %%mm1 \n" // src0[len+i]*win[len+j]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2186 "pfmul %%mm5, %%mm0 \n" // src1[ j]*win[len+i]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2187 "pfadd %%mm3, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2188 "pfsub %%mm0, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2189 "pswapd %%mm2, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2190 "movq %%mm1, (%2,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2191 "movq %%mm2, (%2,%1) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2192 "sub $8, %1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2193 "add $8, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2194 "jl 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2195 "femms \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2196 :"+r"(i), "+r"(j)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2197 :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2198 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2199 }else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2200 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2201 ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2202 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2203
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2204 static void vector_fmul_window_sse(float *dst, const float *src0, const float *src1,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2205 const float *win, float add_bias, int len){
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2206 #if HAVE_6REGS
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2207 if(add_bias == 0){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2208 x86_reg i = -len*4;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2209 x86_reg j = len*4-16;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2210 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2211 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2212 "movaps (%5,%1), %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2213 "movaps (%5,%0), %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2214 "movaps (%4,%1), %%xmm5 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2215 "movaps (%3,%0), %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2216 "shufps $0x1b, %%xmm1, %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2217 "shufps $0x1b, %%xmm5, %%xmm5 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2218 "movaps %%xmm0, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2219 "movaps %%xmm1, %%xmm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2220 "mulps %%xmm4, %%xmm2 \n" // src0[len+i]*win[len+i]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2221 "mulps %%xmm5, %%xmm3 \n" // src1[ j]*win[len+j]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2222 "mulps %%xmm4, %%xmm1 \n" // src0[len+i]*win[len+j]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2223 "mulps %%xmm5, %%xmm0 \n" // src1[ j]*win[len+i]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2224 "addps %%xmm3, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2225 "subps %%xmm0, %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2226 "shufps $0x1b, %%xmm2, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2227 "movaps %%xmm1, (%2,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2228 "movaps %%xmm2, (%2,%1) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2229 "sub $16, %1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2230 "add $16, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2231 "jl 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2232 :"+r"(i), "+r"(j)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2233 :"r"(dst+len), "r"(src0+len), "r"(src1), "r"(win+len)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2234 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2235 }else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2236 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2237 ff_vector_fmul_window_c(dst, src0, src1, win, add_bias, len);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2238 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2239
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2240 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
2241 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2242 x86_reg i = -4*len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2243 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2244 "movss %3, %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2245 "shufps $0, %%xmm4, %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2246 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2247 "cvtpi2ps (%2,%0), %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2248 "cvtpi2ps 8(%2,%0), %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2249 "cvtpi2ps 16(%2,%0), %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2250 "cvtpi2ps 24(%2,%0), %%xmm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2251 "movlhps %%xmm1, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2252 "movlhps %%xmm3, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2253 "mulps %%xmm4, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2254 "mulps %%xmm4, %%xmm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2255 "movaps %%xmm0, (%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2256 "movaps %%xmm2, 16(%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2257 "add $32, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2258 "jl 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2259 :"+r"(i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2260 :"r"(dst+len), "r"(src+len), "m"(mul)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2261 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2262 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2263
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2264 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
2265 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2266 x86_reg i = -4*len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2267 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2268 "movss %3, %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2269 "shufps $0, %%xmm4, %%xmm4 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2270 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2271 "cvtdq2ps (%2,%0), %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2272 "cvtdq2ps 16(%2,%0), %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2273 "mulps %%xmm4, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2274 "mulps %%xmm4, %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2275 "movaps %%xmm0, (%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2276 "movaps %%xmm1, 16(%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2277 "add $32, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2278 "jl 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2279 :"+r"(i)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2280 :"r"(dst+len), "r"(src+len), "m"(mul)
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
10105
7775f6627612 Mark parameter src of vector_clipf() as const
vitor
parents: 10104
diff changeset
2284 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
2285 int len)
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2286 {
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2287 x86_reg i = (len-16)*4;
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2288 __asm__ volatile(
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2289 "movss %3, %%xmm4 \n"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2290 "movss %4, %%xmm5 \n"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2291 "shufps $0, %%xmm4, %%xmm4 \n"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2292 "shufps $0, %%xmm5, %%xmm5 \n"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2293 "1: \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2294 "movaps (%2,%0), %%xmm0 \n\t" // 3/1 on intel
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2295 "movaps 16(%2,%0), %%xmm1 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2296 "movaps 32(%2,%0), %%xmm2 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2297 "movaps 48(%2,%0), %%xmm3 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2298 "maxps %%xmm4, %%xmm0 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2299 "maxps %%xmm4, %%xmm1 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2300 "maxps %%xmm4, %%xmm2 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2301 "maxps %%xmm4, %%xmm3 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2302 "minps %%xmm5, %%xmm0 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2303 "minps %%xmm5, %%xmm1 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2304 "minps %%xmm5, %%xmm2 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2305 "minps %%xmm5, %%xmm3 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2306 "movaps %%xmm0, (%1,%0) \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2307 "movaps %%xmm1, 16(%1,%0) \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2308 "movaps %%xmm2, 32(%1,%0) \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2309 "movaps %%xmm3, 48(%1,%0) \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2310 "sub $64, %0 \n\t"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2311 "jge 1b \n\t"
10107
3b61bc6ce377 Mark "i" parameter of vector_clipf_sse() as early-clobber
vitor
parents: 10105
diff changeset
2312 :"+&r"(i)
10104
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2313 :"r"(dst), "r"(src), "m"(min), "m"(max)
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2314 :"memory"
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2315 );
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2316 }
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 9995
diff changeset
2317
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2318 static void float_to_int16_3dnow(int16_t *dst, const float *src, long len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2319 x86_reg reglen = len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2320 // not bit-exact: pf2id uses different rounding than C and SSE
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2321 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2322 "add %0 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2323 "lea (%2,%0,2) , %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2324 "add %0 , %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2325 "neg %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2326 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2327 "pf2id (%2,%0,2) , %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2328 "pf2id 8(%2,%0,2) , %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2329 "pf2id 16(%2,%0,2) , %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2330 "pf2id 24(%2,%0,2) , %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2331 "packssdw %%mm1 , %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2332 "packssdw %%mm3 , %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2333 "movq %%mm0 , (%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2334 "movq %%mm2 , 8(%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2335 "add $16 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2336 " js 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2337 "femms \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2338 :"+r"(reglen), "+r"(dst), "+r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2339 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2340 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2341 static void float_to_int16_sse(int16_t *dst, const float *src, long len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2342 x86_reg reglen = len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2343 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2344 "add %0 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2345 "lea (%2,%0,2) , %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2346 "add %0 , %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2347 "neg %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2348 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2349 "cvtps2pi (%2,%0,2) , %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2350 "cvtps2pi 8(%2,%0,2) , %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2351 "cvtps2pi 16(%2,%0,2) , %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2352 "cvtps2pi 24(%2,%0,2) , %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2353 "packssdw %%mm1 , %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2354 "packssdw %%mm3 , %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2355 "movq %%mm0 , (%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2356 "movq %%mm2 , 8(%1,%0) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2357 "add $16 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2358 " js 1b \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2359 "emms \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2360 :"+r"(reglen), "+r"(dst), "+r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2361 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2362 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2363
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2364 static void float_to_int16_sse2(int16_t *dst, const float *src, long len){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2365 x86_reg reglen = len;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2366 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2367 "add %0 , %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2368 "lea (%2,%0,2) , %2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2369 "add %0 , %1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2370 "neg %0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2371 "1: \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2372 "cvtps2dq (%2,%0,2) , %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2373 "cvtps2dq 16(%2,%0,2) , %%xmm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2374 "packssdw %%xmm1 , %%xmm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2375 "movdqa %%xmm0 , (%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 :"+r"(reglen), "+r"(dst), "+r"(src)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2379 );
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 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
2383 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
2384 void ff_float_to_int16_interleave6_3dn2(int16_t *dst, const float **src, int len);
10644
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2385 int32_t ff_scalarproduct_int16_mmx2(int16_t *v1, int16_t *v2, int order, int shift);
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2386 int32_t ff_scalarproduct_int16_sse2(int16_t *v1, int16_t *v2, int order, int shift);
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2387 int32_t ff_scalarproduct_and_madd_int16_mmx2(int16_t *v1, int16_t *v2, int16_t *v3, int order, int mul);
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2388 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, int16_t *v2, int16_t *v3, int order, int mul);
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2389 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, int16_t *v2, int16_t *v3, int order, int mul);
10431
546b7ebeaf07 huffyuv: add some const qualifiers
lorenm
parents: 10430
diff changeset
2390 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
2391 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
2392 int ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src, int w, int left);
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2393 void ff_x264_deblock_v_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2394 void ff_x264_deblock_h_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0);
10645
6f958f237d7d r20739 broke compilation on systems without yasm
lorenm
parents: 10644
diff changeset
2395 void ff_x264_deblock_h_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta);
6f958f237d7d r20739 broke compilation on systems without yasm
lorenm
parents: 10644
diff changeset
2396 void ff_x264_deblock_v_luma_intra_sse2(uint8_t *pix, int stride, int alpha, int beta);
6f958f237d7d r20739 broke compilation on systems without yasm
lorenm
parents: 10644
diff changeset
2397 void ff_x264_deblock_h_luma_intra_sse2(uint8_t *pix, int stride, int alpha, int beta);
6f958f237d7d r20739 broke compilation on systems without yasm
lorenm
parents: 10644
diff changeset
2398
6f958f237d7d r20739 broke compilation on systems without yasm
lorenm
parents: 10644
diff changeset
2399 #if HAVE_YASM && ARCH_X86_32
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2400 void ff_x264_deblock_v8_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2401 static void ff_x264_deblock_v_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2402 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2403 ff_x264_deblock_v8_luma_intra_mmxext(pix+0, stride, alpha, beta);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2404 ff_x264_deblock_v8_luma_intra_mmxext(pix+8, stride, alpha, beta);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2405 }
10645
6f958f237d7d r20739 broke compilation on systems without yasm
lorenm
parents: 10644
diff changeset
2406 #elif !HAVE_YASM
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2407 #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
2408 #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
2409 #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
2410 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2411 #define ff_float_to_int16_interleave6_sse2 ff_float_to_int16_interleave6_sse
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2412
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2413 #define FLOAT_TO_INT16_INTERLEAVE(cpu, body) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2414 /* 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
2415 static av_noinline void float_to_int16_interleave_misc_##cpu(int16_t *dst, const float **src, long len, int channels){\
10961
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10766
diff changeset
2416 DECLARE_ALIGNED_16(int16_t, tmp)[len];\
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2417 int i,j,c;\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2418 for(c=0; c<channels; c++){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2419 float_to_int16_##cpu(tmp, src[c], len);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2420 for(i=0, j=c; i<len; i++, j+=channels)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2421 dst[j] = tmp[i];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2422 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2423 }\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2424 \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2425 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
2426 if(channels==1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2427 float_to_int16_##cpu(dst, src[0], len);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2428 else if(channels==2){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2429 x86_reg reglen = len; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2430 const float *src0 = src[0];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2431 const float *src1 = src[1];\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2432 __asm__ volatile(\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2433 "shl $2, %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2434 "add %0, %1 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2435 "add %0, %2 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2436 "add %0, %3 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2437 "neg %0 \n"\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2438 body\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2439 :"+r"(reglen), "+r"(dst), "+r"(src0), "+r"(src1)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2440 );\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2441 }else if(channels==6){\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2442 ff_float_to_int16_interleave6_##cpu(dst, src, len);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2443 }else\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2444 float_to_int16_interleave_misc_##cpu(dst, src, len, channels);\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2445 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2446
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2447 FLOAT_TO_INT16_INTERLEAVE(3dnow,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2448 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2449 "pf2id (%2,%0), %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2450 "pf2id 8(%2,%0), %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2451 "pf2id (%3,%0), %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2452 "pf2id 8(%3,%0), %%mm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2453 "packssdw %%mm1, %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2454 "packssdw %%mm3, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2455 "movq %%mm0, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2456 "punpcklwd %%mm2, %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2457 "punpckhwd %%mm2, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2458 "movq %%mm0, (%1,%0)\n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2459 "movq %%mm1, 8(%1,%0)\n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2460 "add $16, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2461 "js 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2462 "femms \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2463 )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2464
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2465 FLOAT_TO_INT16_INTERLEAVE(sse,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2466 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2467 "cvtps2pi (%2,%0), %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2468 "cvtps2pi 8(%2,%0), %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2469 "cvtps2pi (%3,%0), %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2470 "cvtps2pi 8(%3,%0), %%mm3 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2471 "packssdw %%mm1, %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2472 "packssdw %%mm3, %%mm2 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2473 "movq %%mm0, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2474 "punpcklwd %%mm2, %%mm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2475 "punpckhwd %%mm2, %%mm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2476 "movq %%mm0, (%1,%0)\n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2477 "movq %%mm1, 8(%1,%0)\n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2478 "add $16, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2479 "js 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2480 "emms \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2481 )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2482
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2483 FLOAT_TO_INT16_INTERLEAVE(sse2,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2484 "1: \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2485 "cvtps2dq (%2,%0), %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2486 "cvtps2dq (%3,%0), %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2487 "packssdw %%xmm1, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2488 "movhlps %%xmm0, %%xmm1 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2489 "punpcklwd %%xmm1, %%xmm0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2490 "movdqa %%xmm0, (%1,%0) \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2491 "add $16, %0 \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2492 "js 1b \n"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2493 )
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2494
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2495 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
2496 if(channels==6)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2497 ff_float_to_int16_interleave6_3dn2(dst, src, len);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2498 else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2499 float_to_int16_interleave_3dnow(dst, src, len, channels);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2500 }
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 void ff_snow_horizontal_compose97i_sse2(IDWTELEM *b, int width);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2504 void ff_snow_horizontal_compose97i_mmx(IDWTELEM *b, int width);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2505 void ff_snow_vertical_compose97i_sse2(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2506 void ff_snow_vertical_compose97i_mmx(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2507 void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2508 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2509 void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2510 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2511
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2512
10964
abb3b23bda35 Implement an sse version of scalarproduct_float().
alexc
parents: 10961
diff changeset
2513 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
2514
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2515 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2516 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2517 mm_flags = mm_support();
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2518
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2519 if (avctx->dsp_mask) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2520 if (avctx->dsp_mask & FF_MM_FORCE)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2521 mm_flags |= (avctx->dsp_mask & 0xffff);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2522 else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2523 mm_flags &= ~(avctx->dsp_mask & 0xffff);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2524 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2525
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2526 #if 0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2527 av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2528 if (mm_flags & FF_MM_MMX)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2529 av_log(avctx, AV_LOG_INFO, " mmx");
9342
7f594601d5e9 Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents: 9341
diff changeset
2530 if (mm_flags & FF_MM_MMX2)
7f594601d5e9 Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents: 9341
diff changeset
2531 av_log(avctx, AV_LOG_INFO, " mmx2");
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2532 if (mm_flags & FF_MM_3DNOW)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2533 av_log(avctx, AV_LOG_INFO, " 3dnow");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2534 if (mm_flags & FF_MM_SSE)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2535 av_log(avctx, AV_LOG_INFO, " sse");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2536 if (mm_flags & FF_MM_SSE2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2537 av_log(avctx, AV_LOG_INFO, " sse2");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2538 av_log(avctx, AV_LOG_INFO, "\n");
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2539 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2540
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2541 if (mm_flags & FF_MM_MMX) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2542 const int idct_algo= avctx->idct_algo;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2543
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2544 if(avctx->lowres==0){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2545 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2546 c->idct_put= ff_simple_idct_put_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2547 c->idct_add= ff_simple_idct_add_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2548 c->idct = ff_simple_idct_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2549 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM;
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2550 #if CONFIG_GPL
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2551 }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){
9342
7f594601d5e9 Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents: 9341
diff changeset
2552 if(mm_flags & FF_MM_MMX2){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2553 c->idct_put= ff_libmpeg2mmx2_idct_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2554 c->idct_add= ff_libmpeg2mmx2_idct_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2555 c->idct = ff_mmxext_idct;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2556 }else{
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2557 c->idct_put= ff_libmpeg2mmx_idct_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2558 c->idct_add= ff_libmpeg2mmx_idct_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2559 c->idct = ff_mmx_idct;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2560 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2561 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2562 #endif
9975
d6d7e8d4a04d Do not redundantly check for both CONFIG_THEORA_DECODER and CONFIG_VP3_DECODER.
diego
parents: 9959
diff changeset
2563 }else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER) &&
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2564 idct_algo==FF_IDCT_VP3){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2565 if(mm_flags & FF_MM_SSE2){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2566 c->idct_put= ff_vp3_idct_put_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2567 c->idct_add= ff_vp3_idct_add_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2568 c->idct = ff_vp3_idct_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2569 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2570 }else{
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2571 c->idct_put= ff_vp3_idct_put_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2572 c->idct_add= ff_vp3_idct_add_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2573 c->idct = ff_vp3_idct_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2574 c->idct_permutation_type= FF_PARTTRANS_IDCT_PERM;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2575 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2576 }else if(idct_algo==FF_IDCT_CAVS){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2577 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2578 }else if(idct_algo==FF_IDCT_XVIDMMX){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2579 if(mm_flags & FF_MM_SSE2){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2580 c->idct_put= ff_idct_xvid_sse2_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2581 c->idct_add= ff_idct_xvid_sse2_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2582 c->idct = ff_idct_xvid_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2583 c->idct_permutation_type= FF_SSE2_IDCT_PERM;
9342
7f594601d5e9 Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents: 9341
diff changeset
2584 }else if(mm_flags & FF_MM_MMX2){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2585 c->idct_put= ff_idct_xvid_mmx2_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2586 c->idct_add= ff_idct_xvid_mmx2_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2587 c->idct = ff_idct_xvid_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2588 }else{
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2589 c->idct_put= ff_idct_xvid_mmx_put;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2590 c->idct_add= ff_idct_xvid_mmx_add;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2591 c->idct = ff_idct_xvid_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2592 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2593 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2594 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2595
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2596 c->put_pixels_clamped = put_pixels_clamped_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2597 c->put_signed_pixels_clamped = put_signed_pixels_clamped_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2598 c->add_pixels_clamped = add_pixels_clamped_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2599 c->clear_block = clear_block_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2600 c->clear_blocks = clear_blocks_mmx;
10766
78c2be62260a Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents: 10749
diff changeset
2601 if ((mm_flags & FF_MM_SSE) &&
78c2be62260a Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents: 10749
diff changeset
2602 !(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)){
78c2be62260a Fix XvMC. XvMCCreateBlocks() may not allocate 16-byte aligned blocks,
gb
parents: 10749
diff changeset
2603 /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
9861
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
2604 c->clear_block = clear_block_sse;
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
2605 c->clear_blocks = clear_blocks_sse;
89270a3bc4a0 SSE version of clear_blocks
darkshikari
parents: 9445
diff changeset
2606 }
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2607
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2608 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2609 c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2610 c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2611 c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_ ## CPU; \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2612 c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2613
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2614 SET_HPEL_FUNCS(put, 0, 16, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2615 SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2616 SET_HPEL_FUNCS(avg, 0, 16, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2617 SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2618 SET_HPEL_FUNCS(put, 1, 8, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2619 SET_HPEL_FUNCS(put_no_rnd, 1, 8, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2620 SET_HPEL_FUNCS(avg, 1, 8, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2621 SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2622
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2623 c->gmc= gmc_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2624
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2625 c->add_bytes= add_bytes_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2626 c->add_bytes_l2= add_bytes_l2_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2627
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2628 c->draw_edges = draw_edges_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2629
10749
5cca4b6c459d Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents: 10645
diff changeset
2630 if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2631 c->h263_v_loop_filter= h263_v_loop_filter_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2632 c->h263_h_loop_filter= h263_h_loop_filter_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2633 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2634 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_mmx_rnd;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2635 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_mmx;
9439
ef3a7b711cc0 Rename put_no_rnd_h264_chroma* to reflect its usage in VC1 only
conrad
parents: 9342
diff changeset
2636 c->put_no_rnd_vc1_chroma_pixels_tab[0]= put_vc1_chroma_mc8_mmx_nornd;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2637
8519
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2638 c->put_rv40_chroma_pixels_tab[0]= put_rv40_chroma_mc8_mmx;
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2639 c->put_rv40_chroma_pixels_tab[1]= put_rv40_chroma_mc4_mmx;
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2640
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2641 c->h264_idct_dc_add=
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2642 c->h264_idct_add= ff_h264_idct_add_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2643 c->h264_idct8_dc_add=
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2644 c->h264_idct8_add= ff_h264_idct8_add_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2645
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2646 c->h264_idct_add16 = ff_h264_idct_add16_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2647 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2648 c->h264_idct_add8 = ff_h264_idct_add8_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2649 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2650
8817
492f8911992c add MMX version of vp6_filter_diag
aurel
parents: 8816
diff changeset
2651 if (CONFIG_VP6_DECODER) {
492f8911992c add MMX version of vp6_filter_diag
aurel
parents: 8816
diff changeset
2652 c->vp6_filter_diag4 = ff_vp6_filter_diag4_mmx;
492f8911992c add MMX version of vp6_filter_diag
aurel
parents: 8816
diff changeset
2653 }
492f8911992c add MMX version of vp6_filter_diag
aurel
parents: 8816
diff changeset
2654
9342
7f594601d5e9 Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents: 9341
diff changeset
2655 if (mm_flags & FF_MM_MMX2) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2656 c->prefetch = prefetch_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2657
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2658 c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2659 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2660
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2661 c->avg_pixels_tab[0][0] = avg_pixels16_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2662 c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2663 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2664
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2665 c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2666 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2667
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2668 c->avg_pixels_tab[1][0] = avg_pixels8_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2669 c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2670 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2671
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2672 c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2673 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2674 c->h264_idct_add16 = ff_h264_idct_add16_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2675 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2676 c->h264_idct_add8 = ff_h264_idct_add8_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2677 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2678
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2679 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2680 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2681 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2682 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2683 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2684 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2685 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2686
9975
d6d7e8d4a04d Do not redundantly check for both CONFIG_THEORA_DECODER and CONFIG_VP3_DECODER.
diego
parents: 9959
diff changeset
2687 if (CONFIG_VP3_DECODER) {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2688 c->vp3_v_loop_filter= ff_vp3_v_loop_filter_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2689 c->vp3_h_loop_filter= ff_vp3_h_loop_filter_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2690 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2691 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2692
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
8519
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2730 c->avg_rv40_chroma_pixels_tab[0]= avg_rv40_chroma_mc8_mmx2;
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2731 c->avg_rv40_chroma_pixels_tab[1]= avg_rv40_chroma_mc4_mmx2;
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2732
9440
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
2733 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= avg_vc1_chroma_mc8_mmx2_nornd;
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
2734
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2735 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_mmx2_rnd;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2736 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2737 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2738 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2739 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2740 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2741 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2742 c->h264_h_loop_filter_chroma= h264_h_loop_filter_chroma_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2743 c->h264_v_loop_filter_chroma_intra= h264_v_loop_filter_chroma_intra_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2744 c->h264_h_loop_filter_chroma_intra= h264_h_loop_filter_chroma_intra_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2745 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2746
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2747 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2748 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2749 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2750 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2751 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2752 c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2753 c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2754 c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2755
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2756 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2757 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2758 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2759 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2760 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2761 c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2762 c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2763 c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2764
8760
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2765 #if HAVE_YASM
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2766 c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmx2;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2767 #endif
8798
a5c8210814d7 Add check whether the compiler/assembler supports 10 or more operands.
diego
parents: 8760
diff changeset
2768 #if HAVE_7REGS && HAVE_TEN_OPERANDS
8760
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2769 if( mm_flags&FF_MM_3DNOW )
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2770 c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2771 #endif
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8668
diff changeset
2772
8596
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
2773 if (CONFIG_CAVS_DECODER)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2774 ff_cavsdsp_init_mmx2(c, avctx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2775
9995
3141f69e3905 Do not check for both CONFIG_VC1_DECODER and CONFIG_WMV3_DECODER,
diego
parents: 9975
diff changeset
2776 if (CONFIG_VC1_DECODER)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2777 ff_vc1dsp_init_mmx(c, avctx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2778
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2779 c->add_png_paeth_prediction= add_png_paeth_prediction_mmx2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2780 } else if (mm_flags & FF_MM_3DNOW) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2781 c->prefetch = prefetch_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2782
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2783 c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2784 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2785
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2786 c->avg_pixels_tab[0][0] = avg_pixels16_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2787 c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2788 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2789
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2790 c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2791 c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2792
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2793 c->avg_pixels_tab[1][0] = avg_pixels8_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2794 c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2795 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2796
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2797 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2798 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2799 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2800 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2801 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2802 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2803 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2804 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2805
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2806 SET_QPEL_FUNCS(put_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2807 SET_QPEL_FUNCS(put_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2808 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2809 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2810 SET_QPEL_FUNCS(avg_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2811 SET_QPEL_FUNCS(avg_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2812
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2813 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2814 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2815 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2816 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2817 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2818 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2819
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2820 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2821 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2822 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2823 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, 3dnow);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2824
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2825 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_3dnow_rnd;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2826 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2827
8519
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2828 c->avg_rv40_chroma_pixels_tab[0]= avg_rv40_chroma_mc8_3dnow;
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2829 c->avg_rv40_chroma_pixels_tab[1]= avg_rv40_chroma_mc4_3dnow;
cc64e1343397 Use H264 MMX chroma functions to accelerate RV40 decoding.
cehoyos
parents: 8510
diff changeset
2830
8596
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
2831 if (CONFIG_CAVS_DECODER)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2832 ff_cavsdsp_init_3dnow(c, avctx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2833 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2834
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2835
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2836 #define H264_QPEL_FUNCS(x, y, CPU)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2837 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
2838 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
2839 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
2840 c->avg_h264_qpel_pixels_tab[1][x+y*4] = avg_h264_qpel8_mc##x##y##_##CPU;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2841 if((mm_flags & FF_MM_SSE2) && !(mm_flags & FF_MM_3DNOW)){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2842 // these functions are slower than mmx on AMD, but faster on Intel
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2843 c->put_pixels_tab[0][0] = put_pixels16_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2844 c->avg_pixels_tab[0][0] = avg_pixels16_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2845 H264_QPEL_FUNCS(0, 0, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2846 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2847 if(mm_flags & FF_MM_SSE2){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2848 c->h264_idct8_add = ff_h264_idct8_add_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2849 c->h264_idct8_add4= ff_h264_idct8_add4_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2850
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2851 H264_QPEL_FUNCS(0, 1, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2852 H264_QPEL_FUNCS(0, 2, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2853 H264_QPEL_FUNCS(0, 3, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2854 H264_QPEL_FUNCS(1, 1, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2855 H264_QPEL_FUNCS(1, 2, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2856 H264_QPEL_FUNCS(1, 3, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2857 H264_QPEL_FUNCS(2, 1, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2858 H264_QPEL_FUNCS(2, 2, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2859 H264_QPEL_FUNCS(2, 3, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2860 H264_QPEL_FUNCS(3, 1, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2861 H264_QPEL_FUNCS(3, 2, sse2);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2862 H264_QPEL_FUNCS(3, 3, sse2);
8818
17cc6df384a6 add SSE2 version of vp6_filter_diag
aurel
parents: 8817
diff changeset
2863
17cc6df384a6 add SSE2 version of vp6_filter_diag
aurel
parents: 8817
diff changeset
2864 if (CONFIG_VP6_DECODER) {
17cc6df384a6 add SSE2 version of vp6_filter_diag
aurel
parents: 8817
diff changeset
2865 c->vp6_filter_diag4 = ff_vp6_filter_diag4_sse2;
17cc6df384a6 add SSE2 version of vp6_filter_diag
aurel
parents: 8817
diff changeset
2866 }
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2867 }
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2868 #if HAVE_SSSE3
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2869 if(mm_flags & FF_MM_SSSE3){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2870 H264_QPEL_FUNCS(1, 0, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2871 H264_QPEL_FUNCS(1, 1, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2872 H264_QPEL_FUNCS(1, 2, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2873 H264_QPEL_FUNCS(1, 3, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2874 H264_QPEL_FUNCS(2, 0, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2875 H264_QPEL_FUNCS(2, 1, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2876 H264_QPEL_FUNCS(2, 2, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2877 H264_QPEL_FUNCS(2, 3, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2878 H264_QPEL_FUNCS(3, 0, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2879 H264_QPEL_FUNCS(3, 1, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2880 H264_QPEL_FUNCS(3, 2, ssse3);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2881 H264_QPEL_FUNCS(3, 3, ssse3);
9439
ef3a7b711cc0 Rename put_no_rnd_h264_chroma* to reflect its usage in VC1 only
conrad
parents: 9342
diff changeset
2882 c->put_no_rnd_vc1_chroma_pixels_tab[0]= put_vc1_chroma_mc8_ssse3_nornd;
9440
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
2883 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= avg_vc1_chroma_mc8_ssse3_nornd;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2884 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_ssse3_rnd;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2885 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_ssse3_rnd;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2886 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_ssse3;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2887 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_ssse3;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2888 c->add_png_paeth_prediction= add_png_paeth_prediction_ssse3;
10430
12c8175d6db5 simd add_hfyu_left_prediction
lorenm
parents: 10301
diff changeset
2889 #if HAVE_YASM
12c8175d6db5 simd add_hfyu_left_prediction
lorenm
parents: 10301
diff changeset
2890 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
12c8175d6db5 simd add_hfyu_left_prediction
lorenm
parents: 10301
diff changeset
2891 if (mm_flags & FF_MM_SSE4) // not really sse4, just slow on Conroe
12c8175d6db5 simd add_hfyu_left_prediction
lorenm
parents: 10301
diff changeset
2892 c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
12c8175d6db5 simd add_hfyu_left_prediction
lorenm
parents: 10301
diff changeset
2893 #endif
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2894 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2895 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2896
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2897 #if CONFIG_GPL && HAVE_YASM
9342
7f594601d5e9 Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents: 9341
diff changeset
2898 if (mm_flags & FF_MM_MMX2){
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2899 #if ARCH_X86_32
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2900 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_mmxext;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2901 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_mmxext;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2902 #endif
8510
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents: 8499
diff changeset
2903 if( mm_flags&FF_MM_SSE2 ){
9959
6972b493f41f Icc 11.1 still does not align the stack pointer, disable some x264 functions.
cehoyos
parents: 9861
diff changeset
2904 #if ARCH_X86_64 || !defined(__ICC) || __ICC > 1110
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2905 c->h264_v_loop_filter_luma = ff_x264_deblock_v_luma_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2906 c->h264_h_loop_filter_luma = ff_x264_deblock_h_luma_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2907 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2908 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_sse2;
8510
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents: 8499
diff changeset
2909 #endif
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents: 8499
diff changeset
2910 c->h264_idct_add16 = ff_h264_idct_add16_sse2;
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents: 8499
diff changeset
2911 c->h264_idct_add8 = ff_h264_idct_add8_sse2;
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents: 8499
diff changeset
2912 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2913 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2914 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2915 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2916
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2917 #if CONFIG_SNOW_DECODER
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2918 if(mm_flags & FF_MM_SSE2 & 0){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2919 c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2;
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2920 #if HAVE_7REGS
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2921 c->vertical_compose97i = ff_snow_vertical_compose97i_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2922 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2923 c->inner_add_yblock = ff_snow_inner_add_yblock_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2924 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2925 else{
9342
7f594601d5e9 Rename FF_MM_MMXEXT to FF_MM_MMX2, for both clarity and consistency
stefano
parents: 9341
diff changeset
2926 if(mm_flags & FF_MM_MMX2){
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2927 c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx;
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8527
diff changeset
2928 #if HAVE_7REGS
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2929 c->vertical_compose97i = ff_snow_vertical_compose97i_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2930 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2931 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2932 c->inner_add_yblock = ff_snow_inner_add_yblock_mmx;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2933 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2934 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2935
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2936 if(mm_flags & FF_MM_3DNOW){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2937 c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2938 c->vector_fmul = vector_fmul_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2939 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2940 c->float_to_int16 = float_to_int16_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2941 c->float_to_int16_interleave = float_to_int16_interleave_3dnow;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2942 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2943 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2944 if(mm_flags & FF_MM_3DNOWEXT){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2945 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2946 c->vector_fmul_window = vector_fmul_window_3dnow2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2947 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2948 c->float_to_int16_interleave = float_to_int16_interleave_3dn2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2949 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2950 }
10633
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2951 if(mm_flags & FF_MM_MMX2){
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2952 #if HAVE_YASM
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2953 c->scalarproduct_int16 = ff_scalarproduct_int16_mmx2;
10644
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2954 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
2955 #endif
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2956 }
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2957 if(mm_flags & FF_MM_SSE){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2958 c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2959 c->ac3_downmix = ac3_downmix_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2960 c->vector_fmul = vector_fmul_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2961 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
2962 c->vector_fmul_add = vector_fmul_add_sse;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2963 c->vector_fmul_window = vector_fmul_window_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2964 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
2965 c->vector_clipf = vector_clipf_sse;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2966 c->float_to_int16 = float_to_int16_sse;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2967 c->float_to_int16_interleave = float_to_int16_interleave_sse;
10964
abb3b23bda35 Implement an sse version of scalarproduct_float().
alexc
parents: 10961
diff changeset
2968 #if HAVE_YASM
abb3b23bda35 Implement an sse version of scalarproduct_float().
alexc
parents: 10961
diff changeset
2969 c->scalarproduct_float = ff_scalarproduct_float_sse;
abb3b23bda35 Implement an sse version of scalarproduct_float().
alexc
parents: 10961
diff changeset
2970 #endif
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2971 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2972 if(mm_flags & FF_MM_3DNOW)
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10107
diff changeset
2973 c->vector_fmul_add = vector_fmul_add_3dnow; // faster than sse
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2974 if(mm_flags & FF_MM_SSE2){
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2975 c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2976 c->float_to_int16 = float_to_int16_sse2;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2977 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
2978 #if HAVE_YASM
66242b8fbd32 port ape dsp functions from sse2 to mmx
lorenm
parents: 10431
diff changeset
2979 c->scalarproduct_int16 = ff_scalarproduct_int16_sse2;
10644
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2980 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
2981 #endif
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2982 }
10644
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2983 if((mm_flags & FF_MM_SSSE3) && !(mm_flags & (FF_MM_SSE42|FF_MM_3DNOW)) && HAVE_YASM) // cachesplit
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10633
diff changeset
2984 c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2985 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2986
8596
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
2987 if (CONFIG_ENCODERS)
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2988 dsputilenc_init_mmx(c, avctx);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2989
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2990 #if 0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2991 // for speed testing
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2992 get_pixels = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2993 put_pixels_clamped = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2994 add_pixels_clamped = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2995
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2996 pix_abs16x16 = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2997 pix_abs16x16_x2 = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2998 pix_abs16x16_y2 = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2999 pix_abs16x16_xy2 = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3000
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3001 put_pixels_tab[0] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3002 put_pixels_tab[1] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3003 put_pixels_tab[2] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3004 put_pixels_tab[3] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3005
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3006 put_no_rnd_pixels_tab[0] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3007 put_no_rnd_pixels_tab[1] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3008 put_no_rnd_pixels_tab[2] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3009 put_no_rnd_pixels_tab[3] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3010
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3011 avg_pixels_tab[0] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3012 avg_pixels_tab[1] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3013 avg_pixels_tab[2] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3014 avg_pixels_tab[3] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3015
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3016 avg_no_rnd_pixels_tab[0] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3017 avg_no_rnd_pixels_tab[1] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3018 avg_no_rnd_pixels_tab[2] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3019 avg_no_rnd_pixels_tab[3] = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3020
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3021 //av_fdct = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3022 //ff_idct = just_return;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3023 #endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3024 }