Mercurial > libavcodec.hg
annotate i386/dsputil_mmx.c @ 6234:5d82b4e8a7f3 libavcodec
const
author | michael |
---|---|
date | Fri, 01 Feb 2008 04:14:04 +0000 |
parents | 166bef5cad01 |
children | ffb2a7b80d6d |
rev | line source |
---|---|
0 | 1 /* |
2 * MMX optimized DSP utils | |
429 | 3 * Copyright (c) 2000, 2001 Fabrice Bellard. |
1739
07a484280a82
copyright year update of the files i touched and remembered, things look annoyingly unmaintained otherwise
michael
parents:
1729
diff
changeset
|
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at> |
0 | 5 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3932
diff
changeset
|
6 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3932
diff
changeset
|
7 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3932
diff
changeset
|
8 * FFmpeg is free software; you can redistribute it and/or |
429 | 9 * modify it under the terms of the GNU Lesser General Public |
10 * License as published by the Free Software Foundation; either | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3932
diff
changeset
|
11 * version 2.1 of the License, or (at your option) any later version. |
0 | 12 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3932
diff
changeset
|
13 * FFmpeg is distributed in the hope that it will be useful, |
0 | 14 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
429 | 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
16 * Lesser General Public License for more details. | |
0 | 17 * |
429 | 18 * You should have received a copy of the GNU Lesser General Public |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3932
diff
changeset
|
19 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
0 | 21 * |
22 * MMX optimization by Nick Kurshev <nickols_k@mail.ru> | |
23 */ | |
24 | |
5010
d5ba514e3f4a
Add libavcodec to compiler include flags in order to simplify header
diego
parents:
5007
diff
changeset
|
25 #include "dsputil.h" |
5946
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
26 #include "dsputil_mmx.h" |
5010
d5ba514e3f4a
Add libavcodec to compiler include flags in order to simplify header
diego
parents:
5007
diff
changeset
|
27 #include "simple_idct.h" |
d5ba514e3f4a
Add libavcodec to compiler include flags in order to simplify header
diego
parents:
5007
diff
changeset
|
28 #include "mpegvideo.h" |
3398
e0927bc44a10
Move REG_* macros from libavcodec/i386/mmx.h to libavutil/x86_cpu.h
lucabe
parents:
3250
diff
changeset
|
29 #include "x86_cpu.h" |
1984
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
30 #include "mmx.h" |
5014
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
31 #include "vp3dsp_mmx.h" |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
32 #include "vp3dsp_sse2.h" |
5277
7b3fcb7c61ce
Avoid linking with h263.c functions when the relevant codecs
aurel
parents:
5255
diff
changeset
|
33 #include "h263.h" |
0 | 34 |
1729 | 35 //#undef NDEBUG |
36 //#include <assert.h> | |
37 | |
2868 | 38 extern void ff_idct_xvid_mmx(short *block); |
39 extern void ff_idct_xvid_mmx2(short *block); | |
1647 | 40 |
4197 | 41 int mm_flags; /* multimedia extension flags */ |
936 | 42 |
0 | 43 /* pixel operations */ |
5947 | 44 DECLARE_ALIGNED_8 (const uint64_t, ff_bone) = 0x0101010101010101ULL; |
45 DECLARE_ALIGNED_8 (const uint64_t, ff_wtwo) = 0x0002000200020002ULL; | |
5946
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
46 |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
47 DECLARE_ALIGNED_16(const uint64_t, ff_pdw_80000000[2]) = |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
48 {0x8000000080000000ULL, 0x8000000080000000ULL}; |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
49 |
5946
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
50 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_3 ) = 0x0003000300030003ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
51 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_4 ) = 0x0004000400040004ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
52 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_5 ) = 0x0005000500050005ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
53 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_8 ) = 0x0008000800080008ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
54 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_15 ) = 0x000F000F000F000FULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
55 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_16 ) = 0x0010001000100010ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
56 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_20 ) = 0x0014001400140014ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
57 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_32 ) = 0x0020002000200020ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
58 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_42 ) = 0x002A002A002A002AULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
59 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_64 ) = 0x0040004000400040ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
60 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_96 ) = 0x0060006000600060ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
61 DECLARE_ALIGNED_16(const uint64_t, ff_pw_128) = 0x0080008000800080ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
62 |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
63 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_1 ) = 0x0101010101010101ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
64 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_3 ) = 0x0303030303030303ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
65 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_7 ) = 0x0707070707070707ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
66 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_3F ) = 0x3F3F3F3F3F3F3F3FULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
67 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_A1 ) = 0xA1A1A1A1A1A1A1A1ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
68 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_FC ) = 0xFCFCFCFCFCFCFCFCULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
69 |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
70 DECLARE_ALIGNED_16(const double, ff_pd_1[2]) = { 1.0, 1.0 }; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
71 DECLARE_ALIGNED_16(const double, ff_pd_2[2]) = { 2.0, 2.0 }; |
5737 | 72 |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
73 #define JUMPALIGN() __asm __volatile (ASMALIGN(3)::) |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
74 #define MOVQ_ZERO(regd) __asm __volatile ("pxor %%" #regd ", %%" #regd ::) |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
75 |
448 | 76 #define MOVQ_WONE(regd) \ |
77 __asm __volatile ( \ | |
78 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ | |
79 "psrlw $15, %%" #regd ::) | |
80 | |
81 #define MOVQ_BFE(regd) \ | |
82 __asm __volatile ( \ | |
83 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\ | |
84 "paddb %%" #regd ", %%" #regd " \n\t" ::) | |
85 | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
86 #ifndef PIC |
5947 | 87 #define MOVQ_BONE(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_bone)) |
88 #define MOVQ_WTWO(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_wtwo)) | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
89 #else |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
90 // for shared library it's better to use this way for accessing constants |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
91 // pcmpeqd -> -1 |
448 | 92 #define MOVQ_BONE(regd) \ |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
93 __asm __volatile ( \ |
448 | 94 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
95 "psrlw $15, %%" #regd " \n\t" \ | |
96 "packuswb %%" #regd ", %%" #regd " \n\t" ::) | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
97 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
98 #define MOVQ_WTWO(regd) \ |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
99 __asm __volatile ( \ |
448 | 100 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
101 "psrlw $15, %%" #regd " \n\t" \ | |
102 "psllw $1, %%" #regd " \n\t"::) | |
387 | 103 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
104 #endif |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
105 |
448 | 106 // using regr as temporary and for the output result |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
107 // first argument is unmodifed and second is trashed |
471 | 108 // regfe is supposed to contain 0xfefefefefefefefe |
109 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \ | |
2979 | 110 "movq " #rega ", " #regr " \n\t"\ |
111 "pand " #regb ", " #regr " \n\t"\ | |
112 "pxor " #rega ", " #regb " \n\t"\ | |
113 "pand " #regfe "," #regb " \n\t"\ | |
114 "psrlq $1, " #regb " \n\t"\ | |
115 "paddb " #regb ", " #regr " \n\t" | |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
116 |
471 | 117 #define PAVGB_MMX(rega, regb, regr, regfe) \ |
2979 | 118 "movq " #rega ", " #regr " \n\t"\ |
119 "por " #regb ", " #regr " \n\t"\ | |
120 "pxor " #rega ", " #regb " \n\t"\ | |
121 "pand " #regfe "," #regb " \n\t"\ | |
122 "psrlq $1, " #regb " \n\t"\ | |
123 "psubb " #regb ", " #regr " \n\t" | |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
124 |
471 | 125 // mm6 is supposed to contain 0xfefefefefefefefe |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
126 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \ |
2979 | 127 "movq " #rega ", " #regr " \n\t"\ |
128 "movq " #regc ", " #regp " \n\t"\ | |
129 "pand " #regb ", " #regr " \n\t"\ | |
130 "pand " #regd ", " #regp " \n\t"\ | |
131 "pxor " #rega ", " #regb " \n\t"\ | |
132 "pxor " #regc ", " #regd " \n\t"\ | |
133 "pand %%mm6, " #regb " \n\t"\ | |
134 "pand %%mm6, " #regd " \n\t"\ | |
135 "psrlq $1, " #regb " \n\t"\ | |
136 "psrlq $1, " #regd " \n\t"\ | |
137 "paddb " #regb ", " #regr " \n\t"\ | |
138 "paddb " #regd ", " #regp " \n\t" | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
139 |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
140 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \ |
2979 | 141 "movq " #rega ", " #regr " \n\t"\ |
142 "movq " #regc ", " #regp " \n\t"\ | |
143 "por " #regb ", " #regr " \n\t"\ | |
144 "por " #regd ", " #regp " \n\t"\ | |
145 "pxor " #rega ", " #regb " \n\t"\ | |
146 "pxor " #regc ", " #regd " \n\t"\ | |
147 "pand %%mm6, " #regb " \n\t"\ | |
148 "pand %%mm6, " #regd " \n\t"\ | |
149 "psrlq $1, " #regd " \n\t"\ | |
150 "psrlq $1, " #regb " \n\t"\ | |
151 "psubb " #regb ", " #regr " \n\t"\ | |
152 "psubb " #regd ", " #regp " \n\t" | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
153 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
154 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
155 /* MMX no rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
156 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx |
448 | 157 #define SET_RND MOVQ_WONE |
2979 | 158 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX_NO_RND(a, b, c, d, e, f) |
159 #define PAVGB(a, b, c, e) PAVGB_MMX_NO_RND(a, b, c, e) | |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
160 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
161 #include "dsputil_mmx_rnd.h" |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
162 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
163 #undef DEF |
448 | 164 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
165 #undef PAVGBP |
471 | 166 #undef PAVGB |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
167 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
168 /* MMX rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
169 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
170 #define DEF(x, y) x ## _ ## y ##_mmx |
448 | 171 #define SET_RND MOVQ_WTWO |
2979 | 172 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f) |
173 #define PAVGB(a, b, c, e) PAVGB_MMX(a, b, c, e) | |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
174 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
175 #include "dsputil_mmx_rnd.h" |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
176 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
177 #undef DEF |
448 | 178 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
179 #undef PAVGBP |
471 | 180 #undef PAVGB |
387 | 181 |
0 | 182 /***********************************/ |
183 /* 3Dnow specific */ | |
184 | |
185 #define DEF(x) x ## _3dnow | |
186 #define PAVGB "pavgusb" | |
187 | |
188 #include "dsputil_mmx_avg.h" | |
189 | |
190 #undef DEF | |
191 #undef PAVGB | |
192 | |
193 /***********************************/ | |
194 /* MMX2 specific */ | |
195 | |
386 | 196 #define DEF(x) x ## _mmx2 |
0 | 197 |
198 /* Introduced only in MMX2 set */ | |
199 #define PAVGB "pavgb" | |
200 | |
201 #include "dsputil_mmx_avg.h" | |
202 | |
203 #undef DEF | |
204 #undef PAVGB | |
205 | |
206 /***********************************/ | |
207 /* standard MMX */ | |
208 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
209 #ifdef CONFIG_ENCODERS |
1064 | 210 static void get_pixels_mmx(DCTELEM *block, const uint8_t *pixels, int line_size) |
0 | 211 { |
386 | 212 asm volatile( |
2979 | 213 "mov $-128, %%"REG_a" \n\t" |
214 "pxor %%mm7, %%mm7 \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
215 ASMALIGN(4) |
2979 | 216 "1: \n\t" |
217 "movq (%0), %%mm0 \n\t" | |
218 "movq (%0, %2), %%mm2 \n\t" | |
219 "movq %%mm0, %%mm1 \n\t" | |
220 "movq %%mm2, %%mm3 \n\t" | |
221 "punpcklbw %%mm7, %%mm0 \n\t" | |
222 "punpckhbw %%mm7, %%mm1 \n\t" | |
223 "punpcklbw %%mm7, %%mm2 \n\t" | |
224 "punpckhbw %%mm7, %%mm3 \n\t" | |
225 "movq %%mm0, (%1, %%"REG_a") \n\t" | |
226 "movq %%mm1, 8(%1, %%"REG_a") \n\t" | |
227 "movq %%mm2, 16(%1, %%"REG_a") \n\t" | |
228 "movq %%mm3, 24(%1, %%"REG_a") \n\t" | |
229 "add %3, %0 \n\t" | |
230 "add $32, %%"REG_a" \n\t" | |
231 "js 1b \n\t" | |
386 | 232 : "+r" (pixels) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
233 : "r" (block+64), "r" ((long)line_size), "r" ((long)line_size*2) |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
234 : "%"REG_a |
386 | 235 ); |
0 | 236 } |
237 | |
1064 | 238 static inline void diff_pixels_mmx(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride) |
324 | 239 { |
240 asm volatile( | |
2979 | 241 "pxor %%mm7, %%mm7 \n\t" |
242 "mov $-128, %%"REG_a" \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
243 ASMALIGN(4) |
2979 | 244 "1: \n\t" |
245 "movq (%0), %%mm0 \n\t" | |
246 "movq (%1), %%mm2 \n\t" | |
247 "movq %%mm0, %%mm1 \n\t" | |
248 "movq %%mm2, %%mm3 \n\t" | |
249 "punpcklbw %%mm7, %%mm0 \n\t" | |
250 "punpckhbw %%mm7, %%mm1 \n\t" | |
251 "punpcklbw %%mm7, %%mm2 \n\t" | |
252 "punpckhbw %%mm7, %%mm3 \n\t" | |
253 "psubw %%mm2, %%mm0 \n\t" | |
254 "psubw %%mm3, %%mm1 \n\t" | |
255 "movq %%mm0, (%2, %%"REG_a") \n\t" | |
256 "movq %%mm1, 8(%2, %%"REG_a") \n\t" | |
257 "add %3, %0 \n\t" | |
258 "add %3, %1 \n\t" | |
259 "add $16, %%"REG_a" \n\t" | |
260 "jnz 1b \n\t" | |
324 | 261 : "+r" (s1), "+r" (s2) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
262 : "r" (block+64), "r" ((long)stride) |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
263 : "%"REG_a |
324 | 264 ); |
265 } | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
266 #endif //CONFIG_ENCODERS |
324 | 267 |
1064 | 268 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 269 { |
270 const DCTELEM *p; | |
1064 | 271 uint8_t *pix; |
0 | 272 |
273 /* read the pixels */ | |
274 p = block; | |
275 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
276 /* unrolled loop */ |
2979 | 277 __asm __volatile( |
278 "movq %3, %%mm0 \n\t" | |
279 "movq 8%3, %%mm1 \n\t" | |
280 "movq 16%3, %%mm2 \n\t" | |
281 "movq 24%3, %%mm3 \n\t" | |
282 "movq 32%3, %%mm4 \n\t" | |
283 "movq 40%3, %%mm5 \n\t" | |
284 "movq 48%3, %%mm6 \n\t" | |
285 "movq 56%3, %%mm7 \n\t" | |
286 "packuswb %%mm1, %%mm0 \n\t" | |
287 "packuswb %%mm3, %%mm2 \n\t" | |
288 "packuswb %%mm5, %%mm4 \n\t" | |
289 "packuswb %%mm7, %%mm6 \n\t" | |
290 "movq %%mm0, (%0) \n\t" | |
291 "movq %%mm2, (%0, %1) \n\t" | |
292 "movq %%mm4, (%0, %1, 2) \n\t" | |
293 "movq %%mm6, (%0, %2) \n\t" | |
294 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "m"(*p) | |
295 :"memory"); | |
0 | 296 pix += line_size*4; |
297 p += 32; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
298 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
299 // if here would be an exact copy of the code above |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
300 // compiler would generate some very strange code |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
301 // thus using "r" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
302 __asm __volatile( |
2979 | 303 "movq (%3), %%mm0 \n\t" |
304 "movq 8(%3), %%mm1 \n\t" | |
305 "movq 16(%3), %%mm2 \n\t" | |
306 "movq 24(%3), %%mm3 \n\t" | |
307 "movq 32(%3), %%mm4 \n\t" | |
308 "movq 40(%3), %%mm5 \n\t" | |
309 "movq 48(%3), %%mm6 \n\t" | |
310 "movq 56(%3), %%mm7 \n\t" | |
311 "packuswb %%mm1, %%mm0 \n\t" | |
312 "packuswb %%mm3, %%mm2 \n\t" | |
313 "packuswb %%mm5, %%mm4 \n\t" | |
314 "packuswb %%mm7, %%mm6 \n\t" | |
315 "movq %%mm0, (%0) \n\t" | |
316 "movq %%mm2, (%0, %1) \n\t" | |
317 "movq %%mm4, (%0, %1, 2) \n\t" | |
318 "movq %%mm6, (%0, %2) \n\t" | |
319 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "r"(p) | |
320 :"memory"); | |
0 | 321 } |
322 | |
3089 | 323 static DECLARE_ALIGNED_8(const unsigned char, vector128[8]) = |
1985
b2bc62fdecc0
move the 0x80 vector outside of the function, thus saving the compiler
melanson
parents:
1984
diff
changeset
|
324 { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 }; |
b2bc62fdecc0
move the 0x80 vector outside of the function, thus saving the compiler
melanson
parents:
1984
diff
changeset
|
325 |
1984
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
326 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
327 { |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
328 int i; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
329 |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
330 movq_m2r(*vector128, mm1); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
331 for (i = 0; i < 8; i++) { |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
332 movq_m2r(*(block), mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
333 packsswb_m2r(*(block + 4), mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
334 block += 8; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
335 paddb_r2r(mm1, mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
336 movq_r2m(mm0, *pixels); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
337 pixels += line_size; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
338 } |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
339 } |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
340 |
1064 | 341 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 342 { |
343 const DCTELEM *p; | |
1064 | 344 uint8_t *pix; |
0 | 345 int i; |
346 | |
347 /* read the pixels */ | |
348 p = block; | |
349 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
350 MOVQ_ZERO(mm7); |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
351 i = 4; |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
352 do { |
2979 | 353 __asm __volatile( |
354 "movq (%2), %%mm0 \n\t" | |
355 "movq 8(%2), %%mm1 \n\t" | |
356 "movq 16(%2), %%mm2 \n\t" | |
357 "movq 24(%2), %%mm3 \n\t" | |
358 "movq %0, %%mm4 \n\t" | |
359 "movq %1, %%mm6 \n\t" | |
360 "movq %%mm4, %%mm5 \n\t" | |
361 "punpcklbw %%mm7, %%mm4 \n\t" | |
362 "punpckhbw %%mm7, %%mm5 \n\t" | |
363 "paddsw %%mm4, %%mm0 \n\t" | |
364 "paddsw %%mm5, %%mm1 \n\t" | |
365 "movq %%mm6, %%mm5 \n\t" | |
366 "punpcklbw %%mm7, %%mm6 \n\t" | |
367 "punpckhbw %%mm7, %%mm5 \n\t" | |
368 "paddsw %%mm6, %%mm2 \n\t" | |
369 "paddsw %%mm5, %%mm3 \n\t" | |
370 "packuswb %%mm1, %%mm0 \n\t" | |
371 "packuswb %%mm3, %%mm2 \n\t" | |
372 "movq %%mm0, %0 \n\t" | |
373 "movq %%mm2, %1 \n\t" | |
374 :"+m"(*pix), "+m"(*(pix+line_size)) | |
375 :"r"(p) | |
376 :"memory"); | |
0 | 377 pix += line_size*2; |
378 p += 16; | |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
379 } while (--i); |
0 | 380 } |
381 | |
2209 | 382 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
383 { | |
384 __asm __volatile( | |
2979 | 385 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
386 ASMALIGN(3) |
2979 | 387 "1: \n\t" |
388 "movd (%1), %%mm0 \n\t" | |
389 "movd (%1, %3), %%mm1 \n\t" | |
390 "movd %%mm0, (%2) \n\t" | |
391 "movd %%mm1, (%2, %3) \n\t" | |
392 "add %%"REG_a", %1 \n\t" | |
393 "add %%"REG_a", %2 \n\t" | |
394 "movd (%1), %%mm0 \n\t" | |
395 "movd (%1, %3), %%mm1 \n\t" | |
396 "movd %%mm0, (%2) \n\t" | |
397 "movd %%mm1, (%2, %3) \n\t" | |
398 "add %%"REG_a", %1 \n\t" | |
399 "add %%"REG_a", %2 \n\t" | |
400 "subl $4, %0 \n\t" | |
401 "jnz 1b \n\t" | |
402 : "+g"(h), "+r" (pixels), "+r" (block) | |
403 : "r"((long)line_size) | |
404 : "%"REG_a, "memory" | |
405 ); | |
2209 | 406 } |
407 | |
1064 | 408 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 409 { |
471 | 410 __asm __volatile( |
2979 | 411 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
412 ASMALIGN(3) |
2979 | 413 "1: \n\t" |
414 "movq (%1), %%mm0 \n\t" | |
415 "movq (%1, %3), %%mm1 \n\t" | |
416 "movq %%mm0, (%2) \n\t" | |
417 "movq %%mm1, (%2, %3) \n\t" | |
418 "add %%"REG_a", %1 \n\t" | |
419 "add %%"REG_a", %2 \n\t" | |
420 "movq (%1), %%mm0 \n\t" | |
421 "movq (%1, %3), %%mm1 \n\t" | |
422 "movq %%mm0, (%2) \n\t" | |
423 "movq %%mm1, (%2, %3) \n\t" | |
424 "add %%"REG_a", %1 \n\t" | |
425 "add %%"REG_a", %2 \n\t" | |
426 "subl $4, %0 \n\t" | |
427 "jnz 1b \n\t" | |
428 : "+g"(h), "+r" (pixels), "+r" (block) | |
429 : "r"((long)line_size) | |
430 : "%"REG_a, "memory" | |
431 ); | |
0 | 432 } |
433 | |
1064 | 434 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
651 | 435 { |
436 __asm __volatile( | |
2979 | 437 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
438 ASMALIGN(3) |
2979 | 439 "1: \n\t" |
440 "movq (%1), %%mm0 \n\t" | |
441 "movq 8(%1), %%mm4 \n\t" | |
442 "movq (%1, %3), %%mm1 \n\t" | |
443 "movq 8(%1, %3), %%mm5 \n\t" | |
444 "movq %%mm0, (%2) \n\t" | |
445 "movq %%mm4, 8(%2) \n\t" | |
446 "movq %%mm1, (%2, %3) \n\t" | |
447 "movq %%mm5, 8(%2, %3) \n\t" | |
448 "add %%"REG_a", %1 \n\t" | |
449 "add %%"REG_a", %2 \n\t" | |
450 "movq (%1), %%mm0 \n\t" | |
451 "movq 8(%1), %%mm4 \n\t" | |
452 "movq (%1, %3), %%mm1 \n\t" | |
453 "movq 8(%1, %3), %%mm5 \n\t" | |
454 "movq %%mm0, (%2) \n\t" | |
455 "movq %%mm4, 8(%2) \n\t" | |
456 "movq %%mm1, (%2, %3) \n\t" | |
457 "movq %%mm5, 8(%2, %3) \n\t" | |
458 "add %%"REG_a", %1 \n\t" | |
459 "add %%"REG_a", %2 \n\t" | |
460 "subl $4, %0 \n\t" | |
461 "jnz 1b \n\t" | |
462 : "+g"(h), "+r" (pixels), "+r" (block) | |
463 : "r"((long)line_size) | |
464 : "%"REG_a, "memory" | |
465 ); | |
651 | 466 } |
467 | |
296 | 468 static void clear_blocks_mmx(DCTELEM *blocks) |
469 { | |
471 | 470 __asm __volatile( |
2979 | 471 "pxor %%mm7, %%mm7 \n\t" |
472 "mov $-128*6, %%"REG_a" \n\t" | |
473 "1: \n\t" | |
474 "movq %%mm7, (%0, %%"REG_a") \n\t" | |
475 "movq %%mm7, 8(%0, %%"REG_a") \n\t" | |
476 "movq %%mm7, 16(%0, %%"REG_a") \n\t" | |
477 "movq %%mm7, 24(%0, %%"REG_a") \n\t" | |
478 "add $32, %%"REG_a" \n\t" | |
479 " js 1b \n\t" | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
480 : : "r" (((uint8_t *)blocks)+128*6) |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
481 : "%"REG_a |
296 | 482 ); |
483 } | |
484 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
485 #ifdef CONFIG_ENCODERS |
1064 | 486 static int pix_sum16_mmx(uint8_t * pix, int line_size){ |
688 | 487 const int h=16; |
488 int sum; | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
489 long index= -line_size*h; |
688 | 490 |
491 __asm __volatile( | |
2979 | 492 "pxor %%mm7, %%mm7 \n\t" |
493 "pxor %%mm6, %%mm6 \n\t" | |
494 "1: \n\t" | |
495 "movq (%2, %1), %%mm0 \n\t" | |
496 "movq (%2, %1), %%mm1 \n\t" | |
497 "movq 8(%2, %1), %%mm2 \n\t" | |
498 "movq 8(%2, %1), %%mm3 \n\t" | |
499 "punpcklbw %%mm7, %%mm0 \n\t" | |
500 "punpckhbw %%mm7, %%mm1 \n\t" | |
501 "punpcklbw %%mm7, %%mm2 \n\t" | |
502 "punpckhbw %%mm7, %%mm3 \n\t" | |
503 "paddw %%mm0, %%mm1 \n\t" | |
504 "paddw %%mm2, %%mm3 \n\t" | |
505 "paddw %%mm1, %%mm3 \n\t" | |
506 "paddw %%mm3, %%mm6 \n\t" | |
507 "add %3, %1 \n\t" | |
508 " js 1b \n\t" | |
509 "movq %%mm6, %%mm5 \n\t" | |
510 "psrlq $32, %%mm6 \n\t" | |
511 "paddw %%mm5, %%mm6 \n\t" | |
512 "movq %%mm6, %%mm5 \n\t" | |
513 "psrlq $16, %%mm6 \n\t" | |
514 "paddw %%mm5, %%mm6 \n\t" | |
515 "movd %%mm6, %0 \n\t" | |
516 "andl $0xFFFF, %0 \n\t" | |
688 | 517 : "=&r" (sum), "+r" (index) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
518 : "r" (pix - index), "r" ((long)line_size) |
688 | 519 ); |
520 | |
521 return sum; | |
522 } | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
523 #endif //CONFIG_ENCODERS |
688 | 524 |
866 | 525 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w){ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
526 long i=0; |
866 | 527 asm volatile( |
2979 | 528 "1: \n\t" |
529 "movq (%1, %0), %%mm0 \n\t" | |
530 "movq (%2, %0), %%mm1 \n\t" | |
531 "paddb %%mm0, %%mm1 \n\t" | |
532 "movq %%mm1, (%2, %0) \n\t" | |
533 "movq 8(%1, %0), %%mm0 \n\t" | |
534 "movq 8(%2, %0), %%mm1 \n\t" | |
535 "paddb %%mm0, %%mm1 \n\t" | |
536 "movq %%mm1, 8(%2, %0) \n\t" | |
537 "add $16, %0 \n\t" | |
538 "cmp %3, %0 \n\t" | |
539 " jb 1b \n\t" | |
866 | 540 : "+r" (i) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
541 : "r"(src), "r"(dst), "r"((long)w-15) |
866 | 542 ); |
543 for(; i<w; i++) | |
544 dst[i+0] += src[i+0]; | |
545 } | |
546 | |
1648 | 547 #define H263_LOOP_FILTER \ |
2979 | 548 "pxor %%mm7, %%mm7 \n\t"\ |
549 "movq %0, %%mm0 \n\t"\ | |
550 "movq %0, %%mm1 \n\t"\ | |
551 "movq %3, %%mm2 \n\t"\ | |
552 "movq %3, %%mm3 \n\t"\ | |
553 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
554 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
555 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
556 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
557 "psubw %%mm2, %%mm0 \n\t"\ | |
558 "psubw %%mm3, %%mm1 \n\t"\ | |
559 "movq %1, %%mm2 \n\t"\ | |
560 "movq %1, %%mm3 \n\t"\ | |
561 "movq %2, %%mm4 \n\t"\ | |
562 "movq %2, %%mm5 \n\t"\ | |
563 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
564 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
565 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
566 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
567 "psubw %%mm2, %%mm4 \n\t"\ | |
568 "psubw %%mm3, %%mm5 \n\t"\ | |
569 "psllw $2, %%mm4 \n\t"\ | |
570 "psllw $2, %%mm5 \n\t"\ | |
571 "paddw %%mm0, %%mm4 \n\t"\ | |
572 "paddw %%mm1, %%mm5 \n\t"\ | |
573 "pxor %%mm6, %%mm6 \n\t"\ | |
574 "pcmpgtw %%mm4, %%mm6 \n\t"\ | |
575 "pcmpgtw %%mm5, %%mm7 \n\t"\ | |
576 "pxor %%mm6, %%mm4 \n\t"\ | |
577 "pxor %%mm7, %%mm5 \n\t"\ | |
578 "psubw %%mm6, %%mm4 \n\t"\ | |
579 "psubw %%mm7, %%mm5 \n\t"\ | |
580 "psrlw $3, %%mm4 \n\t"\ | |
581 "psrlw $3, %%mm5 \n\t"\ | |
582 "packuswb %%mm5, %%mm4 \n\t"\ | |
583 "packsswb %%mm7, %%mm6 \n\t"\ | |
584 "pxor %%mm7, %%mm7 \n\t"\ | |
585 "movd %4, %%mm2 \n\t"\ | |
586 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
587 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
588 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
589 "psubusb %%mm4, %%mm2 \n\t"\ | |
590 "movq %%mm2, %%mm3 \n\t"\ | |
591 "psubusb %%mm4, %%mm3 \n\t"\ | |
592 "psubb %%mm3, %%mm2 \n\t"\ | |
593 "movq %1, %%mm3 \n\t"\ | |
594 "movq %2, %%mm4 \n\t"\ | |
595 "pxor %%mm6, %%mm3 \n\t"\ | |
596 "pxor %%mm6, %%mm4 \n\t"\ | |
597 "paddusb %%mm2, %%mm3 \n\t"\ | |
598 "psubusb %%mm2, %%mm4 \n\t"\ | |
599 "pxor %%mm6, %%mm3 \n\t"\ | |
600 "pxor %%mm6, %%mm4 \n\t"\ | |
601 "paddusb %%mm2, %%mm2 \n\t"\ | |
602 "packsswb %%mm1, %%mm0 \n\t"\ | |
603 "pcmpgtb %%mm0, %%mm7 \n\t"\ | |
604 "pxor %%mm7, %%mm0 \n\t"\ | |
605 "psubb %%mm7, %%mm0 \n\t"\ | |
606 "movq %%mm0, %%mm1 \n\t"\ | |
607 "psubusb %%mm2, %%mm0 \n\t"\ | |
608 "psubb %%mm0, %%mm1 \n\t"\ | |
609 "pand %5, %%mm1 \n\t"\ | |
610 "psrlw $2, %%mm1 \n\t"\ | |
611 "pxor %%mm7, %%mm1 \n\t"\ | |
612 "psubb %%mm7, %%mm1 \n\t"\ | |
613 "movq %0, %%mm5 \n\t"\ | |
614 "movq %3, %%mm6 \n\t"\ | |
615 "psubb %%mm1, %%mm5 \n\t"\ | |
616 "paddb %%mm1, %%mm6 \n\t" | |
1648 | 617 |
1647 | 618 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale){ |
5394
e9a6215f4e3a
help some gcc version to optimize out those functions
aurel
parents:
5278
diff
changeset
|
619 if(ENABLE_ANY_H263) { |
1647 | 620 const int strength= ff_h263_loop_filter_strength[qscale]; |
621 | |
622 asm volatile( | |
2967 | 623 |
1648 | 624 H263_LOOP_FILTER |
2967 | 625 |
2979 | 626 "movq %%mm3, %1 \n\t" |
627 "movq %%mm4, %2 \n\t" | |
628 "movq %%mm5, %0 \n\t" | |
629 "movq %%mm6, %3 \n\t" | |
1647 | 630 : "+m" (*(uint64_t*)(src - 2*stride)), |
631 "+m" (*(uint64_t*)(src - 1*stride)), | |
632 "+m" (*(uint64_t*)(src + 0*stride)), | |
633 "+m" (*(uint64_t*)(src + 1*stride)) | |
634 : "g" (2*strength), "m"(ff_pb_FC) | |
635 ); | |
5394
e9a6215f4e3a
help some gcc version to optimize out those functions
aurel
parents:
5278
diff
changeset
|
636 } |
1647 | 637 } |
638 | |
1648 | 639 static inline void transpose4x4(uint8_t *dst, uint8_t *src, int dst_stride, int src_stride){ |
640 asm volatile( //FIXME could save 1 instruction if done as 8x4 ... | |
2979 | 641 "movd %4, %%mm0 \n\t" |
642 "movd %5, %%mm1 \n\t" | |
643 "movd %6, %%mm2 \n\t" | |
644 "movd %7, %%mm3 \n\t" | |
645 "punpcklbw %%mm1, %%mm0 \n\t" | |
646 "punpcklbw %%mm3, %%mm2 \n\t" | |
647 "movq %%mm0, %%mm1 \n\t" | |
648 "punpcklwd %%mm2, %%mm0 \n\t" | |
649 "punpckhwd %%mm2, %%mm1 \n\t" | |
650 "movd %%mm0, %0 \n\t" | |
651 "punpckhdq %%mm0, %%mm0 \n\t" | |
652 "movd %%mm0, %1 \n\t" | |
653 "movd %%mm1, %2 \n\t" | |
654 "punpckhdq %%mm1, %%mm1 \n\t" | |
655 "movd %%mm1, %3 \n\t" | |
2967 | 656 |
1648 | 657 : "=m" (*(uint32_t*)(dst + 0*dst_stride)), |
658 "=m" (*(uint32_t*)(dst + 1*dst_stride)), | |
659 "=m" (*(uint32_t*)(dst + 2*dst_stride)), | |
660 "=m" (*(uint32_t*)(dst + 3*dst_stride)) | |
661 : "m" (*(uint32_t*)(src + 0*src_stride)), | |
662 "m" (*(uint32_t*)(src + 1*src_stride)), | |
663 "m" (*(uint32_t*)(src + 2*src_stride)), | |
664 "m" (*(uint32_t*)(src + 3*src_stride)) | |
665 ); | |
666 } | |
667 | |
668 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale){ | |
5394
e9a6215f4e3a
help some gcc version to optimize out those functions
aurel
parents:
5278
diff
changeset
|
669 if(ENABLE_ANY_H263) { |
1648 | 670 const int strength= ff_h263_loop_filter_strength[qscale]; |
6181 | 671 DECLARE_ALIGNED(8, uint64_t, temp[4]); |
1648 | 672 uint8_t *btemp= (uint8_t*)temp; |
2967 | 673 |
1648 | 674 src -= 2; |
675 | |
676 transpose4x4(btemp , src , 8, stride); | |
677 transpose4x4(btemp+4, src + 4*stride, 8, stride); | |
678 asm volatile( | |
679 H263_LOOP_FILTER // 5 3 4 6 | |
2967 | 680 |
1648 | 681 : "+m" (temp[0]), |
682 "+m" (temp[1]), | |
683 "+m" (temp[2]), | |
684 "+m" (temp[3]) | |
685 : "g" (2*strength), "m"(ff_pb_FC) | |
686 ); | |
687 | |
688 asm volatile( | |
2979 | 689 "movq %%mm5, %%mm1 \n\t" |
690 "movq %%mm4, %%mm0 \n\t" | |
691 "punpcklbw %%mm3, %%mm5 \n\t" | |
692 "punpcklbw %%mm6, %%mm4 \n\t" | |
693 "punpckhbw %%mm3, %%mm1 \n\t" | |
694 "punpckhbw %%mm6, %%mm0 \n\t" | |
695 "movq %%mm5, %%mm3 \n\t" | |
696 "movq %%mm1, %%mm6 \n\t" | |
697 "punpcklwd %%mm4, %%mm5 \n\t" | |
698 "punpcklwd %%mm0, %%mm1 \n\t" | |
699 "punpckhwd %%mm4, %%mm3 \n\t" | |
700 "punpckhwd %%mm0, %%mm6 \n\t" | |
701 "movd %%mm5, (%0) \n\t" | |
702 "punpckhdq %%mm5, %%mm5 \n\t" | |
703 "movd %%mm5, (%0,%2) \n\t" | |
704 "movd %%mm3, (%0,%2,2) \n\t" | |
705 "punpckhdq %%mm3, %%mm3 \n\t" | |
706 "movd %%mm3, (%0,%3) \n\t" | |
707 "movd %%mm1, (%1) \n\t" | |
708 "punpckhdq %%mm1, %%mm1 \n\t" | |
709 "movd %%mm1, (%1,%2) \n\t" | |
710 "movd %%mm6, (%1,%2,2) \n\t" | |
711 "punpckhdq %%mm6, %%mm6 \n\t" | |
712 "movd %%mm6, (%1,%3) \n\t" | |
2505
86e2b1424801
optimization and gcc 4.0 bug workaround patch by (Martin Drab >drab kepler.fjfi.cvut cz<)
michael
parents:
2293
diff
changeset
|
713 :: "r" (src), |
86e2b1424801
optimization and gcc 4.0 bug workaround patch by (Martin Drab >drab kepler.fjfi.cvut cz<)
michael
parents:
2293
diff
changeset
|
714 "r" (src + 4*stride), |
86e2b1424801
optimization and gcc 4.0 bug workaround patch by (Martin Drab >drab kepler.fjfi.cvut cz<)
michael
parents:
2293
diff
changeset
|
715 "r" ((long) stride ), |
86e2b1424801
optimization and gcc 4.0 bug workaround patch by (Martin Drab >drab kepler.fjfi.cvut cz<)
michael
parents:
2293
diff
changeset
|
716 "r" ((long)(3*stride)) |
1648 | 717 ); |
5394
e9a6215f4e3a
help some gcc version to optimize out those functions
aurel
parents:
5278
diff
changeset
|
718 } |
1648 | 719 } |
720 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
721 #ifdef CONFIG_ENCODERS |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
722 static int pix_norm1_mmx(uint8_t *pix, int line_size) { |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
723 int tmp; |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
724 asm volatile ( |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
725 "movl $16,%%ecx\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
726 "pxor %%mm0,%%mm0\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
727 "pxor %%mm7,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
728 "1:\n" |
2979 | 729 "movq (%0),%%mm2\n" /* mm2 = pix[0-7] */ |
730 "movq 8(%0),%%mm3\n" /* mm3 = pix[8-15] */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
731 |
2979 | 732 "movq %%mm2,%%mm1\n" /* mm1 = mm2 = pix[0-7] */ |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
733 |
2979 | 734 "punpckhbw %%mm0,%%mm1\n" /* mm1 = [pix4-7] */ |
735 "punpcklbw %%mm0,%%mm2\n" /* mm2 = [pix0-3] */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
736 |
2979 | 737 "movq %%mm3,%%mm4\n" /* mm4 = mm3 = pix[8-15] */ |
738 "punpckhbw %%mm0,%%mm3\n" /* mm3 = [pix12-15] */ | |
739 "punpcklbw %%mm0,%%mm4\n" /* mm4 = [pix8-11] */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
740 |
2979 | 741 "pmaddwd %%mm1,%%mm1\n" /* mm1 = (pix0^2+pix1^2,pix2^2+pix3^2) */ |
742 "pmaddwd %%mm2,%%mm2\n" /* mm2 = (pix4^2+pix5^2,pix6^2+pix7^2) */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
743 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
744 "pmaddwd %%mm3,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
745 "pmaddwd %%mm4,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
746 |
2979 | 747 "paddd %%mm1,%%mm2\n" /* mm2 = (pix0^2+pix1^2+pix4^2+pix5^2, |
748 pix2^2+pix3^2+pix6^2+pix7^2) */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
749 "paddd %%mm3,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
750 "paddd %%mm2,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
751 |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
752 "add %2, %0\n" |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
753 "paddd %%mm4,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
754 "dec %%ecx\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
755 "jnz 1b\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
756 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
757 "movq %%mm7,%%mm1\n" |
2979 | 758 "psrlq $32, %%mm7\n" /* shift hi dword to lo */ |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
759 "paddd %%mm7,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
760 "movd %%mm1,%1\n" |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
761 : "+r" (pix), "=r"(tmp) : "r" ((long)line_size) : "%ecx" ); |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
762 return tmp; |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
763 } |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
764 |
2067 | 765 static int sse8_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
766 int tmp; | |
767 asm volatile ( | |
768 "movl %4,%%ecx\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
769 "shr $1,%%ecx\n" |
2979 | 770 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */ |
771 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */ | |
2067 | 772 "1:\n" |
2979 | 773 "movq (%0),%%mm1\n" /* mm1 = pix1[0][0-7] */ |
774 "movq (%1),%%mm2\n" /* mm2 = pix2[0][0-7] */ | |
775 "movq (%0,%3),%%mm3\n" /* mm3 = pix1[1][0-7] */ | |
776 "movq (%1,%3),%%mm4\n" /* mm4 = pix2[1][0-7] */ | |
2067 | 777 |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
778 /* todo: mm1-mm2, mm3-mm4 */ |
5963 | 779 /* algo: subtract mm1 from mm2 with saturation and vice versa */ |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
780 /* OR the results to get absolute difference */ |
2067 | 781 "movq %%mm1,%%mm5\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
782 "movq %%mm3,%%mm6\n" |
2067 | 783 "psubusb %%mm2,%%mm1\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
784 "psubusb %%mm4,%%mm3\n" |
2067 | 785 "psubusb %%mm5,%%mm2\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
786 "psubusb %%mm6,%%mm4\n" |
2067 | 787 |
788 "por %%mm1,%%mm2\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
789 "por %%mm3,%%mm4\n" |
2067 | 790 |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
791 /* now convert to 16-bit vectors so we can square them */ |
2067 | 792 "movq %%mm2,%%mm1\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
793 "movq %%mm4,%%mm3\n" |
2067 | 794 |
795 "punpckhbw %%mm0,%%mm2\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
796 "punpckhbw %%mm0,%%mm4\n" |
2979 | 797 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */ |
798 "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */ | |
2067 | 799 |
800 "pmaddwd %%mm2,%%mm2\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
801 "pmaddwd %%mm4,%%mm4\n" |
2067 | 802 "pmaddwd %%mm1,%%mm1\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
803 "pmaddwd %%mm3,%%mm3\n" |
2067 | 804 |
2979 | 805 "lea (%0,%3,2), %0\n" /* pix1 += 2*line_size */ |
806 "lea (%1,%3,2), %1\n" /* pix2 += 2*line_size */ | |
2067 | 807 |
808 "paddd %%mm2,%%mm1\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
809 "paddd %%mm4,%%mm3\n" |
2067 | 810 "paddd %%mm1,%%mm7\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
811 "paddd %%mm3,%%mm7\n" |
2067 | 812 |
813 "decl %%ecx\n" | |
814 "jnz 1b\n" | |
815 | |
816 "movq %%mm7,%%mm1\n" | |
2979 | 817 "psrlq $32, %%mm7\n" /* shift hi dword to lo */ |
2067 | 818 "paddd %%mm7,%%mm1\n" |
819 "movd %%mm1,%2\n" | |
2967 | 820 : "+r" (pix1), "+r" (pix2), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
821 : "r" ((long)line_size) , "m" (h) |
2067 | 822 : "%ecx"); |
823 return tmp; | |
824 } | |
825 | |
1708 | 826 static int sse16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
827 int tmp; |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
828 asm volatile ( |
1708 | 829 "movl %4,%%ecx\n" |
2979 | 830 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */ |
831 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
832 "1:\n" |
2979 | 833 "movq (%0),%%mm1\n" /* mm1 = pix1[0-7] */ |
834 "movq (%1),%%mm2\n" /* mm2 = pix2[0-7] */ | |
835 "movq 8(%0),%%mm3\n" /* mm3 = pix1[8-15] */ | |
836 "movq 8(%1),%%mm4\n" /* mm4 = pix2[8-15] */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
837 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
838 /* todo: mm1-mm2, mm3-mm4 */ |
5963 | 839 /* algo: subtract mm1 from mm2 with saturation and vice versa */ |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
840 /* OR the results to get absolute difference */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
841 "movq %%mm1,%%mm5\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
842 "movq %%mm3,%%mm6\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
843 "psubusb %%mm2,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
844 "psubusb %%mm4,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
845 "psubusb %%mm5,%%mm2\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
846 "psubusb %%mm6,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
847 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
848 "por %%mm1,%%mm2\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
849 "por %%mm3,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
850 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
851 /* now convert to 16-bit vectors so we can square them */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
852 "movq %%mm2,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
853 "movq %%mm4,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
854 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
855 "punpckhbw %%mm0,%%mm2\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
856 "punpckhbw %%mm0,%%mm4\n" |
2979 | 857 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */ |
858 "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
859 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
860 "pmaddwd %%mm2,%%mm2\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
861 "pmaddwd %%mm4,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
862 "pmaddwd %%mm1,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
863 "pmaddwd %%mm3,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
864 |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
865 "add %3,%0\n" |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
866 "add %3,%1\n" |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
867 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
868 "paddd %%mm2,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
869 "paddd %%mm4,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
870 "paddd %%mm1,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
871 "paddd %%mm3,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
872 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
873 "decl %%ecx\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
874 "jnz 1b\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
875 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
876 "movq %%mm7,%%mm1\n" |
2979 | 877 "psrlq $32, %%mm7\n" /* shift hi dword to lo */ |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
878 "paddd %%mm7,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
879 "movd %%mm1,%2\n" |
2967 | 880 : "+r" (pix1), "+r" (pix2), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
881 : "r" ((long)line_size) , "m" (h) |
1708 | 882 : "%ecx"); |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
883 return tmp; |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
884 } |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
885 |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
886 static int sse16_sse2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
887 int tmp; |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
888 asm volatile ( |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
889 "shr $1,%2\n" |
2979 | 890 "pxor %%xmm0,%%xmm0\n" /* mm0 = 0 */ |
891 "pxor %%xmm7,%%xmm7\n" /* mm7 holds the sum */ | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
892 "1:\n" |
2979 | 893 "movdqu (%0),%%xmm1\n" /* mm1 = pix1[0][0-15] */ |
894 "movdqu (%1),%%xmm2\n" /* mm2 = pix2[0][0-15] */ | |
895 "movdqu (%0,%4),%%xmm3\n" /* mm3 = pix1[1][0-15] */ | |
896 "movdqu (%1,%4),%%xmm4\n" /* mm4 = pix2[1][0-15] */ | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
897 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
898 /* todo: mm1-mm2, mm3-mm4 */ |
5963 | 899 /* algo: subtract mm1 from mm2 with saturation and vice versa */ |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
900 /* OR the results to get absolute difference */ |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
901 "movdqa %%xmm1,%%xmm5\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
902 "movdqa %%xmm3,%%xmm6\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
903 "psubusb %%xmm2,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
904 "psubusb %%xmm4,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
905 "psubusb %%xmm5,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
906 "psubusb %%xmm6,%%xmm4\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
907 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
908 "por %%xmm1,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
909 "por %%xmm3,%%xmm4\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
910 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
911 /* now convert to 16-bit vectors so we can square them */ |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
912 "movdqa %%xmm2,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
913 "movdqa %%xmm4,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
914 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
915 "punpckhbw %%xmm0,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
916 "punpckhbw %%xmm0,%%xmm4\n" |
2979 | 917 "punpcklbw %%xmm0,%%xmm1\n" /* mm1 now spread over (mm1,mm2) */ |
918 "punpcklbw %%xmm0,%%xmm3\n" /* mm4 now spread over (mm3,mm4) */ | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
919 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
920 "pmaddwd %%xmm2,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
921 "pmaddwd %%xmm4,%%xmm4\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
922 "pmaddwd %%xmm1,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
923 "pmaddwd %%xmm3,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
924 |
2979 | 925 "lea (%0,%4,2), %0\n" /* pix1 += 2*line_size */ |
926 "lea (%1,%4,2), %1\n" /* pix2 += 2*line_size */ | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
927 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
928 "paddd %%xmm2,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
929 "paddd %%xmm4,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
930 "paddd %%xmm1,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
931 "paddd %%xmm3,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
932 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
933 "decl %2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
934 "jnz 1b\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
935 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
936 "movdqa %%xmm7,%%xmm1\n" |
2979 | 937 "psrldq $8, %%xmm7\n" /* shift hi qword to lo */ |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
938 "paddd %%xmm1,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
939 "movdqa %%xmm7,%%xmm1\n" |
2979 | 940 "psrldq $4, %%xmm7\n" /* shift hi dword to lo */ |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
941 "paddd %%xmm1,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
942 "movd %%xmm7,%3\n" |
2967 | 943 : "+r" (pix1), "+r" (pix2), "+r"(h), "=r"(tmp) |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
944 : "r" ((long)line_size)); |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
945 return tmp; |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
946 } |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
947 |
2067 | 948 static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) { |
949 int tmp; | |
950 asm volatile ( | |
951 "movl %3,%%ecx\n" | |
952 "pxor %%mm7,%%mm7\n" | |
953 "pxor %%mm6,%%mm6\n" | |
2967 | 954 |
2067 | 955 "movq (%0),%%mm0\n" |
956 "movq %%mm0, %%mm1\n" | |
957 "psllq $8, %%mm0\n" | |
958 "psrlq $8, %%mm1\n" | |
959 "psrlq $8, %%mm0\n" | |
960 "movq %%mm0, %%mm2\n" | |
961 "movq %%mm1, %%mm3\n" | |
962 "punpcklbw %%mm7,%%mm0\n" | |
963 "punpcklbw %%mm7,%%mm1\n" | |
964 "punpckhbw %%mm7,%%mm2\n" | |
965 "punpckhbw %%mm7,%%mm3\n" | |
966 "psubw %%mm1, %%mm0\n" | |
967 "psubw %%mm3, %%mm2\n" | |
2967 | 968 |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
969 "add %2,%0\n" |
2967 | 970 |
2067 | 971 "movq (%0),%%mm4\n" |
972 "movq %%mm4, %%mm1\n" | |
973 "psllq $8, %%mm4\n" | |
974 "psrlq $8, %%mm1\n" | |
975 "psrlq $8, %%mm4\n" | |
976 "movq %%mm4, %%mm5\n" | |
977 "movq %%mm1, %%mm3\n" | |
978 "punpcklbw %%mm7,%%mm4\n" | |
979 "punpcklbw %%mm7,%%mm1\n" | |
980 "punpckhbw %%mm7,%%mm5\n" | |
981 "punpckhbw %%mm7,%%mm3\n" | |
982 "psubw %%mm1, %%mm4\n" | |
983 "psubw %%mm3, %%mm5\n" | |
984 "psubw %%mm4, %%mm0\n" | |
985 "psubw %%mm5, %%mm2\n" | |
986 "pxor %%mm3, %%mm3\n" | |
987 "pxor %%mm1, %%mm1\n" | |
988 "pcmpgtw %%mm0, %%mm3\n\t" | |
989 "pcmpgtw %%mm2, %%mm1\n\t" | |
990 "pxor %%mm3, %%mm0\n" | |
991 "pxor %%mm1, %%mm2\n" | |
2967 | 992 "psubw %%mm3, %%mm0\n" |
2067 | 993 "psubw %%mm1, %%mm2\n" |
994 "paddw %%mm0, %%mm2\n" | |
995 "paddw %%mm2, %%mm6\n" | |
996 | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
997 "add %2,%0\n" |
2067 | 998 "1:\n" |
2967 | 999 |
2067 | 1000 "movq (%0),%%mm0\n" |
1001 "movq %%mm0, %%mm1\n" | |
1002 "psllq $8, %%mm0\n" | |
1003 "psrlq $8, %%mm1\n" | |
1004 "psrlq $8, %%mm0\n" | |
1005 "movq %%mm0, %%mm2\n" | |
1006 "movq %%mm1, %%mm3\n" | |
1007 "punpcklbw %%mm7,%%mm0\n" | |
1008 "punpcklbw %%mm7,%%mm1\n" | |
1009 "punpckhbw %%mm7,%%mm2\n" | |
1010 "punpckhbw %%mm7,%%mm3\n" | |
1011 "psubw %%mm1, %%mm0\n" | |
1012 "psubw %%mm3, %%mm2\n" | |
1013 "psubw %%mm0, %%mm4\n" | |
1014 "psubw %%mm2, %%mm5\n" | |
1015 "pxor %%mm3, %%mm3\n" | |
1016 "pxor %%mm1, %%mm1\n" | |
1017 "pcmpgtw %%mm4, %%mm3\n\t" | |
1018 "pcmpgtw %%mm5, %%mm1\n\t" | |
1019 "pxor %%mm3, %%mm4\n" | |
1020 "pxor %%mm1, %%mm5\n" | |
2967 | 1021 "psubw %%mm3, %%mm4\n" |
2067 | 1022 "psubw %%mm1, %%mm5\n" |
1023 "paddw %%mm4, %%mm5\n" | |
1024 "paddw %%mm5, %%mm6\n" | |
2967 | 1025 |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1026 "add %2,%0\n" |
2967 | 1027 |
2067 | 1028 "movq (%0),%%mm4\n" |
1029 "movq %%mm4, %%mm1\n" | |
1030 "psllq $8, %%mm4\n" | |
1031 "psrlq $8, %%mm1\n" | |
1032 "psrlq $8, %%mm4\n" | |
1033 "movq %%mm4, %%mm5\n" | |
1034 "movq %%mm1, %%mm3\n" | |
1035 "punpcklbw %%mm7,%%mm4\n" | |
1036 "punpcklbw %%mm7,%%mm1\n" | |
1037 "punpckhbw %%mm7,%%mm5\n" | |
1038 "punpckhbw %%mm7,%%mm3\n" | |
1039 "psubw %%mm1, %%mm4\n" | |
1040 "psubw %%mm3, %%mm5\n" | |
1041 "psubw %%mm4, %%mm0\n" | |
1042 "psubw %%mm5, %%mm2\n" | |
1043 "pxor %%mm3, %%mm3\n" | |
1044 "pxor %%mm1, %%mm1\n" | |
1045 "pcmpgtw %%mm0, %%mm3\n\t" | |
1046 "pcmpgtw %%mm2, %%mm1\n\t" | |
1047 "pxor %%mm3, %%mm0\n" | |
1048 "pxor %%mm1, %%mm2\n" | |
2967 | 1049 "psubw %%mm3, %%mm0\n" |
2067 | 1050 "psubw %%mm1, %%mm2\n" |
1051 "paddw %%mm0, %%mm2\n" | |
1052 "paddw %%mm2, %%mm6\n" | |
1053 | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1054 "add %2,%0\n" |
2067 | 1055 "subl $2, %%ecx\n" |
1056 " jnz 1b\n" | |
1057 | |
1058 "movq %%mm6, %%mm0\n" | |
1059 "punpcklwd %%mm7,%%mm0\n" | |
1060 "punpckhwd %%mm7,%%mm6\n" | |
1061 "paddd %%mm0, %%mm6\n" | |
2967 | 1062 |
2067 | 1063 "movq %%mm6,%%mm0\n" |
1064 "psrlq $32, %%mm6\n" | |
1065 "paddd %%mm6,%%mm0\n" | |
1066 "movd %%mm0,%1\n" | |
2967 | 1067 : "+r" (pix1), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1068 : "r" ((long)line_size) , "g" (h-2) |
2067 | 1069 : "%ecx"); |
1070 return tmp; | |
1071 } | |
1072 | |
1073 static int hf_noise16_mmx(uint8_t * pix1, int line_size, int h) { | |
1074 int tmp; | |
1075 uint8_t * pix= pix1; | |
1076 asm volatile ( | |
1077 "movl %3,%%ecx\n" | |
1078 "pxor %%mm7,%%mm7\n" | |
1079 "pxor %%mm6,%%mm6\n" | |
2967 | 1080 |
2067 | 1081 "movq (%0),%%mm0\n" |
1082 "movq 1(%0),%%mm1\n" | |
1083 "movq %%mm0, %%mm2\n" | |
1084 "movq %%mm1, %%mm3\n" | |
1085 "punpcklbw %%mm7,%%mm0\n" | |
1086 "punpcklbw %%mm7,%%mm1\n" | |
1087 "punpckhbw %%mm7,%%mm2\n" | |
1088 "punpckhbw %%mm7,%%mm3\n" | |
1089 "psubw %%mm1, %%mm0\n" | |
1090 "psubw %%mm3, %%mm2\n" | |
2967 | 1091 |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1092 "add %2,%0\n" |
2967 | 1093 |
2067 | 1094 "movq (%0),%%mm4\n" |
1095 "movq 1(%0),%%mm1\n" | |
1096 "movq %%mm4, %%mm5\n" | |
1097 "movq %%mm1, %%mm3\n" | |
1098 "punpcklbw %%mm7,%%mm4\n" | |
1099 "punpcklbw %%mm7,%%mm1\n" | |
1100 "punpckhbw %%mm7,%%mm5\n" | |
1101 "punpckhbw %%mm7,%%mm3\n" | |
1102 "psubw %%mm1, %%mm4\n" | |
1103 "psubw %%mm3, %%mm5\n" | |
1104 "psubw %%mm4, %%mm0\n" | |
1105 "psubw %%mm5, %%mm2\n" | |
1106 "pxor %%mm3, %%mm3\n" | |
1107 "pxor %%mm1, %%mm1\n" | |
1108 "pcmpgtw %%mm0, %%mm3\n\t" | |
1109 "pcmpgtw %%mm2, %%mm1\n\t" | |
1110 "pxor %%mm3, %%mm0\n" | |
1111 "pxor %%mm1, %%mm2\n" | |
2967 | 1112 "psubw %%mm3, %%mm0\n" |
2067 | 1113 "psubw %%mm1, %%mm2\n" |
1114 "paddw %%mm0, %%mm2\n" | |
1115 "paddw %%mm2, %%mm6\n" | |
1116 | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1117 "add %2,%0\n" |
2067 | 1118 "1:\n" |
2967 | 1119 |
2067 | 1120 "movq (%0),%%mm0\n" |
1121 "movq 1(%0),%%mm1\n" | |
1122 "movq %%mm0, %%mm2\n" | |
1123 "movq %%mm1, %%mm3\n" | |
1124 "punpcklbw %%mm7,%%mm0\n" | |
1125 "punpcklbw %%mm7,%%mm1\n" | |
1126 "punpckhbw %%mm7,%%mm2\n" | |
1127 "punpckhbw %%mm7,%%mm3\n" | |
1128 "psubw %%mm1, %%mm0\n" | |
1129 "psubw %%mm3, %%mm2\n" | |
1130 "psubw %%mm0, %%mm4\n" | |
1131 "psubw %%mm2, %%mm5\n" | |
1132 "pxor %%mm3, %%mm3\n" | |
1133 "pxor %%mm1, %%mm1\n" | |
1134 "pcmpgtw %%mm4, %%mm3\n\t" | |
1135 "pcmpgtw %%mm5, %%mm1\n\t" | |
1136 "pxor %%mm3, %%mm4\n" | |
1137 "pxor %%mm1, %%mm5\n" | |
1138 "psubw %%mm3, %%mm4\n" | |
1139 "psubw %%mm1, %%mm5\n" | |
1140 "paddw %%mm4, %%mm5\n" | |
1141 "paddw %%mm5, %%mm6\n" | |
2967 | 1142 |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1143 "add %2,%0\n" |
2967 | 1144 |
2067 | 1145 "movq (%0),%%mm4\n" |
1146 "movq 1(%0),%%mm1\n" | |
1147 "movq %%mm4, %%mm5\n" | |
1148 "movq %%mm1, %%mm3\n" | |
1149 "punpcklbw %%mm7,%%mm4\n" | |
1150 "punpcklbw %%mm7,%%mm1\n" | |
1151 "punpckhbw %%mm7,%%mm5\n" | |
1152 "punpckhbw %%mm7,%%mm3\n" | |
1153 "psubw %%mm1, %%mm4\n" | |
1154 "psubw %%mm3, %%mm5\n" | |
1155 "psubw %%mm4, %%mm0\n" | |
1156 "psubw %%mm5, %%mm2\n" | |
1157 "pxor %%mm3, %%mm3\n" | |
1158 "pxor %%mm1, %%mm1\n" | |
1159 "pcmpgtw %%mm0, %%mm3\n\t" | |
1160 "pcmpgtw %%mm2, %%mm1\n\t" | |
1161 "pxor %%mm3, %%mm0\n" | |
1162 "pxor %%mm1, %%mm2\n" | |
2967 | 1163 "psubw %%mm3, %%mm0\n" |
2067 | 1164 "psubw %%mm1, %%mm2\n" |
1165 "paddw %%mm0, %%mm2\n" | |
1166 "paddw %%mm2, %%mm6\n" | |
1167 | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1168 "add %2,%0\n" |
2067 | 1169 "subl $2, %%ecx\n" |
1170 " jnz 1b\n" | |
1171 | |
1172 "movq %%mm6, %%mm0\n" | |
1173 "punpcklwd %%mm7,%%mm0\n" | |
1174 "punpckhwd %%mm7,%%mm6\n" | |
1175 "paddd %%mm0, %%mm6\n" | |
2967 | 1176 |
2067 | 1177 "movq %%mm6,%%mm0\n" |
1178 "psrlq $32, %%mm6\n" | |
1179 "paddd %%mm6,%%mm0\n" | |
1180 "movd %%mm0,%1\n" | |
2967 | 1181 : "+r" (pix1), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1182 : "r" ((long)line_size) , "g" (h-2) |
2067 | 1183 : "%ecx"); |
1184 return tmp + hf_noise8_mmx(pix+8, line_size, h); | |
1185 } | |
1186 | |
2864
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1187 static int nsse16_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1188 MpegEncContext *c = p; |
2940 | 1189 int score1, score2; |
1190 | |
1191 if(c) score1 = c->dsp.sse[0](c, pix1, pix2, line_size, h); | |
1192 else score1 = sse16_mmx(c, pix1, pix2, line_size, h); | |
1193 score2= hf_noise16_mmx(pix1, line_size, h) - hf_noise16_mmx(pix2, line_size, h); | |
2067 | 1194 |
4001 | 1195 if(c) return score1 + FFABS(score2)*c->avctx->nsse_weight; |
1196 else return score1 + FFABS(score2)*8; | |
2067 | 1197 } |
1198 | |
2864
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1199 static int nsse8_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1200 MpegEncContext *c = p; |
2067 | 1201 int score1= sse8_mmx(c, pix1, pix2, line_size, h); |
1202 int score2= hf_noise8_mmx(pix1, line_size, h) - hf_noise8_mmx(pix2, line_size, h); | |
1203 | |
4001 | 1204 if(c) return score1 + FFABS(score2)*c->avctx->nsse_weight; |
1205 else return score1 + FFABS(score2)*8; | |
2067 | 1206 } |
1207 | |
1729 | 1208 static int vsad_intra16_mmx(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) { |
1209 int tmp; | |
2967 | 1210 |
1729 | 1211 assert( (((int)pix) & 7) == 0); |
1212 assert((line_size &7) ==0); | |
2967 | 1213 |
1729 | 1214 #define SUM(in0, in1, out0, out1) \ |
1215 "movq (%0), %%mm2\n"\ | |
1216 "movq 8(%0), %%mm3\n"\ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1217 "add %2,%0\n"\ |
1729 | 1218 "movq %%mm2, " #out0 "\n"\ |
1219 "movq %%mm3, " #out1 "\n"\ | |
1220 "psubusb " #in0 ", %%mm2\n"\ | |
1221 "psubusb " #in1 ", %%mm3\n"\ | |
1222 "psubusb " #out0 ", " #in0 "\n"\ | |
1223 "psubusb " #out1 ", " #in1 "\n"\ | |
1224 "por %%mm2, " #in0 "\n"\ | |
1225 "por %%mm3, " #in1 "\n"\ | |
1226 "movq " #in0 ", %%mm2\n"\ | |
1227 "movq " #in1 ", %%mm3\n"\ | |
1228 "punpcklbw %%mm7, " #in0 "\n"\ | |
1229 "punpcklbw %%mm7, " #in1 "\n"\ | |
1230 "punpckhbw %%mm7, %%mm2\n"\ | |
1231 "punpckhbw %%mm7, %%mm3\n"\ | |
1232 "paddw " #in1 ", " #in0 "\n"\ | |
1233 "paddw %%mm3, %%mm2\n"\ | |
1234 "paddw %%mm2, " #in0 "\n"\ | |
1235 "paddw " #in0 ", %%mm6\n" | |
1236 | |
2967 | 1237 |
1729 | 1238 asm volatile ( |
1239 "movl %3,%%ecx\n" | |
1240 "pxor %%mm6,%%mm6\n" | |
1241 "pxor %%mm7,%%mm7\n" | |
1242 "movq (%0),%%mm0\n" | |
1243 "movq 8(%0),%%mm1\n" | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1244 "add %2,%0\n" |
1729 | 1245 "subl $2, %%ecx\n" |
1246 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1247 "1:\n" | |
2967 | 1248 |
1729 | 1249 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1250 |
1729 | 1251 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1252 |
1729 | 1253 "subl $2, %%ecx\n" |
1254 "jnz 1b\n" | |
1255 | |
1256 "movq %%mm6,%%mm0\n" | |
1257 "psrlq $32, %%mm6\n" | |
1258 "paddw %%mm6,%%mm0\n" | |
1259 "movq %%mm0,%%mm6\n" | |
1260 "psrlq $16, %%mm0\n" | |
1261 "paddw %%mm6,%%mm0\n" | |
1262 "movd %%mm0,%1\n" | |
2967 | 1263 : "+r" (pix), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1264 : "r" ((long)line_size) , "m" (h) |
1729 | 1265 : "%ecx"); |
1266 return tmp & 0xFFFF; | |
1267 } | |
1268 #undef SUM | |
1269 | |
1270 static int vsad_intra16_mmx2(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) { | |
1271 int tmp; | |
2967 | 1272 |
1729 | 1273 assert( (((int)pix) & 7) == 0); |
1274 assert((line_size &7) ==0); | |
2967 | 1275 |
1729 | 1276 #define SUM(in0, in1, out0, out1) \ |
1277 "movq (%0), " #out0 "\n"\ | |
1278 "movq 8(%0), " #out1 "\n"\ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1279 "add %2,%0\n"\ |
1729 | 1280 "psadbw " #out0 ", " #in0 "\n"\ |
1281 "psadbw " #out1 ", " #in1 "\n"\ | |
1282 "paddw " #in1 ", " #in0 "\n"\ | |
1283 "paddw " #in0 ", %%mm6\n" | |
1284 | |
1285 asm volatile ( | |
1286 "movl %3,%%ecx\n" | |
1287 "pxor %%mm6,%%mm6\n" | |
1288 "pxor %%mm7,%%mm7\n" | |
1289 "movq (%0),%%mm0\n" | |
1290 "movq 8(%0),%%mm1\n" | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1291 "add %2,%0\n" |
1729 | 1292 "subl $2, %%ecx\n" |
1293 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1294 "1:\n" | |
2967 | 1295 |
1729 | 1296 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1297 |
1729 | 1298 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1299 |
1729 | 1300 "subl $2, %%ecx\n" |
1301 "jnz 1b\n" | |
1302 | |
1303 "movd %%mm6,%1\n" | |
2967 | 1304 : "+r" (pix), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1305 : "r" ((long)line_size) , "m" (h) |
1729 | 1306 : "%ecx"); |
1307 return tmp; | |
1308 } | |
1309 #undef SUM | |
1310 | |
1311 static int vsad16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { | |
1312 int tmp; | |
2967 | 1313 |
1729 | 1314 assert( (((int)pix1) & 7) == 0); |
1315 assert( (((int)pix2) & 7) == 0); | |
1316 assert((line_size &7) ==0); | |
2967 | 1317 |
1729 | 1318 #define SUM(in0, in1, out0, out1) \ |
1319 "movq (%0),%%mm2\n"\ | |
1320 "movq (%1)," #out0 "\n"\ | |
1321 "movq 8(%0),%%mm3\n"\ | |
1322 "movq 8(%1)," #out1 "\n"\ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1323 "add %3,%0\n"\ |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1324 "add %3,%1\n"\ |
1729 | 1325 "psubb " #out0 ", %%mm2\n"\ |
1326 "psubb " #out1 ", %%mm3\n"\ | |
1327 "pxor %%mm7, %%mm2\n"\ | |
1328 "pxor %%mm7, %%mm3\n"\ | |
1329 "movq %%mm2, " #out0 "\n"\ | |
1330 "movq %%mm3, " #out1 "\n"\ | |
1331 "psubusb " #in0 ", %%mm2\n"\ | |
1332 "psubusb " #in1 ", %%mm3\n"\ | |
1333 "psubusb " #out0 ", " #in0 "\n"\ | |
1334 "psubusb " #out1 ", " #in1 "\n"\ | |
1335 "por %%mm2, " #in0 "\n"\ | |
1336 "por %%mm3, " #in1 "\n"\ | |
1337 "movq " #in0 ", %%mm2\n"\ | |
1338 "movq " #in1 ", %%mm3\n"\ | |
1339 "punpcklbw %%mm7, " #in0 "\n"\ | |
1340 "punpcklbw %%mm7, " #in1 "\n"\ | |
1341 "punpckhbw %%mm7, %%mm2\n"\ | |
1342 "punpckhbw %%mm7, %%mm3\n"\ | |
1343 "paddw " #in1 ", " #in0 "\n"\ | |
1344 "paddw %%mm3, %%mm2\n"\ | |
1345 "paddw %%mm2, " #in0 "\n"\ | |
1346 "paddw " #in0 ", %%mm6\n" | |
1347 | |
2967 | 1348 |
1729 | 1349 asm volatile ( |
1350 "movl %4,%%ecx\n" | |
1351 "pxor %%mm6,%%mm6\n" | |
1352 "pcmpeqw %%mm7,%%mm7\n" | |
1353 "psllw $15, %%mm7\n" | |
1354 "packsswb %%mm7, %%mm7\n" | |
1355 "movq (%0),%%mm0\n" | |
1356 "movq (%1),%%mm2\n" | |
1357 "movq 8(%0),%%mm1\n" | |
1358 "movq 8(%1),%%mm3\n" | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1359 "add %3,%0\n" |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1360 "add %3,%1\n" |
1729 | 1361 "subl $2, %%ecx\n" |
1362 "psubb %%mm2, %%mm0\n" | |
1363 "psubb %%mm3, %%mm1\n" | |
1364 "pxor %%mm7, %%mm0\n" | |
1365 "pxor %%mm7, %%mm1\n" | |
1366 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1367 "1:\n" | |
2967 | 1368 |
1729 | 1369 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1370 |
1729 | 1371 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1372 |
1729 | 1373 "subl $2, %%ecx\n" |
1374 "jnz 1b\n" | |
1375 | |
1376 "movq %%mm6,%%mm0\n" | |
1377 "psrlq $32, %%mm6\n" | |
1378 "paddw %%mm6,%%mm0\n" | |
1379 "movq %%mm0,%%mm6\n" | |
1380 "psrlq $16, %%mm0\n" | |
1381 "paddw %%mm6,%%mm0\n" | |
1382 "movd %%mm0,%2\n" | |
2967 | 1383 : "+r" (pix1), "+r" (pix2), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1384 : "r" ((long)line_size) , "m" (h) |
1729 | 1385 : "%ecx"); |
1386 return tmp & 0x7FFF; | |
1387 } | |
1388 #undef SUM | |
1389 | |
1390 static int vsad16_mmx2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { | |
1391 int tmp; | |
2967 | 1392 |
1729 | 1393 assert( (((int)pix1) & 7) == 0); |
1394 assert( (((int)pix2) & 7) == 0); | |
1395 assert((line_size &7) ==0); | |
2967 | 1396 |
1729 | 1397 #define SUM(in0, in1, out0, out1) \ |
1398 "movq (%0)," #out0 "\n"\ | |
1399 "movq (%1),%%mm2\n"\ | |
1400 "movq 8(%0)," #out1 "\n"\ | |
1401 "movq 8(%1),%%mm3\n"\ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1402 "add %3,%0\n"\ |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1403 "add %3,%1\n"\ |
1729 | 1404 "psubb %%mm2, " #out0 "\n"\ |
1405 "psubb %%mm3, " #out1 "\n"\ | |
1406 "pxor %%mm7, " #out0 "\n"\ | |
1407 "pxor %%mm7, " #out1 "\n"\ | |
1408 "psadbw " #out0 ", " #in0 "\n"\ | |
1409 "psadbw " #out1 ", " #in1 "\n"\ | |
1410 "paddw " #in1 ", " #in0 "\n"\ | |
1411 "paddw " #in0 ", %%mm6\n" | |
1412 | |
1413 asm volatile ( | |
1414 "movl %4,%%ecx\n" | |
1415 "pxor %%mm6,%%mm6\n" | |
1416 "pcmpeqw %%mm7,%%mm7\n" | |
1417 "psllw $15, %%mm7\n" | |
1418 "packsswb %%mm7, %%mm7\n" | |
1419 "movq (%0),%%mm0\n" | |
1420 "movq (%1),%%mm2\n" | |
1421 "movq 8(%0),%%mm1\n" | |
1422 "movq 8(%1),%%mm3\n" | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1423 "add %3,%0\n" |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1424 "add %3,%1\n" |
1729 | 1425 "subl $2, %%ecx\n" |
1426 "psubb %%mm2, %%mm0\n" | |
1427 "psubb %%mm3, %%mm1\n" | |
1428 "pxor %%mm7, %%mm0\n" | |
1429 "pxor %%mm7, %%mm1\n" | |
1430 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1431 "1:\n" | |
2967 | 1432 |
1729 | 1433 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1434 |
1729 | 1435 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1436 |
1729 | 1437 "subl $2, %%ecx\n" |
1438 "jnz 1b\n" | |
1439 | |
1440 "movd %%mm6,%2\n" | |
2967 | 1441 : "+r" (pix1), "+r" (pix2), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1442 : "r" ((long)line_size) , "m" (h) |
1729 | 1443 : "%ecx"); |
1444 return tmp; | |
1445 } | |
1446 #undef SUM | |
1447 | |
866 | 1448 static void diff_bytes_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1449 long i=0; |
866 | 1450 asm volatile( |
2979 | 1451 "1: \n\t" |
1452 "movq (%2, %0), %%mm0 \n\t" | |
1453 "movq (%1, %0), %%mm1 \n\t" | |
1454 "psubb %%mm0, %%mm1 \n\t" | |
1455 "movq %%mm1, (%3, %0) \n\t" | |
1456 "movq 8(%2, %0), %%mm0 \n\t" | |
1457 "movq 8(%1, %0), %%mm1 \n\t" | |
1458 "psubb %%mm0, %%mm1 \n\t" | |
1459 "movq %%mm1, 8(%3, %0) \n\t" | |
1460 "add $16, %0 \n\t" | |
1461 "cmp %4, %0 \n\t" | |
1462 " jb 1b \n\t" | |
866 | 1463 : "+r" (i) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1464 : "r"(src1), "r"(src2), "r"(dst), "r"((long)w-15) |
866 | 1465 ); |
1466 for(; i<w; i++) | |
1467 dst[i+0] = src1[i+0]-src2[i+0]; | |
1468 } | |
1527 | 1469 |
1470 static void sub_hfyu_median_prediction_mmx2(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top){ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1471 long i=0; |
1527 | 1472 uint8_t l, lt; |
2967 | 1473 |
1527 | 1474 asm volatile( |
2979 | 1475 "1: \n\t" |
1476 "movq -1(%1, %0), %%mm0 \n\t" // LT | |
1477 "movq (%1, %0), %%mm1 \n\t" // T | |
1478 "movq -1(%2, %0), %%mm2 \n\t" // L | |
1479 "movq (%2, %0), %%mm3 \n\t" // X | |
1480 "movq %%mm2, %%mm4 \n\t" // L | |
1481 "psubb %%mm0, %%mm2 \n\t" | |
1482 "paddb %%mm1, %%mm2 \n\t" // L + T - LT | |
1483 "movq %%mm4, %%mm5 \n\t" // L | |
1484 "pmaxub %%mm1, %%mm4 \n\t" // max(T, L) | |
1485 "pminub %%mm5, %%mm1 \n\t" // min(T, L) | |
1486 "pminub %%mm2, %%mm4 \n\t" | |
1487 "pmaxub %%mm1, %%mm4 \n\t" | |
1488 "psubb %%mm4, %%mm3 \n\t" // dst - pred | |
1489 "movq %%mm3, (%3, %0) \n\t" | |
1490 "add $8, %0 \n\t" | |
1491 "cmp %4, %0 \n\t" | |
1492 " jb 1b \n\t" | |
1527 | 1493 : "+r" (i) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1494 : "r"(src1), "r"(src2), "r"(dst), "r"((long)w) |
1527 | 1495 ); |
1496 | |
1497 l= *left; | |
1498 lt= *left_top; | |
2967 | 1499 |
1527 | 1500 dst[0]= src2[0] - mid_pred(l, src1[0], (l + src1[0] - lt)&0xFF); |
2967 | 1501 |
1527 | 1502 *left_top= src1[w-1]; |
1503 *left = src2[w-1]; | |
1504 } | |
1505 | |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1506 #define DIFF_PIXELS_1(m,a,t,p1,p2)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1507 "mov"#m" "#p1", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1508 "mov"#m" "#p2", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1509 "punpcklbw "#a", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1510 "punpcklbw "#a", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1511 "psubw "#t", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1512 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1513 #define DIFF_PIXELS_8(m0,m1,mm,p1,p2,stride,temp) {\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1514 uint8_t *p1b=p1, *p2b=p2;\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1515 asm volatile(\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1516 DIFF_PIXELS_1(m0, mm##0, mm##7, (%1), (%2))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1517 DIFF_PIXELS_1(m0, mm##1, mm##7, (%1,%3), (%2,%3))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1518 DIFF_PIXELS_1(m0, mm##2, mm##7, (%1,%3,2), (%2,%3,2))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1519 "add %4, %1 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1520 "add %4, %2 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1521 DIFF_PIXELS_1(m0, mm##3, mm##7, (%1), (%2))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1522 DIFF_PIXELS_1(m0, mm##4, mm##7, (%1,%3), (%2,%3))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1523 DIFF_PIXELS_1(m0, mm##5, mm##7, (%1,%3,2), (%2,%3,2))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1524 DIFF_PIXELS_1(m0, mm##6, mm##7, (%1,%4), (%2,%4))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1525 "mov"#m1" "#mm"0, %0 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1526 DIFF_PIXELS_1(m0, mm##7, mm##0, (%1,%3,4), (%2,%3,4))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1527 "mov"#m1" %0, "#mm"0 \n\t"\ |
5912
f75ee7ea171b
tring to workaround gcc 2.95 bug which causes random failures
michael
parents:
5737
diff
changeset
|
1528 : "+m"(temp), "+r"(p1b), "+r"(p2b)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1529 : "r"((long)stride), "r"((long)stride*3)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1530 );\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1531 } |
5912
f75ee7ea171b
tring to workaround gcc 2.95 bug which causes random failures
michael
parents:
5737
diff
changeset
|
1532 //the "+m"(temp) is needed as gcc 2.95 sometimes fails to compile "=m"(temp) |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1533 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1534 #define DIFF_PIXELS_4x8(p1,p2,stride,temp) DIFF_PIXELS_8(d, q, %%mm, p1, p2, stride, temp) |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1535 #define DIFF_PIXELS_8x8(p1,p2,stride,temp) DIFF_PIXELS_8(q, dqa, %%xmm, p1, p2, stride, temp) |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1536 |
1153 | 1537 #define LBUTTERFLY2(a1,b1,a2,b2)\ |
2979 | 1538 "paddw " #b1 ", " #a1 " \n\t"\ |
1539 "paddw " #b2 ", " #a2 " \n\t"\ | |
1540 "paddw " #b1 ", " #b1 " \n\t"\ | |
1541 "paddw " #b2 ", " #b2 " \n\t"\ | |
1542 "psubw " #a1 ", " #b1 " \n\t"\ | |
1543 "psubw " #a2 ", " #b2 " \n\t" | |
866 | 1544 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1545 #define HADAMARD8(m0, m1, m2, m3, m4, m5, m6, m7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1546 LBUTTERFLY2(m0, m1, m2, m3)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1547 LBUTTERFLY2(m4, m5, m6, m7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1548 LBUTTERFLY2(m0, m2, m1, m3)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1549 LBUTTERFLY2(m4, m6, m5, m7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1550 LBUTTERFLY2(m0, m4, m1, m5)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1551 LBUTTERFLY2(m2, m6, m3, m7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1552 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1553 #define HADAMARD48 HADAMARD8(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm6, %%mm7) |
936 | 1554 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1555 #define MMABS_MMX(a,z)\ |
2979 | 1556 "pxor " #z ", " #z " \n\t"\ |
1557 "pcmpgtw " #a ", " #z " \n\t"\ | |
1558 "pxor " #z ", " #a " \n\t"\ | |
1559 "psubw " #z ", " #a " \n\t" | |
936 | 1560 |
1153 | 1561 #define MMABS_MMX2(a,z)\ |
2979 | 1562 "pxor " #z ", " #z " \n\t"\ |
1563 "psubw " #a ", " #z " \n\t"\ | |
1564 "pmaxsw " #z ", " #a " \n\t" | |
1153 | 1565 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1566 #define MMABS_SSSE3(a,z)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1567 "pabsw " #a ", " #a " \n\t" |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1568 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1569 #define MMABS_SUM(a,z, sum)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1570 MMABS(a,z)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1571 "paddusw " #a ", " #sum " \n\t" |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1572 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1573 #define MMABS_SUM_8x8_NOSPILL\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1574 MMABS(%%xmm0, %%xmm8)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1575 MMABS(%%xmm1, %%xmm9)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1576 MMABS_SUM(%%xmm2, %%xmm8, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1577 MMABS_SUM(%%xmm3, %%xmm9, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1578 MMABS_SUM(%%xmm4, %%xmm8, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1579 MMABS_SUM(%%xmm5, %%xmm9, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1580 MMABS_SUM(%%xmm6, %%xmm8, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1581 MMABS_SUM(%%xmm7, %%xmm9, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1582 "paddusw %%xmm1, %%xmm0 \n\t" |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1583 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1584 #ifdef ARCH_X86_64 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1585 #define MMABS_SUM_8x8_SSE2 MMABS_SUM_8x8_NOSPILL |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1586 #else |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1587 #define MMABS_SUM_8x8_SSE2\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1588 "movdqa %%xmm7, (%1) \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1589 MMABS(%%xmm0, %%xmm7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1590 MMABS(%%xmm1, %%xmm7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1591 MMABS_SUM(%%xmm2, %%xmm7, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1592 MMABS_SUM(%%xmm3, %%xmm7, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1593 MMABS_SUM(%%xmm4, %%xmm7, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1594 MMABS_SUM(%%xmm5, %%xmm7, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1595 MMABS_SUM(%%xmm6, %%xmm7, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1596 "movdqa (%1), %%xmm2 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1597 MMABS_SUM(%%xmm2, %%xmm7, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1598 "paddusw %%xmm1, %%xmm0 \n\t" |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1599 #endif |
2967 | 1600 |
936 | 1601 #define LOAD4(o, a, b, c, d)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1602 "movq "#o"(%1), "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1603 "movq "#o"+8(%1), "#b" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1604 "movq "#o"+16(%1), "#c" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1605 "movq "#o"+24(%1), "#d" \n\t"\ |
936 | 1606 |
1607 #define STORE4(o, a, b, c, d)\ | |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1608 "movq "#a", "#o"(%1) \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1609 "movq "#b", "#o"+8(%1) \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1610 "movq "#c", "#o"+16(%1) \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1611 "movq "#d", "#o"+24(%1) \n\t"\ |
936 | 1612 |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1613 /* FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get up to |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1614 * about 100k on extreme inputs. But that's very unlikely to occur in natural video, |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1615 * and it's even more unlikely to not have any alternative mvs/modes with lower cost. */ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1616 #define HSUM_MMX(a, t, dst)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1617 "movq "#a", "#t" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1618 "psrlq $32, "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1619 "paddusw "#t", "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1620 "movq "#a", "#t" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1621 "psrlq $16, "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1622 "paddusw "#t", "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1623 "movd "#a", "#dst" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1624 |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1625 #define HSUM_MMX2(a, t, dst)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1626 "pshufw $0x0E, "#a", "#t" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1627 "paddusw "#t", "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1628 "pshufw $0x01, "#a", "#t" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1629 "paddusw "#t", "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1630 "movd "#a", "#dst" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1631 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1632 #define HSUM_SSE2(a, t, dst)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1633 "movhlps "#a", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1634 "paddusw "#t", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1635 "pshuflw $0x0E, "#a", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1636 "paddusw "#t", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1637 "pshuflw $0x01, "#a", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1638 "paddusw "#t", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1639 "movd "#a", "#dst" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1640 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1641 #define HADAMARD8_DIFF_MMX(cpu) \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1642 static int hadamard8_diff_##cpu(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1643 DECLARE_ALIGNED_8(uint64_t, temp[13]);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1644 int sum;\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1645 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1646 assert(h==8);\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1647 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1648 DIFF_PIXELS_4x8(src1, src2, stride, temp[0]);\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1649 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1650 asm volatile(\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1651 HADAMARD48\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1652 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1653 "movq %%mm7, 96(%1) \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1654 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1655 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1656 STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1657 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1658 "movq 96(%1), %%mm7 \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1659 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1660 STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1661 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1662 : "=r" (sum)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1663 : "r"(temp)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1664 );\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1665 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1666 DIFF_PIXELS_4x8(src1+4, src2+4, stride, temp[4]);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1667 \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1668 asm volatile(\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1669 HADAMARD48\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1670 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1671 "movq %%mm7, 96(%1) \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1672 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1673 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1674 STORE4(32, %%mm0, %%mm3, %%mm7, %%mm2)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1675 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1676 "movq 96(%1), %%mm7 \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1677 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1678 "movq %%mm7, %%mm5 \n\t"/*FIXME remove*/\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1679 "movq %%mm6, %%mm7 \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1680 "movq %%mm0, %%mm6 \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1681 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1682 LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1683 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1684 HADAMARD48\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1685 "movq %%mm7, 64(%1) \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1686 MMABS(%%mm0, %%mm7)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1687 MMABS(%%mm1, %%mm7)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1688 MMABS_SUM(%%mm2, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1689 MMABS_SUM(%%mm3, %%mm7, %%mm1)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1690 MMABS_SUM(%%mm4, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1691 MMABS_SUM(%%mm5, %%mm7, %%mm1)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1692 MMABS_SUM(%%mm6, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1693 "movq 64(%1), %%mm2 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1694 MMABS_SUM(%%mm2, %%mm7, %%mm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1695 "paddusw %%mm1, %%mm0 \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1696 "movq %%mm0, 64(%1) \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1697 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1698 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1699 LOAD4(32, %%mm4, %%mm5, %%mm6, %%mm7)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1700 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1701 HADAMARD48\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1702 "movq %%mm7, (%1) \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1703 MMABS(%%mm0, %%mm7)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1704 MMABS(%%mm1, %%mm7)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1705 MMABS_SUM(%%mm2, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1706 MMABS_SUM(%%mm3, %%mm7, %%mm1)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1707 MMABS_SUM(%%mm4, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1708 MMABS_SUM(%%mm5, %%mm7, %%mm1)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1709 MMABS_SUM(%%mm6, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1710 "movq (%1), %%mm2 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1711 MMABS_SUM(%%mm2, %%mm7, %%mm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1712 "paddusw 64(%1), %%mm0 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1713 "paddusw %%mm1, %%mm0 \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1714 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1715 HSUM(%%mm0, %%mm1, %0)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1716 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1717 : "=r" (sum)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1718 : "r"(temp)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1719 );\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1720 return sum&0xFFFF;\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1721 }\ |
6056
558c1fd0ee72
Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents:
6030
diff
changeset
|
1722 WRAPPER8_16_SQ(hadamard8_diff_##cpu, hadamard8_diff16_##cpu) |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1723 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1724 #define HADAMARD8_DIFF_SSE2(cpu) \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1725 static int hadamard8_diff_##cpu(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1726 DECLARE_ALIGNED_16(uint64_t, temp[4]);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1727 int sum;\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1728 \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1729 assert(h==8);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1730 \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1731 DIFF_PIXELS_8x8(src1, src2, stride, temp[0]);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1732 \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1733 asm volatile(\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1734 HADAMARD8(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm6, %%xmm7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1735 TRANSPOSE8(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm6, %%xmm7, (%1))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1736 HADAMARD8(%%xmm0, %%xmm5, %%xmm7, %%xmm3, %%xmm6, %%xmm4, %%xmm2, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1737 MMABS_SUM_8x8\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1738 HSUM_SSE2(%%xmm0, %%xmm1, %0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1739 : "=r" (sum)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1740 : "r"(temp)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1741 );\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1742 return sum&0xFFFF;\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1743 }\ |
6056
558c1fd0ee72
Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents:
6030
diff
changeset
|
1744 WRAPPER8_16_SQ(hadamard8_diff_##cpu, hadamard8_diff16_##cpu) |
936 | 1745 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1746 #define MMABS(a,z) MMABS_MMX(a,z) |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1747 #define HSUM(a,t,dst) HSUM_MMX(a,t,dst) |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1748 HADAMARD8_DIFF_MMX(mmx) |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1749 #undef MMABS |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1750 #undef HSUM |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1751 |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1752 #define MMABS(a,z) MMABS_MMX2(a,z) |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1753 #define MMABS_SUM_8x8 MMABS_SUM_8x8_SSE2 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1754 #define HSUM(a,t,dst) HSUM_MMX2(a,t,dst) |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1755 HADAMARD8_DIFF_MMX(mmx2) |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1756 HADAMARD8_DIFF_SSE2(sse2) |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1757 #undef MMABS |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1758 #undef MMABS_SUM_8x8 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1759 #undef HSUM |
1153 | 1760 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1761 #ifdef HAVE_SSSE3 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1762 #define MMABS(a,z) MMABS_SSSE3(a,z) |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1763 #define MMABS_SUM_8x8 MMABS_SUM_8x8_NOSPILL |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1764 HADAMARD8_DIFF_SSE2(ssse3) |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1765 #undef MMABS |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1766 #undef MMABS_SUM_8x8 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1767 #endif |
4749 | 1768 |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1769 #define DCT_SAD4(m,mm,o)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1770 "mov"#m" "#o"+ 0(%1), "#mm"2 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1771 "mov"#m" "#o"+16(%1), "#mm"3 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1772 "mov"#m" "#o"+32(%1), "#mm"4 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1773 "mov"#m" "#o"+48(%1), "#mm"5 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1774 MMABS_SUM(mm##2, mm##6, mm##0)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1775 MMABS_SUM(mm##3, mm##7, mm##1)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1776 MMABS_SUM(mm##4, mm##6, mm##0)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1777 MMABS_SUM(mm##5, mm##7, mm##1)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1778 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1779 #define DCT_SAD_MMX\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1780 "pxor %%mm0, %%mm0 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1781 "pxor %%mm1, %%mm1 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1782 DCT_SAD4(q, %%mm, 0)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1783 DCT_SAD4(q, %%mm, 8)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1784 DCT_SAD4(q, %%mm, 64)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1785 DCT_SAD4(q, %%mm, 72)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1786 "paddusw %%mm1, %%mm0 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1787 HSUM(%%mm0, %%mm1, %0) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1788 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1789 #define DCT_SAD_SSE2\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1790 "pxor %%xmm0, %%xmm0 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1791 "pxor %%xmm1, %%xmm1 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1792 DCT_SAD4(dqa, %%xmm, 0)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1793 DCT_SAD4(dqa, %%xmm, 64)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1794 "paddusw %%xmm1, %%xmm0 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1795 HSUM(%%xmm0, %%xmm1, %0) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1796 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1797 #define DCT_SAD_FUNC(cpu) \ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1798 static int sum_abs_dctelem_##cpu(DCTELEM *block){\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1799 int sum;\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1800 asm volatile(\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1801 DCT_SAD\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1802 :"=r"(sum)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1803 :"r"(block)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1804 );\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1805 return sum&0xFFFF;\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1806 } |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1807 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1808 #define DCT_SAD DCT_SAD_MMX |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1809 #define HSUM(a,t,dst) HSUM_MMX(a,t,dst) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1810 #define MMABS(a,z) MMABS_MMX(a,z) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1811 DCT_SAD_FUNC(mmx) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1812 #undef MMABS |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1813 #undef HSUM |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1814 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1815 #define HSUM(a,t,dst) HSUM_MMX2(a,t,dst) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1816 #define MMABS(a,z) MMABS_MMX2(a,z) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1817 DCT_SAD_FUNC(mmx2) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1818 #undef HSUM |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1819 #undef DCT_SAD |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1820 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1821 #define DCT_SAD DCT_SAD_SSE2 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1822 #define HSUM(a,t,dst) HSUM_SSE2(a,t,dst) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1823 DCT_SAD_FUNC(sse2) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1824 #undef MMABS |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1825 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1826 #ifdef HAVE_SSSE3 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1827 #define MMABS(a,z) MMABS_SSSE3(a,z) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1828 DCT_SAD_FUNC(ssse3) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1829 #undef MMABS |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1830 #endif |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1831 #undef HSUM |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1832 #undef DCT_SAD |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1833 |
5255 | 1834 static int ssd_int8_vs_int16_mmx(const int8_t *pix1, const int16_t *pix2, int size){ |
4749 | 1835 int sum; |
1836 long i=size; | |
1837 asm volatile( | |
1838 "pxor %%mm4, %%mm4 \n" | |
1839 "1: \n" | |
1840 "sub $8, %0 \n" | |
1841 "movq (%2,%0), %%mm2 \n" | |
1842 "movq (%3,%0,2), %%mm0 \n" | |
1843 "movq 8(%3,%0,2), %%mm1 \n" | |
1844 "punpckhbw %%mm2, %%mm3 \n" | |
1845 "punpcklbw %%mm2, %%mm2 \n" | |
1846 "psraw $8, %%mm3 \n" | |
1847 "psraw $8, %%mm2 \n" | |
1848 "psubw %%mm3, %%mm1 \n" | |
1849 "psubw %%mm2, %%mm0 \n" | |
1850 "pmaddwd %%mm1, %%mm1 \n" | |
1851 "pmaddwd %%mm0, %%mm0 \n" | |
1852 "paddd %%mm1, %%mm4 \n" | |
1853 "paddd %%mm0, %%mm4 \n" | |
1854 "jg 1b \n" | |
1855 "movq %%mm4, %%mm3 \n" | |
1856 "psrlq $32, %%mm3 \n" | |
1857 "paddd %%mm3, %%mm4 \n" | |
1858 "movd %%mm4, %1 \n" | |
1859 :"+r"(i), "=r"(sum) | |
1860 :"r"(pix1), "r"(pix2) | |
1861 ); | |
1862 return sum; | |
1863 } | |
1864 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
1865 #endif //CONFIG_ENCODERS |
866 | 1866 |
959 | 1867 #define put_no_rnd_pixels8_mmx(a,b,c,d) put_pixels8_mmx(a,b,c,d) |
1868 #define put_no_rnd_pixels16_mmx(a,b,c,d) put_pixels16_mmx(a,b,c,d) | |
1869 | |
954 | 1870 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\ |
2979 | 1871 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\ |
1872 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\ | |
1873 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\ | |
1874 "movq "#in7", " #m3 " \n\t" /* d */\ | |
1875 "movq "#in0", %%mm5 \n\t" /* D */\ | |
1876 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\ | |
1877 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\ | |
1878 "movq "#in1", %%mm5 \n\t" /* C */\ | |
1879 "movq "#in2", %%mm6 \n\t" /* B */\ | |
1880 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\ | |
1881 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\ | |
1882 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\ | |
1883 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\ | |
1884 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\ | |
1885 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\ | |
1886 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\ | |
1887 "psraw $5, %%mm5 \n\t"\ | |
1888 "packuswb %%mm5, %%mm5 \n\t"\ | |
954 | 1889 OP(%%mm5, out, %%mm7, d) |
1890 | |
959 | 1891 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\ |
1057 | 1892 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
954 | 1893 uint64_t temp;\ |
1894 \ | |
1895 asm volatile(\ | |
2979 | 1896 "pxor %%mm7, %%mm7 \n\t"\ |
1897 "1: \n\t"\ | |
1898 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
1899 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
1900 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
1901 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
1902 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
1903 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
1904 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
1905 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
1906 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
1907 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
1908 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
1909 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
1910 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
1911 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
1912 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
1913 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
1914 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
1915 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1916 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
1917 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
1918 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
1919 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
1920 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
1921 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
1922 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
1923 "paddw %6, %%mm6 \n\t"\ | |
1924 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1925 "psraw $5, %%mm0 \n\t"\ | |
1926 "movq %%mm0, %5 \n\t"\ | |
954 | 1927 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ |
1928 \ | |
2979 | 1929 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\ |
1930 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\ | |
1931 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\ | |
1932 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\ | |
1933 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\ | |
1934 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\ | |
1935 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\ | |
1936 "paddw %%mm0, %%mm2 \n\t" /* b */\ | |
1937 "paddw %%mm5, %%mm3 \n\t" /* c */\ | |
1938 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
1939 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
1940 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\ | |
1941 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\ | |
1942 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\ | |
1943 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\ | |
1944 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
1945 "paddw %%mm2, %%mm1 \n\t" /* a */\ | |
1946 "paddw %%mm6, %%mm4 \n\t" /* d */\ | |
1947 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
1948 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\ | |
1949 "paddw %6, %%mm1 \n\t"\ | |
1950 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\ | |
1951 "psraw $5, %%mm3 \n\t"\ | |
1952 "movq %5, %%mm1 \n\t"\ | |
1953 "packuswb %%mm3, %%mm1 \n\t"\ | |
959 | 1954 OP_MMX2(%%mm1, (%1),%%mm4, q)\ |
954 | 1955 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\ |
1956 \ | |
2979 | 1957 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\ |
1958 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\ | |
1959 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\ | |
1960 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\ | |
1961 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\ | |
1962 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\ | |
1963 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\ | |
1964 "paddw %%mm1, %%mm5 \n\t" /* b */\ | |
1965 "paddw %%mm4, %%mm0 \n\t" /* c */\ | |
1966 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1967 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\ | |
1968 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\ | |
1969 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\ | |
1970 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\ | |
1971 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\ | |
1972 "paddw %%mm3, %%mm2 \n\t" /* d */\ | |
1973 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\ | |
1974 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\ | |
1975 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\ | |
1976 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\ | |
1977 "paddw %%mm2, %%mm6 \n\t" /* a */\ | |
1978 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\ | |
1979 "paddw %6, %%mm0 \n\t"\ | |
1980 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1981 "psraw $5, %%mm0 \n\t"\ | |
954 | 1982 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\ |
1983 \ | |
2979 | 1984 "paddw %%mm5, %%mm3 \n\t" /* a */\ |
1985 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\ | |
1986 "paddw %%mm4, %%mm6 \n\t" /* b */\ | |
1987 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\ | |
1988 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\ | |
1989 "paddw %%mm1, %%mm4 \n\t" /* c */\ | |
1990 "paddw %%mm2, %%mm5 \n\t" /* d */\ | |
1991 "paddw %%mm6, %%mm6 \n\t" /* 2b */\ | |
1992 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\ | |
1993 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\ | |
1994 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\ | |
1995 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\ | |
1996 "paddw %6, %%mm4 \n\t"\ | |
1997 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\ | |
1998 "psraw $5, %%mm4 \n\t"\ | |
1999 "packuswb %%mm4, %%mm0 \n\t"\ | |
959 | 2000 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\ |
954 | 2001 \ |
2979 | 2002 "add %3, %0 \n\t"\ |
2003 "add %4, %1 \n\t"\ | |
2004 "decl %2 \n\t"\ | |
2005 " jnz 1b \n\t"\ | |
967 | 2006 : "+a"(src), "+c"(dst), "+m"(h)\ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
2007 : "d"((long)srcStride), "S"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
966 | 2008 : "memory"\ |
954 | 2009 );\ |
2010 }\ | |
2011 \ | |
2012 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
2013 int i;\ | |
2014 int16_t temp[16];\ | |
2015 /* quick HACK, XXX FIXME MUST be optimized */\ | |
2016 for(i=0; i<h; i++)\ | |
2017 {\ | |
2018 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
2019 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
2020 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
2021 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
2022 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
2023 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\ | |
2024 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\ | |
2025 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\ | |
2026 temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\ | |
2027 temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\ | |
2028 temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\ | |
2029 temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\ | |
2030 temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\ | |
2031 temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\ | |
2032 temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\ | |
2033 temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\ | |
2034 asm volatile(\ | |
2979 | 2035 "movq (%0), %%mm0 \n\t"\ |
2036 "movq 8(%0), %%mm1 \n\t"\ | |
2037 "paddw %2, %%mm0 \n\t"\ | |
2038 "paddw %2, %%mm1 \n\t"\ | |
2039 "psraw $5, %%mm0 \n\t"\ | |
2040 "psraw $5, %%mm1 \n\t"\ | |
2041 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 2042 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ |
2979 | 2043 "movq 16(%0), %%mm0 \n\t"\ |
2044 "movq 24(%0), %%mm1 \n\t"\ | |
2045 "paddw %2, %%mm0 \n\t"\ | |
2046 "paddw %2, %%mm1 \n\t"\ | |
2047 "psraw $5, %%mm0 \n\t"\ | |
2048 "psraw $5, %%mm1 \n\t"\ | |
2049 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 2050 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\ |
954 | 2051 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ |
966 | 2052 : "memory"\ |
954 | 2053 );\ |
2054 dst+=dstStride;\ | |
2055 src+=srcStride;\ | |
2056 }\ | |
2057 }\ | |
2058 \ | |
1057 | 2059 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
959 | 2060 uint64_t temp;\ |
2061 \ | |
2062 asm volatile(\ | |
2979 | 2063 "pxor %%mm7, %%mm7 \n\t"\ |
2064 "1: \n\t"\ | |
2065 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
2066 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
2067 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
2068 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
2069 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
2070 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
2071 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
2072 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
2073 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
2074 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
2075 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
2076 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
2077 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
2078 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
2079 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
2080 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
2081 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
2082 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
2083 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
2084 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
2085 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
2086 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
2087 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
2088 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
2089 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
2090 "paddw %6, %%mm6 \n\t"\ | |
2091 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
2092 "psraw $5, %%mm0 \n\t"\ | |
959 | 2093 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ |
2094 \ | |
2979 | 2095 "movd 5(%0), %%mm5 \n\t" /* FGHI */\ |
2096 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\ | |
2097 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\ | |
2098 "paddw %%mm5, %%mm1 \n\t" /* a */\ | |
2099 "paddw %%mm6, %%mm2 \n\t" /* b */\ | |
2100 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\ | |
2101 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\ | |
2102 "paddw %%mm6, %%mm3 \n\t" /* c */\ | |
2103 "paddw %%mm5, %%mm4 \n\t" /* d */\ | |
2104 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
2105 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
2106 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
2107 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
2108 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\ | |
2109 "paddw %6, %%mm1 \n\t"\ | |
2110 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\ | |
2111 "psraw $5, %%mm3 \n\t"\ | |
2112 "packuswb %%mm3, %%mm0 \n\t"\ | |
959 | 2113 OP_MMX2(%%mm0, (%1), %%mm4, q)\ |
2114 \ | |
2979 | 2115 "add %3, %0 \n\t"\ |
2116 "add %4, %1 \n\t"\ | |
2117 "decl %2 \n\t"\ | |
2118 " jnz 1b \n\t"\ | |
967 | 2119 : "+a"(src), "+c"(dst), "+m"(h)\ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
2120 : "S"((long)srcStride), "D"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
966 | 2121 : "memory"\ |
959 | 2122 );\ |
2123 }\ | |
2124 \ | |
2125 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
2126 int i;\ | |
2127 int16_t temp[8];\ | |
2128 /* quick HACK, XXX FIXME MUST be optimized */\ | |
2129 for(i=0; i<h; i++)\ | |
2130 {\ | |
2131 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
2132 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
2133 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
2134 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
2135 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
2136 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\ | |
2137 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\ | |
2138 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\ | |
2139 asm volatile(\ | |
2979 | 2140 "movq (%0), %%mm0 \n\t"\ |
2141 "movq 8(%0), %%mm1 \n\t"\ | |
2142 "paddw %2, %%mm0 \n\t"\ | |
2143 "paddw %2, %%mm1 \n\t"\ | |
2144 "psraw $5, %%mm0 \n\t"\ | |
2145 "psraw $5, %%mm1 \n\t"\ | |
2146 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 2147 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ |
2148 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ | |
966 | 2149 :"memory"\ |
959 | 2150 );\ |
2151 dst+=dstStride;\ | |
2152 src+=srcStride;\ | |
2153 }\ | |
2154 } | |
2155 | |
2156 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\ | |
2157 \ | |
2158 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
954 | 2159 uint64_t temp[17*4];\ |
2160 uint64_t *temp_ptr= temp;\ | |
2161 int count= 17;\ | |
2162 \ | |
2163 /*FIXME unroll */\ | |
2164 asm volatile(\ | |
2979 | 2165 "pxor %%mm7, %%mm7 \n\t"\ |
2166 "1: \n\t"\ | |
2167 "movq (%0), %%mm0 \n\t"\ | |
2168 "movq (%0), %%mm1 \n\t"\ | |
2169 "movq 8(%0), %%mm2 \n\t"\ | |
2170 "movq 8(%0), %%mm3 \n\t"\ | |
2171 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2172 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
2173 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
2174 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
2175 "movq %%mm0, (%1) \n\t"\ | |
2176 "movq %%mm1, 17*8(%1) \n\t"\ | |
2177 "movq %%mm2, 2*17*8(%1) \n\t"\ | |
2178 "movq %%mm3, 3*17*8(%1) \n\t"\ | |
2179 "add $8, %1 \n\t"\ | |
2180 "add %3, %0 \n\t"\ | |
2181 "decl %2 \n\t"\ | |
2182 " jnz 1b \n\t"\ | |
954 | 2183 : "+r" (src), "+r" (temp_ptr), "+r"(count)\ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
2184 : "r" ((long)srcStride)\ |
966 | 2185 : "memory"\ |
954 | 2186 );\ |
2187 \ | |
2188 temp_ptr= temp;\ | |
2189 count=4;\ | |
2190 \ | |
2191 /*FIXME reorder for speed */\ | |
2192 asm volatile(\ | |
2979 | 2193 /*"pxor %%mm7, %%mm7 \n\t"*/\ |
2194 "1: \n\t"\ | |
2195 "movq (%0), %%mm0 \n\t"\ | |
2196 "movq 8(%0), %%mm1 \n\t"\ | |
2197 "movq 16(%0), %%mm2 \n\t"\ | |
2198 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 2199 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
2200 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
2979 | 2201 "add %4, %1 \n\t"\ |
961 | 2202 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 2203 \ |
961 | 2204 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
2979 | 2205 "add %4, %1 \n\t"\ |
961 | 2206 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
2207 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\ | |
2979 | 2208 "add %4, %1 \n\t"\ |
961 | 2209 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\ |
2210 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\ | |
2979 | 2211 "add %4, %1 \n\t"\ |
961 | 2212 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\ |
2213 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\ | |
2979 | 2214 "add %4, %1 \n\t"\ |
961 | 2215 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\ |
2216 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\ | |
2979 | 2217 "add %4, %1 \n\t"\ |
961 | 2218 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\ |
954 | 2219 \ |
961 | 2220 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\ |
2979 | 2221 "add %4, %1 \n\t" \ |
961 | 2222 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\ |
2223 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\ | |
954 | 2224 \ |
2979 | 2225 "add $136, %0 \n\t"\ |
2226 "add %6, %1 \n\t"\ | |
2227 "decl %2 \n\t"\ | |
2228 " jnz 1b \n\t"\ | |
958
9bb668034ecf
slowdown / gcc 2.95.* bug workaround (this should be reversed as soon as gcc 2.95.* support is droped)
michaelni
parents:
954
diff
changeset
|
2229 \ |
967 | 2230 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
2231 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-14*(long)dstStride)\ |
966 | 2232 :"memory"\ |
954 | 2233 );\ |
2234 }\ | |
2235 \ | |
1057 | 2236 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2209 | 2237 uint64_t temp[9*2];\ |
954 | 2238 uint64_t *temp_ptr= temp;\ |
2239 int count= 9;\ | |
2240 \ | |
2241 /*FIXME unroll */\ | |
2242 asm volatile(\ | |
2979 | 2243 "pxor %%mm7, %%mm7 \n\t"\ |
2244 "1: \n\t"\ | |
2245 "movq (%0), %%mm0 \n\t"\ | |
2246 "movq (%0), %%mm1 \n\t"\ | |
2247 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2248 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
2249 "movq %%mm0, (%1) \n\t"\ | |
2250 "movq %%mm1, 9*8(%1) \n\t"\ | |
2251 "add $8, %1 \n\t"\ | |
2252 "add %3, %0 \n\t"\ | |
2253 "decl %2 \n\t"\ | |
2254 " jnz 1b \n\t"\ | |
954 | 2255 : "+r" (src), "+r" (temp_ptr), "+r"(count)\ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
2256 : "r" ((long)srcStride)\ |
966 | 2257 : "memory"\ |
954 | 2258 );\ |
2259 \ | |
2260 temp_ptr= temp;\ | |
2261 count=2;\ | |
2262 \ | |
2263 /*FIXME reorder for speed */\ | |
2264 asm volatile(\ | |
2979 | 2265 /*"pxor %%mm7, %%mm7 \n\t"*/\ |
2266 "1: \n\t"\ | |
2267 "movq (%0), %%mm0 \n\t"\ | |
2268 "movq 8(%0), %%mm1 \n\t"\ | |
2269 "movq 16(%0), %%mm2 \n\t"\ | |
2270 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 2271 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
2272 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
2979 | 2273 "add %4, %1 \n\t"\ |
961 | 2274 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 2275 \ |
961 | 2276 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
2979 | 2277 "add %4, %1 \n\t"\ |
961 | 2278 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
954 | 2279 \ |
961 | 2280 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\ |
2979 | 2281 "add %4, %1 \n\t"\ |
961 | 2282 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\ |
2283 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\ | |
954 | 2284 \ |
2979 | 2285 "add $72, %0 \n\t"\ |
2286 "add %6, %1 \n\t"\ | |
2287 "decl %2 \n\t"\ | |
2288 " jnz 1b \n\t"\ | |
954 | 2289 \ |
961 | 2290 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
2291 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-6*(long)dstStride)\ |
966 | 2292 : "memory"\ |
2293 );\ | |
959 | 2294 }\ |
954 | 2295 \ |
1064 | 2296 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2297 OPNAME ## pixels8_mmx(dst, src, stride, 8);\ |
954 | 2298 }\ |
2299 \ | |
1064 | 2300 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2301 uint64_t temp[8];\ |
954 | 2302 uint8_t * const half= (uint8_t*)temp;\ |
2303 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2304 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ |
954 | 2305 }\ |
2306 \ | |
1064 | 2307 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2308 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\ |
2309 }\ | |
2310 \ | |
1064 | 2311 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2312 uint64_t temp[8];\ |
954 | 2313 uint8_t * const half= (uint8_t*)temp;\ |
2314 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2315 OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\ |
954 | 2316 }\ |
2317 \ | |
1064 | 2318 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2319 uint64_t temp[8];\ |
954 | 2320 uint8_t * const half= (uint8_t*)temp;\ |
959 | 2321 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2322 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ |
954 | 2323 }\ |
2324 \ | |
1064 | 2325 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2326 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 2327 }\ |
2328 \ | |
1064 | 2329 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2330 uint64_t temp[8];\ |
954 | 2331 uint8_t * const half= (uint8_t*)temp;\ |
959 | 2332 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2333 OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\ |
954 | 2334 }\ |
1064 | 2335 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2336 uint64_t half[8 + 9];\ |
2337 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2338 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2339 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2340 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
959 | 2341 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2342 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 2343 }\ |
1064 | 2344 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2345 uint64_t half[8 + 9];\ |
2346 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2347 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2348 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2349 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
959 | 2350 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2351 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 2352 }\ |
1064 | 2353 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2354 uint64_t half[8 + 9];\ |
2355 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2356 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2357 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2358 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
959 | 2359 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2360 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 2361 }\ |
1064 | 2362 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2363 uint64_t half[8 + 9];\ |
2364 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2365 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2366 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2367 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
959 | 2368 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2369 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 2370 }\ |
1064 | 2371 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2372 uint64_t half[8 + 9];\ |
954 | 2373 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
2374 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2375 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 2376 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2377 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 2378 }\ |
1064 | 2379 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2380 uint64_t half[8 + 9];\ |
954 | 2381 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
2382 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2383 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 2384 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2385 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 2386 }\ |
1064 | 2387 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2388 uint64_t half[8 + 9];\ |
2389 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2390 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2391 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
984 | 2392 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 2393 }\ |
1064 | 2394 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2395 uint64_t half[8 + 9];\ |
2396 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2397 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2398 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
984 | 2399 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 2400 }\ |
1064 | 2401 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2402 uint64_t half[9];\ |
954 | 2403 uint8_t * const halfH= ((uint8_t*)half);\ |
2404 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 2405 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 2406 }\ |
1064 | 2407 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2408 OPNAME ## pixels16_mmx(dst, src, stride, 16);\ |
954 | 2409 }\ |
2410 \ | |
1064 | 2411 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2412 uint64_t temp[32];\ |
2413 uint8_t * const half= (uint8_t*)temp;\ | |
2414 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2415 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ |
954 | 2416 }\ |
2417 \ | |
1064 | 2418 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2419 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\ |
2420 }\ | |
2421 \ | |
1064 | 2422 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2423 uint64_t temp[32];\ |
2424 uint8_t * const half= (uint8_t*)temp;\ | |
2425 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2426 OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\ |
954 | 2427 }\ |
2428 \ | |
1064 | 2429 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2430 uint64_t temp[32];\ |
2431 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 2432 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2433 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ |
954 | 2434 }\ |
2435 \ | |
1064 | 2436 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2437 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 2438 }\ |
2439 \ | |
1064 | 2440 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2441 uint64_t temp[32];\ |
2442 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 2443 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2444 OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\ |
954 | 2445 }\ |
1064 | 2446 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2447 uint64_t half[16*2 + 17*2];\ |
2448 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2449 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2450 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2451 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
959 | 2452 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2453 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 2454 }\ |
1064 | 2455 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2456 uint64_t half[16*2 + 17*2];\ |
2457 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2458 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2459 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2460 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
959 | 2461 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2462 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 2463 }\ |
1064 | 2464 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2465 uint64_t half[16*2 + 17*2];\ |
2466 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2467 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2468 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2469 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
959 | 2470 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2471 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 2472 }\ |
1064 | 2473 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2474 uint64_t half[16*2 + 17*2];\ |
2475 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2476 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2477 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2478 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
959 | 2479 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2480 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 2481 }\ |
1064 | 2482 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2483 uint64_t half[16*2 + 17*2];\ |
2484 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2485 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2486 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 2487 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2488 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 2489 }\ |
1064 | 2490 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2491 uint64_t half[16*2 + 17*2];\ |
2492 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2493 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2494 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 2495 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2496 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 2497 }\ |
1064 | 2498 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2499 uint64_t half[17*2];\ |
2500 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2501 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2502 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
984 | 2503 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 2504 }\ |
1064 | 2505 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2506 uint64_t half[17*2];\ |
2507 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2508 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2509 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
984 | 2510 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 2511 }\ |
1064 | 2512 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2513 uint64_t half[17*2];\ |
2514 uint8_t * const halfH= ((uint8_t*)half);\ | |
2515 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 2516 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 2517 } |
2518 | |
2979 | 2519 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t" |
959 | 2520 #define AVG_3DNOW_OP(a,b,temp, size) \ |
2979 | 2521 "mov" #size " " #b ", " #temp " \n\t"\ |
2522 "pavgusb " #temp ", " #a " \n\t"\ | |
2523 "mov" #size " " #a ", " #b " \n\t" | |
959 | 2524 #define AVG_MMX2_OP(a,b,temp, size) \ |
2979 | 2525 "mov" #size " " #b ", " #temp " \n\t"\ |
2526 "pavgb " #temp ", " #a " \n\t"\ | |
2527 "mov" #size " " #a ", " #b " \n\t" | |
959 | 2528 |
2529 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP) | |
2530 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP) | |
2531 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP) | |
2532 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow) | |
2533 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow) | |
2534 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow) | |
954 | 2535 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2) |
959 | 2536 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2) |
954 | 2537 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2) |
2538 | |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2539 /***********************************/ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2540 /* bilinear qpel: not compliant to any spec, only for -lavdopts fast */ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2541 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2542 #define QPEL_2TAP_XY(OPNAME, SIZE, MMX, XY, HPEL)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2543 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2544 OPNAME ## pixels ## SIZE ## HPEL(dst, src, stride, SIZE);\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2545 } |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2546 #define QPEL_2TAP_L3(OPNAME, SIZE, MMX, XY, S0, S1, S2)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2547 static void OPNAME ## 2tap_qpel ## SIZE ## _mc ## XY ## _ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2548 OPNAME ## 2tap_qpel ## SIZE ## _l3_ ## MMX(dst, src+S0, stride, SIZE, S1, S2);\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2549 } |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2550 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2551 #define QPEL_2TAP(OPNAME, SIZE, MMX)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2552 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 20, _x2_ ## MMX)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2553 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 02, _y2_ ## MMX)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2554 QPEL_2TAP_XY(OPNAME, SIZE, MMX, 22, _xy2_mmx)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2555 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc00_ ## MMX =\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2556 OPNAME ## qpel ## SIZE ## _mc00_ ## MMX;\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2557 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc21_ ## MMX =\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2558 OPNAME ## 2tap_qpel ## SIZE ## _mc20_ ## MMX;\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2559 static const qpel_mc_func OPNAME ## 2tap_qpel ## SIZE ## _mc12_ ## MMX =\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2560 OPNAME ## 2tap_qpel ## SIZE ## _mc02_ ## MMX;\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2561 static void OPNAME ## 2tap_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2562 OPNAME ## pixels ## SIZE ## _y2_ ## MMX(dst, src+1, stride, SIZE);\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2563 }\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2564 static void OPNAME ## 2tap_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2565 OPNAME ## pixels ## SIZE ## _x2_ ## MMX(dst, src+stride, stride, SIZE);\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2566 }\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2567 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 10, 0, 1, 0)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2568 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 30, 1, -1, 0)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2569 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 01, 0, stride, 0)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2570 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 03, stride, -stride, 0)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2571 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 11, 0, stride, 1)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2572 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 31, 1, stride, -1)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2573 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 13, stride, -stride, 1)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2574 QPEL_2TAP_L3(OPNAME, SIZE, MMX, 33, stride+1, -stride, -1)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2575 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2576 QPEL_2TAP(put_, 16, mmx2) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2577 QPEL_2TAP(avg_, 16, mmx2) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2578 QPEL_2TAP(put_, 8, mmx2) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2579 QPEL_2TAP(avg_, 8, mmx2) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2580 QPEL_2TAP(put_, 16, 3dnow) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2581 QPEL_2TAP(avg_, 16, 3dnow) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2582 QPEL_2TAP(put_, 8, 3dnow) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2583 QPEL_2TAP(avg_, 8, 3dnow) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2584 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2585 |
393 | 2586 #if 0 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2587 static void just_return() { return; } |
393 | 2588 #endif |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2589 |
954 | 2590 #define SET_QPEL_FUNC(postfix1, postfix2) \ |
2591 c->put_ ## postfix1 = put_ ## postfix2;\ | |
2592 c->put_no_rnd_ ## postfix1 = put_no_rnd_ ## postfix2;\ | |
2593 c->avg_ ## postfix1 = avg_ ## postfix2; | |
1092 | 2594 |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2595 static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy, |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2596 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height){ |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2597 const int w = 8; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2598 const int ix = ox>>(16+shift); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2599 const int iy = oy>>(16+shift); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2600 const int oxs = ox>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2601 const int oys = oy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2602 const int dxxs = dxx>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2603 const int dxys = dxy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2604 const int dyxs = dyx>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2605 const int dyys = dyy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2606 const uint16_t r4[4] = {r,r,r,r}; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2607 const uint16_t dxy4[4] = {dxys,dxys,dxys,dxys}; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2608 const uint16_t dyy4[4] = {dyys,dyys,dyys,dyys}; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2609 const uint64_t shift2 = 2*shift; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2610 uint8_t edge_buf[(h+1)*stride]; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2611 int x, y; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2612 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2613 const int dxw = (dxx-(1<<(16+shift)))*(w-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2614 const int dyh = (dyy-(1<<(16+shift)))*(h-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2615 const int dxh = dxy*(h-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2616 const int dyw = dyx*(w-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2617 if( // non-constant fullpel offset (3% of blocks) |
6196
166bef5cad01
add parenthesis, fix warning: i386/dsputil_mmx.c:2618: warning: suggest parentheses around arithmetic in operand of |
bcoudurier
parents:
6195
diff
changeset
|
2618 ((ox^(ox+dxw)) | (ox^(ox+dxh)) | (ox^(ox+dxw+dxh)) | |
166bef5cad01
add parenthesis, fix warning: i386/dsputil_mmx.c:2618: warning: suggest parentheses around arithmetic in operand of |
bcoudurier
parents:
6195
diff
changeset
|
2619 (oy^(oy+dyw)) | (oy^(oy+dyh)) | (oy^(oy+dyw+dyh))) >> (16+shift) |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2620 // uses more than 16 bits of subpel mv (only at huge resolution) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2621 || (dxx|dxy|dyx|dyy)&15 ) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2622 { |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2623 //FIXME could still use mmx for some of the rows |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2624 ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r, width, height); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2625 return; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2626 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2627 |
3250 | 2628 src += ix + iy*stride; |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2629 if( (unsigned)ix >= width-w || |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2630 (unsigned)iy >= height-h ) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2631 { |
3250 | 2632 ff_emulated_edge_mc(edge_buf, src, stride, w+1, h+1, ix, iy, width, height); |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2633 src = edge_buf; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2634 } |
3250 | 2635 |
2636 asm volatile( | |
2637 "movd %0, %%mm6 \n\t" | |
2638 "pxor %%mm7, %%mm7 \n\t" | |
2639 "punpcklwd %%mm6, %%mm6 \n\t" | |
2640 "punpcklwd %%mm6, %%mm6 \n\t" | |
2641 :: "r"(1<<shift) | |
2642 ); | |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2643 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2644 for(x=0; x<w; x+=4){ |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2645 uint16_t dx4[4] = { oxs - dxys + dxxs*(x+0), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2646 oxs - dxys + dxxs*(x+1), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2647 oxs - dxys + dxxs*(x+2), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2648 oxs - dxys + dxxs*(x+3) }; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2649 uint16_t dy4[4] = { oys - dyys + dyxs*(x+0), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2650 oys - dyys + dyxs*(x+1), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2651 oys - dyys + dyxs*(x+2), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2652 oys - dyys + dyxs*(x+3) }; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2653 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2654 for(y=0; y<h; y++){ |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2655 asm volatile( |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2656 "movq %0, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2657 "movq %1, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2658 "paddw %2, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2659 "paddw %3, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2660 "movq %%mm4, %0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2661 "movq %%mm5, %1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2662 "psrlw $12, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2663 "psrlw $12, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2664 : "+m"(*dx4), "+m"(*dy4) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2665 : "m"(*dxy4), "m"(*dyy4) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2666 ); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2667 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2668 asm volatile( |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2669 "movq %%mm6, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2670 "movq %%mm6, %%mm1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2671 "psubw %%mm4, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2672 "psubw %%mm5, %%mm1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2673 "movq %%mm2, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2674 "movq %%mm4, %%mm3 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2675 "pmullw %%mm1, %%mm0 \n\t" // (s-dx)*(s-dy) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2676 "pmullw %%mm5, %%mm3 \n\t" // dx*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2677 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2678 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2679 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2680 "movd %4, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2681 "movd %3, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2682 "punpcklbw %%mm7, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2683 "punpcklbw %%mm7, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2684 "pmullw %%mm5, %%mm3 \n\t" // src[1,1] * dx*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2685 "pmullw %%mm4, %%mm2 \n\t" // src[0,1] * (s-dx)*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2686 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2687 "movd %2, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2688 "movd %1, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2689 "punpcklbw %%mm7, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2690 "punpcklbw %%mm7, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2691 "pmullw %%mm5, %%mm1 \n\t" // src[1,0] * dx*(s-dy) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2692 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy) |
3250 | 2693 "paddw %5, %%mm1 \n\t" |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2694 "paddw %%mm3, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2695 "paddw %%mm1, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2696 "paddw %%mm2, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2697 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2698 "psrlw %6, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2699 "packuswb %%mm0, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2700 "movd %%mm0, %0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2701 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2702 : "=m"(dst[x+y*stride]) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2703 : "m"(src[0]), "m"(src[1]), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2704 "m"(src[stride]), "m"(src[stride+1]), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2705 "m"(*r4), "m"(shift2) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2706 ); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2707 src += stride; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2708 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2709 src += 4-h*stride; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2710 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2711 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2712 |
3777 | 2713 #ifdef CONFIG_ENCODERS |
5024
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2714 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2715 #define PHADDD(a, t)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2716 "movq "#a", "#t" \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2717 "psrlq $32, "#a" \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2718 "paddd "#t", "#a" \n\t" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2719 /* |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2720 pmulhw: dst[0-15]=(src[0-15]*dst[0-15])[16-31] |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2721 pmulhrw: dst[0-15]=(src[0-15]*dst[0-15] + 0x8000)[16-31] |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2722 pmulhrsw: dst[0-15]=(src[0-15]*dst[0-15] + 0x4000)[15-30] |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2723 */ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2724 #define PMULHRW(x, y, s, o)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2725 "pmulhw " #s ", "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2726 "pmulhw " #s ", "#y " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2727 "paddw " #o ", "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2728 "paddw " #o ", "#y " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2729 "psraw $1, "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2730 "psraw $1, "#y " \n\t" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2731 #define DEF(x) x ## _mmx |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2732 #define SET_RND MOVQ_WONE |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2733 #define SCALE_OFFSET 1 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2734 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2735 #include "dsputil_mmx_qns.h" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2736 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2737 #undef DEF |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2738 #undef SET_RND |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2739 #undef SCALE_OFFSET |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2740 #undef PMULHRW |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2741 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2742 #define DEF(x) x ## _3dnow |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2743 #define SET_RND(x) |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2744 #define SCALE_OFFSET 0 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2745 #define PMULHRW(x, y, s, o)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2746 "pmulhrw " #s ", "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2747 "pmulhrw " #s ", "#y " \n\t" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2748 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2749 #include "dsputil_mmx_qns.h" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2750 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2751 #undef DEF |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2752 #undef SET_RND |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2753 #undef SCALE_OFFSET |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2754 #undef PMULHRW |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2755 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2756 #ifdef HAVE_SSSE3 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2757 #undef PHADDD |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2758 #define DEF(x) x ## _ssse3 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2759 #define SET_RND(x) |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2760 #define SCALE_OFFSET -1 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2761 #define PHADDD(a, t)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2762 "pshufw $0x0E, "#a", "#t" \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2763 "paddd "#t", "#a" \n\t" /* faster than phaddd on core2 */ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2764 #define PMULHRW(x, y, s, o)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2765 "pmulhrsw " #s ", "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2766 "pmulhrsw " #s ", "#y " \n\t" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2767 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2768 #include "dsputil_mmx_qns.h" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2769 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2770 #undef DEF |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2771 #undef SET_RND |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2772 #undef SCALE_OFFSET |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2773 #undef PMULHRW |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2774 #undef PHADDD |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2775 #endif //HAVE_SSSE3 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2776 |
3777 | 2777 #endif /* CONFIG_ENCODERS */ |
2754 | 2778 |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2779 #define PREFETCH(name, op) \ |
4172 | 2780 static void name(void *mem, int stride, int h){\ |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2781 const uint8_t *p= mem;\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2782 do{\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2783 asm volatile(#op" %0" :: "m"(*p));\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2784 p+= stride;\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2785 }while(--h);\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2786 } |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2787 PREFETCH(prefetch_mmx2, prefetcht0) |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2788 PREFETCH(prefetch_3dnow, prefetch) |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2789 #undef PREFETCH |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2790 |
2754 | 2791 #include "h264dsp_mmx.c" |
2967 | 2792 |
6009 | 2793 /* CAVS specific */ |
3524 | 2794 void ff_cavsdsp_init_mmx2(DSPContext* c, AVCodecContext *avctx); |
2795 | |
2796 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2797 put_pixels8_mmx(dst, src, stride, 8); | |
2798 } | |
2799 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2800 avg_pixels8_mmx(dst, src, stride, 8); | |
2801 } | |
2802 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2803 put_pixels16_mmx(dst, src, stride, 16); | |
2804 } | |
2805 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2806 avg_pixels16_mmx(dst, src, stride, 16); | |
2807 } | |
2808 | |
6030 | 2809 /* FLAC specific */ |
2810 void ff_flac_compute_autocorr_sse2(const int32_t *data, int len, int lag, | |
2811 double *autoc); | |
2812 | |
5948 | 2813 /* VC1 specific */ |
2814 void ff_vc1dsp_init_mmx(DSPContext* dsp, AVCodecContext *avctx); | |
2815 | |
2816 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) { | |
2817 put_pixels8_mmx(dst, src, stride, 8); | |
2818 } | |
2819 | |
1092 | 2820 /* external functions, from idct_mmx.c */ |
2821 void ff_mmx_idct(DCTELEM *block); | |
2822 void ff_mmxext_idct(DCTELEM *block); | |
2823 | |
2824 /* XXX: those functions should be suppressed ASAP when all IDCTs are | |
2825 converted */ | |
4020
723818b5de0f
Put libmpeg2 IDCT functions under CONFIG_GPL, fixes link failure
diego
parents:
4001
diff
changeset
|
2826 #ifdef CONFIG_GPL |
1092 | 2827 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block) |
2828 { | |
2829 ff_mmx_idct (block); | |
2830 put_pixels_clamped_mmx(block, dest, line_size); | |
2831 } | |
2832 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2833 { | |
2834 ff_mmx_idct (block); | |
2835 add_pixels_clamped_mmx(block, dest, line_size); | |
2836 } | |
2837 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block) | |
2838 { | |
2839 ff_mmxext_idct (block); | |
2840 put_pixels_clamped_mmx(block, dest, line_size); | |
2841 } | |
2842 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2843 { | |
2844 ff_mmxext_idct (block); | |
2845 add_pixels_clamped_mmx(block, dest, line_size); | |
2846 } | |
4020
723818b5de0f
Put libmpeg2 IDCT functions under CONFIG_GPL, fixes link failure
diego
parents:
4001
diff
changeset
|
2847 #endif |
2868 | 2848 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block) |
2849 { | |
2850 ff_idct_xvid_mmx (block); | |
2851 put_pixels_clamped_mmx(block, dest, line_size); | |
2852 } | |
2853 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2854 { | |
2855 ff_idct_xvid_mmx (block); | |
2856 add_pixels_clamped_mmx(block, dest, line_size); | |
2857 } | |
2858 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block) | |
2859 { | |
2860 ff_idct_xvid_mmx2 (block); | |
2861 put_pixels_clamped_mmx(block, dest, line_size); | |
2862 } | |
2863 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2864 { | |
2865 ff_idct_xvid_mmx2 (block); | |
2866 add_pixels_clamped_mmx(block, dest, line_size); | |
2867 } | |
2967 | 2868 |
3541
3fbddeb13686
10l, vorbis_inverse_coupling_sse() was really 3dnow
lorenm
parents:
3536
diff
changeset
|
2869 static void vorbis_inverse_coupling_3dnow(float *mag, float *ang, int blocksize) |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2870 { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2871 int i; |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2872 asm volatile("pxor %%mm7, %%mm7":); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2873 for(i=0; i<blocksize; i+=2) { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2874 asm volatile( |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2875 "movq %0, %%mm0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2876 "movq %1, %%mm1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2877 "movq %%mm0, %%mm2 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2878 "movq %%mm1, %%mm3 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2879 "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2880 "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2881 "pslld $31, %%mm2 \n\t" // keep only the sign bit |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2882 "pxor %%mm2, %%mm1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2883 "movq %%mm3, %%mm4 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2884 "pand %%mm1, %%mm3 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2885 "pandn %%mm1, %%mm4 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2886 "pfadd %%mm0, %%mm3 \n\t" // a = m + ((a<0) & (a ^ sign(m))) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2887 "pfsub %%mm4, %%mm0 \n\t" // m = m + ((a>0) & (a ^ sign(m))) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2888 "movq %%mm3, %1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2889 "movq %%mm0, %0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2890 :"+m"(mag[i]), "+m"(ang[i]) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2891 ::"memory" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2892 ); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2893 } |
3561 | 2894 asm volatile("femms"); |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2895 } |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2896 static void vorbis_inverse_coupling_sse(float *mag, float *ang, int blocksize) |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2897 { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2898 int i; |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2899 |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2900 asm volatile( |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2901 "movaps %0, %%xmm5 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2902 ::"m"(ff_pdw_80000000[0]) |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2903 ); |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2904 for(i=0; i<blocksize; i+=4) { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2905 asm volatile( |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2906 "movaps %0, %%xmm0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2907 "movaps %1, %%xmm1 \n\t" |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2908 "xorps %%xmm2, %%xmm2 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2909 "xorps %%xmm3, %%xmm3 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2910 "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2911 "cmpleps %%xmm1, %%xmm3 \n\t" // a <= 0.0 |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2912 "andps %%xmm5, %%xmm2 \n\t" // keep only the sign bit |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2913 "xorps %%xmm2, %%xmm1 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2914 "movaps %%xmm3, %%xmm4 \n\t" |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2915 "andps %%xmm1, %%xmm3 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2916 "andnps %%xmm1, %%xmm4 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2917 "addps %%xmm0, %%xmm3 \n\t" // a = m + ((a<0) & (a ^ sign(m))) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2918 "subps %%xmm4, %%xmm0 \n\t" // m = m + ((a>0) & (a ^ sign(m))) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2919 "movaps %%xmm3, %1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2920 "movaps %%xmm0, %0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2921 :"+m"(mag[i]), "+m"(ang[i]) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2922 ::"memory" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2923 ); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2924 } |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2925 } |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2926 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2927 static void vector_fmul_3dnow(float *dst, const float *src, int len){ |
3574 | 2928 long i = (len-4)*4; |
2929 asm volatile( | |
2930 "1: \n\t" | |
2931 "movq (%1,%0), %%mm0 \n\t" | |
2932 "movq 8(%1,%0), %%mm1 \n\t" | |
2933 "pfmul (%2,%0), %%mm0 \n\t" | |
2934 "pfmul 8(%2,%0), %%mm1 \n\t" | |
2935 "movq %%mm0, (%1,%0) \n\t" | |
2936 "movq %%mm1, 8(%1,%0) \n\t" | |
2937 "sub $16, %0 \n\t" | |
2938 "jge 1b \n\t" | |
2939 "femms \n\t" | |
2940 :"+r"(i) | |
2941 :"r"(dst), "r"(src) | |
2942 :"memory" | |
2943 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2944 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2945 static void vector_fmul_sse(float *dst, const float *src, int len){ |
3574 | 2946 long i = (len-8)*4; |
2947 asm volatile( | |
2948 "1: \n\t" | |
2949 "movaps (%1,%0), %%xmm0 \n\t" | |
2950 "movaps 16(%1,%0), %%xmm1 \n\t" | |
2951 "mulps (%2,%0), %%xmm0 \n\t" | |
2952 "mulps 16(%2,%0), %%xmm1 \n\t" | |
2953 "movaps %%xmm0, (%1,%0) \n\t" | |
2954 "movaps %%xmm1, 16(%1,%0) \n\t" | |
2955 "sub $32, %0 \n\t" | |
2956 "jge 1b \n\t" | |
2957 :"+r"(i) | |
2958 :"r"(dst), "r"(src) | |
2959 :"memory" | |
2960 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2961 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2962 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2963 static void vector_fmul_reverse_3dnow2(float *dst, const float *src0, const float *src1, int len){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2964 long i = len*4-16; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2965 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2966 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2967 "pswapd 8(%1), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2968 "pswapd (%1), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2969 "pfmul (%3,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2970 "pfmul 8(%3,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2971 "movq %%mm0, (%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2972 "movq %%mm1, 8(%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2973 "add $16, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2974 "sub $16, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2975 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2976 :"+r"(i), "+r"(src1) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2977 :"r"(dst), "r"(src0) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2978 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2979 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2980 } |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2981 static void vector_fmul_reverse_sse(float *dst, const float *src0, const float *src1, int len){ |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2982 long i = len*4-32; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2983 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2984 "1: \n\t" |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2985 "movaps 16(%1), %%xmm0 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2986 "movaps (%1), %%xmm1 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2987 "shufps $0x1b, %%xmm0, %%xmm0 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2988 "shufps $0x1b, %%xmm1, %%xmm1 \n\t" |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2989 "mulps (%3,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2990 "mulps 16(%3,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2991 "movaps %%xmm0, (%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2992 "movaps %%xmm1, 16(%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2993 "add $32, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2994 "sub $32, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2995 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2996 :"+r"(i), "+r"(src1) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2997 :"r"(dst), "r"(src0) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2998 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2999 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3000 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3001 static void vector_fmul_add_add_3dnow(float *dst, const float *src0, const float *src1, |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3002 const float *src2, int src3, int len, int step){ |
3574 | 3003 long i = (len-4)*4; |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3004 if(step == 2 && src3 == 0){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3005 dst += (len-4)*2; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3006 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3007 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3008 "movq (%2,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3009 "movq 8(%2,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3010 "pfmul (%3,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3011 "pfmul 8(%3,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3012 "pfadd (%4,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3013 "pfadd 8(%4,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3014 "movd %%mm0, (%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3015 "movd %%mm1, 16(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3016 "psrlq $32, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3017 "psrlq $32, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3018 "movd %%mm0, 8(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3019 "movd %%mm1, 24(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3020 "sub $32, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3021 "sub $16, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3022 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3023 :"+r"(i), "+r"(dst) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3024 :"r"(src0), "r"(src1), "r"(src2) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3025 :"memory" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3026 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3027 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3028 else if(step == 1 && src3 == 0){ |
3574 | 3029 asm volatile( |
3030 "1: \n\t" | |
3031 "movq (%2,%0), %%mm0 \n\t" | |
3032 "movq 8(%2,%0), %%mm1 \n\t" | |
3033 "pfmul (%3,%0), %%mm0 \n\t" | |
3034 "pfmul 8(%3,%0), %%mm1 \n\t" | |
3035 "pfadd (%4,%0), %%mm0 \n\t" | |
3036 "pfadd 8(%4,%0), %%mm1 \n\t" | |
3037 "movq %%mm0, (%1,%0) \n\t" | |
3038 "movq %%mm1, 8(%1,%0) \n\t" | |
3039 "sub $16, %0 \n\t" | |
3040 "jge 1b \n\t" | |
3041 :"+r"(i) | |
3042 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
3043 :"memory" | |
3044 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3045 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3046 else |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3047 ff_vector_fmul_add_add_c(dst, src0, src1, src2, src3, len, step); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3048 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3049 } |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3050 static void vector_fmul_add_add_sse(float *dst, const float *src0, const float *src1, |
3574 | 3051 const float *src2, int src3, int len, int step){ |
3052 long i = (len-8)*4; | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3053 if(step == 2 && src3 == 0){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3054 dst += (len-8)*2; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3055 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3056 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3057 "movaps (%2,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3058 "movaps 16(%2,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3059 "mulps (%3,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3060 "mulps 16(%3,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3061 "addps (%4,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3062 "addps 16(%4,%0), %%xmm1 \n\t" |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3063 "movss %%xmm0, (%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3064 "movss %%xmm1, 32(%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3065 "movhlps %%xmm0, %%xmm2 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3066 "movhlps %%xmm1, %%xmm3 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3067 "movss %%xmm2, 16(%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3068 "movss %%xmm3, 48(%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3069 "shufps $0xb1, %%xmm0, %%xmm0 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3070 "shufps $0xb1, %%xmm1, %%xmm1 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3071 "movss %%xmm0, 8(%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3072 "movss %%xmm1, 40(%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3073 "movhlps %%xmm0, %%xmm2 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3074 "movhlps %%xmm1, %%xmm3 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3075 "movss %%xmm2, 24(%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3076 "movss %%xmm3, 56(%1) \n\t" |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3077 "sub $64, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3078 "sub $32, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3079 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3080 :"+r"(i), "+r"(dst) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3081 :"r"(src0), "r"(src1), "r"(src2) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3082 :"memory" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3083 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3084 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3085 else if(step == 1 && src3 == 0){ |
3574 | 3086 asm volatile( |
3087 "1: \n\t" | |
3088 "movaps (%2,%0), %%xmm0 \n\t" | |
3089 "movaps 16(%2,%0), %%xmm1 \n\t" | |
3090 "mulps (%3,%0), %%xmm0 \n\t" | |
3091 "mulps 16(%3,%0), %%xmm1 \n\t" | |
3092 "addps (%4,%0), %%xmm0 \n\t" | |
3093 "addps 16(%4,%0), %%xmm1 \n\t" | |
3094 "movaps %%xmm0, (%1,%0) \n\t" | |
3095 "movaps %%xmm1, 16(%1,%0) \n\t" | |
3096 "sub $32, %0 \n\t" | |
3097 "jge 1b \n\t" | |
3098 :"+r"(i) | |
3099 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
3100 :"memory" | |
3101 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3102 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3103 else |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3104 ff_vector_fmul_add_add_c(dst, src0, src1, src2, src3, len, step); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3105 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3106 |
4172 | 3107 static void float_to_int16_3dnow(int16_t *dst, const float *src, int len){ |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3108 // not bit-exact: pf2id uses different rounding than C and SSE |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3109 int i; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3110 for(i=0; i<len; i+=4) { |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3111 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3112 "pf2id %1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3113 "pf2id %2, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3114 "packssdw %%mm1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3115 "movq %%mm0, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3116 :"=m"(dst[i]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3117 :"m"(src[i]), "m"(src[i+2]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3118 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3119 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3120 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3121 } |
4172 | 3122 static void float_to_int16_sse(int16_t *dst, const float *src, int len){ |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3123 int i; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3124 for(i=0; i<len; i+=4) { |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3125 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3126 "cvtps2pi %1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3127 "cvtps2pi %2, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3128 "packssdw %%mm1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3129 "movq %%mm0, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3130 :"=m"(dst[i]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3131 :"m"(src[i]), "m"(src[i+2]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3132 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3133 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3134 asm volatile("emms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3135 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3136 |
6195
5f704e9cb518
fix prototypes, remove warning: i386/dsputil_mmx.c:3594: warning: assignment from incompatible pointer type
bcoudurier
parents:
6181
diff
changeset
|
3137 extern void ff_snow_horizontal_compose97i_sse2(IDWTELEM *b, int width); |
5f704e9cb518
fix prototypes, remove warning: i386/dsputil_mmx.c:3594: warning: assignment from incompatible pointer type
bcoudurier
parents:
6181
diff
changeset
|
3138 extern void ff_snow_horizontal_compose97i_mmx(IDWTELEM *b, int width); |
5f704e9cb518
fix prototypes, remove warning: i386/dsputil_mmx.c:3594: warning: assignment from incompatible pointer type
bcoudurier
parents:
6181
diff
changeset
|
3139 extern void ff_snow_vertical_compose97i_sse2(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width); |
5f704e9cb518
fix prototypes, remove warning: i386/dsputil_mmx.c:3594: warning: assignment from incompatible pointer type
bcoudurier
parents:
6181
diff
changeset
|
3140 extern void ff_snow_vertical_compose97i_mmx(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width); |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
4197
diff
changeset
|
3141 extern void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
3142 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8); |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
4197
diff
changeset
|
3143 extern void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
3144 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8); |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3145 |
1092 | 3146 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx) |
0 | 3147 { |
4197 | 3148 mm_flags = mm_support(); |
1115 | 3149 |
1122 | 3150 if (avctx->dsp_mask) { |
2979 | 3151 if (avctx->dsp_mask & FF_MM_FORCE) |
4197 | 3152 mm_flags |= (avctx->dsp_mask & 0xffff); |
2979 | 3153 else |
4197 | 3154 mm_flags &= ~(avctx->dsp_mask & 0xffff); |
1122 | 3155 } |
1115 | 3156 |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
3157 #if 0 |
1868 | 3158 av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:"); |
4197 | 3159 if (mm_flags & MM_MMX) |
1868 | 3160 av_log(avctx, AV_LOG_INFO, " mmx"); |
4197 | 3161 if (mm_flags & MM_MMXEXT) |
1868 | 3162 av_log(avctx, AV_LOG_INFO, " mmxext"); |
4197 | 3163 if (mm_flags & MM_3DNOW) |
1868 | 3164 av_log(avctx, AV_LOG_INFO, " 3dnow"); |
4197 | 3165 if (mm_flags & MM_SSE) |
1868 | 3166 av_log(avctx, AV_LOG_INFO, " sse"); |
4197 | 3167 if (mm_flags & MM_SSE2) |
1868 | 3168 av_log(avctx, AV_LOG_INFO, " sse2"); |
3169 av_log(avctx, AV_LOG_INFO, "\n"); | |
0 | 3170 #endif |
3171 | |
4197 | 3172 if (mm_flags & MM_MMX) { |
1092 | 3173 const int idct_algo= avctx->idct_algo; |
3174 | |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1186
diff
changeset
|
3175 #ifdef CONFIG_ENCODERS |
2024
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1985
diff
changeset
|
3176 const int dct_algo = avctx->dct_algo; |
1565 | 3177 if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){ |
4197 | 3178 if(mm_flags & MM_SSE2){ |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
3179 c->fdct = ff_fdct_sse2; |
4197 | 3180 }else if(mm_flags & MM_MMXEXT){ |
1565 | 3181 c->fdct = ff_fdct_mmx2; |
3182 }else{ | |
3183 c->fdct = ff_fdct_mmx; | |
3184 } | |
3185 } | |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1186
diff
changeset
|
3186 #endif //CONFIG_ENCODERS |
2256 | 3187 if(avctx->lowres==0){ |
3188 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){ | |
3189 c->idct_put= ff_simple_idct_put_mmx; | |
3190 c->idct_add= ff_simple_idct_add_mmx; | |
3191 c->idct = ff_simple_idct_mmx; | |
3192 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM; | |
3717
ea9fe1c9d126
Remove the LGPL exception clause as discussed on ffmpeg-devel
diego
parents:
3712
diff
changeset
|
3193 #ifdef CONFIG_GPL |
2256 | 3194 }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){ |
4197 | 3195 if(mm_flags & MM_MMXEXT){ |
2256 | 3196 c->idct_put= ff_libmpeg2mmx2_idct_put; |
3197 c->idct_add= ff_libmpeg2mmx2_idct_add; | |
3198 c->idct = ff_mmxext_idct; | |
3199 }else{ | |
3200 c->idct_put= ff_libmpeg2mmx_idct_put; | |
3201 c->idct_add= ff_libmpeg2mmx_idct_add; | |
3202 c->idct = ff_mmx_idct; | |
3203 } | |
3204 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM; | |
3717
ea9fe1c9d126
Remove the LGPL exception clause as discussed on ffmpeg-devel
diego
parents:
3712
diff
changeset
|
3205 #endif |
5007 | 3206 }else if((ENABLE_VP3_DECODER || ENABLE_VP5_DECODER || ENABLE_VP6_DECODER) && |
3207 idct_algo==FF_IDCT_VP3 && | |
3721
2000e401593d
disable vp3 mmx idct for theora files to avoid artifacts
aurel
parents:
3717
diff
changeset
|
3208 avctx->codec->id!=CODEC_ID_THEORA && |
3712
f7f75f718efb
Enables back the mmx/sse optimized version of the vp3 idct.
aurel
parents:
3666
diff
changeset
|
3209 !(avctx->flags & CODEC_FLAG_BITEXACT)){ |
4197 | 3210 if(mm_flags & MM_SSE2){ |
2696
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3211 c->idct_put= ff_vp3_idct_put_sse2; |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3212 c->idct_add= ff_vp3_idct_add_sse2; |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3213 c->idct = ff_vp3_idct_sse2; |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3214 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM; |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3215 }else{ |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3216 ff_vp3_dsp_init_mmx(); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3217 c->idct_put= ff_vp3_idct_put_mmx; |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3218 c->idct_add= ff_vp3_idct_add_mmx; |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3219 c->idct = ff_vp3_idct_mmx; |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3220 c->idct_permutation_type= FF_PARTTRANS_IDCT_PERM; |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3221 } |
3524 | 3222 }else if(idct_algo==FF_IDCT_CAVS){ |
3223 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM; | |
2868 | 3224 }else if(idct_algo==FF_IDCT_XVIDMMX){ |
4197 | 3225 if(mm_flags & MM_MMXEXT){ |
2868 | 3226 c->idct_put= ff_idct_xvid_mmx2_put; |
3227 c->idct_add= ff_idct_xvid_mmx2_add; | |
3228 c->idct = ff_idct_xvid_mmx2; | |
3229 }else{ | |
3230 c->idct_put= ff_idct_xvid_mmx_put; | |
3231 c->idct_add= ff_idct_xvid_mmx_add; | |
3232 c->idct = ff_idct_xvid_mmx; | |
3233 } | |
1092 | 3234 } |
3235 } | |
1868 | 3236 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3237 #ifdef CONFIG_ENCODERS |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3238 c->get_pixels = get_pixels_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3239 c->diff_pixels = diff_pixels_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3240 #endif //CONFIG_ENCODERS |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3241 c->put_pixels_clamped = put_pixels_clamped_mmx; |
1984
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
3242 c->put_signed_pixels_clamped = put_signed_pixels_clamped_mmx; |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3243 c->add_pixels_clamped = add_pixels_clamped_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3244 c->clear_blocks = clear_blocks_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3245 #ifdef CONFIG_ENCODERS |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3246 c->pix_sum = pix_sum16_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3247 #endif //CONFIG_ENCODERS |
415 | 3248 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3249 c->put_pixels_tab[0][0] = put_pixels16_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3250 c->put_pixels_tab[0][1] = put_pixels16_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3251 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3252 c->put_pixels_tab[0][3] = put_pixels16_xy2_mmx; |
0 | 3253 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3254 c->put_no_rnd_pixels_tab[0][0] = put_pixels16_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3255 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3256 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3257 c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_mmx; |
651 | 3258 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3259 c->avg_pixels_tab[0][0] = avg_pixels16_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3260 c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3261 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3262 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx; |
415 | 3263 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3264 c->avg_no_rnd_pixels_tab[0][0] = avg_no_rnd_pixels16_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3265 c->avg_no_rnd_pixels_tab[0][1] = avg_no_rnd_pixels16_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3266 c->avg_no_rnd_pixels_tab[0][2] = avg_no_rnd_pixels16_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3267 c->avg_no_rnd_pixels_tab[0][3] = avg_no_rnd_pixels16_xy2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3268 |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3269 c->put_pixels_tab[1][0] = put_pixels8_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3270 c->put_pixels_tab[1][1] = put_pixels8_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3271 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3272 c->put_pixels_tab[1][3] = put_pixels8_xy2_mmx; |
0 | 3273 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3274 c->put_no_rnd_pixels_tab[1][0] = put_pixels8_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3275 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3276 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3277 c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_mmx; |
651 | 3278 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3279 c->avg_pixels_tab[1][0] = avg_pixels8_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3280 c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3281 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3282 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx; |
651 | 3283 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3284 c->avg_no_rnd_pixels_tab[1][0] = avg_no_rnd_pixels8_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3285 c->avg_no_rnd_pixels_tab[1][1] = avg_no_rnd_pixels8_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3286 c->avg_no_rnd_pixels_tab[1][2] = avg_no_rnd_pixels8_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3287 c->avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels8_xy2_mmx; |
2967 | 3288 |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
3289 c->gmc= gmc_mmx; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
3290 |
866 | 3291 c->add_bytes= add_bytes_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3292 #ifdef CONFIG_ENCODERS |
866 | 3293 c->diff_bytes= diff_bytes_mmx; |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
3294 c->sum_abs_dctelem= sum_abs_dctelem_mmx; |
2967 | 3295 |
936 | 3296 c->hadamard8_diff[0]= hadamard8_diff16_mmx; |
3297 c->hadamard8_diff[1]= hadamard8_diff_mmx; | |
2967 | 3298 |
2979 | 3299 c->pix_norm1 = pix_norm1_mmx; |
4197 | 3300 c->sse[0] = (mm_flags & MM_SSE2) ? sse16_sse2 : sse16_mmx; |
2979 | 3301 c->sse[1] = sse8_mmx; |
1729 | 3302 c->vsad[4]= vsad_intra16_mmx; |
3303 | |
2979 | 3304 c->nsse[0] = nsse16_mmx; |
3305 c->nsse[1] = nsse8_mmx; | |
1729 | 3306 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
3307 c->vsad[0] = vsad16_mmx; | |
3308 } | |
2967 | 3309 |
1784 | 3310 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
3311 c->try_8x8basis= try_8x8basis_mmx; | |
3312 } | |
3313 c->add_8x8basis= add_8x8basis_mmx; | |
2967 | 3314 |
4749 | 3315 c->ssd_int8_vs_int16 = ssd_int8_vs_int16_mmx; |
3316 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3317 #endif //CONFIG_ENCODERS |
1647 | 3318 |
5277
7b3fcb7c61ce
Avoid linking with h263.c functions when the relevant codecs
aurel
parents:
5255
diff
changeset
|
3319 if (ENABLE_ANY_H263) { |
5278 | 3320 c->h263_v_loop_filter= h263_v_loop_filter_mmx; |
3321 c->h263_h_loop_filter= h263_h_loop_filter_mmx; | |
5277
7b3fcb7c61ce
Avoid linking with h263.c functions when the relevant codecs
aurel
parents:
5255
diff
changeset
|
3322 } |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
3323 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_mmx_rnd; |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
3324 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_mmx; |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
3325 c->put_no_rnd_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_mmx_nornd; |
2967 | 3326 |
3173 | 3327 c->h264_idct_dc_add= |
3328 c->h264_idct_add= ff_h264_idct_add_mmx; | |
3174 | 3329 c->h264_idct8_dc_add= |
3330 c->h264_idct8_add= ff_h264_idct8_add_mmx; | |
3173 | 3331 |
4197 | 3332 if (mm_flags & MM_MMXEXT) { |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3333 c->prefetch = prefetch_mmx2; |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3334 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3335 c->put_pixels_tab[0][1] = put_pixels16_x2_mmx2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3336 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2; |
651 | 3337 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3338 c->avg_pixels_tab[0][0] = avg_pixels16_mmx2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3339 c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3340 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2; |
415 | 3341 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3342 c->put_pixels_tab[1][1] = put_pixels8_x2_mmx2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3343 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2; |
651 | 3344 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3345 c->avg_pixels_tab[1][0] = avg_pixels8_mmx2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3346 c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3347 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2; |
1092 | 3348 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3349 #ifdef CONFIG_ENCODERS |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
3350 c->sum_abs_dctelem= sum_abs_dctelem_mmx2; |
1153 | 3351 c->hadamard8_diff[0]= hadamard8_diff16_mmx2; |
3352 c->hadamard8_diff[1]= hadamard8_diff_mmx2; | |
1729 | 3353 c->vsad[4]= vsad_intra16_mmx2; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3354 #endif //CONFIG_ENCODERS |
1153 | 3355 |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3089
diff
changeset
|
3356 c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2; |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3089
diff
changeset
|
3357 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2; |
2745 | 3358 |
1092 | 3359 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
3360 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2; | |
3361 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2; | |
3362 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2; | |
3363 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2; | |
3364 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2; | |
3365 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2; | |
1772
8cd5257195c9
vsad16_mmx2 only applies if encoders are turned on
melanson
parents:
1765
diff
changeset
|
3366 #ifdef CONFIG_ENCODERS |
1729 | 3367 c->vsad[0] = vsad16_mmx2; |
1772
8cd5257195c9
vsad16_mmx2 only applies if encoders are turned on
melanson
parents:
1765
diff
changeset
|
3368 #endif //CONFIG_ENCODERS |
1092 | 3369 } |
959 | 3370 |
961 | 3371 #if 1 |
954 | 3372 SET_QPEL_FUNC(qpel_pixels_tab[0][ 0], qpel16_mc00_mmx2) |
3373 SET_QPEL_FUNC(qpel_pixels_tab[0][ 1], qpel16_mc10_mmx2) | |
3374 SET_QPEL_FUNC(qpel_pixels_tab[0][ 2], qpel16_mc20_mmx2) | |
3375 SET_QPEL_FUNC(qpel_pixels_tab[0][ 3], qpel16_mc30_mmx2) | |
3376 SET_QPEL_FUNC(qpel_pixels_tab[0][ 4], qpel16_mc01_mmx2) | |
3377 SET_QPEL_FUNC(qpel_pixels_tab[0][ 5], qpel16_mc11_mmx2) | |
3378 SET_QPEL_FUNC(qpel_pixels_tab[0][ 6], qpel16_mc21_mmx2) | |
3379 SET_QPEL_FUNC(qpel_pixels_tab[0][ 7], qpel16_mc31_mmx2) | |
3380 SET_QPEL_FUNC(qpel_pixels_tab[0][ 8], qpel16_mc02_mmx2) | |
3381 SET_QPEL_FUNC(qpel_pixels_tab[0][ 9], qpel16_mc12_mmx2) | |
3382 SET_QPEL_FUNC(qpel_pixels_tab[0][10], qpel16_mc22_mmx2) | |
3383 SET_QPEL_FUNC(qpel_pixels_tab[0][11], qpel16_mc32_mmx2) | |
3384 SET_QPEL_FUNC(qpel_pixels_tab[0][12], qpel16_mc03_mmx2) | |
3385 SET_QPEL_FUNC(qpel_pixels_tab[0][13], qpel16_mc13_mmx2) | |
3386 SET_QPEL_FUNC(qpel_pixels_tab[0][14], qpel16_mc23_mmx2) | |
3387 SET_QPEL_FUNC(qpel_pixels_tab[0][15], qpel16_mc33_mmx2) | |
3388 SET_QPEL_FUNC(qpel_pixels_tab[1][ 0], qpel8_mc00_mmx2) | |
3389 SET_QPEL_FUNC(qpel_pixels_tab[1][ 1], qpel8_mc10_mmx2) | |
3390 SET_QPEL_FUNC(qpel_pixels_tab[1][ 2], qpel8_mc20_mmx2) | |
3391 SET_QPEL_FUNC(qpel_pixels_tab[1][ 3], qpel8_mc30_mmx2) | |
3392 SET_QPEL_FUNC(qpel_pixels_tab[1][ 4], qpel8_mc01_mmx2) | |
3393 SET_QPEL_FUNC(qpel_pixels_tab[1][ 5], qpel8_mc11_mmx2) | |
3394 SET_QPEL_FUNC(qpel_pixels_tab[1][ 6], qpel8_mc21_mmx2) | |
3395 SET_QPEL_FUNC(qpel_pixels_tab[1][ 7], qpel8_mc31_mmx2) | |
3396 SET_QPEL_FUNC(qpel_pixels_tab[1][ 8], qpel8_mc02_mmx2) | |
3397 SET_QPEL_FUNC(qpel_pixels_tab[1][ 9], qpel8_mc12_mmx2) | |
3398 SET_QPEL_FUNC(qpel_pixels_tab[1][10], qpel8_mc22_mmx2) | |
3399 SET_QPEL_FUNC(qpel_pixels_tab[1][11], qpel8_mc32_mmx2) | |
3400 SET_QPEL_FUNC(qpel_pixels_tab[1][12], qpel8_mc03_mmx2) | |
3401 SET_QPEL_FUNC(qpel_pixels_tab[1][13], qpel8_mc13_mmx2) | |
3402 SET_QPEL_FUNC(qpel_pixels_tab[1][14], qpel8_mc23_mmx2) | |
3403 SET_QPEL_FUNC(qpel_pixels_tab[1][15], qpel8_mc33_mmx2) | |
961 | 3404 #endif |
1527 | 3405 |
2209 | 3406 //FIXME 3dnow too |
3407 #define dspfunc(PFX, IDX, NUM) \ | |
3408 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_mmx2; \ | |
3409 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_mmx2; \ | |
3410 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_mmx2; \ | |
3411 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_mmx2; \ | |
3412 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_mmx2; \ | |
3413 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_mmx2; \ | |
3414 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_mmx2; \ | |
3415 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_mmx2; \ | |
3416 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_mmx2; \ | |
3417 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_mmx2; \ | |
3418 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_mmx2; \ | |
3419 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_mmx2; \ | |
3420 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_mmx2; \ | |
3421 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_mmx2; \ | |
3422 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_mmx2; \ | |
3423 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_mmx2 | |
3424 | |
3425 dspfunc(put_h264_qpel, 0, 16); | |
3426 dspfunc(put_h264_qpel, 1, 8); | |
3427 dspfunc(put_h264_qpel, 2, 4); | |
3428 dspfunc(avg_h264_qpel, 0, 16); | |
3429 dspfunc(avg_h264_qpel, 1, 8); | |
3430 dspfunc(avg_h264_qpel, 2, 4); | |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3431 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3432 dspfunc(put_2tap_qpel, 0, 16); |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3433 dspfunc(put_2tap_qpel, 1, 8); |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3434 dspfunc(avg_2tap_qpel, 0, 16); |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3435 dspfunc(avg_2tap_qpel, 1, 8); |
2209 | 3436 #undef dspfunc |
3437 | |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
3438 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_mmx2_rnd; |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
3439 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_mmx2; |
3213 | 3440 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_mmx2; |
3441 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_mmx2; | |
2633 | 3442 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_mmx2; |
3443 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_mmx2; | |
3444 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_mmx2; | |
3445 c->h264_h_loop_filter_chroma= h264_h_loop_filter_chroma_mmx2; | |
2707
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
3446 c->h264_v_loop_filter_chroma_intra= h264_v_loop_filter_chroma_intra_mmx2; |
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
3447 c->h264_h_loop_filter_chroma_intra= h264_h_loop_filter_chroma_intra_mmx2; |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3576
diff
changeset
|
3448 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2; |
2633 | 3449 |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3450 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3451 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3452 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3453 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3454 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3455 c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3456 c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3457 c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3458 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3459 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3460 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3461 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3462 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3463 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3464 c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3465 c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3466 c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3467 |
5949 | 3468 if (ENABLE_CAVS_DECODER) |
5950 | 3469 ff_cavsdsp_init_mmx2(c, avctx); |
5949 | 3470 |
3471 if (ENABLE_VC1_DECODER || ENABLE_WMV3_DECODER) | |
5950 | 3472 ff_vc1dsp_init_mmx(c, avctx); |
5933 | 3473 |
1686
68abbec33289
Here are juste two added #ifdef CONFIG_ENCODERS to allow
michael
parents:
1648
diff
changeset
|
3474 #ifdef CONFIG_ENCODERS |
1527 | 3475 c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_mmx2; |
1686
68abbec33289
Here are juste two added #ifdef CONFIG_ENCODERS to allow
michael
parents:
1648
diff
changeset
|
3476 #endif //CONFIG_ENCODERS |
4197 | 3477 } else if (mm_flags & MM_3DNOW) { |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3478 c->prefetch = prefetch_3dnow; |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3479 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3480 c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3481 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow; |
393 | 3482 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3483 c->avg_pixels_tab[0][0] = avg_pixels16_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3484 c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3485 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow; |
651 | 3486 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3487 c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3488 c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3489 |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3490 c->avg_pixels_tab[1][0] = avg_pixels8_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3491 c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3492 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow; |
1092 | 3493 |
3494 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
3495 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow; | |
3496 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow; | |
3497 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow; | |
3498 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow; | |
3499 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow; | |
3500 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow; | |
3501 } | |
984 | 3502 |
954 | 3503 SET_QPEL_FUNC(qpel_pixels_tab[0][ 0], qpel16_mc00_3dnow) |
3504 SET_QPEL_FUNC(qpel_pixels_tab[0][ 1], qpel16_mc10_3dnow) | |
3505 SET_QPEL_FUNC(qpel_pixels_tab[0][ 2], qpel16_mc20_3dnow) | |
3506 SET_QPEL_FUNC(qpel_pixels_tab[0][ 3], qpel16_mc30_3dnow) | |
3507 SET_QPEL_FUNC(qpel_pixels_tab[0][ 4], qpel16_mc01_3dnow) | |
3508 SET_QPEL_FUNC(qpel_pixels_tab[0][ 5], qpel16_mc11_3dnow) | |
3509 SET_QPEL_FUNC(qpel_pixels_tab[0][ 6], qpel16_mc21_3dnow) | |
3510 SET_QPEL_FUNC(qpel_pixels_tab[0][ 7], qpel16_mc31_3dnow) | |
3511 SET_QPEL_FUNC(qpel_pixels_tab[0][ 8], qpel16_mc02_3dnow) | |
3512 SET_QPEL_FUNC(qpel_pixels_tab[0][ 9], qpel16_mc12_3dnow) | |
3513 SET_QPEL_FUNC(qpel_pixels_tab[0][10], qpel16_mc22_3dnow) | |
3514 SET_QPEL_FUNC(qpel_pixels_tab[0][11], qpel16_mc32_3dnow) | |
3515 SET_QPEL_FUNC(qpel_pixels_tab[0][12], qpel16_mc03_3dnow) | |
3516 SET_QPEL_FUNC(qpel_pixels_tab[0][13], qpel16_mc13_3dnow) | |
3517 SET_QPEL_FUNC(qpel_pixels_tab[0][14], qpel16_mc23_3dnow) | |
3518 SET_QPEL_FUNC(qpel_pixels_tab[0][15], qpel16_mc33_3dnow) | |
3519 SET_QPEL_FUNC(qpel_pixels_tab[1][ 0], qpel8_mc00_3dnow) | |
3520 SET_QPEL_FUNC(qpel_pixels_tab[1][ 1], qpel8_mc10_3dnow) | |
3521 SET_QPEL_FUNC(qpel_pixels_tab[1][ 2], qpel8_mc20_3dnow) | |
3522 SET_QPEL_FUNC(qpel_pixels_tab[1][ 3], qpel8_mc30_3dnow) | |
3523 SET_QPEL_FUNC(qpel_pixels_tab[1][ 4], qpel8_mc01_3dnow) | |
3524 SET_QPEL_FUNC(qpel_pixels_tab[1][ 5], qpel8_mc11_3dnow) | |
3525 SET_QPEL_FUNC(qpel_pixels_tab[1][ 6], qpel8_mc21_3dnow) | |
3526 SET_QPEL_FUNC(qpel_pixels_tab[1][ 7], qpel8_mc31_3dnow) | |
3527 SET_QPEL_FUNC(qpel_pixels_tab[1][ 8], qpel8_mc02_3dnow) | |
3528 SET_QPEL_FUNC(qpel_pixels_tab[1][ 9], qpel8_mc12_3dnow) | |
3529 SET_QPEL_FUNC(qpel_pixels_tab[1][10], qpel8_mc22_3dnow) | |
3530 SET_QPEL_FUNC(qpel_pixels_tab[1][11], qpel8_mc32_3dnow) | |
3531 SET_QPEL_FUNC(qpel_pixels_tab[1][12], qpel8_mc03_3dnow) | |
3532 SET_QPEL_FUNC(qpel_pixels_tab[1][13], qpel8_mc13_3dnow) | |
3533 SET_QPEL_FUNC(qpel_pixels_tab[1][14], qpel8_mc23_3dnow) | |
3534 SET_QPEL_FUNC(qpel_pixels_tab[1][15], qpel8_mc33_3dnow) | |
2209 | 3535 |
3536 #define dspfunc(PFX, IDX, NUM) \ | |
3537 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_3dnow; \ | |
3538 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_3dnow; \ | |
3539 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_3dnow; \ | |
3540 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_3dnow; \ | |
3541 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_3dnow; \ | |
3542 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_3dnow; \ | |
3543 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_3dnow; \ | |
3544 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_3dnow; \ | |
3545 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_3dnow; \ | |
3546 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_3dnow; \ | |
3547 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_3dnow; \ | |
3548 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_3dnow; \ | |
3549 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_3dnow; \ | |
3550 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_3dnow; \ | |
3551 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_3dnow; \ | |
3552 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_3dnow | |
3553 | |
3554 dspfunc(put_h264_qpel, 0, 16); | |
3555 dspfunc(put_h264_qpel, 1, 8); | |
3556 dspfunc(put_h264_qpel, 2, 4); | |
3557 dspfunc(avg_h264_qpel, 0, 16); | |
3558 dspfunc(avg_h264_qpel, 1, 8); | |
3559 dspfunc(avg_h264_qpel, 2, 4); | |
2732 | 3560 |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3561 dspfunc(put_2tap_qpel, 0, 16); |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3562 dspfunc(put_2tap_qpel, 1, 8); |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3563 dspfunc(avg_2tap_qpel, 0, 16); |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3564 dspfunc(avg_2tap_qpel, 1, 8); |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
3565 |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
3566 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_3dnow_rnd; |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
3567 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_3dnow; |
0 | 3568 } |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3569 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3570 #ifdef CONFIG_ENCODERS |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3571 if(mm_flags & MM_SSE2){ |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
3572 c->sum_abs_dctelem= sum_abs_dctelem_sse2; |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3573 c->hadamard8_diff[0]= hadamard8_diff16_sse2; |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3574 c->hadamard8_diff[1]= hadamard8_diff_sse2; |
6030 | 3575 if (ENABLE_FLAC_ENCODER) |
3576 c->flac_compute_autocorr = ff_flac_compute_autocorr_sse2; | |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3577 } |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3578 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3579 #ifdef HAVE_SSSE3 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3580 if(mm_flags & MM_SSSE3){ |
5024
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3581 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3582 c->try_8x8basis= try_8x8basis_ssse3; |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3583 } |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3584 c->add_8x8basis= add_8x8basis_ssse3; |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
3585 c->sum_abs_dctelem= sum_abs_dctelem_ssse3; |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3586 c->hadamard8_diff[0]= hadamard8_diff16_ssse3; |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3587 c->hadamard8_diff[1]= hadamard8_diff_ssse3; |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3588 } |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3589 #endif |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3590 #endif |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3591 |
4589
30261f4ed12d
Fix wrong conditional, Snow decoding, not encoding, was SIMD-accelerated.
diego
parents:
4436
diff
changeset
|
3592 #ifdef CONFIG_SNOW_DECODER |
5591 | 3593 if(mm_flags & MM_SSE2 & 0){ |
3210 | 3594 c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2; |
5602 | 3595 #ifdef HAVE_7REGS |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3596 c->vertical_compose97i = ff_snow_vertical_compose97i_sse2; |
5601
b26025b9586d
workaround gcc bug, untested as my gcc is not complaining
michael
parents:
5594
diff
changeset
|
3597 #endif |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
3598 c->inner_add_yblock = ff_snow_inner_add_yblock_sse2; |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3599 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3600 else{ |
5594
384629ebcb93
avoid overflow in the 3rd lifting step, this now needs mmx2 at minimum
michael
parents:
5591
diff
changeset
|
3601 if(mm_flags & MM_MMXEXT){ |
3210 | 3602 c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx; |
5602 | 3603 #ifdef HAVE_7REGS |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3604 c->vertical_compose97i = ff_snow_vertical_compose97i_mmx; |
5601
b26025b9586d
workaround gcc bug, untested as my gcc is not complaining
michael
parents:
5594
diff
changeset
|
3605 #endif |
5594
384629ebcb93
avoid overflow in the 3rd lifting step, this now needs mmx2 at minimum
michael
parents:
5591
diff
changeset
|
3606 } |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
3607 c->inner_add_yblock = ff_snow_inner_add_yblock_mmx; |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3608 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3609 #endif |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
3610 |
4197 | 3611 if(mm_flags & MM_3DNOW){ |
5024
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3612 #ifdef CONFIG_ENCODERS |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3613 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3614 c->try_8x8basis= try_8x8basis_3dnow; |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3615 } |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3616 c->add_8x8basis= add_8x8basis_3dnow; |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3617 #endif //CONFIG_ENCODERS |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3618 c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3619 c->vector_fmul = vector_fmul_3dnow; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3620 if(!(avctx->flags & CODEC_FLAG_BITEXACT)) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3621 c->float_to_int16 = float_to_int16_3dnow; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3622 } |
4197 | 3623 if(mm_flags & MM_3DNOWEXT) |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3624 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2; |
4197 | 3625 if(mm_flags & MM_SSE){ |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
3626 c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse; |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3627 c->vector_fmul = vector_fmul_sse; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3628 c->float_to_int16 = float_to_int16_sse; |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3629 c->vector_fmul_reverse = vector_fmul_reverse_sse; |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3630 c->vector_fmul_add_add = vector_fmul_add_add_sse; |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3631 } |
4197 | 3632 if(mm_flags & MM_3DNOW) |
3574 | 3633 c->vector_fmul_add_add = vector_fmul_add_add_3dnow; // faster than sse |
0 | 3634 } |
2967 | 3635 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3636 #ifdef CONFIG_ENCODERS |
1092 | 3637 dsputil_init_pix_mmx(c, avctx); |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3638 #endif //CONFIG_ENCODERS |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3639 #if 0 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3640 // for speed testing |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3641 get_pixels = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3642 put_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3643 add_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3644 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3645 pix_abs16x16 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3646 pix_abs16x16_x2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3647 pix_abs16x16_y2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3648 pix_abs16x16_xy2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3649 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3650 put_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3651 put_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3652 put_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3653 put_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3654 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3655 put_no_rnd_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3656 put_no_rnd_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3657 put_no_rnd_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3658 put_no_rnd_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3659 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3660 avg_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3661 avg_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3662 avg_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3663 avg_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3664 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3665 avg_no_rnd_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3666 avg_no_rnd_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3667 avg_no_rnd_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3668 avg_no_rnd_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3669 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3670 //av_fdct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3671 //ff_idct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3672 #endif |
0 | 3673 } |