Mercurial > libavcodec.hg
annotate i386/dsputil_mmx.c @ 6408:7af599600f2d libavcodec
One variable less. No real change to the actual code.
author | michael |
---|---|
date | Tue, 26 Feb 2008 15:19:37 +0000 |
parents | 9a736918fd90 |
children | 5154ab444372 |
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; |
6331 | 52 DECLARE_ALIGNED_16(const xmm_t, ff_pw_5 ) = {0x0005000500050005ULL, 0x0005000500050005ULL}; |
5946
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; |
6331 | 55 DECLARE_ALIGNED_16(const xmm_t, ff_pw_16 ) = {0x0010001000100010ULL, 0x0010001000100010ULL}; |
5946
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; |
6329
5969caa9190d
clean up an ugliness introduced in r11826. this syntax will require fewer changes when adding future sse2 code.
lorenm
parents:
6327
diff
changeset
|
57 DECLARE_ALIGNED_16(const xmm_t, ff_pw_32 ) = {0x0020002000200020ULL, 0x0020002000200020ULL}; |
5946
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; |
6333 | 61 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_128) = 0x0080008000800080ULL; |
6384 | 62 DECLARE_ALIGNED_8 (const uint64_t, ff_pw_255) = 0x00ff00ff00ff00ffULL; |
5946
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
63 |
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_1 ) = 0x0101010101010101ULL; |
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_3 ) = 0x0303030303030303ULL; |
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_7 ) = 0x0707070707070707ULL; |
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_3F ) = 0x3F3F3F3F3F3F3F3FULL; |
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_A1 ) = 0xA1A1A1A1A1A1A1A1ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
69 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
|
70 |
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_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
|
72 DECLARE_ALIGNED_16(const double, ff_pd_2[2]) = { 2.0, 2.0 }; |
5737 | 73 |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
74 #define JUMPALIGN() asm volatile (ASMALIGN(3)::) |
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
75 #define MOVQ_ZERO(regd) asm volatile ("pxor %%" #regd ", %%" #regd ::) |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
76 |
448 | 77 #define MOVQ_BFE(regd) \ |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
78 asm volatile ( \ |
448 | 79 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\ |
80 "paddb %%" #regd ", %%" #regd " \n\t" ::) | |
81 | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
82 #ifndef PIC |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
83 #define MOVQ_BONE(regd) asm volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_bone)) |
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
84 #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
|
85 #else |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
86 // 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
|
87 // pcmpeqd -> -1 |
448 | 88 #define MOVQ_BONE(regd) \ |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
89 asm volatile ( \ |
448 | 90 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
91 "psrlw $15, %%" #regd " \n\t" \ | |
92 "packuswb %%" #regd ", %%" #regd " \n\t" ::) | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
93 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
94 #define MOVQ_WTWO(regd) \ |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
95 asm volatile ( \ |
448 | 96 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
97 "psrlw $15, %%" #regd " \n\t" \ | |
98 "psllw $1, %%" #regd " \n\t"::) | |
387 | 99 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
100 #endif |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
101 |
448 | 102 // 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
|
103 // first argument is unmodifed and second is trashed |
471 | 104 // regfe is supposed to contain 0xfefefefefefefefe |
105 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \ | |
2979 | 106 "movq " #rega ", " #regr " \n\t"\ |
107 "pand " #regb ", " #regr " \n\t"\ | |
108 "pxor " #rega ", " #regb " \n\t"\ | |
109 "pand " #regfe "," #regb " \n\t"\ | |
110 "psrlq $1, " #regb " \n\t"\ | |
111 "paddb " #regb ", " #regr " \n\t" | |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
112 |
471 | 113 #define PAVGB_MMX(rega, regb, regr, regfe) \ |
2979 | 114 "movq " #rega ", " #regr " \n\t"\ |
115 "por " #regb ", " #regr " \n\t"\ | |
116 "pxor " #rega ", " #regb " \n\t"\ | |
117 "pand " #regfe "," #regb " \n\t"\ | |
118 "psrlq $1, " #regb " \n\t"\ | |
119 "psubb " #regb ", " #regr " \n\t" | |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
120 |
471 | 121 // mm6 is supposed to contain 0xfefefefefefefefe |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
122 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \ |
2979 | 123 "movq " #rega ", " #regr " \n\t"\ |
124 "movq " #regc ", " #regp " \n\t"\ | |
125 "pand " #regb ", " #regr " \n\t"\ | |
126 "pand " #regd ", " #regp " \n\t"\ | |
127 "pxor " #rega ", " #regb " \n\t"\ | |
128 "pxor " #regc ", " #regd " \n\t"\ | |
129 "pand %%mm6, " #regb " \n\t"\ | |
130 "pand %%mm6, " #regd " \n\t"\ | |
131 "psrlq $1, " #regb " \n\t"\ | |
132 "psrlq $1, " #regd " \n\t"\ | |
133 "paddb " #regb ", " #regr " \n\t"\ | |
134 "paddb " #regd ", " #regp " \n\t" | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
135 |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
136 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \ |
2979 | 137 "movq " #rega ", " #regr " \n\t"\ |
138 "movq " #regc ", " #regp " \n\t"\ | |
139 "por " #regb ", " #regr " \n\t"\ | |
140 "por " #regd ", " #regp " \n\t"\ | |
141 "pxor " #rega ", " #regb " \n\t"\ | |
142 "pxor " #regc ", " #regd " \n\t"\ | |
143 "pand %%mm6, " #regb " \n\t"\ | |
144 "pand %%mm6, " #regd " \n\t"\ | |
145 "psrlq $1, " #regd " \n\t"\ | |
146 "psrlq $1, " #regb " \n\t"\ | |
147 "psubb " #regb ", " #regr " \n\t"\ | |
148 "psubb " #regd ", " #regp " \n\t" | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
149 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
150 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
151 /* MMX no rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
152 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx |
448 | 153 #define SET_RND MOVQ_WONE |
2979 | 154 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX_NO_RND(a, b, c, d, e, f) |
155 #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
|
156 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
157 #include "dsputil_mmx_rnd.h" |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
158 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
159 #undef DEF |
448 | 160 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
161 #undef PAVGBP |
471 | 162 #undef PAVGB |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
163 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
164 /* MMX rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
165 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
166 #define DEF(x, y) x ## _ ## y ##_mmx |
448 | 167 #define SET_RND MOVQ_WTWO |
2979 | 168 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f) |
169 #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
|
170 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
171 #include "dsputil_mmx_rnd.h" |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
172 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
173 #undef DEF |
448 | 174 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
175 #undef PAVGBP |
471 | 176 #undef PAVGB |
387 | 177 |
0 | 178 /***********************************/ |
179 /* 3Dnow specific */ | |
180 | |
181 #define DEF(x) x ## _3dnow | |
182 #define PAVGB "pavgusb" | |
183 | |
184 #include "dsputil_mmx_avg.h" | |
185 | |
186 #undef DEF | |
187 #undef PAVGB | |
188 | |
189 /***********************************/ | |
190 /* MMX2 specific */ | |
191 | |
386 | 192 #define DEF(x) x ## _mmx2 |
0 | 193 |
194 /* Introduced only in MMX2 set */ | |
195 #define PAVGB "pavgb" | |
196 | |
197 #include "dsputil_mmx_avg.h" | |
198 | |
199 #undef DEF | |
200 #undef PAVGB | |
201 | |
6327 | 202 #define put_no_rnd_pixels16_mmx put_pixels16_mmx |
203 #define put_no_rnd_pixels8_mmx put_pixels8_mmx | |
6321 | 204 #define put_pixels16_mmx2 put_pixels16_mmx |
205 #define put_pixels8_mmx2 put_pixels8_mmx | |
206 #define put_pixels4_mmx2 put_pixels4_mmx | |
207 #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx | |
208 #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx | |
209 #define put_pixels16_3dnow put_pixels16_mmx | |
210 #define put_pixels8_3dnow put_pixels8_mmx | |
211 #define put_pixels4_3dnow put_pixels4_mmx | |
212 #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx | |
213 #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx | |
214 | |
0 | 215 /***********************************/ |
216 /* standard MMX */ | |
217 | |
1064 | 218 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 219 { |
220 const DCTELEM *p; | |
1064 | 221 uint8_t *pix; |
0 | 222 |
223 /* read the pixels */ | |
224 p = block; | |
225 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
226 /* unrolled loop */ |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
227 asm volatile( |
2979 | 228 "movq %3, %%mm0 \n\t" |
229 "movq 8%3, %%mm1 \n\t" | |
230 "movq 16%3, %%mm2 \n\t" | |
231 "movq 24%3, %%mm3 \n\t" | |
232 "movq 32%3, %%mm4 \n\t" | |
233 "movq 40%3, %%mm5 \n\t" | |
234 "movq 48%3, %%mm6 \n\t" | |
235 "movq 56%3, %%mm7 \n\t" | |
236 "packuswb %%mm1, %%mm0 \n\t" | |
237 "packuswb %%mm3, %%mm2 \n\t" | |
238 "packuswb %%mm5, %%mm4 \n\t" | |
239 "packuswb %%mm7, %%mm6 \n\t" | |
240 "movq %%mm0, (%0) \n\t" | |
241 "movq %%mm2, (%0, %1) \n\t" | |
242 "movq %%mm4, (%0, %1, 2) \n\t" | |
243 "movq %%mm6, (%0, %2) \n\t" | |
244 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "m"(*p) | |
245 :"memory"); | |
0 | 246 pix += line_size*4; |
247 p += 32; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
248 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
249 // 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
|
250 // compiler would generate some very strange code |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
251 // thus using "r" |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
252 asm volatile( |
2979 | 253 "movq (%3), %%mm0 \n\t" |
254 "movq 8(%3), %%mm1 \n\t" | |
255 "movq 16(%3), %%mm2 \n\t" | |
256 "movq 24(%3), %%mm3 \n\t" | |
257 "movq 32(%3), %%mm4 \n\t" | |
258 "movq 40(%3), %%mm5 \n\t" | |
259 "movq 48(%3), %%mm6 \n\t" | |
260 "movq 56(%3), %%mm7 \n\t" | |
261 "packuswb %%mm1, %%mm0 \n\t" | |
262 "packuswb %%mm3, %%mm2 \n\t" | |
263 "packuswb %%mm5, %%mm4 \n\t" | |
264 "packuswb %%mm7, %%mm6 \n\t" | |
265 "movq %%mm0, (%0) \n\t" | |
266 "movq %%mm2, (%0, %1) \n\t" | |
267 "movq %%mm4, (%0, %1, 2) \n\t" | |
268 "movq %%mm6, (%0, %2) \n\t" | |
269 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "r"(p) | |
270 :"memory"); | |
0 | 271 } |
272 | |
3089 | 273 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
|
274 { 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
|
275 |
1984
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
276 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
|
277 { |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
278 int i; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
279 |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
280 movq_m2r(*vector128, mm1); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
281 for (i = 0; i < 8; i++) { |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
282 movq_m2r(*(block), mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
283 packsswb_m2r(*(block + 4), mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
284 block += 8; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
285 paddb_r2r(mm1, mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
286 movq_r2m(mm0, *pixels); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
287 pixels += line_size; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
288 } |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
289 } |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
290 |
1064 | 291 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 292 { |
293 const DCTELEM *p; | |
1064 | 294 uint8_t *pix; |
0 | 295 int i; |
296 | |
297 /* read the pixels */ | |
298 p = block; | |
299 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
300 MOVQ_ZERO(mm7); |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
301 i = 4; |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
302 do { |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
303 asm volatile( |
2979 | 304 "movq (%2), %%mm0 \n\t" |
305 "movq 8(%2), %%mm1 \n\t" | |
306 "movq 16(%2), %%mm2 \n\t" | |
307 "movq 24(%2), %%mm3 \n\t" | |
308 "movq %0, %%mm4 \n\t" | |
309 "movq %1, %%mm6 \n\t" | |
310 "movq %%mm4, %%mm5 \n\t" | |
311 "punpcklbw %%mm7, %%mm4 \n\t" | |
312 "punpckhbw %%mm7, %%mm5 \n\t" | |
313 "paddsw %%mm4, %%mm0 \n\t" | |
314 "paddsw %%mm5, %%mm1 \n\t" | |
315 "movq %%mm6, %%mm5 \n\t" | |
316 "punpcklbw %%mm7, %%mm6 \n\t" | |
317 "punpckhbw %%mm7, %%mm5 \n\t" | |
318 "paddsw %%mm6, %%mm2 \n\t" | |
319 "paddsw %%mm5, %%mm3 \n\t" | |
320 "packuswb %%mm1, %%mm0 \n\t" | |
321 "packuswb %%mm3, %%mm2 \n\t" | |
322 "movq %%mm0, %0 \n\t" | |
323 "movq %%mm2, %1 \n\t" | |
324 :"+m"(*pix), "+m"(*(pix+line_size)) | |
325 :"r"(p) | |
326 :"memory"); | |
0 | 327 pix += line_size*2; |
328 p += 16; | |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
329 } while (--i); |
0 | 330 } |
331 | |
2209 | 332 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
333 { | |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
334 asm volatile( |
2979 | 335 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
336 ASMALIGN(3) |
2979 | 337 "1: \n\t" |
338 "movd (%1), %%mm0 \n\t" | |
339 "movd (%1, %3), %%mm1 \n\t" | |
340 "movd %%mm0, (%2) \n\t" | |
341 "movd %%mm1, (%2, %3) \n\t" | |
342 "add %%"REG_a", %1 \n\t" | |
343 "add %%"REG_a", %2 \n\t" | |
344 "movd (%1), %%mm0 \n\t" | |
345 "movd (%1, %3), %%mm1 \n\t" | |
346 "movd %%mm0, (%2) \n\t" | |
347 "movd %%mm1, (%2, %3) \n\t" | |
348 "add %%"REG_a", %1 \n\t" | |
349 "add %%"REG_a", %2 \n\t" | |
350 "subl $4, %0 \n\t" | |
351 "jnz 1b \n\t" | |
352 : "+g"(h), "+r" (pixels), "+r" (block) | |
353 : "r"((long)line_size) | |
354 : "%"REG_a, "memory" | |
355 ); | |
2209 | 356 } |
357 | |
1064 | 358 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 359 { |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
360 asm volatile( |
2979 | 361 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
362 ASMALIGN(3) |
2979 | 363 "1: \n\t" |
364 "movq (%1), %%mm0 \n\t" | |
365 "movq (%1, %3), %%mm1 \n\t" | |
366 "movq %%mm0, (%2) \n\t" | |
367 "movq %%mm1, (%2, %3) \n\t" | |
368 "add %%"REG_a", %1 \n\t" | |
369 "add %%"REG_a", %2 \n\t" | |
370 "movq (%1), %%mm0 \n\t" | |
371 "movq (%1, %3), %%mm1 \n\t" | |
372 "movq %%mm0, (%2) \n\t" | |
373 "movq %%mm1, (%2, %3) \n\t" | |
374 "add %%"REG_a", %1 \n\t" | |
375 "add %%"REG_a", %2 \n\t" | |
376 "subl $4, %0 \n\t" | |
377 "jnz 1b \n\t" | |
378 : "+g"(h), "+r" (pixels), "+r" (block) | |
379 : "r"((long)line_size) | |
380 : "%"REG_a, "memory" | |
381 ); | |
0 | 382 } |
383 | |
1064 | 384 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
651 | 385 { |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
386 asm volatile( |
2979 | 387 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
388 ASMALIGN(3) |
2979 | 389 "1: \n\t" |
390 "movq (%1), %%mm0 \n\t" | |
391 "movq 8(%1), %%mm4 \n\t" | |
392 "movq (%1, %3), %%mm1 \n\t" | |
393 "movq 8(%1, %3), %%mm5 \n\t" | |
394 "movq %%mm0, (%2) \n\t" | |
395 "movq %%mm4, 8(%2) \n\t" | |
396 "movq %%mm1, (%2, %3) \n\t" | |
397 "movq %%mm5, 8(%2, %3) \n\t" | |
398 "add %%"REG_a", %1 \n\t" | |
399 "add %%"REG_a", %2 \n\t" | |
400 "movq (%1), %%mm0 \n\t" | |
401 "movq 8(%1), %%mm4 \n\t" | |
402 "movq (%1, %3), %%mm1 \n\t" | |
403 "movq 8(%1, %3), %%mm5 \n\t" | |
404 "movq %%mm0, (%2) \n\t" | |
405 "movq %%mm4, 8(%2) \n\t" | |
406 "movq %%mm1, (%2, %3) \n\t" | |
407 "movq %%mm5, 8(%2, %3) \n\t" | |
408 "add %%"REG_a", %1 \n\t" | |
409 "add %%"REG_a", %2 \n\t" | |
410 "subl $4, %0 \n\t" | |
411 "jnz 1b \n\t" | |
412 : "+g"(h), "+r" (pixels), "+r" (block) | |
413 : "r"((long)line_size) | |
414 : "%"REG_a, "memory" | |
415 ); | |
651 | 416 } |
417 | |
6331 | 418 static void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
419 { | |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
420 asm volatile( |
6331 | 421 "1: \n\t" |
422 "movdqu (%1), %%xmm0 \n\t" | |
423 "movdqu (%1,%3), %%xmm1 \n\t" | |
424 "movdqu (%1,%3,2), %%xmm2 \n\t" | |
425 "movdqu (%1,%4), %%xmm3 \n\t" | |
426 "movdqa %%xmm0, (%2) \n\t" | |
427 "movdqa %%xmm1, (%2,%3) \n\t" | |
428 "movdqa %%xmm2, (%2,%3,2) \n\t" | |
429 "movdqa %%xmm3, (%2,%4) \n\t" | |
430 "subl $4, %0 \n\t" | |
431 "lea (%1,%3,4), %1 \n\t" | |
432 "lea (%2,%3,4), %2 \n\t" | |
433 "jnz 1b \n\t" | |
434 : "+g"(h), "+r" (pixels), "+r" (block) | |
435 : "r"((long)line_size), "r"(3L*line_size) | |
436 : "memory" | |
437 ); | |
438 } | |
439 | |
440 static void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
441 { | |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
442 asm volatile( |
6331 | 443 "1: \n\t" |
444 "movdqu (%1), %%xmm0 \n\t" | |
445 "movdqu (%1,%3), %%xmm1 \n\t" | |
446 "movdqu (%1,%3,2), %%xmm2 \n\t" | |
447 "movdqu (%1,%4), %%xmm3 \n\t" | |
448 "pavgb (%2), %%xmm0 \n\t" | |
449 "pavgb (%2,%3), %%xmm1 \n\t" | |
450 "pavgb (%2,%3,2), %%xmm2 \n\t" | |
451 "pavgb (%2,%4), %%xmm3 \n\t" | |
452 "movdqa %%xmm0, (%2) \n\t" | |
453 "movdqa %%xmm1, (%2,%3) \n\t" | |
454 "movdqa %%xmm2, (%2,%3,2) \n\t" | |
455 "movdqa %%xmm3, (%2,%4) \n\t" | |
456 "subl $4, %0 \n\t" | |
457 "lea (%1,%3,4), %1 \n\t" | |
458 "lea (%2,%3,4), %2 \n\t" | |
459 "jnz 1b \n\t" | |
460 : "+g"(h), "+r" (pixels), "+r" (block) | |
461 : "r"((long)line_size), "r"(3L*line_size) | |
462 : "memory" | |
463 ); | |
464 } | |
465 | |
296 | 466 static void clear_blocks_mmx(DCTELEM *blocks) |
467 { | |
6391
3dc36ec2dcad
__asm __volatile -> asm volatile, improves code consistency and works
reimar
parents:
6384
diff
changeset
|
468 asm volatile( |
2979 | 469 "pxor %%mm7, %%mm7 \n\t" |
470 "mov $-128*6, %%"REG_a" \n\t" | |
471 "1: \n\t" | |
472 "movq %%mm7, (%0, %%"REG_a") \n\t" | |
473 "movq %%mm7, 8(%0, %%"REG_a") \n\t" | |
474 "movq %%mm7, 16(%0, %%"REG_a") \n\t" | |
475 "movq %%mm7, 24(%0, %%"REG_a") \n\t" | |
476 "add $32, %%"REG_a" \n\t" | |
477 " 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
|
478 : : "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
|
479 : "%"REG_a |
296 | 480 ); |
481 } | |
482 | |
866 | 483 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
|
484 long i=0; |
866 | 485 asm volatile( |
2979 | 486 "1: \n\t" |
487 "movq (%1, %0), %%mm0 \n\t" | |
488 "movq (%2, %0), %%mm1 \n\t" | |
489 "paddb %%mm0, %%mm1 \n\t" | |
490 "movq %%mm1, (%2, %0) \n\t" | |
491 "movq 8(%1, %0), %%mm0 \n\t" | |
492 "movq 8(%2, %0), %%mm1 \n\t" | |
493 "paddb %%mm0, %%mm1 \n\t" | |
494 "movq %%mm1, 8(%2, %0) \n\t" | |
495 "add $16, %0 \n\t" | |
496 "cmp %3, %0 \n\t" | |
497 " jb 1b \n\t" | |
866 | 498 : "+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
|
499 : "r"(src), "r"(dst), "r"((long)w-15) |
866 | 500 ); |
501 for(; i<w; i++) | |
502 dst[i+0] += src[i+0]; | |
503 } | |
504 | |
6384 | 505 static void add_bytes_l2_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){ |
506 long i=0; | |
507 asm volatile( | |
508 "1: \n\t" | |
509 "movq (%2, %0), %%mm0 \n\t" | |
510 "movq 8(%2, %0), %%mm1 \n\t" | |
511 "paddb (%3, %0), %%mm0 \n\t" | |
512 "paddb 8(%3, %0), %%mm1 \n\t" | |
513 "movq %%mm0, (%1, %0) \n\t" | |
514 "movq %%mm1, 8(%1, %0) \n\t" | |
515 "add $16, %0 \n\t" | |
516 "cmp %4, %0 \n\t" | |
517 " jb 1b \n\t" | |
518 : "+r" (i) | |
519 : "r"(dst), "r"(src1), "r"(src2), "r"((long)w-15) | |
520 ); | |
521 for(; i<w; i++) | |
522 dst[i] = src1[i] + src2[i]; | |
523 } | |
524 | |
1648 | 525 #define H263_LOOP_FILTER \ |
2979 | 526 "pxor %%mm7, %%mm7 \n\t"\ |
527 "movq %0, %%mm0 \n\t"\ | |
528 "movq %0, %%mm1 \n\t"\ | |
529 "movq %3, %%mm2 \n\t"\ | |
530 "movq %3, %%mm3 \n\t"\ | |
531 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
532 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
533 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
534 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
535 "psubw %%mm2, %%mm0 \n\t"\ | |
536 "psubw %%mm3, %%mm1 \n\t"\ | |
537 "movq %1, %%mm2 \n\t"\ | |
538 "movq %1, %%mm3 \n\t"\ | |
539 "movq %2, %%mm4 \n\t"\ | |
540 "movq %2, %%mm5 \n\t"\ | |
541 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
542 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
543 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
544 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
545 "psubw %%mm2, %%mm4 \n\t"\ | |
546 "psubw %%mm3, %%mm5 \n\t"\ | |
547 "psllw $2, %%mm4 \n\t"\ | |
548 "psllw $2, %%mm5 \n\t"\ | |
549 "paddw %%mm0, %%mm4 \n\t"\ | |
550 "paddw %%mm1, %%mm5 \n\t"\ | |
551 "pxor %%mm6, %%mm6 \n\t"\ | |
552 "pcmpgtw %%mm4, %%mm6 \n\t"\ | |
553 "pcmpgtw %%mm5, %%mm7 \n\t"\ | |
554 "pxor %%mm6, %%mm4 \n\t"\ | |
555 "pxor %%mm7, %%mm5 \n\t"\ | |
556 "psubw %%mm6, %%mm4 \n\t"\ | |
557 "psubw %%mm7, %%mm5 \n\t"\ | |
558 "psrlw $3, %%mm4 \n\t"\ | |
559 "psrlw $3, %%mm5 \n\t"\ | |
560 "packuswb %%mm5, %%mm4 \n\t"\ | |
561 "packsswb %%mm7, %%mm6 \n\t"\ | |
562 "pxor %%mm7, %%mm7 \n\t"\ | |
563 "movd %4, %%mm2 \n\t"\ | |
564 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
565 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
566 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
567 "psubusb %%mm4, %%mm2 \n\t"\ | |
568 "movq %%mm2, %%mm3 \n\t"\ | |
569 "psubusb %%mm4, %%mm3 \n\t"\ | |
570 "psubb %%mm3, %%mm2 \n\t"\ | |
571 "movq %1, %%mm3 \n\t"\ | |
572 "movq %2, %%mm4 \n\t"\ | |
573 "pxor %%mm6, %%mm3 \n\t"\ | |
574 "pxor %%mm6, %%mm4 \n\t"\ | |
575 "paddusb %%mm2, %%mm3 \n\t"\ | |
576 "psubusb %%mm2, %%mm4 \n\t"\ | |
577 "pxor %%mm6, %%mm3 \n\t"\ | |
578 "pxor %%mm6, %%mm4 \n\t"\ | |
579 "paddusb %%mm2, %%mm2 \n\t"\ | |
580 "packsswb %%mm1, %%mm0 \n\t"\ | |
581 "pcmpgtb %%mm0, %%mm7 \n\t"\ | |
582 "pxor %%mm7, %%mm0 \n\t"\ | |
583 "psubb %%mm7, %%mm0 \n\t"\ | |
584 "movq %%mm0, %%mm1 \n\t"\ | |
585 "psubusb %%mm2, %%mm0 \n\t"\ | |
586 "psubb %%mm0, %%mm1 \n\t"\ | |
587 "pand %5, %%mm1 \n\t"\ | |
588 "psrlw $2, %%mm1 \n\t"\ | |
589 "pxor %%mm7, %%mm1 \n\t"\ | |
590 "psubb %%mm7, %%mm1 \n\t"\ | |
591 "movq %0, %%mm5 \n\t"\ | |
592 "movq %3, %%mm6 \n\t"\ | |
593 "psubb %%mm1, %%mm5 \n\t"\ | |
594 "paddb %%mm1, %%mm6 \n\t" | |
1648 | 595 |
1647 | 596 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
|
597 if(ENABLE_ANY_H263) { |
1647 | 598 const int strength= ff_h263_loop_filter_strength[qscale]; |
599 | |
600 asm volatile( | |
2967 | 601 |
1648 | 602 H263_LOOP_FILTER |
2967 | 603 |
2979 | 604 "movq %%mm3, %1 \n\t" |
605 "movq %%mm4, %2 \n\t" | |
606 "movq %%mm5, %0 \n\t" | |
607 "movq %%mm6, %3 \n\t" | |
1647 | 608 : "+m" (*(uint64_t*)(src - 2*stride)), |
609 "+m" (*(uint64_t*)(src - 1*stride)), | |
610 "+m" (*(uint64_t*)(src + 0*stride)), | |
611 "+m" (*(uint64_t*)(src + 1*stride)) | |
612 : "g" (2*strength), "m"(ff_pb_FC) | |
613 ); | |
5394
e9a6215f4e3a
help some gcc version to optimize out those functions
aurel
parents:
5278
diff
changeset
|
614 } |
1647 | 615 } |
616 | |
1648 | 617 static inline void transpose4x4(uint8_t *dst, uint8_t *src, int dst_stride, int src_stride){ |
618 asm volatile( //FIXME could save 1 instruction if done as 8x4 ... | |
2979 | 619 "movd %4, %%mm0 \n\t" |
620 "movd %5, %%mm1 \n\t" | |
621 "movd %6, %%mm2 \n\t" | |
622 "movd %7, %%mm3 \n\t" | |
623 "punpcklbw %%mm1, %%mm0 \n\t" | |
624 "punpcklbw %%mm3, %%mm2 \n\t" | |
625 "movq %%mm0, %%mm1 \n\t" | |
626 "punpcklwd %%mm2, %%mm0 \n\t" | |
627 "punpckhwd %%mm2, %%mm1 \n\t" | |
628 "movd %%mm0, %0 \n\t" | |
629 "punpckhdq %%mm0, %%mm0 \n\t" | |
630 "movd %%mm0, %1 \n\t" | |
631 "movd %%mm1, %2 \n\t" | |
632 "punpckhdq %%mm1, %%mm1 \n\t" | |
633 "movd %%mm1, %3 \n\t" | |
2967 | 634 |
1648 | 635 : "=m" (*(uint32_t*)(dst + 0*dst_stride)), |
636 "=m" (*(uint32_t*)(dst + 1*dst_stride)), | |
637 "=m" (*(uint32_t*)(dst + 2*dst_stride)), | |
638 "=m" (*(uint32_t*)(dst + 3*dst_stride)) | |
639 : "m" (*(uint32_t*)(src + 0*src_stride)), | |
640 "m" (*(uint32_t*)(src + 1*src_stride)), | |
641 "m" (*(uint32_t*)(src + 2*src_stride)), | |
642 "m" (*(uint32_t*)(src + 3*src_stride)) | |
643 ); | |
644 } | |
645 | |
646 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
|
647 if(ENABLE_ANY_H263) { |
1648 | 648 const int strength= ff_h263_loop_filter_strength[qscale]; |
6181 | 649 DECLARE_ALIGNED(8, uint64_t, temp[4]); |
1648 | 650 uint8_t *btemp= (uint8_t*)temp; |
2967 | 651 |
1648 | 652 src -= 2; |
653 | |
654 transpose4x4(btemp , src , 8, stride); | |
655 transpose4x4(btemp+4, src + 4*stride, 8, stride); | |
656 asm volatile( | |
657 H263_LOOP_FILTER // 5 3 4 6 | |
2967 | 658 |
1648 | 659 : "+m" (temp[0]), |
660 "+m" (temp[1]), | |
661 "+m" (temp[2]), | |
662 "+m" (temp[3]) | |
663 : "g" (2*strength), "m"(ff_pb_FC) | |
664 ); | |
665 | |
666 asm volatile( | |
2979 | 667 "movq %%mm5, %%mm1 \n\t" |
668 "movq %%mm4, %%mm0 \n\t" | |
669 "punpcklbw %%mm3, %%mm5 \n\t" | |
670 "punpcklbw %%mm6, %%mm4 \n\t" | |
671 "punpckhbw %%mm3, %%mm1 \n\t" | |
672 "punpckhbw %%mm6, %%mm0 \n\t" | |
673 "movq %%mm5, %%mm3 \n\t" | |
674 "movq %%mm1, %%mm6 \n\t" | |
675 "punpcklwd %%mm4, %%mm5 \n\t" | |
676 "punpcklwd %%mm0, %%mm1 \n\t" | |
677 "punpckhwd %%mm4, %%mm3 \n\t" | |
678 "punpckhwd %%mm0, %%mm6 \n\t" | |
679 "movd %%mm5, (%0) \n\t" | |
680 "punpckhdq %%mm5, %%mm5 \n\t" | |
681 "movd %%mm5, (%0,%2) \n\t" | |
682 "movd %%mm3, (%0,%2,2) \n\t" | |
683 "punpckhdq %%mm3, %%mm3 \n\t" | |
684 "movd %%mm3, (%0,%3) \n\t" | |
685 "movd %%mm1, (%1) \n\t" | |
686 "punpckhdq %%mm1, %%mm1 \n\t" | |
687 "movd %%mm1, (%1,%2) \n\t" | |
688 "movd %%mm6, (%1,%2,2) \n\t" | |
689 "punpckhdq %%mm6, %%mm6 \n\t" | |
690 "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
|
691 :: "r" (src), |
86e2b1424801
optimization and gcc 4.0 bug workaround patch by (Martin Drab >drab kepler.fjfi.cvut cz<)
michael
parents:
2293
diff
changeset
|
692 "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
|
693 "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
|
694 "r" ((long)(3*stride)) |
1648 | 695 ); |
5394
e9a6215f4e3a
help some gcc version to optimize out those functions
aurel
parents:
5278
diff
changeset
|
696 } |
1648 | 697 } |
698 | |
6384 | 699 #define PAETH(cpu, abs3)\ |
700 void add_png_paeth_prediction_##cpu(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp)\ | |
701 {\ | |
702 long i = -bpp;\ | |
703 long end = w-3;\ | |
704 asm volatile(\ | |
705 "pxor %%mm7, %%mm7 \n"\ | |
706 "movd (%1,%0), %%mm0 \n"\ | |
707 "movd (%2,%0), %%mm1 \n"\ | |
708 "punpcklbw %%mm7, %%mm0 \n"\ | |
709 "punpcklbw %%mm7, %%mm1 \n"\ | |
710 "add %4, %0 \n"\ | |
711 "1: \n"\ | |
712 "movq %%mm1, %%mm2 \n"\ | |
713 "movd (%2,%0), %%mm1 \n"\ | |
714 "movq %%mm2, %%mm3 \n"\ | |
715 "punpcklbw %%mm7, %%mm1 \n"\ | |
716 "movq %%mm2, %%mm4 \n"\ | |
717 "psubw %%mm1, %%mm3 \n"\ | |
718 "psubw %%mm0, %%mm4 \n"\ | |
719 "movq %%mm3, %%mm5 \n"\ | |
720 "paddw %%mm4, %%mm5 \n"\ | |
721 abs3\ | |
722 "movq %%mm4, %%mm6 \n"\ | |
723 "pminsw %%mm5, %%mm6 \n"\ | |
724 "pcmpgtw %%mm6, %%mm3 \n"\ | |
725 "pcmpgtw %%mm5, %%mm4 \n"\ | |
726 "movq %%mm4, %%mm6 \n"\ | |
727 "pand %%mm3, %%mm4 \n"\ | |
728 "pandn %%mm3, %%mm6 \n"\ | |
729 "pandn %%mm0, %%mm3 \n"\ | |
730 "movd (%3,%0), %%mm0 \n"\ | |
731 "pand %%mm1, %%mm6 \n"\ | |
732 "pand %%mm4, %%mm2 \n"\ | |
733 "punpcklbw %%mm7, %%mm0 \n"\ | |
734 "movq %6, %%mm5 \n"\ | |
735 "paddw %%mm6, %%mm0 \n"\ | |
736 "paddw %%mm2, %%mm3 \n"\ | |
737 "paddw %%mm3, %%mm0 \n"\ | |
738 "pand %%mm5, %%mm0 \n"\ | |
739 "movq %%mm0, %%mm3 \n"\ | |
740 "packuswb %%mm3, %%mm3 \n"\ | |
741 "movd %%mm3, (%1,%0) \n"\ | |
742 "add %4, %0 \n"\ | |
743 "cmp %5, %0 \n"\ | |
744 "jle 1b \n"\ | |
745 :"+r"(i)\ | |
746 :"r"(dst), "r"(top), "r"(src), "r"((long)bpp), "g"(end),\ | |
747 "m"(ff_pw_255)\ | |
748 :"memory"\ | |
749 );\ | |
750 } | |
751 | |
752 #define ABS3_MMX2\ | |
753 "psubw %%mm5, %%mm7 \n"\ | |
754 "pmaxsw %%mm7, %%mm5 \n"\ | |
755 "pxor %%mm6, %%mm6 \n"\ | |
756 "pxor %%mm7, %%mm7 \n"\ | |
757 "psubw %%mm3, %%mm6 \n"\ | |
758 "psubw %%mm4, %%mm7 \n"\ | |
759 "pmaxsw %%mm6, %%mm3 \n"\ | |
760 "pmaxsw %%mm7, %%mm4 \n"\ | |
761 "pxor %%mm7, %%mm7 \n" | |
762 | |
763 #define ABS3_SSSE3\ | |
764 "pabsw %%mm3, %%mm3 \n"\ | |
765 "pabsw %%mm4, %%mm4 \n"\ | |
766 "pabsw %%mm5, %%mm5 \n" | |
767 | |
768 PAETH(mmx2, ABS3_MMX2) | |
769 #ifdef HAVE_SSSE3 | |
770 PAETH(ssse3, ABS3_SSSE3) | |
771 #endif | |
772 | |
954 | 773 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\ |
2979 | 774 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\ |
775 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\ | |
776 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\ | |
777 "movq "#in7", " #m3 " \n\t" /* d */\ | |
778 "movq "#in0", %%mm5 \n\t" /* D */\ | |
779 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\ | |
780 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\ | |
781 "movq "#in1", %%mm5 \n\t" /* C */\ | |
782 "movq "#in2", %%mm6 \n\t" /* B */\ | |
783 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\ | |
784 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\ | |
785 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\ | |
786 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\ | |
787 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\ | |
788 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\ | |
789 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\ | |
790 "psraw $5, %%mm5 \n\t"\ | |
791 "packuswb %%mm5, %%mm5 \n\t"\ | |
954 | 792 OP(%%mm5, out, %%mm7, d) |
793 | |
959 | 794 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\ |
1057 | 795 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
954 | 796 uint64_t temp;\ |
797 \ | |
798 asm volatile(\ | |
2979 | 799 "pxor %%mm7, %%mm7 \n\t"\ |
800 "1: \n\t"\ | |
801 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
802 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
803 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
804 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
805 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
806 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
807 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
808 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
809 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
810 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
811 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
812 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
813 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
814 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
815 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
816 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
817 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
818 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
819 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
820 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
821 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
822 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
823 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
824 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
825 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
826 "paddw %6, %%mm6 \n\t"\ | |
827 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
828 "psraw $5, %%mm0 \n\t"\ | |
829 "movq %%mm0, %5 \n\t"\ | |
954 | 830 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ |
831 \ | |
2979 | 832 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\ |
833 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\ | |
834 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\ | |
835 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\ | |
836 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\ | |
837 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\ | |
838 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\ | |
839 "paddw %%mm0, %%mm2 \n\t" /* b */\ | |
840 "paddw %%mm5, %%mm3 \n\t" /* c */\ | |
841 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
842 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
843 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\ | |
844 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\ | |
845 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\ | |
846 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\ | |
847 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
848 "paddw %%mm2, %%mm1 \n\t" /* a */\ | |
849 "paddw %%mm6, %%mm4 \n\t" /* d */\ | |
850 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
851 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\ | |
852 "paddw %6, %%mm1 \n\t"\ | |
853 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\ | |
854 "psraw $5, %%mm3 \n\t"\ | |
855 "movq %5, %%mm1 \n\t"\ | |
856 "packuswb %%mm3, %%mm1 \n\t"\ | |
959 | 857 OP_MMX2(%%mm1, (%1),%%mm4, q)\ |
954 | 858 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\ |
859 \ | |
2979 | 860 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\ |
861 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\ | |
862 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\ | |
863 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\ | |
864 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\ | |
865 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\ | |
866 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\ | |
867 "paddw %%mm1, %%mm5 \n\t" /* b */\ | |
868 "paddw %%mm4, %%mm0 \n\t" /* c */\ | |
869 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
870 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\ | |
871 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\ | |
872 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\ | |
873 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\ | |
874 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\ | |
875 "paddw %%mm3, %%mm2 \n\t" /* d */\ | |
876 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\ | |
877 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\ | |
878 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\ | |
879 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\ | |
880 "paddw %%mm2, %%mm6 \n\t" /* a */\ | |
881 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\ | |
882 "paddw %6, %%mm0 \n\t"\ | |
883 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
884 "psraw $5, %%mm0 \n\t"\ | |
954 | 885 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\ |
886 \ | |
2979 | 887 "paddw %%mm5, %%mm3 \n\t" /* a */\ |
888 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\ | |
889 "paddw %%mm4, %%mm6 \n\t" /* b */\ | |
890 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\ | |
891 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\ | |
892 "paddw %%mm1, %%mm4 \n\t" /* c */\ | |
893 "paddw %%mm2, %%mm5 \n\t" /* d */\ | |
894 "paddw %%mm6, %%mm6 \n\t" /* 2b */\ | |
895 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\ | |
896 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\ | |
897 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\ | |
898 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\ | |
899 "paddw %6, %%mm4 \n\t"\ | |
900 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\ | |
901 "psraw $5, %%mm4 \n\t"\ | |
902 "packuswb %%mm4, %%mm0 \n\t"\ | |
959 | 903 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\ |
954 | 904 \ |
2979 | 905 "add %3, %0 \n\t"\ |
906 "add %4, %1 \n\t"\ | |
907 "decl %2 \n\t"\ | |
908 " jnz 1b \n\t"\ | |
6335
950811a14eb3
put loop counter in a register if possible. makes some of the qpel functions 3% faster.
lorenm
parents:
6333
diff
changeset
|
909 : "+a"(src), "+c"(dst), "+g"(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
|
910 : "d"((long)srcStride), "S"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
966 | 911 : "memory"\ |
954 | 912 );\ |
913 }\ | |
914 \ | |
915 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
916 int i;\ | |
917 int16_t temp[16];\ | |
918 /* quick HACK, XXX FIXME MUST be optimized */\ | |
919 for(i=0; i<h; i++)\ | |
920 {\ | |
921 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
922 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
923 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
924 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
925 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
926 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\ | |
927 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\ | |
928 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\ | |
929 temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\ | |
930 temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\ | |
931 temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\ | |
932 temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\ | |
933 temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\ | |
934 temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\ | |
935 temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\ | |
936 temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\ | |
937 asm volatile(\ | |
2979 | 938 "movq (%0), %%mm0 \n\t"\ |
939 "movq 8(%0), %%mm1 \n\t"\ | |
940 "paddw %2, %%mm0 \n\t"\ | |
941 "paddw %2, %%mm1 \n\t"\ | |
942 "psraw $5, %%mm0 \n\t"\ | |
943 "psraw $5, %%mm1 \n\t"\ | |
944 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 945 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ |
2979 | 946 "movq 16(%0), %%mm0 \n\t"\ |
947 "movq 24(%0), %%mm1 \n\t"\ | |
948 "paddw %2, %%mm0 \n\t"\ | |
949 "paddw %2, %%mm1 \n\t"\ | |
950 "psraw $5, %%mm0 \n\t"\ | |
951 "psraw $5, %%mm1 \n\t"\ | |
952 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 953 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\ |
954 | 954 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ |
966 | 955 : "memory"\ |
954 | 956 );\ |
957 dst+=dstStride;\ | |
958 src+=srcStride;\ | |
959 }\ | |
960 }\ | |
961 \ | |
1057 | 962 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
959 | 963 uint64_t temp;\ |
964 \ | |
965 asm volatile(\ | |
2979 | 966 "pxor %%mm7, %%mm7 \n\t"\ |
967 "1: \n\t"\ | |
968 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
969 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
970 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
971 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
972 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
973 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
974 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
975 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
976 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
977 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
978 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
979 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
980 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
981 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
982 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
983 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
984 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
985 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
986 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
987 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
988 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
989 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
990 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
991 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
992 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
993 "paddw %6, %%mm6 \n\t"\ | |
994 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
995 "psraw $5, %%mm0 \n\t"\ | |
959 | 996 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ |
997 \ | |
2979 | 998 "movd 5(%0), %%mm5 \n\t" /* FGHI */\ |
999 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\ | |
1000 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\ | |
1001 "paddw %%mm5, %%mm1 \n\t" /* a */\ | |
1002 "paddw %%mm6, %%mm2 \n\t" /* b */\ | |
1003 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\ | |
1004 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\ | |
1005 "paddw %%mm6, %%mm3 \n\t" /* c */\ | |
1006 "paddw %%mm5, %%mm4 \n\t" /* d */\ | |
1007 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
1008 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
1009 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
1010 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
1011 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\ | |
1012 "paddw %6, %%mm1 \n\t"\ | |
1013 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\ | |
1014 "psraw $5, %%mm3 \n\t"\ | |
1015 "packuswb %%mm3, %%mm0 \n\t"\ | |
959 | 1016 OP_MMX2(%%mm0, (%1), %%mm4, q)\ |
1017 \ | |
2979 | 1018 "add %3, %0 \n\t"\ |
1019 "add %4, %1 \n\t"\ | |
1020 "decl %2 \n\t"\ | |
1021 " jnz 1b \n\t"\ | |
6335
950811a14eb3
put loop counter in a register if possible. makes some of the qpel functions 3% faster.
lorenm
parents:
6333
diff
changeset
|
1022 : "+a"(src), "+c"(dst), "+g"(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
|
1023 : "S"((long)srcStride), "D"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
966 | 1024 : "memory"\ |
959 | 1025 );\ |
1026 }\ | |
1027 \ | |
1028 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1029 int i;\ | |
1030 int16_t temp[8];\ | |
1031 /* quick HACK, XXX FIXME MUST be optimized */\ | |
1032 for(i=0; i<h; i++)\ | |
1033 {\ | |
1034 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
1035 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
1036 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
1037 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
1038 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
1039 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\ | |
1040 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\ | |
1041 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\ | |
1042 asm volatile(\ | |
2979 | 1043 "movq (%0), %%mm0 \n\t"\ |
1044 "movq 8(%0), %%mm1 \n\t"\ | |
1045 "paddw %2, %%mm0 \n\t"\ | |
1046 "paddw %2, %%mm1 \n\t"\ | |
1047 "psraw $5, %%mm0 \n\t"\ | |
1048 "psraw $5, %%mm1 \n\t"\ | |
1049 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 1050 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ |
1051 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ | |
966 | 1052 :"memory"\ |
959 | 1053 );\ |
1054 dst+=dstStride;\ | |
1055 src+=srcStride;\ | |
1056 }\ | |
1057 } | |
1058 | |
1059 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\ | |
1060 \ | |
1061 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
954 | 1062 uint64_t temp[17*4];\ |
1063 uint64_t *temp_ptr= temp;\ | |
1064 int count= 17;\ | |
1065 \ | |
1066 /*FIXME unroll */\ | |
1067 asm volatile(\ | |
2979 | 1068 "pxor %%mm7, %%mm7 \n\t"\ |
1069 "1: \n\t"\ | |
1070 "movq (%0), %%mm0 \n\t"\ | |
1071 "movq (%0), %%mm1 \n\t"\ | |
1072 "movq 8(%0), %%mm2 \n\t"\ | |
1073 "movq 8(%0), %%mm3 \n\t"\ | |
1074 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1075 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1076 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1077 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1078 "movq %%mm0, (%1) \n\t"\ | |
1079 "movq %%mm1, 17*8(%1) \n\t"\ | |
1080 "movq %%mm2, 2*17*8(%1) \n\t"\ | |
1081 "movq %%mm3, 3*17*8(%1) \n\t"\ | |
1082 "add $8, %1 \n\t"\ | |
1083 "add %3, %0 \n\t"\ | |
1084 "decl %2 \n\t"\ | |
1085 " jnz 1b \n\t"\ | |
954 | 1086 : "+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
|
1087 : "r" ((long)srcStride)\ |
966 | 1088 : "memory"\ |
954 | 1089 );\ |
1090 \ | |
1091 temp_ptr= temp;\ | |
1092 count=4;\ | |
1093 \ | |
1094 /*FIXME reorder for speed */\ | |
1095 asm volatile(\ | |
2979 | 1096 /*"pxor %%mm7, %%mm7 \n\t"*/\ |
1097 "1: \n\t"\ | |
1098 "movq (%0), %%mm0 \n\t"\ | |
1099 "movq 8(%0), %%mm1 \n\t"\ | |
1100 "movq 16(%0), %%mm2 \n\t"\ | |
1101 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 1102 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
1103 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
2979 | 1104 "add %4, %1 \n\t"\ |
961 | 1105 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 1106 \ |
961 | 1107 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
2979 | 1108 "add %4, %1 \n\t"\ |
961 | 1109 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
1110 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\ | |
2979 | 1111 "add %4, %1 \n\t"\ |
961 | 1112 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\ |
1113 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\ | |
2979 | 1114 "add %4, %1 \n\t"\ |
961 | 1115 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\ |
1116 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\ | |
2979 | 1117 "add %4, %1 \n\t"\ |
961 | 1118 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\ |
1119 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\ | |
2979 | 1120 "add %4, %1 \n\t"\ |
961 | 1121 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\ |
954 | 1122 \ |
961 | 1123 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\ |
2979 | 1124 "add %4, %1 \n\t" \ |
961 | 1125 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\ |
1126 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\ | |
954 | 1127 \ |
2979 | 1128 "add $136, %0 \n\t"\ |
1129 "add %6, %1 \n\t"\ | |
1130 "decl %2 \n\t"\ | |
1131 " 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
|
1132 \ |
967 | 1133 : "+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
|
1134 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-14*(long)dstStride)\ |
966 | 1135 :"memory"\ |
954 | 1136 );\ |
1137 }\ | |
1138 \ | |
1057 | 1139 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2209 | 1140 uint64_t temp[9*2];\ |
954 | 1141 uint64_t *temp_ptr= temp;\ |
1142 int count= 9;\ | |
1143 \ | |
1144 /*FIXME unroll */\ | |
1145 asm volatile(\ | |
2979 | 1146 "pxor %%mm7, %%mm7 \n\t"\ |
1147 "1: \n\t"\ | |
1148 "movq (%0), %%mm0 \n\t"\ | |
1149 "movq (%0), %%mm1 \n\t"\ | |
1150 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1151 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1152 "movq %%mm0, (%1) \n\t"\ | |
1153 "movq %%mm1, 9*8(%1) \n\t"\ | |
1154 "add $8, %1 \n\t"\ | |
1155 "add %3, %0 \n\t"\ | |
1156 "decl %2 \n\t"\ | |
1157 " jnz 1b \n\t"\ | |
954 | 1158 : "+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
|
1159 : "r" ((long)srcStride)\ |
966 | 1160 : "memory"\ |
954 | 1161 );\ |
1162 \ | |
1163 temp_ptr= temp;\ | |
1164 count=2;\ | |
1165 \ | |
1166 /*FIXME reorder for speed */\ | |
1167 asm volatile(\ | |
2979 | 1168 /*"pxor %%mm7, %%mm7 \n\t"*/\ |
1169 "1: \n\t"\ | |
1170 "movq (%0), %%mm0 \n\t"\ | |
1171 "movq 8(%0), %%mm1 \n\t"\ | |
1172 "movq 16(%0), %%mm2 \n\t"\ | |
1173 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 1174 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
1175 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
2979 | 1176 "add %4, %1 \n\t"\ |
961 | 1177 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 1178 \ |
961 | 1179 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
2979 | 1180 "add %4, %1 \n\t"\ |
961 | 1181 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
954 | 1182 \ |
961 | 1183 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\ |
2979 | 1184 "add %4, %1 \n\t"\ |
961 | 1185 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\ |
1186 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\ | |
954 | 1187 \ |
2979 | 1188 "add $72, %0 \n\t"\ |
1189 "add %6, %1 \n\t"\ | |
1190 "decl %2 \n\t"\ | |
1191 " jnz 1b \n\t"\ | |
954 | 1192 \ |
961 | 1193 : "+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
|
1194 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-6*(long)dstStride)\ |
966 | 1195 : "memory"\ |
1196 );\ | |
959 | 1197 }\ |
954 | 1198 \ |
1064 | 1199 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
6321 | 1200 OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);\ |
954 | 1201 }\ |
1202 \ | |
1064 | 1203 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1204 uint64_t temp[8];\ |
954 | 1205 uint8_t * const half= (uint8_t*)temp;\ |
1206 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
|
1207 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ |
954 | 1208 }\ |
1209 \ | |
1064 | 1210 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1211 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\ |
1212 }\ | |
1213 \ | |
1064 | 1214 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1215 uint64_t temp[8];\ |
954 | 1216 uint8_t * const half= (uint8_t*)temp;\ |
1217 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
|
1218 OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\ |
954 | 1219 }\ |
1220 \ | |
1064 | 1221 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1222 uint64_t temp[8];\ |
954 | 1223 uint8_t * const half= (uint8_t*)temp;\ |
959 | 1224 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
|
1225 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ |
954 | 1226 }\ |
1227 \ | |
1064 | 1228 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 1229 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 1230 }\ |
1231 \ | |
1064 | 1232 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1233 uint64_t temp[8];\ |
954 | 1234 uint8_t * const half= (uint8_t*)temp;\ |
959 | 1235 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
|
1236 OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\ |
954 | 1237 }\ |
1064 | 1238 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1239 uint64_t half[8 + 9];\ |
1240 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1241 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1242 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
|
1243 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
959 | 1244 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
|
1245 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 1246 }\ |
1064 | 1247 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1248 uint64_t half[8 + 9];\ |
1249 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1250 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1251 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
|
1252 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
959 | 1253 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
|
1254 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 1255 }\ |
1064 | 1256 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1257 uint64_t half[8 + 9];\ |
1258 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1259 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1260 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
|
1261 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
959 | 1262 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
|
1263 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 1264 }\ |
1064 | 1265 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1266 uint64_t half[8 + 9];\ |
1267 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1268 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1269 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
|
1270 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
959 | 1271 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
|
1272 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 1273 }\ |
1064 | 1274 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1275 uint64_t half[8 + 9];\ |
954 | 1276 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
1277 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1278 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 1279 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
|
1280 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 1281 }\ |
1064 | 1282 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1283 uint64_t half[8 + 9];\ |
954 | 1284 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
1285 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1286 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 1287 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
|
1288 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 1289 }\ |
1064 | 1290 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1291 uint64_t half[8 + 9];\ |
1292 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 1293 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
|
1294 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
984 | 1295 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 1296 }\ |
1064 | 1297 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1298 uint64_t half[8 + 9];\ |
1299 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 1300 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
|
1301 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
984 | 1302 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 1303 }\ |
1064 | 1304 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1305 uint64_t half[9];\ |
954 | 1306 uint8_t * const halfH= ((uint8_t*)half);\ |
1307 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 1308 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 1309 }\ |
1064 | 1310 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
6321 | 1311 OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);\ |
954 | 1312 }\ |
1313 \ | |
1064 | 1314 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1315 uint64_t temp[32];\ |
1316 uint8_t * const half= (uint8_t*)temp;\ | |
1317 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
|
1318 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ |
954 | 1319 }\ |
1320 \ | |
1064 | 1321 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1322 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\ |
1323 }\ | |
1324 \ | |
1064 | 1325 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1326 uint64_t temp[32];\ |
1327 uint8_t * const half= (uint8_t*)temp;\ | |
1328 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
|
1329 OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\ |
954 | 1330 }\ |
1331 \ | |
1064 | 1332 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1333 uint64_t temp[32];\ |
1334 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 1335 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
|
1336 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ |
954 | 1337 }\ |
1338 \ | |
1064 | 1339 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 1340 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 1341 }\ |
1342 \ | |
1064 | 1343 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1344 uint64_t temp[32];\ |
1345 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 1346 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
|
1347 OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\ |
954 | 1348 }\ |
1064 | 1349 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1350 uint64_t half[16*2 + 17*2];\ |
1351 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1352 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1353 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
|
1354 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
959 | 1355 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
|
1356 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 1357 }\ |
1064 | 1358 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1359 uint64_t half[16*2 + 17*2];\ |
1360 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1361 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1362 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
|
1363 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
959 | 1364 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
|
1365 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 1366 }\ |
1064 | 1367 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1368 uint64_t half[16*2 + 17*2];\ |
1369 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1370 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1371 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
|
1372 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
959 | 1373 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
|
1374 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 1375 }\ |
1064 | 1376 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1377 uint64_t half[16*2 + 17*2];\ |
1378 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1379 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1380 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
|
1381 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
959 | 1382 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
|
1383 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 1384 }\ |
1064 | 1385 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1386 uint64_t half[16*2 + 17*2];\ |
1387 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1388 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1389 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 1390 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
|
1391 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 1392 }\ |
1064 | 1393 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1394 uint64_t half[16*2 + 17*2];\ |
1395 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1396 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1397 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 1398 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
|
1399 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 1400 }\ |
1064 | 1401 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1402 uint64_t half[17*2];\ |
1403 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 1404 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
|
1405 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
984 | 1406 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 1407 }\ |
1064 | 1408 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1409 uint64_t half[17*2];\ |
1410 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 1411 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
|
1412 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
984 | 1413 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 1414 }\ |
1064 | 1415 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1416 uint64_t half[17*2];\ |
1417 uint8_t * const halfH= ((uint8_t*)half);\ | |
1418 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 1419 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 1420 } |
1421 | |
2979 | 1422 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t" |
959 | 1423 #define AVG_3DNOW_OP(a,b,temp, size) \ |
2979 | 1424 "mov" #size " " #b ", " #temp " \n\t"\ |
1425 "pavgusb " #temp ", " #a " \n\t"\ | |
1426 "mov" #size " " #a ", " #b " \n\t" | |
959 | 1427 #define AVG_MMX2_OP(a,b,temp, size) \ |
2979 | 1428 "mov" #size " " #b ", " #temp " \n\t"\ |
1429 "pavgb " #temp ", " #a " \n\t"\ | |
1430 "mov" #size " " #a ", " #b " \n\t" | |
959 | 1431 |
1432 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP) | |
1433 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP) | |
1434 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP) | |
1435 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow) | |
1436 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow) | |
1437 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow) | |
954 | 1438 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2) |
959 | 1439 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2) |
954 | 1440 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2) |
1441 | |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1442 /***********************************/ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1443 /* 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
|
1444 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1445 #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
|
1446 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
|
1447 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
|
1448 } |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1449 #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
|
1450 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
|
1451 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
|
1452 } |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1453 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1454 #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
|
1455 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
|
1456 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
|
1457 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
|
1458 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
|
1459 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
|
1460 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
|
1461 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
|
1462 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
|
1463 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
|
1464 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
|
1465 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
|
1466 }\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1467 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
|
1468 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
|
1469 }\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1470 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
|
1471 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
|
1472 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
|
1473 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
|
1474 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
|
1475 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
|
1476 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
|
1477 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
|
1478 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1479 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
|
1480 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
|
1481 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
|
1482 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
|
1483 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
|
1484 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
|
1485 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
|
1486 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
|
1487 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
1488 |
393 | 1489 #if 0 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1490 static void just_return() { return; } |
393 | 1491 #endif |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1492 |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1493 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
|
1494 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
|
1495 const int w = 8; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1496 const int ix = ox>>(16+shift); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1497 const int iy = oy>>(16+shift); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1498 const int oxs = ox>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1499 const int oys = oy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1500 const int dxxs = dxx>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1501 const int dxys = dxy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1502 const int dyxs = dyx>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1503 const int dyys = dyy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1504 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
|
1505 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
|
1506 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
|
1507 const uint64_t shift2 = 2*shift; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1508 uint8_t edge_buf[(h+1)*stride]; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1509 int x, y; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1510 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1511 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
|
1512 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
|
1513 const int dxh = dxy*(h-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1514 const int dyw = dyx*(w-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1515 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
|
1516 ((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
|
1517 (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
|
1518 // 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
|
1519 || (dxx|dxy|dyx|dyy)&15 ) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1520 { |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1521 //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
|
1522 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
|
1523 return; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1524 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1525 |
3250 | 1526 src += ix + iy*stride; |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1527 if( (unsigned)ix >= width-w || |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1528 (unsigned)iy >= height-h ) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1529 { |
3250 | 1530 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
|
1531 src = edge_buf; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1532 } |
3250 | 1533 |
1534 asm volatile( | |
1535 "movd %0, %%mm6 \n\t" | |
1536 "pxor %%mm7, %%mm7 \n\t" | |
1537 "punpcklwd %%mm6, %%mm6 \n\t" | |
1538 "punpcklwd %%mm6, %%mm6 \n\t" | |
1539 :: "r"(1<<shift) | |
1540 ); | |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1541 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1542 for(x=0; x<w; x+=4){ |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1543 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
|
1544 oxs - dxys + dxxs*(x+1), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1545 oxs - dxys + dxxs*(x+2), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1546 oxs - dxys + dxxs*(x+3) }; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1547 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
|
1548 oys - dyys + dyxs*(x+1), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1549 oys - dyys + dyxs*(x+2), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1550 oys - dyys + dyxs*(x+3) }; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1551 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1552 for(y=0; y<h; y++){ |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1553 asm volatile( |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1554 "movq %0, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1555 "movq %1, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1556 "paddw %2, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1557 "paddw %3, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1558 "movq %%mm4, %0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1559 "movq %%mm5, %1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1560 "psrlw $12, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1561 "psrlw $12, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1562 : "+m"(*dx4), "+m"(*dy4) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1563 : "m"(*dxy4), "m"(*dyy4) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1564 ); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1565 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1566 asm volatile( |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1567 "movq %%mm6, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1568 "movq %%mm6, %%mm1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1569 "psubw %%mm4, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1570 "psubw %%mm5, %%mm1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1571 "movq %%mm2, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1572 "movq %%mm4, %%mm3 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1573 "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
|
1574 "pmullw %%mm5, %%mm3 \n\t" // dx*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1575 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1576 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1577 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1578 "movd %4, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1579 "movd %3, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1580 "punpcklbw %%mm7, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1581 "punpcklbw %%mm7, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1582 "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
|
1583 "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
|
1584 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1585 "movd %2, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1586 "movd %1, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1587 "punpcklbw %%mm7, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1588 "punpcklbw %%mm7, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1589 "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
|
1590 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy) |
3250 | 1591 "paddw %5, %%mm1 \n\t" |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1592 "paddw %%mm3, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1593 "paddw %%mm1, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1594 "paddw %%mm2, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1595 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1596 "psrlw %6, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1597 "packuswb %%mm0, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1598 "movd %%mm0, %0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1599 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1600 : "=m"(dst[x+y*stride]) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1601 : "m"(src[0]), "m"(src[1]), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1602 "m"(src[stride]), "m"(src[stride+1]), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1603 "m"(*r4), "m"(shift2) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1604 ); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1605 src += stride; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1606 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1607 src += 4-h*stride; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1608 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1609 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
1610 |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1611 #define PREFETCH(name, op) \ |
4172 | 1612 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
|
1613 const uint8_t *p= mem;\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1614 do{\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1615 asm volatile(#op" %0" :: "m"(*p));\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1616 p+= stride;\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1617 }while(--h);\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1618 } |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1619 PREFETCH(prefetch_mmx2, prefetcht0) |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1620 PREFETCH(prefetch_3dnow, prefetch) |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1621 #undef PREFETCH |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
1622 |
2754 | 1623 #include "h264dsp_mmx.c" |
2967 | 1624 |
6009 | 1625 /* CAVS specific */ |
3524 | 1626 void ff_cavsdsp_init_mmx2(DSPContext* c, AVCodecContext *avctx); |
1627 | |
1628 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1629 put_pixels8_mmx(dst, src, stride, 8); | |
1630 } | |
1631 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1632 avg_pixels8_mmx(dst, src, stride, 8); | |
1633 } | |
1634 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1635 put_pixels16_mmx(dst, src, stride, 16); | |
1636 } | |
1637 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
1638 avg_pixels16_mmx(dst, src, stride, 16); | |
1639 } | |
1640 | |
5948 | 1641 /* VC1 specific */ |
1642 void ff_vc1dsp_init_mmx(DSPContext* dsp, AVCodecContext *avctx); | |
1643 | |
1644 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) { | |
1645 put_pixels8_mmx(dst, src, stride, 8); | |
1646 } | |
1647 | |
1092 | 1648 /* external functions, from idct_mmx.c */ |
1649 void ff_mmx_idct(DCTELEM *block); | |
1650 void ff_mmxext_idct(DCTELEM *block); | |
1651 | |
1652 /* XXX: those functions should be suppressed ASAP when all IDCTs are | |
1653 converted */ | |
4020
723818b5de0f
Put libmpeg2 IDCT functions under CONFIG_GPL, fixes link failure
diego
parents:
4001
diff
changeset
|
1654 #ifdef CONFIG_GPL |
1092 | 1655 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block) |
1656 { | |
1657 ff_mmx_idct (block); | |
1658 put_pixels_clamped_mmx(block, dest, line_size); | |
1659 } | |
1660 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1661 { | |
1662 ff_mmx_idct (block); | |
1663 add_pixels_clamped_mmx(block, dest, line_size); | |
1664 } | |
1665 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1666 { | |
1667 ff_mmxext_idct (block); | |
1668 put_pixels_clamped_mmx(block, dest, line_size); | |
1669 } | |
1670 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1671 { | |
1672 ff_mmxext_idct (block); | |
1673 add_pixels_clamped_mmx(block, dest, line_size); | |
1674 } | |
4020
723818b5de0f
Put libmpeg2 IDCT functions under CONFIG_GPL, fixes link failure
diego
parents:
4001
diff
changeset
|
1675 #endif |
2868 | 1676 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block) |
1677 { | |
1678 ff_idct_xvid_mmx (block); | |
1679 put_pixels_clamped_mmx(block, dest, line_size); | |
1680 } | |
1681 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1682 { | |
1683 ff_idct_xvid_mmx (block); | |
1684 add_pixels_clamped_mmx(block, dest, line_size); | |
1685 } | |
1686 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1687 { | |
1688 ff_idct_xvid_mmx2 (block); | |
1689 put_pixels_clamped_mmx(block, dest, line_size); | |
1690 } | |
1691 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1692 { | |
1693 ff_idct_xvid_mmx2 (block); | |
1694 add_pixels_clamped_mmx(block, dest, line_size); | |
1695 } | |
2967 | 1696 |
3541
3fbddeb13686
10l, vorbis_inverse_coupling_sse() was really 3dnow
lorenm
parents:
3536
diff
changeset
|
1697 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
|
1698 { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1699 int i; |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1700 asm volatile("pxor %%mm7, %%mm7":); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1701 for(i=0; i<blocksize; i+=2) { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1702 asm volatile( |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1703 "movq %0, %%mm0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1704 "movq %1, %%mm1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1705 "movq %%mm0, %%mm2 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1706 "movq %%mm1, %%mm3 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1707 "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1708 "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1709 "pslld $31, %%mm2 \n\t" // keep only the sign bit |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1710 "pxor %%mm2, %%mm1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1711 "movq %%mm3, %%mm4 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1712 "pand %%mm1, %%mm3 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1713 "pandn %%mm1, %%mm4 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1714 "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
|
1715 "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
|
1716 "movq %%mm3, %1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1717 "movq %%mm0, %0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1718 :"+m"(mag[i]), "+m"(ang[i]) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1719 ::"memory" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1720 ); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1721 } |
3561 | 1722 asm volatile("femms"); |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1723 } |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1724 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
|
1725 { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1726 int i; |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1727 |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1728 asm volatile( |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1729 "movaps %0, %%xmm5 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1730 ::"m"(ff_pdw_80000000[0]) |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1731 ); |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1732 for(i=0; i<blocksize; i+=4) { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1733 asm volatile( |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1734 "movaps %0, %%xmm0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1735 "movaps %1, %%xmm1 \n\t" |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1736 "xorps %%xmm2, %%xmm2 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1737 "xorps %%xmm3, %%xmm3 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1738 "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1739 "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
|
1740 "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
|
1741 "xorps %%xmm2, %%xmm1 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1742 "movaps %%xmm3, %%xmm4 \n\t" |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1743 "andps %%xmm1, %%xmm3 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
1744 "andnps %%xmm1, %%xmm4 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1745 "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
|
1746 "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
|
1747 "movaps %%xmm3, %1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1748 "movaps %%xmm0, %0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1749 :"+m"(mag[i]), "+m"(ang[i]) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1750 ::"memory" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1751 ); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1752 } |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1753 } |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
1754 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1755 static void vector_fmul_3dnow(float *dst, const float *src, int len){ |
3574 | 1756 long i = (len-4)*4; |
1757 asm volatile( | |
1758 "1: \n\t" | |
1759 "movq (%1,%0), %%mm0 \n\t" | |
1760 "movq 8(%1,%0), %%mm1 \n\t" | |
1761 "pfmul (%2,%0), %%mm0 \n\t" | |
1762 "pfmul 8(%2,%0), %%mm1 \n\t" | |
1763 "movq %%mm0, (%1,%0) \n\t" | |
1764 "movq %%mm1, 8(%1,%0) \n\t" | |
1765 "sub $16, %0 \n\t" | |
1766 "jge 1b \n\t" | |
1767 "femms \n\t" | |
1768 :"+r"(i) | |
1769 :"r"(dst), "r"(src) | |
1770 :"memory" | |
1771 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1772 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1773 static void vector_fmul_sse(float *dst, const float *src, int len){ |
3574 | 1774 long i = (len-8)*4; |
1775 asm volatile( | |
1776 "1: \n\t" | |
1777 "movaps (%1,%0), %%xmm0 \n\t" | |
1778 "movaps 16(%1,%0), %%xmm1 \n\t" | |
1779 "mulps (%2,%0), %%xmm0 \n\t" | |
1780 "mulps 16(%2,%0), %%xmm1 \n\t" | |
1781 "movaps %%xmm0, (%1,%0) \n\t" | |
1782 "movaps %%xmm1, 16(%1,%0) \n\t" | |
1783 "sub $32, %0 \n\t" | |
1784 "jge 1b \n\t" | |
1785 :"+r"(i) | |
1786 :"r"(dst), "r"(src) | |
1787 :"memory" | |
1788 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1789 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1790 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1791 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
|
1792 long i = len*4-16; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1793 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1794 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1795 "pswapd 8(%1), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1796 "pswapd (%1), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1797 "pfmul (%3,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1798 "pfmul 8(%3,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1799 "movq %%mm0, (%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1800 "movq %%mm1, 8(%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1801 "add $16, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1802 "sub $16, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1803 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1804 :"+r"(i), "+r"(src1) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1805 :"r"(dst), "r"(src0) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1806 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1807 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1808 } |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
1809 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
|
1810 long i = len*4-32; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1811 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1812 "1: \n\t" |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
1813 "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
|
1814 "movaps (%1), %%xmm1 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
1815 "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
|
1816 "shufps $0x1b, %%xmm1, %%xmm1 \n\t" |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1817 "mulps (%3,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1818 "mulps 16(%3,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1819 "movaps %%xmm0, (%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1820 "movaps %%xmm1, 16(%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1821 "add $32, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1822 "sub $32, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1823 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1824 :"+r"(i), "+r"(src1) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1825 :"r"(dst), "r"(src0) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1826 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1827 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1828 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1829 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
|
1830 const float *src2, int src3, int len, int step){ |
3574 | 1831 long i = (len-4)*4; |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1832 if(step == 2 && src3 == 0){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1833 dst += (len-4)*2; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1834 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1835 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1836 "movq (%2,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1837 "movq 8(%2,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1838 "pfmul (%3,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1839 "pfmul 8(%3,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1840 "pfadd (%4,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1841 "pfadd 8(%4,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1842 "movd %%mm0, (%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1843 "movd %%mm1, 16(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1844 "psrlq $32, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1845 "psrlq $32, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1846 "movd %%mm0, 8(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1847 "movd %%mm1, 24(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1848 "sub $32, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1849 "sub $16, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1850 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1851 :"+r"(i), "+r"(dst) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1852 :"r"(src0), "r"(src1), "r"(src2) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1853 :"memory" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1854 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1855 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1856 else if(step == 1 && src3 == 0){ |
3574 | 1857 asm volatile( |
1858 "1: \n\t" | |
1859 "movq (%2,%0), %%mm0 \n\t" | |
1860 "movq 8(%2,%0), %%mm1 \n\t" | |
1861 "pfmul (%3,%0), %%mm0 \n\t" | |
1862 "pfmul 8(%3,%0), %%mm1 \n\t" | |
1863 "pfadd (%4,%0), %%mm0 \n\t" | |
1864 "pfadd 8(%4,%0), %%mm1 \n\t" | |
1865 "movq %%mm0, (%1,%0) \n\t" | |
1866 "movq %%mm1, 8(%1,%0) \n\t" | |
1867 "sub $16, %0 \n\t" | |
1868 "jge 1b \n\t" | |
1869 :"+r"(i) | |
1870 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
1871 :"memory" | |
1872 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1873 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1874 else |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1875 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
|
1876 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1877 } |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
1878 static void vector_fmul_add_add_sse(float *dst, const float *src0, const float *src1, |
3574 | 1879 const float *src2, int src3, int len, int step){ |
1880 long i = (len-8)*4; | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1881 if(step == 2 && src3 == 0){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1882 dst += (len-8)*2; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1883 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1884 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1885 "movaps (%2,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1886 "movaps 16(%2,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1887 "mulps (%3,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1888 "mulps 16(%3,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1889 "addps (%4,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1890 "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
|
1891 "movss %%xmm0, (%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
1892 "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
|
1893 "movhlps %%xmm0, %%xmm2 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
1894 "movhlps %%xmm1, %%xmm3 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
1895 "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
|
1896 "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
|
1897 "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
|
1898 "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
|
1899 "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
|
1900 "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
|
1901 "movhlps %%xmm0, %%xmm2 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
1902 "movhlps %%xmm1, %%xmm3 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
1903 "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
|
1904 "movss %%xmm3, 56(%1) \n\t" |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1905 "sub $64, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1906 "sub $32, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1907 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1908 :"+r"(i), "+r"(dst) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1909 :"r"(src0), "r"(src1), "r"(src2) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1910 :"memory" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1911 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1912 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1913 else if(step == 1 && src3 == 0){ |
3574 | 1914 asm volatile( |
1915 "1: \n\t" | |
1916 "movaps (%2,%0), %%xmm0 \n\t" | |
1917 "movaps 16(%2,%0), %%xmm1 \n\t" | |
1918 "mulps (%3,%0), %%xmm0 \n\t" | |
1919 "mulps 16(%3,%0), %%xmm1 \n\t" | |
1920 "addps (%4,%0), %%xmm0 \n\t" | |
1921 "addps 16(%4,%0), %%xmm1 \n\t" | |
1922 "movaps %%xmm0, (%1,%0) \n\t" | |
1923 "movaps %%xmm1, 16(%1,%0) \n\t" | |
1924 "sub $32, %0 \n\t" | |
1925 "jge 1b \n\t" | |
1926 :"+r"(i) | |
1927 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
1928 :"memory" | |
1929 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1930 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1931 else |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1932 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
|
1933 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1934 |
4172 | 1935 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
|
1936 // 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
|
1937 int i; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1938 for(i=0; i<len; i+=4) { |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1939 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1940 "pf2id %1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1941 "pf2id %2, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1942 "packssdw %%mm1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1943 "movq %%mm0, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1944 :"=m"(dst[i]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1945 :"m"(src[i]), "m"(src[i+2]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1946 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1947 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1948 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1949 } |
4172 | 1950 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
|
1951 int i; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1952 for(i=0; i<len; i+=4) { |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1953 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1954 "cvtps2pi %1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1955 "cvtps2pi %2, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1956 "packssdw %%mm1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1957 "movq %%mm0, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1958 :"=m"(dst[i]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1959 :"m"(src[i]), "m"(src[i+2]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1960 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1961 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1962 asm volatile("emms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1963 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
1964 |
6195
5f704e9cb518
fix prototypes, remove warning: i386/dsputil_mmx.c:3594: warning: assignment from incompatible pointer type
bcoudurier
parents:
6181
diff
changeset
|
1965 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
|
1966 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
|
1967 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
|
1968 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
|
1969 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
|
1970 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
|
1971 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
|
1972 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
|
1973 |
1092 | 1974 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx) |
0 | 1975 { |
4197 | 1976 mm_flags = mm_support(); |
1115 | 1977 |
1122 | 1978 if (avctx->dsp_mask) { |
2979 | 1979 if (avctx->dsp_mask & FF_MM_FORCE) |
4197 | 1980 mm_flags |= (avctx->dsp_mask & 0xffff); |
2979 | 1981 else |
4197 | 1982 mm_flags &= ~(avctx->dsp_mask & 0xffff); |
1122 | 1983 } |
1115 | 1984 |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
1985 #if 0 |
1868 | 1986 av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:"); |
4197 | 1987 if (mm_flags & MM_MMX) |
1868 | 1988 av_log(avctx, AV_LOG_INFO, " mmx"); |
4197 | 1989 if (mm_flags & MM_MMXEXT) |
1868 | 1990 av_log(avctx, AV_LOG_INFO, " mmxext"); |
4197 | 1991 if (mm_flags & MM_3DNOW) |
1868 | 1992 av_log(avctx, AV_LOG_INFO, " 3dnow"); |
4197 | 1993 if (mm_flags & MM_SSE) |
1868 | 1994 av_log(avctx, AV_LOG_INFO, " sse"); |
4197 | 1995 if (mm_flags & MM_SSE2) |
1868 | 1996 av_log(avctx, AV_LOG_INFO, " sse2"); |
1997 av_log(avctx, AV_LOG_INFO, "\n"); | |
0 | 1998 #endif |
1999 | |
4197 | 2000 if (mm_flags & MM_MMX) { |
1092 | 2001 const int idct_algo= avctx->idct_algo; |
2002 | |
2256 | 2003 if(avctx->lowres==0){ |
2004 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){ | |
2005 c->idct_put= ff_simple_idct_put_mmx; | |
2006 c->idct_add= ff_simple_idct_add_mmx; | |
2007 c->idct = ff_simple_idct_mmx; | |
2008 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
|
2009 #ifdef CONFIG_GPL |
2256 | 2010 }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){ |
4197 | 2011 if(mm_flags & MM_MMXEXT){ |
2256 | 2012 c->idct_put= ff_libmpeg2mmx2_idct_put; |
2013 c->idct_add= ff_libmpeg2mmx2_idct_add; | |
2014 c->idct = ff_mmxext_idct; | |
2015 }else{ | |
2016 c->idct_put= ff_libmpeg2mmx_idct_put; | |
2017 c->idct_add= ff_libmpeg2mmx_idct_add; | |
2018 c->idct = ff_mmx_idct; | |
2019 } | |
2020 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
|
2021 #endif |
5007 | 2022 }else if((ENABLE_VP3_DECODER || ENABLE_VP5_DECODER || ENABLE_VP6_DECODER) && |
2023 idct_algo==FF_IDCT_VP3 && | |
3721
2000e401593d
disable vp3 mmx idct for theora files to avoid artifacts
aurel
parents:
3717
diff
changeset
|
2024 avctx->codec->id!=CODEC_ID_THEORA && |
3712
f7f75f718efb
Enables back the mmx/sse optimized version of the vp3 idct.
aurel
parents:
3666
diff
changeset
|
2025 !(avctx->flags & CODEC_FLAG_BITEXACT)){ |
4197 | 2026 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
|
2027 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
|
2028 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
|
2029 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
|
2030 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
|
2031 }else{ |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2032 ff_vp3_dsp_init_mmx(); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2033 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
|
2034 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
|
2035 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
|
2036 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
|
2037 } |
3524 | 2038 }else if(idct_algo==FF_IDCT_CAVS){ |
2039 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM; | |
2868 | 2040 }else if(idct_algo==FF_IDCT_XVIDMMX){ |
4197 | 2041 if(mm_flags & MM_MMXEXT){ |
2868 | 2042 c->idct_put= ff_idct_xvid_mmx2_put; |
2043 c->idct_add= ff_idct_xvid_mmx2_add; | |
2044 c->idct = ff_idct_xvid_mmx2; | |
2045 }else{ | |
2046 c->idct_put= ff_idct_xvid_mmx_put; | |
2047 c->idct_add= ff_idct_xvid_mmx_add; | |
2048 c->idct = ff_idct_xvid_mmx; | |
2049 } | |
1092 | 2050 } |
2051 } | |
1868 | 2052 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
2053 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
|
2054 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
|
2055 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
|
2056 c->clear_blocks = clear_blocks_mmx; |
415 | 2057 |
6327 | 2058 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU) \ |
2059 c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _ ## CPU; \ | |
2060 c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_ ## CPU; \ | |
2061 c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_ ## CPU; \ | |
2062 c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU | |
2063 | |
2064 SET_HPEL_FUNCS(put, 0, 16, mmx); | |
2065 SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx); | |
2066 SET_HPEL_FUNCS(avg, 0, 16, mmx); | |
2067 SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx); | |
2068 SET_HPEL_FUNCS(put, 1, 8, mmx); | |
2069 SET_HPEL_FUNCS(put_no_rnd, 1, 8, mmx); | |
2070 SET_HPEL_FUNCS(avg, 1, 8, mmx); | |
2071 SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx); | |
2967 | 2072 |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2073 c->gmc= gmc_mmx; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2074 |
866 | 2075 c->add_bytes= add_bytes_mmx; |
6384 | 2076 c->add_bytes_l2= add_bytes_l2_mmx; |
1647 | 2077 |
5277
7b3fcb7c61ce
Avoid linking with h263.c functions when the relevant codecs
aurel
parents:
5255
diff
changeset
|
2078 if (ENABLE_ANY_H263) { |
5278 | 2079 c->h263_v_loop_filter= h263_v_loop_filter_mmx; |
2080 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
|
2081 } |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
2082 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
|
2083 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
|
2084 c->put_no_rnd_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_mmx_nornd; |
2967 | 2085 |
3173 | 2086 c->h264_idct_dc_add= |
2087 c->h264_idct_add= ff_h264_idct_add_mmx; | |
3174 | 2088 c->h264_idct8_dc_add= |
2089 c->h264_idct8_add= ff_h264_idct8_add_mmx; | |
6320 | 2090 if (mm_flags & MM_SSE2) |
2091 c->h264_idct8_add= ff_h264_idct8_add_sse2; | |
3173 | 2092 |
4197 | 2093 if (mm_flags & MM_MMXEXT) { |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2094 c->prefetch = prefetch_mmx2; |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2095 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
2096 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
|
2097 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2; |
651 | 2098 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
2099 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
|
2100 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
|
2101 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2; |
415 | 2102 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
2103 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
|
2104 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2; |
651 | 2105 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
2106 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
|
2107 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
|
2108 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2; |
1092 | 2109 |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3089
diff
changeset
|
2110 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
|
2111 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2; |
2745 | 2112 |
1092 | 2113 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
2114 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2; | |
2115 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2; | |
2116 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2; | |
2117 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2; | |
2118 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2; | |
2119 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2; | |
2120 } | |
959 | 2121 |
6327 | 2122 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU) \ |
2123 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## SIZE ## _mc00_ ## CPU; \ | |
2124 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## SIZE ## _mc10_ ## CPU; \ | |
2125 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## SIZE ## _mc20_ ## CPU; \ | |
2126 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## SIZE ## _mc30_ ## CPU; \ | |
2127 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## SIZE ## _mc01_ ## CPU; \ | |
2128 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## SIZE ## _mc11_ ## CPU; \ | |
2129 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## SIZE ## _mc21_ ## CPU; \ | |
2130 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## SIZE ## _mc31_ ## CPU; \ | |
2131 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## SIZE ## _mc02_ ## CPU; \ | |
2132 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## SIZE ## _mc12_ ## CPU; \ | |
2133 c->PFX ## _pixels_tab[IDX][10] = PFX ## SIZE ## _mc22_ ## CPU; \ | |
2134 c->PFX ## _pixels_tab[IDX][11] = PFX ## SIZE ## _mc32_ ## CPU; \ | |
2135 c->PFX ## _pixels_tab[IDX][12] = PFX ## SIZE ## _mc03_ ## CPU; \ | |
2136 c->PFX ## _pixels_tab[IDX][13] = PFX ## SIZE ## _mc13_ ## CPU; \ | |
2137 c->PFX ## _pixels_tab[IDX][14] = PFX ## SIZE ## _mc23_ ## CPU; \ | |
2138 c->PFX ## _pixels_tab[IDX][15] = PFX ## SIZE ## _mc33_ ## CPU | |
2139 | |
2140 SET_QPEL_FUNCS(put_qpel, 0, 16, mmx2); | |
2141 SET_QPEL_FUNCS(put_qpel, 1, 8, mmx2); | |
2142 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmx2); | |
2143 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, mmx2); | |
2144 SET_QPEL_FUNCS(avg_qpel, 0, 16, mmx2); | |
2145 SET_QPEL_FUNCS(avg_qpel, 1, 8, mmx2); | |
2146 | |
2147 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmx2); | |
2148 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, mmx2); | |
2149 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, mmx2); | |
2150 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmx2); | |
2151 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, mmx2); | |
2152 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, mmx2); | |
2153 | |
2154 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, mmx2); | |
2155 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, mmx2); | |
2156 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, mmx2); | |
2157 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, mmx2); | |
2209 | 2158 |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
2159 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
|
2160 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_mmx2; |
3213 | 2161 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_mmx2; |
2162 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_mmx2; | |
2633 | 2163 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_mmx2; |
2164 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_mmx2; | |
2165 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_mmx2; | |
2166 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
|
2167 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
|
2168 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
|
2169 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2; |
2633 | 2170 |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
2171 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
|
2172 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
|
2173 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
|
2174 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
|
2175 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
|
2176 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
|
2177 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
|
2178 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
|
2179 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
2180 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
|
2181 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
|
2182 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
|
2183 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
|
2184 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
|
2185 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
|
2186 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
|
2187 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
|
2188 |
5949 | 2189 if (ENABLE_CAVS_DECODER) |
5950 | 2190 ff_cavsdsp_init_mmx2(c, avctx); |
5949 | 2191 |
2192 if (ENABLE_VC1_DECODER || ENABLE_WMV3_DECODER) | |
5950 | 2193 ff_vc1dsp_init_mmx(c, avctx); |
5933 | 2194 |
6384 | 2195 c->add_png_paeth_prediction= add_png_paeth_prediction_mmx2; |
4197 | 2196 } else if (mm_flags & MM_3DNOW) { |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2197 c->prefetch = prefetch_3dnow; |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2198 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
2199 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
|
2200 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow; |
393 | 2201 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
2202 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
|
2203 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
|
2204 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow; |
651 | 2205 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
2206 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
|
2207 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
|
2208 |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
2209 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
|
2210 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
|
2211 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow; |
1092 | 2212 |
2213 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
2214 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow; | |
2215 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow; | |
2216 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow; | |
2217 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow; | |
2218 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow; | |
2219 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow; | |
2220 } | |
984 | 2221 |
6327 | 2222 SET_QPEL_FUNCS(put_qpel, 0, 16, 3dnow); |
2223 SET_QPEL_FUNCS(put_qpel, 1, 8, 3dnow); | |
2224 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow); | |
2225 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, 3dnow); | |
2226 SET_QPEL_FUNCS(avg_qpel, 0, 16, 3dnow); | |
2227 SET_QPEL_FUNCS(avg_qpel, 1, 8, 3dnow); | |
2228 | |
2229 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow); | |
2230 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, 3dnow); | |
2231 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, 3dnow); | |
2232 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow); | |
2233 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, 3dnow); | |
2234 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, 3dnow); | |
2235 | |
2236 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow); | |
2237 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, 3dnow); | |
2238 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow); | |
2239 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, 3dnow); | |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2240 |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
2241 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
|
2242 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_3dnow; |
0 | 2243 } |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
2244 |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2245 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2246 #define H264_QPEL_FUNCS(x, y, CPU)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2247 c->put_h264_qpel_pixels_tab[0][x+y*4] = put_h264_qpel16_mc##x##y##_##CPU;\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2248 c->put_h264_qpel_pixels_tab[1][x+y*4] = put_h264_qpel8_mc##x##y##_##CPU;\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2249 c->avg_h264_qpel_pixels_tab[0][x+y*4] = avg_h264_qpel16_mc##x##y##_##CPU;\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2250 c->avg_h264_qpel_pixels_tab[1][x+y*4] = avg_h264_qpel8_mc##x##y##_##CPU; |
6331 | 2251 if((mm_flags & MM_SSE2) && !(mm_flags & MM_3DNOW)){ |
2252 // these functions are slower than mmx on AMD, but faster on Intel | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2253 /* FIXME works in most codecs, but crashes svq1 due to unaligned chroma |
6331 | 2254 c->put_pixels_tab[0][0] = put_pixels16_sse2; |
2255 c->avg_pixels_tab[0][0] = avg_pixels16_sse2; | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2256 */ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2257 H264_QPEL_FUNCS(0, 0, sse2); |
6331 | 2258 } |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2259 if(mm_flags & MM_SSE2){ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2260 H264_QPEL_FUNCS(0, 1, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2261 H264_QPEL_FUNCS(0, 2, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2262 H264_QPEL_FUNCS(0, 3, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2263 H264_QPEL_FUNCS(1, 1, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2264 H264_QPEL_FUNCS(1, 2, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2265 H264_QPEL_FUNCS(1, 3, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2266 H264_QPEL_FUNCS(2, 1, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2267 H264_QPEL_FUNCS(2, 2, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2268 H264_QPEL_FUNCS(2, 3, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2269 H264_QPEL_FUNCS(3, 1, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2270 H264_QPEL_FUNCS(3, 2, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2271 H264_QPEL_FUNCS(3, 3, sse2); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2272 } |
6331 | 2273 #ifdef HAVE_SSSE3 |
2274 if(mm_flags & MM_SSSE3){ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2275 H264_QPEL_FUNCS(1, 0, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2276 H264_QPEL_FUNCS(1, 1, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2277 H264_QPEL_FUNCS(1, 2, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2278 H264_QPEL_FUNCS(1, 3, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2279 H264_QPEL_FUNCS(2, 0, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2280 H264_QPEL_FUNCS(2, 1, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2281 H264_QPEL_FUNCS(2, 2, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2282 H264_QPEL_FUNCS(2, 3, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2283 H264_QPEL_FUNCS(3, 0, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2284 H264_QPEL_FUNCS(3, 1, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2285 H264_QPEL_FUNCS(3, 2, ssse3); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2286 H264_QPEL_FUNCS(3, 3, ssse3); |
6384 | 2287 c->add_png_paeth_prediction= add_png_paeth_prediction_ssse3; |
6331 | 2288 } |
2289 #endif | |
2290 | |
4589
30261f4ed12d
Fix wrong conditional, Snow decoding, not encoding, was SIMD-accelerated.
diego
parents:
4436
diff
changeset
|
2291 #ifdef CONFIG_SNOW_DECODER |
5591 | 2292 if(mm_flags & MM_SSE2 & 0){ |
3210 | 2293 c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2; |
5602 | 2294 #ifdef HAVE_7REGS |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
2295 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
|
2296 #endif |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
2297 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
|
2298 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
2299 else{ |
5594
384629ebcb93
avoid overflow in the 3rd lifting step, this now needs mmx2 at minimum
michael
parents:
5591
diff
changeset
|
2300 if(mm_flags & MM_MMXEXT){ |
3210 | 2301 c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx; |
5602 | 2302 #ifdef HAVE_7REGS |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
2303 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
|
2304 #endif |
5594
384629ebcb93
avoid overflow in the 3rd lifting step, this now needs mmx2 at minimum
michael
parents:
5591
diff
changeset
|
2305 } |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
2306 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
|
2307 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
2308 #endif |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2309 |
4197 | 2310 if(mm_flags & MM_3DNOW){ |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2311 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
|
2312 c->vector_fmul = vector_fmul_3dnow; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2313 if(!(avctx->flags & CODEC_FLAG_BITEXACT)) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2314 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
|
2315 } |
4197 | 2316 if(mm_flags & MM_3DNOWEXT) |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2317 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2; |
4197 | 2318 if(mm_flags & MM_SSE){ |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2319 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
|
2320 c->vector_fmul = vector_fmul_sse; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2321 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
|
2322 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
|
2323 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
|
2324 } |
4197 | 2325 if(mm_flags & MM_3DNOW) |
3574 | 2326 c->vector_fmul_add_add = vector_fmul_add_add_3dnow; // faster than sse |
0 | 2327 } |
2967 | 2328 |
6403
9a736918fd90
split encoding part of dsputil_mmx into its own file
aurel
parents:
6391
diff
changeset
|
2329 if (ENABLE_ENCODERS) |
9a736918fd90
split encoding part of dsputil_mmx into its own file
aurel
parents:
6391
diff
changeset
|
2330 dsputilenc_init_mmx(c, avctx); |
9a736918fd90
split encoding part of dsputil_mmx into its own file
aurel
parents:
6391
diff
changeset
|
2331 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2332 #if 0 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2333 // for speed testing |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2334 get_pixels = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2335 put_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2336 add_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2337 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2338 pix_abs16x16 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2339 pix_abs16x16_x2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2340 pix_abs16x16_y2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2341 pix_abs16x16_xy2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2342 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2343 put_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2344 put_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2345 put_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2346 put_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2347 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2348 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
|
2349 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
|
2350 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
|
2351 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
|
2352 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2353 avg_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2354 avg_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2355 avg_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2356 avg_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2357 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2358 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
|
2359 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
|
2360 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
|
2361 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
|
2362 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2363 //av_fdct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2364 //ff_idct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2365 #endif |
0 | 2366 } |