Mercurial > libavcodec.hg
annotate i386/dsputil_mmx.c @ 6336:ef3fb5a7e275 libavcodec
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
author | lorenm |
---|---|
date | Wed, 06 Feb 2008 12:32:31 +0000 |
parents | 950811a14eb3 |
children | 0a403ade8c81 |
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; |
5946
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
62 |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
63 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_1 ) = 0x0101010101010101ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
64 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_3 ) = 0x0303030303030303ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
65 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_7 ) = 0x0707070707070707ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
66 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_3F ) = 0x3F3F3F3F3F3F3F3FULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
67 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_A1 ) = 0xA1A1A1A1A1A1A1A1ULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
68 DECLARE_ALIGNED_8 (const uint64_t, ff_pb_FC ) = 0xFCFCFCFCFCFCFCFCULL; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
69 |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
70 DECLARE_ALIGNED_16(const double, ff_pd_1[2]) = { 1.0, 1.0 }; |
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5933
diff
changeset
|
71 DECLARE_ALIGNED_16(const double, ff_pd_2[2]) = { 2.0, 2.0 }; |
5737 | 72 |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
73 #define JUMPALIGN() __asm __volatile (ASMALIGN(3)::) |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
74 #define MOVQ_ZERO(regd) __asm __volatile ("pxor %%" #regd ", %%" #regd ::) |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
75 |
448 | 76 #define MOVQ_WONE(regd) \ |
77 __asm __volatile ( \ | |
78 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ | |
79 "psrlw $15, %%" #regd ::) | |
80 | |
81 #define MOVQ_BFE(regd) \ | |
82 __asm __volatile ( \ | |
83 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\ | |
84 "paddb %%" #regd ", %%" #regd " \n\t" ::) | |
85 | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
86 #ifndef PIC |
5947 | 87 #define MOVQ_BONE(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_bone)) |
88 #define MOVQ_WTWO(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(ff_wtwo)) | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
89 #else |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
90 // for shared library it's better to use this way for accessing constants |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
91 // pcmpeqd -> -1 |
448 | 92 #define MOVQ_BONE(regd) \ |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
93 __asm __volatile ( \ |
448 | 94 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
95 "psrlw $15, %%" #regd " \n\t" \ | |
96 "packuswb %%" #regd ", %%" #regd " \n\t" ::) | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
97 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
98 #define MOVQ_WTWO(regd) \ |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
99 __asm __volatile ( \ |
448 | 100 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
101 "psrlw $15, %%" #regd " \n\t" \ | |
102 "psllw $1, %%" #regd " \n\t"::) | |
387 | 103 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
104 #endif |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
105 |
448 | 106 // using regr as temporary and for the output result |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
107 // first argument is unmodifed and second is trashed |
471 | 108 // regfe is supposed to contain 0xfefefefefefefefe |
109 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \ | |
2979 | 110 "movq " #rega ", " #regr " \n\t"\ |
111 "pand " #regb ", " #regr " \n\t"\ | |
112 "pxor " #rega ", " #regb " \n\t"\ | |
113 "pand " #regfe "," #regb " \n\t"\ | |
114 "psrlq $1, " #regb " \n\t"\ | |
115 "paddb " #regb ", " #regr " \n\t" | |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
116 |
471 | 117 #define PAVGB_MMX(rega, regb, regr, regfe) \ |
2979 | 118 "movq " #rega ", " #regr " \n\t"\ |
119 "por " #regb ", " #regr " \n\t"\ | |
120 "pxor " #rega ", " #regb " \n\t"\ | |
121 "pand " #regfe "," #regb " \n\t"\ | |
122 "psrlq $1, " #regb " \n\t"\ | |
123 "psubb " #regb ", " #regr " \n\t" | |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
124 |
471 | 125 // mm6 is supposed to contain 0xfefefefefefefefe |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
126 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \ |
2979 | 127 "movq " #rega ", " #regr " \n\t"\ |
128 "movq " #regc ", " #regp " \n\t"\ | |
129 "pand " #regb ", " #regr " \n\t"\ | |
130 "pand " #regd ", " #regp " \n\t"\ | |
131 "pxor " #rega ", " #regb " \n\t"\ | |
132 "pxor " #regc ", " #regd " \n\t"\ | |
133 "pand %%mm6, " #regb " \n\t"\ | |
134 "pand %%mm6, " #regd " \n\t"\ | |
135 "psrlq $1, " #regb " \n\t"\ | |
136 "psrlq $1, " #regd " \n\t"\ | |
137 "paddb " #regb ", " #regr " \n\t"\ | |
138 "paddb " #regd ", " #regp " \n\t" | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
139 |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
140 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \ |
2979 | 141 "movq " #rega ", " #regr " \n\t"\ |
142 "movq " #regc ", " #regp " \n\t"\ | |
143 "por " #regb ", " #regr " \n\t"\ | |
144 "por " #regd ", " #regp " \n\t"\ | |
145 "pxor " #rega ", " #regb " \n\t"\ | |
146 "pxor " #regc ", " #regd " \n\t"\ | |
147 "pand %%mm6, " #regb " \n\t"\ | |
148 "pand %%mm6, " #regd " \n\t"\ | |
149 "psrlq $1, " #regd " \n\t"\ | |
150 "psrlq $1, " #regb " \n\t"\ | |
151 "psubb " #regb ", " #regr " \n\t"\ | |
152 "psubb " #regd ", " #regp " \n\t" | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
153 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
154 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
155 /* MMX no rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
156 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx |
448 | 157 #define SET_RND MOVQ_WONE |
2979 | 158 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX_NO_RND(a, b, c, d, e, f) |
159 #define PAVGB(a, b, c, e) PAVGB_MMX_NO_RND(a, b, c, e) | |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
160 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
161 #include "dsputil_mmx_rnd.h" |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
162 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
163 #undef DEF |
448 | 164 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
165 #undef PAVGBP |
471 | 166 #undef PAVGB |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
167 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
168 /* MMX rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
169 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
170 #define DEF(x, y) x ## _ ## y ##_mmx |
448 | 171 #define SET_RND MOVQ_WTWO |
2979 | 172 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f) |
173 #define PAVGB(a, b, c, e) PAVGB_MMX(a, b, c, e) | |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
174 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
175 #include "dsputil_mmx_rnd.h" |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
176 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
177 #undef DEF |
448 | 178 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
179 #undef PAVGBP |
471 | 180 #undef PAVGB |
387 | 181 |
0 | 182 /***********************************/ |
183 /* 3Dnow specific */ | |
184 | |
185 #define DEF(x) x ## _3dnow | |
186 #define PAVGB "pavgusb" | |
187 | |
188 #include "dsputil_mmx_avg.h" | |
189 | |
190 #undef DEF | |
191 #undef PAVGB | |
192 | |
193 /***********************************/ | |
194 /* MMX2 specific */ | |
195 | |
386 | 196 #define DEF(x) x ## _mmx2 |
0 | 197 |
198 /* Introduced only in MMX2 set */ | |
199 #define PAVGB "pavgb" | |
200 | |
201 #include "dsputil_mmx_avg.h" | |
202 | |
203 #undef DEF | |
204 #undef PAVGB | |
205 | |
6327 | 206 #define put_no_rnd_pixels16_mmx put_pixels16_mmx |
207 #define put_no_rnd_pixels8_mmx put_pixels8_mmx | |
6321 | 208 #define put_pixels16_mmx2 put_pixels16_mmx |
209 #define put_pixels8_mmx2 put_pixels8_mmx | |
210 #define put_pixels4_mmx2 put_pixels4_mmx | |
211 #define put_no_rnd_pixels16_mmx2 put_no_rnd_pixels16_mmx | |
212 #define put_no_rnd_pixels8_mmx2 put_no_rnd_pixels8_mmx | |
213 #define put_pixels16_3dnow put_pixels16_mmx | |
214 #define put_pixels8_3dnow put_pixels8_mmx | |
215 #define put_pixels4_3dnow put_pixels4_mmx | |
216 #define put_no_rnd_pixels16_3dnow put_no_rnd_pixels16_mmx | |
217 #define put_no_rnd_pixels8_3dnow put_no_rnd_pixels8_mmx | |
218 | |
0 | 219 /***********************************/ |
220 /* standard MMX */ | |
221 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
222 #ifdef CONFIG_ENCODERS |
1064 | 223 static void get_pixels_mmx(DCTELEM *block, const uint8_t *pixels, int line_size) |
0 | 224 { |
386 | 225 asm volatile( |
2979 | 226 "mov $-128, %%"REG_a" \n\t" |
227 "pxor %%mm7, %%mm7 \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
228 ASMALIGN(4) |
2979 | 229 "1: \n\t" |
230 "movq (%0), %%mm0 \n\t" | |
231 "movq (%0, %2), %%mm2 \n\t" | |
232 "movq %%mm0, %%mm1 \n\t" | |
233 "movq %%mm2, %%mm3 \n\t" | |
234 "punpcklbw %%mm7, %%mm0 \n\t" | |
235 "punpckhbw %%mm7, %%mm1 \n\t" | |
236 "punpcklbw %%mm7, %%mm2 \n\t" | |
237 "punpckhbw %%mm7, %%mm3 \n\t" | |
238 "movq %%mm0, (%1, %%"REG_a") \n\t" | |
239 "movq %%mm1, 8(%1, %%"REG_a") \n\t" | |
240 "movq %%mm2, 16(%1, %%"REG_a") \n\t" | |
241 "movq %%mm3, 24(%1, %%"REG_a") \n\t" | |
242 "add %3, %0 \n\t" | |
243 "add $32, %%"REG_a" \n\t" | |
244 "js 1b \n\t" | |
386 | 245 : "+r" (pixels) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
246 : "r" (block+64), "r" ((long)line_size), "r" ((long)line_size*2) |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
247 : "%"REG_a |
386 | 248 ); |
0 | 249 } |
250 | |
1064 | 251 static inline void diff_pixels_mmx(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride) |
324 | 252 { |
253 asm volatile( | |
2979 | 254 "pxor %%mm7, %%mm7 \n\t" |
255 "mov $-128, %%"REG_a" \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
256 ASMALIGN(4) |
2979 | 257 "1: \n\t" |
258 "movq (%0), %%mm0 \n\t" | |
259 "movq (%1), %%mm2 \n\t" | |
260 "movq %%mm0, %%mm1 \n\t" | |
261 "movq %%mm2, %%mm3 \n\t" | |
262 "punpcklbw %%mm7, %%mm0 \n\t" | |
263 "punpckhbw %%mm7, %%mm1 \n\t" | |
264 "punpcklbw %%mm7, %%mm2 \n\t" | |
265 "punpckhbw %%mm7, %%mm3 \n\t" | |
266 "psubw %%mm2, %%mm0 \n\t" | |
267 "psubw %%mm3, %%mm1 \n\t" | |
268 "movq %%mm0, (%2, %%"REG_a") \n\t" | |
269 "movq %%mm1, 8(%2, %%"REG_a") \n\t" | |
270 "add %3, %0 \n\t" | |
271 "add %3, %1 \n\t" | |
272 "add $16, %%"REG_a" \n\t" | |
273 "jnz 1b \n\t" | |
324 | 274 : "+r" (s1), "+r" (s2) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
275 : "r" (block+64), "r" ((long)stride) |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
276 : "%"REG_a |
324 | 277 ); |
278 } | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
279 #endif //CONFIG_ENCODERS |
324 | 280 |
1064 | 281 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 282 { |
283 const DCTELEM *p; | |
1064 | 284 uint8_t *pix; |
0 | 285 |
286 /* read the pixels */ | |
287 p = block; | |
288 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
289 /* unrolled loop */ |
2979 | 290 __asm __volatile( |
291 "movq %3, %%mm0 \n\t" | |
292 "movq 8%3, %%mm1 \n\t" | |
293 "movq 16%3, %%mm2 \n\t" | |
294 "movq 24%3, %%mm3 \n\t" | |
295 "movq 32%3, %%mm4 \n\t" | |
296 "movq 40%3, %%mm5 \n\t" | |
297 "movq 48%3, %%mm6 \n\t" | |
298 "movq 56%3, %%mm7 \n\t" | |
299 "packuswb %%mm1, %%mm0 \n\t" | |
300 "packuswb %%mm3, %%mm2 \n\t" | |
301 "packuswb %%mm5, %%mm4 \n\t" | |
302 "packuswb %%mm7, %%mm6 \n\t" | |
303 "movq %%mm0, (%0) \n\t" | |
304 "movq %%mm2, (%0, %1) \n\t" | |
305 "movq %%mm4, (%0, %1, 2) \n\t" | |
306 "movq %%mm6, (%0, %2) \n\t" | |
307 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "m"(*p) | |
308 :"memory"); | |
0 | 309 pix += line_size*4; |
310 p += 32; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
311 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
312 // 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
|
313 // compiler would generate some very strange code |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
314 // thus using "r" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
315 __asm __volatile( |
2979 | 316 "movq (%3), %%mm0 \n\t" |
317 "movq 8(%3), %%mm1 \n\t" | |
318 "movq 16(%3), %%mm2 \n\t" | |
319 "movq 24(%3), %%mm3 \n\t" | |
320 "movq 32(%3), %%mm4 \n\t" | |
321 "movq 40(%3), %%mm5 \n\t" | |
322 "movq 48(%3), %%mm6 \n\t" | |
323 "movq 56(%3), %%mm7 \n\t" | |
324 "packuswb %%mm1, %%mm0 \n\t" | |
325 "packuswb %%mm3, %%mm2 \n\t" | |
326 "packuswb %%mm5, %%mm4 \n\t" | |
327 "packuswb %%mm7, %%mm6 \n\t" | |
328 "movq %%mm0, (%0) \n\t" | |
329 "movq %%mm2, (%0, %1) \n\t" | |
330 "movq %%mm4, (%0, %1, 2) \n\t" | |
331 "movq %%mm6, (%0, %2) \n\t" | |
332 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "r"(p) | |
333 :"memory"); | |
0 | 334 } |
335 | |
3089 | 336 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
|
337 { 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
|
338 |
1984
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
339 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
|
340 { |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
341 int i; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
342 |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
343 movq_m2r(*vector128, mm1); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
344 for (i = 0; i < 8; i++) { |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
345 movq_m2r(*(block), mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
346 packsswb_m2r(*(block + 4), mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
347 block += 8; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
348 paddb_r2r(mm1, mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
349 movq_r2m(mm0, *pixels); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
350 pixels += line_size; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
351 } |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
352 } |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
353 |
1064 | 354 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 355 { |
356 const DCTELEM *p; | |
1064 | 357 uint8_t *pix; |
0 | 358 int i; |
359 | |
360 /* read the pixels */ | |
361 p = block; | |
362 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
363 MOVQ_ZERO(mm7); |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
364 i = 4; |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
365 do { |
2979 | 366 __asm __volatile( |
367 "movq (%2), %%mm0 \n\t" | |
368 "movq 8(%2), %%mm1 \n\t" | |
369 "movq 16(%2), %%mm2 \n\t" | |
370 "movq 24(%2), %%mm3 \n\t" | |
371 "movq %0, %%mm4 \n\t" | |
372 "movq %1, %%mm6 \n\t" | |
373 "movq %%mm4, %%mm5 \n\t" | |
374 "punpcklbw %%mm7, %%mm4 \n\t" | |
375 "punpckhbw %%mm7, %%mm5 \n\t" | |
376 "paddsw %%mm4, %%mm0 \n\t" | |
377 "paddsw %%mm5, %%mm1 \n\t" | |
378 "movq %%mm6, %%mm5 \n\t" | |
379 "punpcklbw %%mm7, %%mm6 \n\t" | |
380 "punpckhbw %%mm7, %%mm5 \n\t" | |
381 "paddsw %%mm6, %%mm2 \n\t" | |
382 "paddsw %%mm5, %%mm3 \n\t" | |
383 "packuswb %%mm1, %%mm0 \n\t" | |
384 "packuswb %%mm3, %%mm2 \n\t" | |
385 "movq %%mm0, %0 \n\t" | |
386 "movq %%mm2, %1 \n\t" | |
387 :"+m"(*pix), "+m"(*(pix+line_size)) | |
388 :"r"(p) | |
389 :"memory"); | |
0 | 390 pix += line_size*2; |
391 p += 16; | |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
392 } while (--i); |
0 | 393 } |
394 | |
2209 | 395 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
396 { | |
397 __asm __volatile( | |
2979 | 398 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
399 ASMALIGN(3) |
2979 | 400 "1: \n\t" |
401 "movd (%1), %%mm0 \n\t" | |
402 "movd (%1, %3), %%mm1 \n\t" | |
403 "movd %%mm0, (%2) \n\t" | |
404 "movd %%mm1, (%2, %3) \n\t" | |
405 "add %%"REG_a", %1 \n\t" | |
406 "add %%"REG_a", %2 \n\t" | |
407 "movd (%1), %%mm0 \n\t" | |
408 "movd (%1, %3), %%mm1 \n\t" | |
409 "movd %%mm0, (%2) \n\t" | |
410 "movd %%mm1, (%2, %3) \n\t" | |
411 "add %%"REG_a", %1 \n\t" | |
412 "add %%"REG_a", %2 \n\t" | |
413 "subl $4, %0 \n\t" | |
414 "jnz 1b \n\t" | |
415 : "+g"(h), "+r" (pixels), "+r" (block) | |
416 : "r"((long)line_size) | |
417 : "%"REG_a, "memory" | |
418 ); | |
2209 | 419 } |
420 | |
1064 | 421 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 422 { |
471 | 423 __asm __volatile( |
2979 | 424 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
425 ASMALIGN(3) |
2979 | 426 "1: \n\t" |
427 "movq (%1), %%mm0 \n\t" | |
428 "movq (%1, %3), %%mm1 \n\t" | |
429 "movq %%mm0, (%2) \n\t" | |
430 "movq %%mm1, (%2, %3) \n\t" | |
431 "add %%"REG_a", %1 \n\t" | |
432 "add %%"REG_a", %2 \n\t" | |
433 "movq (%1), %%mm0 \n\t" | |
434 "movq (%1, %3), %%mm1 \n\t" | |
435 "movq %%mm0, (%2) \n\t" | |
436 "movq %%mm1, (%2, %3) \n\t" | |
437 "add %%"REG_a", %1 \n\t" | |
438 "add %%"REG_a", %2 \n\t" | |
439 "subl $4, %0 \n\t" | |
440 "jnz 1b \n\t" | |
441 : "+g"(h), "+r" (pixels), "+r" (block) | |
442 : "r"((long)line_size) | |
443 : "%"REG_a, "memory" | |
444 ); | |
0 | 445 } |
446 | |
1064 | 447 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
651 | 448 { |
449 __asm __volatile( | |
2979 | 450 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
451 ASMALIGN(3) |
2979 | 452 "1: \n\t" |
453 "movq (%1), %%mm0 \n\t" | |
454 "movq 8(%1), %%mm4 \n\t" | |
455 "movq (%1, %3), %%mm1 \n\t" | |
456 "movq 8(%1, %3), %%mm5 \n\t" | |
457 "movq %%mm0, (%2) \n\t" | |
458 "movq %%mm4, 8(%2) \n\t" | |
459 "movq %%mm1, (%2, %3) \n\t" | |
460 "movq %%mm5, 8(%2, %3) \n\t" | |
461 "add %%"REG_a", %1 \n\t" | |
462 "add %%"REG_a", %2 \n\t" | |
463 "movq (%1), %%mm0 \n\t" | |
464 "movq 8(%1), %%mm4 \n\t" | |
465 "movq (%1, %3), %%mm1 \n\t" | |
466 "movq 8(%1, %3), %%mm5 \n\t" | |
467 "movq %%mm0, (%2) \n\t" | |
468 "movq %%mm4, 8(%2) \n\t" | |
469 "movq %%mm1, (%2, %3) \n\t" | |
470 "movq %%mm5, 8(%2, %3) \n\t" | |
471 "add %%"REG_a", %1 \n\t" | |
472 "add %%"REG_a", %2 \n\t" | |
473 "subl $4, %0 \n\t" | |
474 "jnz 1b \n\t" | |
475 : "+g"(h), "+r" (pixels), "+r" (block) | |
476 : "r"((long)line_size) | |
477 : "%"REG_a, "memory" | |
478 ); | |
651 | 479 } |
480 | |
6331 | 481 static void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
482 { | |
483 __asm __volatile( | |
484 "1: \n\t" | |
485 "movdqu (%1), %%xmm0 \n\t" | |
486 "movdqu (%1,%3), %%xmm1 \n\t" | |
487 "movdqu (%1,%3,2), %%xmm2 \n\t" | |
488 "movdqu (%1,%4), %%xmm3 \n\t" | |
489 "movdqa %%xmm0, (%2) \n\t" | |
490 "movdqa %%xmm1, (%2,%3) \n\t" | |
491 "movdqa %%xmm2, (%2,%3,2) \n\t" | |
492 "movdqa %%xmm3, (%2,%4) \n\t" | |
493 "subl $4, %0 \n\t" | |
494 "lea (%1,%3,4), %1 \n\t" | |
495 "lea (%2,%3,4), %2 \n\t" | |
496 "jnz 1b \n\t" | |
497 : "+g"(h), "+r" (pixels), "+r" (block) | |
498 : "r"((long)line_size), "r"(3L*line_size) | |
499 : "memory" | |
500 ); | |
501 } | |
502 | |
503 static void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h) | |
504 { | |
505 __asm __volatile( | |
506 "1: \n\t" | |
507 "movdqu (%1), %%xmm0 \n\t" | |
508 "movdqu (%1,%3), %%xmm1 \n\t" | |
509 "movdqu (%1,%3,2), %%xmm2 \n\t" | |
510 "movdqu (%1,%4), %%xmm3 \n\t" | |
511 "pavgb (%2), %%xmm0 \n\t" | |
512 "pavgb (%2,%3), %%xmm1 \n\t" | |
513 "pavgb (%2,%3,2), %%xmm2 \n\t" | |
514 "pavgb (%2,%4), %%xmm3 \n\t" | |
515 "movdqa %%xmm0, (%2) \n\t" | |
516 "movdqa %%xmm1, (%2,%3) \n\t" | |
517 "movdqa %%xmm2, (%2,%3,2) \n\t" | |
518 "movdqa %%xmm3, (%2,%4) \n\t" | |
519 "subl $4, %0 \n\t" | |
520 "lea (%1,%3,4), %1 \n\t" | |
521 "lea (%2,%3,4), %2 \n\t" | |
522 "jnz 1b \n\t" | |
523 : "+g"(h), "+r" (pixels), "+r" (block) | |
524 : "r"((long)line_size), "r"(3L*line_size) | |
525 : "memory" | |
526 ); | |
527 } | |
528 | |
296 | 529 static void clear_blocks_mmx(DCTELEM *blocks) |
530 { | |
471 | 531 __asm __volatile( |
2979 | 532 "pxor %%mm7, %%mm7 \n\t" |
533 "mov $-128*6, %%"REG_a" \n\t" | |
534 "1: \n\t" | |
535 "movq %%mm7, (%0, %%"REG_a") \n\t" | |
536 "movq %%mm7, 8(%0, %%"REG_a") \n\t" | |
537 "movq %%mm7, 16(%0, %%"REG_a") \n\t" | |
538 "movq %%mm7, 24(%0, %%"REG_a") \n\t" | |
539 "add $32, %%"REG_a" \n\t" | |
540 " 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
|
541 : : "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
|
542 : "%"REG_a |
296 | 543 ); |
544 } | |
545 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
546 #ifdef CONFIG_ENCODERS |
1064 | 547 static int pix_sum16_mmx(uint8_t * pix, int line_size){ |
688 | 548 const int h=16; |
549 int sum; | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
550 long index= -line_size*h; |
688 | 551 |
552 __asm __volatile( | |
2979 | 553 "pxor %%mm7, %%mm7 \n\t" |
554 "pxor %%mm6, %%mm6 \n\t" | |
555 "1: \n\t" | |
556 "movq (%2, %1), %%mm0 \n\t" | |
557 "movq (%2, %1), %%mm1 \n\t" | |
558 "movq 8(%2, %1), %%mm2 \n\t" | |
559 "movq 8(%2, %1), %%mm3 \n\t" | |
560 "punpcklbw %%mm7, %%mm0 \n\t" | |
561 "punpckhbw %%mm7, %%mm1 \n\t" | |
562 "punpcklbw %%mm7, %%mm2 \n\t" | |
563 "punpckhbw %%mm7, %%mm3 \n\t" | |
564 "paddw %%mm0, %%mm1 \n\t" | |
565 "paddw %%mm2, %%mm3 \n\t" | |
566 "paddw %%mm1, %%mm3 \n\t" | |
567 "paddw %%mm3, %%mm6 \n\t" | |
568 "add %3, %1 \n\t" | |
569 " js 1b \n\t" | |
570 "movq %%mm6, %%mm5 \n\t" | |
571 "psrlq $32, %%mm6 \n\t" | |
572 "paddw %%mm5, %%mm6 \n\t" | |
573 "movq %%mm6, %%mm5 \n\t" | |
574 "psrlq $16, %%mm6 \n\t" | |
575 "paddw %%mm5, %%mm6 \n\t" | |
576 "movd %%mm6, %0 \n\t" | |
577 "andl $0xFFFF, %0 \n\t" | |
688 | 578 : "=&r" (sum), "+r" (index) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
579 : "r" (pix - index), "r" ((long)line_size) |
688 | 580 ); |
581 | |
582 return sum; | |
583 } | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
584 #endif //CONFIG_ENCODERS |
688 | 585 |
866 | 586 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
|
587 long i=0; |
866 | 588 asm volatile( |
2979 | 589 "1: \n\t" |
590 "movq (%1, %0), %%mm0 \n\t" | |
591 "movq (%2, %0), %%mm1 \n\t" | |
592 "paddb %%mm0, %%mm1 \n\t" | |
593 "movq %%mm1, (%2, %0) \n\t" | |
594 "movq 8(%1, %0), %%mm0 \n\t" | |
595 "movq 8(%2, %0), %%mm1 \n\t" | |
596 "paddb %%mm0, %%mm1 \n\t" | |
597 "movq %%mm1, 8(%2, %0) \n\t" | |
598 "add $16, %0 \n\t" | |
599 "cmp %3, %0 \n\t" | |
600 " jb 1b \n\t" | |
866 | 601 : "+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
|
602 : "r"(src), "r"(dst), "r"((long)w-15) |
866 | 603 ); |
604 for(; i<w; i++) | |
605 dst[i+0] += src[i+0]; | |
606 } | |
607 | |
1648 | 608 #define H263_LOOP_FILTER \ |
2979 | 609 "pxor %%mm7, %%mm7 \n\t"\ |
610 "movq %0, %%mm0 \n\t"\ | |
611 "movq %0, %%mm1 \n\t"\ | |
612 "movq %3, %%mm2 \n\t"\ | |
613 "movq %3, %%mm3 \n\t"\ | |
614 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
615 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
616 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
617 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
618 "psubw %%mm2, %%mm0 \n\t"\ | |
619 "psubw %%mm3, %%mm1 \n\t"\ | |
620 "movq %1, %%mm2 \n\t"\ | |
621 "movq %1, %%mm3 \n\t"\ | |
622 "movq %2, %%mm4 \n\t"\ | |
623 "movq %2, %%mm5 \n\t"\ | |
624 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
625 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
626 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
627 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
628 "psubw %%mm2, %%mm4 \n\t"\ | |
629 "psubw %%mm3, %%mm5 \n\t"\ | |
630 "psllw $2, %%mm4 \n\t"\ | |
631 "psllw $2, %%mm5 \n\t"\ | |
632 "paddw %%mm0, %%mm4 \n\t"\ | |
633 "paddw %%mm1, %%mm5 \n\t"\ | |
634 "pxor %%mm6, %%mm6 \n\t"\ | |
635 "pcmpgtw %%mm4, %%mm6 \n\t"\ | |
636 "pcmpgtw %%mm5, %%mm7 \n\t"\ | |
637 "pxor %%mm6, %%mm4 \n\t"\ | |
638 "pxor %%mm7, %%mm5 \n\t"\ | |
639 "psubw %%mm6, %%mm4 \n\t"\ | |
640 "psubw %%mm7, %%mm5 \n\t"\ | |
641 "psrlw $3, %%mm4 \n\t"\ | |
642 "psrlw $3, %%mm5 \n\t"\ | |
643 "packuswb %%mm5, %%mm4 \n\t"\ | |
644 "packsswb %%mm7, %%mm6 \n\t"\ | |
645 "pxor %%mm7, %%mm7 \n\t"\ | |
646 "movd %4, %%mm2 \n\t"\ | |
647 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
648 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
649 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
650 "psubusb %%mm4, %%mm2 \n\t"\ | |
651 "movq %%mm2, %%mm3 \n\t"\ | |
652 "psubusb %%mm4, %%mm3 \n\t"\ | |
653 "psubb %%mm3, %%mm2 \n\t"\ | |
654 "movq %1, %%mm3 \n\t"\ | |
655 "movq %2, %%mm4 \n\t"\ | |
656 "pxor %%mm6, %%mm3 \n\t"\ | |
657 "pxor %%mm6, %%mm4 \n\t"\ | |
658 "paddusb %%mm2, %%mm3 \n\t"\ | |
659 "psubusb %%mm2, %%mm4 \n\t"\ | |
660 "pxor %%mm6, %%mm3 \n\t"\ | |
661 "pxor %%mm6, %%mm4 \n\t"\ | |
662 "paddusb %%mm2, %%mm2 \n\t"\ | |
663 "packsswb %%mm1, %%mm0 \n\t"\ | |
664 "pcmpgtb %%mm0, %%mm7 \n\t"\ | |
665 "pxor %%mm7, %%mm0 \n\t"\ | |
666 "psubb %%mm7, %%mm0 \n\t"\ | |
667 "movq %%mm0, %%mm1 \n\t"\ | |
668 "psubusb %%mm2, %%mm0 \n\t"\ | |
669 "psubb %%mm0, %%mm1 \n\t"\ | |
670 "pand %5, %%mm1 \n\t"\ | |
671 "psrlw $2, %%mm1 \n\t"\ | |
672 "pxor %%mm7, %%mm1 \n\t"\ | |
673 "psubb %%mm7, %%mm1 \n\t"\ | |
674 "movq %0, %%mm5 \n\t"\ | |
675 "movq %3, %%mm6 \n\t"\ | |
676 "psubb %%mm1, %%mm5 \n\t"\ | |
677 "paddb %%mm1, %%mm6 \n\t" | |
1648 | 678 |
1647 | 679 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
|
680 if(ENABLE_ANY_H263) { |
1647 | 681 const int strength= ff_h263_loop_filter_strength[qscale]; |
682 | |
683 asm volatile( | |
2967 | 684 |
1648 | 685 H263_LOOP_FILTER |
2967 | 686 |
2979 | 687 "movq %%mm3, %1 \n\t" |
688 "movq %%mm4, %2 \n\t" | |
689 "movq %%mm5, %0 \n\t" | |
690 "movq %%mm6, %3 \n\t" | |
1647 | 691 : "+m" (*(uint64_t*)(src - 2*stride)), |
692 "+m" (*(uint64_t*)(src - 1*stride)), | |
693 "+m" (*(uint64_t*)(src + 0*stride)), | |
694 "+m" (*(uint64_t*)(src + 1*stride)) | |
695 : "g" (2*strength), "m"(ff_pb_FC) | |
696 ); | |
5394
e9a6215f4e3a
help some gcc version to optimize out those functions
aurel
parents:
5278
diff
changeset
|
697 } |
1647 | 698 } |
699 | |
1648 | 700 static inline void transpose4x4(uint8_t *dst, uint8_t *src, int dst_stride, int src_stride){ |
701 asm volatile( //FIXME could save 1 instruction if done as 8x4 ... | |
2979 | 702 "movd %4, %%mm0 \n\t" |
703 "movd %5, %%mm1 \n\t" | |
704 "movd %6, %%mm2 \n\t" | |
705 "movd %7, %%mm3 \n\t" | |
706 "punpcklbw %%mm1, %%mm0 \n\t" | |
707 "punpcklbw %%mm3, %%mm2 \n\t" | |
708 "movq %%mm0, %%mm1 \n\t" | |
709 "punpcklwd %%mm2, %%mm0 \n\t" | |
710 "punpckhwd %%mm2, %%mm1 \n\t" | |
711 "movd %%mm0, %0 \n\t" | |
712 "punpckhdq %%mm0, %%mm0 \n\t" | |
713 "movd %%mm0, %1 \n\t" | |
714 "movd %%mm1, %2 \n\t" | |
715 "punpckhdq %%mm1, %%mm1 \n\t" | |
716 "movd %%mm1, %3 \n\t" | |
2967 | 717 |
1648 | 718 : "=m" (*(uint32_t*)(dst + 0*dst_stride)), |
719 "=m" (*(uint32_t*)(dst + 1*dst_stride)), | |
720 "=m" (*(uint32_t*)(dst + 2*dst_stride)), | |
721 "=m" (*(uint32_t*)(dst + 3*dst_stride)) | |
722 : "m" (*(uint32_t*)(src + 0*src_stride)), | |
723 "m" (*(uint32_t*)(src + 1*src_stride)), | |
724 "m" (*(uint32_t*)(src + 2*src_stride)), | |
725 "m" (*(uint32_t*)(src + 3*src_stride)) | |
726 ); | |
727 } | |
728 | |
729 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
|
730 if(ENABLE_ANY_H263) { |
1648 | 731 const int strength= ff_h263_loop_filter_strength[qscale]; |
6181 | 732 DECLARE_ALIGNED(8, uint64_t, temp[4]); |
1648 | 733 uint8_t *btemp= (uint8_t*)temp; |
2967 | 734 |
1648 | 735 src -= 2; |
736 | |
737 transpose4x4(btemp , src , 8, stride); | |
738 transpose4x4(btemp+4, src + 4*stride, 8, stride); | |
739 asm volatile( | |
740 H263_LOOP_FILTER // 5 3 4 6 | |
2967 | 741 |
1648 | 742 : "+m" (temp[0]), |
743 "+m" (temp[1]), | |
744 "+m" (temp[2]), | |
745 "+m" (temp[3]) | |
746 : "g" (2*strength), "m"(ff_pb_FC) | |
747 ); | |
748 | |
749 asm volatile( | |
2979 | 750 "movq %%mm5, %%mm1 \n\t" |
751 "movq %%mm4, %%mm0 \n\t" | |
752 "punpcklbw %%mm3, %%mm5 \n\t" | |
753 "punpcklbw %%mm6, %%mm4 \n\t" | |
754 "punpckhbw %%mm3, %%mm1 \n\t" | |
755 "punpckhbw %%mm6, %%mm0 \n\t" | |
756 "movq %%mm5, %%mm3 \n\t" | |
757 "movq %%mm1, %%mm6 \n\t" | |
758 "punpcklwd %%mm4, %%mm5 \n\t" | |
759 "punpcklwd %%mm0, %%mm1 \n\t" | |
760 "punpckhwd %%mm4, %%mm3 \n\t" | |
761 "punpckhwd %%mm0, %%mm6 \n\t" | |
762 "movd %%mm5, (%0) \n\t" | |
763 "punpckhdq %%mm5, %%mm5 \n\t" | |
764 "movd %%mm5, (%0,%2) \n\t" | |
765 "movd %%mm3, (%0,%2,2) \n\t" | |
766 "punpckhdq %%mm3, %%mm3 \n\t" | |
767 "movd %%mm3, (%0,%3) \n\t" | |
768 "movd %%mm1, (%1) \n\t" | |
769 "punpckhdq %%mm1, %%mm1 \n\t" | |
770 "movd %%mm1, (%1,%2) \n\t" | |
771 "movd %%mm6, (%1,%2,2) \n\t" | |
772 "punpckhdq %%mm6, %%mm6 \n\t" | |
773 "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
|
774 :: "r" (src), |
86e2b1424801
optimization and gcc 4.0 bug workaround patch by (Martin Drab >drab kepler.fjfi.cvut cz<)
michael
parents:
2293
diff
changeset
|
775 "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
|
776 "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
|
777 "r" ((long)(3*stride)) |
1648 | 778 ); |
5394
e9a6215f4e3a
help some gcc version to optimize out those functions
aurel
parents:
5278
diff
changeset
|
779 } |
1648 | 780 } |
781 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
782 #ifdef CONFIG_ENCODERS |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
783 static int pix_norm1_mmx(uint8_t *pix, int line_size) { |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
784 int tmp; |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
785 asm volatile ( |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
786 "movl $16,%%ecx\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
787 "pxor %%mm0,%%mm0\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
788 "pxor %%mm7,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
789 "1:\n" |
2979 | 790 "movq (%0),%%mm2\n" /* mm2 = pix[0-7] */ |
791 "movq 8(%0),%%mm3\n" /* mm3 = pix[8-15] */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
792 |
2979 | 793 "movq %%mm2,%%mm1\n" /* mm1 = mm2 = pix[0-7] */ |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
794 |
2979 | 795 "punpckhbw %%mm0,%%mm1\n" /* mm1 = [pix4-7] */ |
796 "punpcklbw %%mm0,%%mm2\n" /* mm2 = [pix0-3] */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
797 |
2979 | 798 "movq %%mm3,%%mm4\n" /* mm4 = mm3 = pix[8-15] */ |
799 "punpckhbw %%mm0,%%mm3\n" /* mm3 = [pix12-15] */ | |
800 "punpcklbw %%mm0,%%mm4\n" /* mm4 = [pix8-11] */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
801 |
2979 | 802 "pmaddwd %%mm1,%%mm1\n" /* mm1 = (pix0^2+pix1^2,pix2^2+pix3^2) */ |
803 "pmaddwd %%mm2,%%mm2\n" /* mm2 = (pix4^2+pix5^2,pix6^2+pix7^2) */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
804 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
805 "pmaddwd %%mm3,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
806 "pmaddwd %%mm4,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
807 |
2979 | 808 "paddd %%mm1,%%mm2\n" /* mm2 = (pix0^2+pix1^2+pix4^2+pix5^2, |
809 pix2^2+pix3^2+pix6^2+pix7^2) */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
810 "paddd %%mm3,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
811 "paddd %%mm2,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
812 |
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
|
813 "add %2, %0\n" |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
814 "paddd %%mm4,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
815 "dec %%ecx\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
816 "jnz 1b\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
817 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
818 "movq %%mm7,%%mm1\n" |
2979 | 819 "psrlq $32, %%mm7\n" /* shift hi dword to lo */ |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
820 "paddd %%mm7,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
821 "movd %%mm1,%1\n" |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
822 : "+r" (pix), "=r"(tmp) : "r" ((long)line_size) : "%ecx" ); |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
823 return tmp; |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
824 } |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
825 |
2067 | 826 static int sse8_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
827 int tmp; | |
828 asm volatile ( | |
829 "movl %4,%%ecx\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
830 "shr $1,%%ecx\n" |
2979 | 831 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */ |
832 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */ | |
2067 | 833 "1:\n" |
2979 | 834 "movq (%0),%%mm1\n" /* mm1 = pix1[0][0-7] */ |
835 "movq (%1),%%mm2\n" /* mm2 = pix2[0][0-7] */ | |
836 "movq (%0,%3),%%mm3\n" /* mm3 = pix1[1][0-7] */ | |
837 "movq (%1,%3),%%mm4\n" /* mm4 = pix2[1][0-7] */ | |
2067 | 838 |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
839 /* todo: mm1-mm2, mm3-mm4 */ |
5963 | 840 /* algo: subtract mm1 from mm2 with saturation and vice versa */ |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
841 /* OR the results to get absolute difference */ |
2067 | 842 "movq %%mm1,%%mm5\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
843 "movq %%mm3,%%mm6\n" |
2067 | 844 "psubusb %%mm2,%%mm1\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
845 "psubusb %%mm4,%%mm3\n" |
2067 | 846 "psubusb %%mm5,%%mm2\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
847 "psubusb %%mm6,%%mm4\n" |
2067 | 848 |
849 "por %%mm1,%%mm2\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
850 "por %%mm3,%%mm4\n" |
2067 | 851 |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
852 /* now convert to 16-bit vectors so we can square them */ |
2067 | 853 "movq %%mm2,%%mm1\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
854 "movq %%mm4,%%mm3\n" |
2067 | 855 |
856 "punpckhbw %%mm0,%%mm2\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
857 "punpckhbw %%mm0,%%mm4\n" |
2979 | 858 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */ |
859 "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */ | |
2067 | 860 |
861 "pmaddwd %%mm2,%%mm2\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
862 "pmaddwd %%mm4,%%mm4\n" |
2067 | 863 "pmaddwd %%mm1,%%mm1\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
864 "pmaddwd %%mm3,%%mm3\n" |
2067 | 865 |
2979 | 866 "lea (%0,%3,2), %0\n" /* pix1 += 2*line_size */ |
867 "lea (%1,%3,2), %1\n" /* pix2 += 2*line_size */ | |
2067 | 868 |
869 "paddd %%mm2,%%mm1\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
870 "paddd %%mm4,%%mm3\n" |
2067 | 871 "paddd %%mm1,%%mm7\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
872 "paddd %%mm3,%%mm7\n" |
2067 | 873 |
874 "decl %%ecx\n" | |
875 "jnz 1b\n" | |
876 | |
877 "movq %%mm7,%%mm1\n" | |
2979 | 878 "psrlq $32, %%mm7\n" /* shift hi dword to lo */ |
2067 | 879 "paddd %%mm7,%%mm1\n" |
880 "movd %%mm1,%2\n" | |
2967 | 881 : "+r" (pix1), "+r" (pix2), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
882 : "r" ((long)line_size) , "m" (h) |
2067 | 883 : "%ecx"); |
884 return tmp; | |
885 } | |
886 | |
1708 | 887 static int sse16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
888 int tmp; |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
889 asm volatile ( |
1708 | 890 "movl %4,%%ecx\n" |
2979 | 891 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */ |
892 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
893 "1:\n" |
2979 | 894 "movq (%0),%%mm1\n" /* mm1 = pix1[0-7] */ |
895 "movq (%1),%%mm2\n" /* mm2 = pix2[0-7] */ | |
896 "movq 8(%0),%%mm3\n" /* mm3 = pix1[8-15] */ | |
897 "movq 8(%1),%%mm4\n" /* mm4 = pix2[8-15] */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
898 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
899 /* todo: mm1-mm2, mm3-mm4 */ |
5963 | 900 /* algo: subtract mm1 from mm2 with saturation and vice versa */ |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
901 /* OR the results to get absolute difference */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
902 "movq %%mm1,%%mm5\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
903 "movq %%mm3,%%mm6\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
904 "psubusb %%mm2,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
905 "psubusb %%mm4,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
906 "psubusb %%mm5,%%mm2\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
907 "psubusb %%mm6,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
908 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
909 "por %%mm1,%%mm2\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
910 "por %%mm3,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
911 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
912 /* now convert to 16-bit vectors so we can square them */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
913 "movq %%mm2,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
914 "movq %%mm4,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
915 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
916 "punpckhbw %%mm0,%%mm2\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
917 "punpckhbw %%mm0,%%mm4\n" |
2979 | 918 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */ |
919 "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */ | |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
920 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
921 "pmaddwd %%mm2,%%mm2\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
922 "pmaddwd %%mm4,%%mm4\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
923 "pmaddwd %%mm1,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
924 "pmaddwd %%mm3,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
925 |
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
|
926 "add %3,%0\n" |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
927 "add %3,%1\n" |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
928 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
929 "paddd %%mm2,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
930 "paddd %%mm4,%%mm3\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
931 "paddd %%mm1,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
932 "paddd %%mm3,%%mm7\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
933 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
934 "decl %%ecx\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
935 "jnz 1b\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
936 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
937 "movq %%mm7,%%mm1\n" |
2979 | 938 "psrlq $32, %%mm7\n" /* shift hi dword to lo */ |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
939 "paddd %%mm7,%%mm1\n" |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
940 "movd %%mm1,%2\n" |
2967 | 941 : "+r" (pix1), "+r" (pix2), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
942 : "r" ((long)line_size) , "m" (h) |
1708 | 943 : "%ecx"); |
997
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
944 return tmp; |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
945 } |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
946 |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
947 static int sse16_sse2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
948 int tmp; |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
949 asm volatile ( |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
950 "shr $1,%2\n" |
2979 | 951 "pxor %%xmm0,%%xmm0\n" /* mm0 = 0 */ |
952 "pxor %%xmm7,%%xmm7\n" /* mm7 holds the sum */ | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
953 "1:\n" |
2979 | 954 "movdqu (%0),%%xmm1\n" /* mm1 = pix1[0][0-15] */ |
955 "movdqu (%1),%%xmm2\n" /* mm2 = pix2[0][0-15] */ | |
956 "movdqu (%0,%4),%%xmm3\n" /* mm3 = pix1[1][0-15] */ | |
957 "movdqu (%1,%4),%%xmm4\n" /* mm4 = pix2[1][0-15] */ | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
958 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
959 /* todo: mm1-mm2, mm3-mm4 */ |
5963 | 960 /* algo: subtract mm1 from mm2 with saturation and vice versa */ |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
961 /* OR the results to get absolute difference */ |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
962 "movdqa %%xmm1,%%xmm5\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
963 "movdqa %%xmm3,%%xmm6\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
964 "psubusb %%xmm2,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
965 "psubusb %%xmm4,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
966 "psubusb %%xmm5,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
967 "psubusb %%xmm6,%%xmm4\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
968 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
969 "por %%xmm1,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
970 "por %%xmm3,%%xmm4\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
971 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
972 /* now convert to 16-bit vectors so we can square them */ |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
973 "movdqa %%xmm2,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
974 "movdqa %%xmm4,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
975 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
976 "punpckhbw %%xmm0,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
977 "punpckhbw %%xmm0,%%xmm4\n" |
2979 | 978 "punpcklbw %%xmm0,%%xmm1\n" /* mm1 now spread over (mm1,mm2) */ |
979 "punpcklbw %%xmm0,%%xmm3\n" /* mm4 now spread over (mm3,mm4) */ | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
980 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
981 "pmaddwd %%xmm2,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
982 "pmaddwd %%xmm4,%%xmm4\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
983 "pmaddwd %%xmm1,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
984 "pmaddwd %%xmm3,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
985 |
2979 | 986 "lea (%0,%4,2), %0\n" /* pix1 += 2*line_size */ |
987 "lea (%1,%4,2), %1\n" /* pix2 += 2*line_size */ | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
988 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
989 "paddd %%xmm2,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
990 "paddd %%xmm4,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
991 "paddd %%xmm1,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
992 "paddd %%xmm3,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
993 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
994 "decl %2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
995 "jnz 1b\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
996 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
997 "movdqa %%xmm7,%%xmm1\n" |
2979 | 998 "psrldq $8, %%xmm7\n" /* shift hi qword to lo */ |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
999 "paddd %%xmm1,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
1000 "movdqa %%xmm7,%%xmm1\n" |
2979 | 1001 "psrldq $4, %%xmm7\n" /* shift hi dword to lo */ |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
1002 "paddd %%xmm1,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
1003 "movd %%xmm7,%3\n" |
2967 | 1004 : "+r" (pix1), "+r" (pix2), "+r"(h), "=r"(tmp) |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
1005 : "r" ((long)line_size)); |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
1006 return tmp; |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
1007 } |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
1008 |
2067 | 1009 static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) { |
1010 int tmp; | |
1011 asm volatile ( | |
1012 "movl %3,%%ecx\n" | |
1013 "pxor %%mm7,%%mm7\n" | |
1014 "pxor %%mm6,%%mm6\n" | |
2967 | 1015 |
2067 | 1016 "movq (%0),%%mm0\n" |
1017 "movq %%mm0, %%mm1\n" | |
1018 "psllq $8, %%mm0\n" | |
1019 "psrlq $8, %%mm1\n" | |
1020 "psrlq $8, %%mm0\n" | |
1021 "movq %%mm0, %%mm2\n" | |
1022 "movq %%mm1, %%mm3\n" | |
1023 "punpcklbw %%mm7,%%mm0\n" | |
1024 "punpcklbw %%mm7,%%mm1\n" | |
1025 "punpckhbw %%mm7,%%mm2\n" | |
1026 "punpckhbw %%mm7,%%mm3\n" | |
1027 "psubw %%mm1, %%mm0\n" | |
1028 "psubw %%mm3, %%mm2\n" | |
2967 | 1029 |
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
|
1030 "add %2,%0\n" |
2967 | 1031 |
2067 | 1032 "movq (%0),%%mm4\n" |
1033 "movq %%mm4, %%mm1\n" | |
1034 "psllq $8, %%mm4\n" | |
1035 "psrlq $8, %%mm1\n" | |
1036 "psrlq $8, %%mm4\n" | |
1037 "movq %%mm4, %%mm5\n" | |
1038 "movq %%mm1, %%mm3\n" | |
1039 "punpcklbw %%mm7,%%mm4\n" | |
1040 "punpcklbw %%mm7,%%mm1\n" | |
1041 "punpckhbw %%mm7,%%mm5\n" | |
1042 "punpckhbw %%mm7,%%mm3\n" | |
1043 "psubw %%mm1, %%mm4\n" | |
1044 "psubw %%mm3, %%mm5\n" | |
1045 "psubw %%mm4, %%mm0\n" | |
1046 "psubw %%mm5, %%mm2\n" | |
1047 "pxor %%mm3, %%mm3\n" | |
1048 "pxor %%mm1, %%mm1\n" | |
1049 "pcmpgtw %%mm0, %%mm3\n\t" | |
1050 "pcmpgtw %%mm2, %%mm1\n\t" | |
1051 "pxor %%mm3, %%mm0\n" | |
1052 "pxor %%mm1, %%mm2\n" | |
2967 | 1053 "psubw %%mm3, %%mm0\n" |
2067 | 1054 "psubw %%mm1, %%mm2\n" |
1055 "paddw %%mm0, %%mm2\n" | |
1056 "paddw %%mm2, %%mm6\n" | |
1057 | |
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
|
1058 "add %2,%0\n" |
2067 | 1059 "1:\n" |
2967 | 1060 |
2067 | 1061 "movq (%0),%%mm0\n" |
1062 "movq %%mm0, %%mm1\n" | |
1063 "psllq $8, %%mm0\n" | |
1064 "psrlq $8, %%mm1\n" | |
1065 "psrlq $8, %%mm0\n" | |
1066 "movq %%mm0, %%mm2\n" | |
1067 "movq %%mm1, %%mm3\n" | |
1068 "punpcklbw %%mm7,%%mm0\n" | |
1069 "punpcklbw %%mm7,%%mm1\n" | |
1070 "punpckhbw %%mm7,%%mm2\n" | |
1071 "punpckhbw %%mm7,%%mm3\n" | |
1072 "psubw %%mm1, %%mm0\n" | |
1073 "psubw %%mm3, %%mm2\n" | |
1074 "psubw %%mm0, %%mm4\n" | |
1075 "psubw %%mm2, %%mm5\n" | |
1076 "pxor %%mm3, %%mm3\n" | |
1077 "pxor %%mm1, %%mm1\n" | |
1078 "pcmpgtw %%mm4, %%mm3\n\t" | |
1079 "pcmpgtw %%mm5, %%mm1\n\t" | |
1080 "pxor %%mm3, %%mm4\n" | |
1081 "pxor %%mm1, %%mm5\n" | |
2967 | 1082 "psubw %%mm3, %%mm4\n" |
2067 | 1083 "psubw %%mm1, %%mm5\n" |
1084 "paddw %%mm4, %%mm5\n" | |
1085 "paddw %%mm5, %%mm6\n" | |
2967 | 1086 |
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 "add %2,%0\n" |
2967 | 1088 |
2067 | 1089 "movq (%0),%%mm4\n" |
1090 "movq %%mm4, %%mm1\n" | |
1091 "psllq $8, %%mm4\n" | |
1092 "psrlq $8, %%mm1\n" | |
1093 "psrlq $8, %%mm4\n" | |
1094 "movq %%mm4, %%mm5\n" | |
1095 "movq %%mm1, %%mm3\n" | |
1096 "punpcklbw %%mm7,%%mm4\n" | |
1097 "punpcklbw %%mm7,%%mm1\n" | |
1098 "punpckhbw %%mm7,%%mm5\n" | |
1099 "punpckhbw %%mm7,%%mm3\n" | |
1100 "psubw %%mm1, %%mm4\n" | |
1101 "psubw %%mm3, %%mm5\n" | |
1102 "psubw %%mm4, %%mm0\n" | |
1103 "psubw %%mm5, %%mm2\n" | |
1104 "pxor %%mm3, %%mm3\n" | |
1105 "pxor %%mm1, %%mm1\n" | |
1106 "pcmpgtw %%mm0, %%mm3\n\t" | |
1107 "pcmpgtw %%mm2, %%mm1\n\t" | |
1108 "pxor %%mm3, %%mm0\n" | |
1109 "pxor %%mm1, %%mm2\n" | |
2967 | 1110 "psubw %%mm3, %%mm0\n" |
2067 | 1111 "psubw %%mm1, %%mm2\n" |
1112 "paddw %%mm0, %%mm2\n" | |
1113 "paddw %%mm2, %%mm6\n" | |
1114 | |
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
|
1115 "add %2,%0\n" |
2067 | 1116 "subl $2, %%ecx\n" |
1117 " jnz 1b\n" | |
1118 | |
1119 "movq %%mm6, %%mm0\n" | |
1120 "punpcklwd %%mm7,%%mm0\n" | |
1121 "punpckhwd %%mm7,%%mm6\n" | |
1122 "paddd %%mm0, %%mm6\n" | |
2967 | 1123 |
2067 | 1124 "movq %%mm6,%%mm0\n" |
1125 "psrlq $32, %%mm6\n" | |
1126 "paddd %%mm6,%%mm0\n" | |
1127 "movd %%mm0,%1\n" | |
2967 | 1128 : "+r" (pix1), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1129 : "r" ((long)line_size) , "g" (h-2) |
2067 | 1130 : "%ecx"); |
1131 return tmp; | |
1132 } | |
1133 | |
1134 static int hf_noise16_mmx(uint8_t * pix1, int line_size, int h) { | |
1135 int tmp; | |
1136 uint8_t * pix= pix1; | |
1137 asm volatile ( | |
1138 "movl %3,%%ecx\n" | |
1139 "pxor %%mm7,%%mm7\n" | |
1140 "pxor %%mm6,%%mm6\n" | |
2967 | 1141 |
2067 | 1142 "movq (%0),%%mm0\n" |
1143 "movq 1(%0),%%mm1\n" | |
1144 "movq %%mm0, %%mm2\n" | |
1145 "movq %%mm1, %%mm3\n" | |
1146 "punpcklbw %%mm7,%%mm0\n" | |
1147 "punpcklbw %%mm7,%%mm1\n" | |
1148 "punpckhbw %%mm7,%%mm2\n" | |
1149 "punpckhbw %%mm7,%%mm3\n" | |
1150 "psubw %%mm1, %%mm0\n" | |
1151 "psubw %%mm3, %%mm2\n" | |
2967 | 1152 |
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
|
1153 "add %2,%0\n" |
2967 | 1154 |
2067 | 1155 "movq (%0),%%mm4\n" |
1156 "movq 1(%0),%%mm1\n" | |
1157 "movq %%mm4, %%mm5\n" | |
1158 "movq %%mm1, %%mm3\n" | |
1159 "punpcklbw %%mm7,%%mm4\n" | |
1160 "punpcklbw %%mm7,%%mm1\n" | |
1161 "punpckhbw %%mm7,%%mm5\n" | |
1162 "punpckhbw %%mm7,%%mm3\n" | |
1163 "psubw %%mm1, %%mm4\n" | |
1164 "psubw %%mm3, %%mm5\n" | |
1165 "psubw %%mm4, %%mm0\n" | |
1166 "psubw %%mm5, %%mm2\n" | |
1167 "pxor %%mm3, %%mm3\n" | |
1168 "pxor %%mm1, %%mm1\n" | |
1169 "pcmpgtw %%mm0, %%mm3\n\t" | |
1170 "pcmpgtw %%mm2, %%mm1\n\t" | |
1171 "pxor %%mm3, %%mm0\n" | |
1172 "pxor %%mm1, %%mm2\n" | |
2967 | 1173 "psubw %%mm3, %%mm0\n" |
2067 | 1174 "psubw %%mm1, %%mm2\n" |
1175 "paddw %%mm0, %%mm2\n" | |
1176 "paddw %%mm2, %%mm6\n" | |
1177 | |
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
|
1178 "add %2,%0\n" |
2067 | 1179 "1:\n" |
2967 | 1180 |
2067 | 1181 "movq (%0),%%mm0\n" |
1182 "movq 1(%0),%%mm1\n" | |
1183 "movq %%mm0, %%mm2\n" | |
1184 "movq %%mm1, %%mm3\n" | |
1185 "punpcklbw %%mm7,%%mm0\n" | |
1186 "punpcklbw %%mm7,%%mm1\n" | |
1187 "punpckhbw %%mm7,%%mm2\n" | |
1188 "punpckhbw %%mm7,%%mm3\n" | |
1189 "psubw %%mm1, %%mm0\n" | |
1190 "psubw %%mm3, %%mm2\n" | |
1191 "psubw %%mm0, %%mm4\n" | |
1192 "psubw %%mm2, %%mm5\n" | |
1193 "pxor %%mm3, %%mm3\n" | |
1194 "pxor %%mm1, %%mm1\n" | |
1195 "pcmpgtw %%mm4, %%mm3\n\t" | |
1196 "pcmpgtw %%mm5, %%mm1\n\t" | |
1197 "pxor %%mm3, %%mm4\n" | |
1198 "pxor %%mm1, %%mm5\n" | |
1199 "psubw %%mm3, %%mm4\n" | |
1200 "psubw %%mm1, %%mm5\n" | |
1201 "paddw %%mm4, %%mm5\n" | |
1202 "paddw %%mm5, %%mm6\n" | |
2967 | 1203 |
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
|
1204 "add %2,%0\n" |
2967 | 1205 |
2067 | 1206 "movq (%0),%%mm4\n" |
1207 "movq 1(%0),%%mm1\n" | |
1208 "movq %%mm4, %%mm5\n" | |
1209 "movq %%mm1, %%mm3\n" | |
1210 "punpcklbw %%mm7,%%mm4\n" | |
1211 "punpcklbw %%mm7,%%mm1\n" | |
1212 "punpckhbw %%mm7,%%mm5\n" | |
1213 "punpckhbw %%mm7,%%mm3\n" | |
1214 "psubw %%mm1, %%mm4\n" | |
1215 "psubw %%mm3, %%mm5\n" | |
1216 "psubw %%mm4, %%mm0\n" | |
1217 "psubw %%mm5, %%mm2\n" | |
1218 "pxor %%mm3, %%mm3\n" | |
1219 "pxor %%mm1, %%mm1\n" | |
1220 "pcmpgtw %%mm0, %%mm3\n\t" | |
1221 "pcmpgtw %%mm2, %%mm1\n\t" | |
1222 "pxor %%mm3, %%mm0\n" | |
1223 "pxor %%mm1, %%mm2\n" | |
2967 | 1224 "psubw %%mm3, %%mm0\n" |
2067 | 1225 "psubw %%mm1, %%mm2\n" |
1226 "paddw %%mm0, %%mm2\n" | |
1227 "paddw %%mm2, %%mm6\n" | |
1228 | |
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
|
1229 "add %2,%0\n" |
2067 | 1230 "subl $2, %%ecx\n" |
1231 " jnz 1b\n" | |
1232 | |
1233 "movq %%mm6, %%mm0\n" | |
1234 "punpcklwd %%mm7,%%mm0\n" | |
1235 "punpckhwd %%mm7,%%mm6\n" | |
1236 "paddd %%mm0, %%mm6\n" | |
2967 | 1237 |
2067 | 1238 "movq %%mm6,%%mm0\n" |
1239 "psrlq $32, %%mm6\n" | |
1240 "paddd %%mm6,%%mm0\n" | |
1241 "movd %%mm0,%1\n" | |
2967 | 1242 : "+r" (pix1), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1243 : "r" ((long)line_size) , "g" (h-2) |
2067 | 1244 : "%ecx"); |
1245 return tmp + hf_noise8_mmx(pix+8, line_size, h); | |
1246 } | |
1247 | |
2864
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1248 static int nsse16_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1249 MpegEncContext *c = p; |
2940 | 1250 int score1, score2; |
1251 | |
1252 if(c) score1 = c->dsp.sse[0](c, pix1, pix2, line_size, h); | |
1253 else score1 = sse16_mmx(c, pix1, pix2, line_size, h); | |
1254 score2= hf_noise16_mmx(pix1, line_size, h) - hf_noise16_mmx(pix2, line_size, h); | |
2067 | 1255 |
4001 | 1256 if(c) return score1 + FFABS(score2)*c->avctx->nsse_weight; |
1257 else return score1 + FFABS(score2)*8; | |
2067 | 1258 } |
1259 | |
2864
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1260 static int nsse8_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1261 MpegEncContext *c = p; |
2067 | 1262 int score1= sse8_mmx(c, pix1, pix2, line_size, h); |
1263 int score2= hf_noise8_mmx(pix1, line_size, h) - hf_noise8_mmx(pix2, line_size, h); | |
1264 | |
4001 | 1265 if(c) return score1 + FFABS(score2)*c->avctx->nsse_weight; |
1266 else return score1 + FFABS(score2)*8; | |
2067 | 1267 } |
1268 | |
1729 | 1269 static int vsad_intra16_mmx(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) { |
1270 int tmp; | |
2967 | 1271 |
1729 | 1272 assert( (((int)pix) & 7) == 0); |
1273 assert((line_size &7) ==0); | |
2967 | 1274 |
1729 | 1275 #define SUM(in0, in1, out0, out1) \ |
1276 "movq (%0), %%mm2\n"\ | |
1277 "movq 8(%0), %%mm3\n"\ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1278 "add %2,%0\n"\ |
1729 | 1279 "movq %%mm2, " #out0 "\n"\ |
1280 "movq %%mm3, " #out1 "\n"\ | |
1281 "psubusb " #in0 ", %%mm2\n"\ | |
1282 "psubusb " #in1 ", %%mm3\n"\ | |
1283 "psubusb " #out0 ", " #in0 "\n"\ | |
1284 "psubusb " #out1 ", " #in1 "\n"\ | |
1285 "por %%mm2, " #in0 "\n"\ | |
1286 "por %%mm3, " #in1 "\n"\ | |
1287 "movq " #in0 ", %%mm2\n"\ | |
1288 "movq " #in1 ", %%mm3\n"\ | |
1289 "punpcklbw %%mm7, " #in0 "\n"\ | |
1290 "punpcklbw %%mm7, " #in1 "\n"\ | |
1291 "punpckhbw %%mm7, %%mm2\n"\ | |
1292 "punpckhbw %%mm7, %%mm3\n"\ | |
1293 "paddw " #in1 ", " #in0 "\n"\ | |
1294 "paddw %%mm3, %%mm2\n"\ | |
1295 "paddw %%mm2, " #in0 "\n"\ | |
1296 "paddw " #in0 ", %%mm6\n" | |
1297 | |
2967 | 1298 |
1729 | 1299 asm volatile ( |
1300 "movl %3,%%ecx\n" | |
1301 "pxor %%mm6,%%mm6\n" | |
1302 "pxor %%mm7,%%mm7\n" | |
1303 "movq (%0),%%mm0\n" | |
1304 "movq 8(%0),%%mm1\n" | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1305 "add %2,%0\n" |
1729 | 1306 "subl $2, %%ecx\n" |
1307 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1308 "1:\n" | |
2967 | 1309 |
1729 | 1310 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1311 |
1729 | 1312 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1313 |
1729 | 1314 "subl $2, %%ecx\n" |
1315 "jnz 1b\n" | |
1316 | |
1317 "movq %%mm6,%%mm0\n" | |
1318 "psrlq $32, %%mm6\n" | |
1319 "paddw %%mm6,%%mm0\n" | |
1320 "movq %%mm0,%%mm6\n" | |
1321 "psrlq $16, %%mm0\n" | |
1322 "paddw %%mm6,%%mm0\n" | |
1323 "movd %%mm0,%1\n" | |
2967 | 1324 : "+r" (pix), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1325 : "r" ((long)line_size) , "m" (h) |
1729 | 1326 : "%ecx"); |
1327 return tmp & 0xFFFF; | |
1328 } | |
1329 #undef SUM | |
1330 | |
1331 static int vsad_intra16_mmx2(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) { | |
1332 int tmp; | |
2967 | 1333 |
1729 | 1334 assert( (((int)pix) & 7) == 0); |
1335 assert((line_size &7) ==0); | |
2967 | 1336 |
1729 | 1337 #define SUM(in0, in1, out0, out1) \ |
1338 "movq (%0), " #out0 "\n"\ | |
1339 "movq 8(%0), " #out1 "\n"\ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1340 "add %2,%0\n"\ |
1729 | 1341 "psadbw " #out0 ", " #in0 "\n"\ |
1342 "psadbw " #out1 ", " #in1 "\n"\ | |
1343 "paddw " #in1 ", " #in0 "\n"\ | |
1344 "paddw " #in0 ", %%mm6\n" | |
1345 | |
1346 asm volatile ( | |
1347 "movl %3,%%ecx\n" | |
1348 "pxor %%mm6,%%mm6\n" | |
1349 "pxor %%mm7,%%mm7\n" | |
1350 "movq (%0),%%mm0\n" | |
1351 "movq 8(%0),%%mm1\n" | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1352 "add %2,%0\n" |
1729 | 1353 "subl $2, %%ecx\n" |
1354 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1355 "1:\n" | |
2967 | 1356 |
1729 | 1357 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1358 |
1729 | 1359 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1360 |
1729 | 1361 "subl $2, %%ecx\n" |
1362 "jnz 1b\n" | |
1363 | |
1364 "movd %%mm6,%1\n" | |
2967 | 1365 : "+r" (pix), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1366 : "r" ((long)line_size) , "m" (h) |
1729 | 1367 : "%ecx"); |
1368 return tmp; | |
1369 } | |
1370 #undef SUM | |
1371 | |
1372 static int vsad16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { | |
1373 int tmp; | |
2967 | 1374 |
1729 | 1375 assert( (((int)pix1) & 7) == 0); |
1376 assert( (((int)pix2) & 7) == 0); | |
1377 assert((line_size &7) ==0); | |
2967 | 1378 |
1729 | 1379 #define SUM(in0, in1, out0, out1) \ |
1380 "movq (%0),%%mm2\n"\ | |
1381 "movq (%1)," #out0 "\n"\ | |
1382 "movq 8(%0),%%mm3\n"\ | |
1383 "movq 8(%1)," #out1 "\n"\ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1384 "add %3,%0\n"\ |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1385 "add %3,%1\n"\ |
1729 | 1386 "psubb " #out0 ", %%mm2\n"\ |
1387 "psubb " #out1 ", %%mm3\n"\ | |
1388 "pxor %%mm7, %%mm2\n"\ | |
1389 "pxor %%mm7, %%mm3\n"\ | |
1390 "movq %%mm2, " #out0 "\n"\ | |
1391 "movq %%mm3, " #out1 "\n"\ | |
1392 "psubusb " #in0 ", %%mm2\n"\ | |
1393 "psubusb " #in1 ", %%mm3\n"\ | |
1394 "psubusb " #out0 ", " #in0 "\n"\ | |
1395 "psubusb " #out1 ", " #in1 "\n"\ | |
1396 "por %%mm2, " #in0 "\n"\ | |
1397 "por %%mm3, " #in1 "\n"\ | |
1398 "movq " #in0 ", %%mm2\n"\ | |
1399 "movq " #in1 ", %%mm3\n"\ | |
1400 "punpcklbw %%mm7, " #in0 "\n"\ | |
1401 "punpcklbw %%mm7, " #in1 "\n"\ | |
1402 "punpckhbw %%mm7, %%mm2\n"\ | |
1403 "punpckhbw %%mm7, %%mm3\n"\ | |
1404 "paddw " #in1 ", " #in0 "\n"\ | |
1405 "paddw %%mm3, %%mm2\n"\ | |
1406 "paddw %%mm2, " #in0 "\n"\ | |
1407 "paddw " #in0 ", %%mm6\n" | |
1408 | |
2967 | 1409 |
1729 | 1410 asm volatile ( |
1411 "movl %4,%%ecx\n" | |
1412 "pxor %%mm6,%%mm6\n" | |
1413 "pcmpeqw %%mm7,%%mm7\n" | |
1414 "psllw $15, %%mm7\n" | |
1415 "packsswb %%mm7, %%mm7\n" | |
1416 "movq (%0),%%mm0\n" | |
1417 "movq (%1),%%mm2\n" | |
1418 "movq 8(%0),%%mm1\n" | |
1419 "movq 8(%1),%%mm3\n" | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1420 "add %3,%0\n" |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1421 "add %3,%1\n" |
1729 | 1422 "subl $2, %%ecx\n" |
1423 "psubb %%mm2, %%mm0\n" | |
1424 "psubb %%mm3, %%mm1\n" | |
1425 "pxor %%mm7, %%mm0\n" | |
1426 "pxor %%mm7, %%mm1\n" | |
1427 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1428 "1:\n" | |
2967 | 1429 |
1729 | 1430 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1431 |
1729 | 1432 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1433 |
1729 | 1434 "subl $2, %%ecx\n" |
1435 "jnz 1b\n" | |
1436 | |
1437 "movq %%mm6,%%mm0\n" | |
1438 "psrlq $32, %%mm6\n" | |
1439 "paddw %%mm6,%%mm0\n" | |
1440 "movq %%mm0,%%mm6\n" | |
1441 "psrlq $16, %%mm0\n" | |
1442 "paddw %%mm6,%%mm0\n" | |
1443 "movd %%mm0,%2\n" | |
2967 | 1444 : "+r" (pix1), "+r" (pix2), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1445 : "r" ((long)line_size) , "m" (h) |
1729 | 1446 : "%ecx"); |
1447 return tmp & 0x7FFF; | |
1448 } | |
1449 #undef SUM | |
1450 | |
1451 static int vsad16_mmx2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { | |
1452 int tmp; | |
2967 | 1453 |
1729 | 1454 assert( (((int)pix1) & 7) == 0); |
1455 assert( (((int)pix2) & 7) == 0); | |
1456 assert((line_size &7) ==0); | |
2967 | 1457 |
1729 | 1458 #define SUM(in0, in1, out0, out1) \ |
1459 "movq (%0)," #out0 "\n"\ | |
1460 "movq (%1),%%mm2\n"\ | |
1461 "movq 8(%0)," #out1 "\n"\ | |
1462 "movq 8(%1),%%mm3\n"\ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1463 "add %3,%0\n"\ |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1464 "add %3,%1\n"\ |
1729 | 1465 "psubb %%mm2, " #out0 "\n"\ |
1466 "psubb %%mm3, " #out1 "\n"\ | |
1467 "pxor %%mm7, " #out0 "\n"\ | |
1468 "pxor %%mm7, " #out1 "\n"\ | |
1469 "psadbw " #out0 ", " #in0 "\n"\ | |
1470 "psadbw " #out1 ", " #in1 "\n"\ | |
1471 "paddw " #in1 ", " #in0 "\n"\ | |
1472 "paddw " #in0 ", %%mm6\n" | |
1473 | |
1474 asm volatile ( | |
1475 "movl %4,%%ecx\n" | |
1476 "pxor %%mm6,%%mm6\n" | |
1477 "pcmpeqw %%mm7,%%mm7\n" | |
1478 "psllw $15, %%mm7\n" | |
1479 "packsswb %%mm7, %%mm7\n" | |
1480 "movq (%0),%%mm0\n" | |
1481 "movq (%1),%%mm2\n" | |
1482 "movq 8(%0),%%mm1\n" | |
1483 "movq 8(%1),%%mm3\n" | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1484 "add %3,%0\n" |
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1485 "add %3,%1\n" |
1729 | 1486 "subl $2, %%ecx\n" |
1487 "psubb %%mm2, %%mm0\n" | |
1488 "psubb %%mm3, %%mm1\n" | |
1489 "pxor %%mm7, %%mm0\n" | |
1490 "pxor %%mm7, %%mm1\n" | |
1491 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1492 "1:\n" | |
2967 | 1493 |
1729 | 1494 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1495 |
1729 | 1496 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1497 |
1729 | 1498 "subl $2, %%ecx\n" |
1499 "jnz 1b\n" | |
1500 | |
1501 "movd %%mm6,%2\n" | |
2967 | 1502 : "+r" (pix1), "+r" (pix2), "=r"(tmp) |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1503 : "r" ((long)line_size) , "m" (h) |
1729 | 1504 : "%ecx"); |
1505 return tmp; | |
1506 } | |
1507 #undef SUM | |
1508 | |
866 | 1509 static void diff_bytes_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1510 long i=0; |
866 | 1511 asm volatile( |
2979 | 1512 "1: \n\t" |
1513 "movq (%2, %0), %%mm0 \n\t" | |
1514 "movq (%1, %0), %%mm1 \n\t" | |
1515 "psubb %%mm0, %%mm1 \n\t" | |
1516 "movq %%mm1, (%3, %0) \n\t" | |
1517 "movq 8(%2, %0), %%mm0 \n\t" | |
1518 "movq 8(%1, %0), %%mm1 \n\t" | |
1519 "psubb %%mm0, %%mm1 \n\t" | |
1520 "movq %%mm1, 8(%3, %0) \n\t" | |
1521 "add $16, %0 \n\t" | |
1522 "cmp %4, %0 \n\t" | |
1523 " jb 1b \n\t" | |
866 | 1524 : "+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
|
1525 : "r"(src1), "r"(src2), "r"(dst), "r"((long)w-15) |
866 | 1526 ); |
1527 for(; i<w; i++) | |
1528 dst[i+0] = src1[i+0]-src2[i+0]; | |
1529 } | |
1527 | 1530 |
1531 static void sub_hfyu_median_prediction_mmx2(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top){ | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1532 long i=0; |
1527 | 1533 uint8_t l, lt; |
2967 | 1534 |
1527 | 1535 asm volatile( |
2979 | 1536 "1: \n\t" |
1537 "movq -1(%1, %0), %%mm0 \n\t" // LT | |
1538 "movq (%1, %0), %%mm1 \n\t" // T | |
1539 "movq -1(%2, %0), %%mm2 \n\t" // L | |
1540 "movq (%2, %0), %%mm3 \n\t" // X | |
1541 "movq %%mm2, %%mm4 \n\t" // L | |
1542 "psubb %%mm0, %%mm2 \n\t" | |
1543 "paddb %%mm1, %%mm2 \n\t" // L + T - LT | |
1544 "movq %%mm4, %%mm5 \n\t" // L | |
1545 "pmaxub %%mm1, %%mm4 \n\t" // max(T, L) | |
1546 "pminub %%mm5, %%mm1 \n\t" // min(T, L) | |
1547 "pminub %%mm2, %%mm4 \n\t" | |
1548 "pmaxub %%mm1, %%mm4 \n\t" | |
1549 "psubb %%mm4, %%mm3 \n\t" // dst - pred | |
1550 "movq %%mm3, (%3, %0) \n\t" | |
1551 "add $8, %0 \n\t" | |
1552 "cmp %4, %0 \n\t" | |
1553 " jb 1b \n\t" | |
1527 | 1554 : "+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
|
1555 : "r"(src1), "r"(src2), "r"(dst), "r"((long)w) |
1527 | 1556 ); |
1557 | |
1558 l= *left; | |
1559 lt= *left_top; | |
2967 | 1560 |
1527 | 1561 dst[0]= src2[0] - mid_pred(l, src1[0], (l + src1[0] - lt)&0xFF); |
2967 | 1562 |
1527 | 1563 *left_top= src1[w-1]; |
1564 *left = src2[w-1]; | |
1565 } | |
1566 | |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1567 #define DIFF_PIXELS_1(m,a,t,p1,p2)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1568 "mov"#m" "#p1", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1569 "mov"#m" "#p2", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1570 "punpcklbw "#a", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1571 "punpcklbw "#a", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1572 "psubw "#t", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1573 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1574 #define DIFF_PIXELS_8(m0,m1,mm,p1,p2,stride,temp) {\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1575 uint8_t *p1b=p1, *p2b=p2;\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1576 asm volatile(\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1577 DIFF_PIXELS_1(m0, mm##0, mm##7, (%1), (%2))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1578 DIFF_PIXELS_1(m0, mm##1, mm##7, (%1,%3), (%2,%3))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1579 DIFF_PIXELS_1(m0, mm##2, mm##7, (%1,%3,2), (%2,%3,2))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1580 "add %4, %1 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1581 "add %4, %2 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1582 DIFF_PIXELS_1(m0, mm##3, mm##7, (%1), (%2))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1583 DIFF_PIXELS_1(m0, mm##4, mm##7, (%1,%3), (%2,%3))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1584 DIFF_PIXELS_1(m0, mm##5, mm##7, (%1,%3,2), (%2,%3,2))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1585 DIFF_PIXELS_1(m0, mm##6, mm##7, (%1,%4), (%2,%4))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1586 "mov"#m1" "#mm"0, %0 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1587 DIFF_PIXELS_1(m0, mm##7, mm##0, (%1,%3,4), (%2,%3,4))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1588 "mov"#m1" %0, "#mm"0 \n\t"\ |
5912
f75ee7ea171b
tring to workaround gcc 2.95 bug which causes random failures
michael
parents:
5737
diff
changeset
|
1589 : "+m"(temp), "+r"(p1b), "+r"(p2b)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1590 : "r"((long)stride), "r"((long)stride*3)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1591 );\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1592 } |
5912
f75ee7ea171b
tring to workaround gcc 2.95 bug which causes random failures
michael
parents:
5737
diff
changeset
|
1593 //the "+m"(temp) is needed as gcc 2.95 sometimes fails to compile "=m"(temp) |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1594 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1595 #define DIFF_PIXELS_4x8(p1,p2,stride,temp) DIFF_PIXELS_8(d, q, %%mm, p1, p2, stride, temp) |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1596 #define DIFF_PIXELS_8x8(p1,p2,stride,temp) DIFF_PIXELS_8(q, dqa, %%xmm, p1, p2, stride, temp) |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1597 |
1153 | 1598 #define LBUTTERFLY2(a1,b1,a2,b2)\ |
2979 | 1599 "paddw " #b1 ", " #a1 " \n\t"\ |
1600 "paddw " #b2 ", " #a2 " \n\t"\ | |
1601 "paddw " #b1 ", " #b1 " \n\t"\ | |
1602 "paddw " #b2 ", " #b2 " \n\t"\ | |
1603 "psubw " #a1 ", " #b1 " \n\t"\ | |
1604 "psubw " #a2 ", " #b2 " \n\t" | |
866 | 1605 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1606 #define HADAMARD8(m0, m1, m2, m3, m4, m5, m6, m7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1607 LBUTTERFLY2(m0, m1, m2, m3)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1608 LBUTTERFLY2(m4, m5, m6, m7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1609 LBUTTERFLY2(m0, m2, m1, m3)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1610 LBUTTERFLY2(m4, m6, m5, m7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1611 LBUTTERFLY2(m0, m4, m1, m5)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1612 LBUTTERFLY2(m2, m6, m3, m7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1613 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1614 #define HADAMARD48 HADAMARD8(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm6, %%mm7) |
936 | 1615 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1616 #define MMABS_MMX(a,z)\ |
2979 | 1617 "pxor " #z ", " #z " \n\t"\ |
1618 "pcmpgtw " #a ", " #z " \n\t"\ | |
1619 "pxor " #z ", " #a " \n\t"\ | |
1620 "psubw " #z ", " #a " \n\t" | |
936 | 1621 |
1153 | 1622 #define MMABS_MMX2(a,z)\ |
2979 | 1623 "pxor " #z ", " #z " \n\t"\ |
1624 "psubw " #a ", " #z " \n\t"\ | |
1625 "pmaxsw " #z ", " #a " \n\t" | |
1153 | 1626 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1627 #define MMABS_SSSE3(a,z)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1628 "pabsw " #a ", " #a " \n\t" |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1629 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1630 #define MMABS_SUM(a,z, sum)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1631 MMABS(a,z)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1632 "paddusw " #a ", " #sum " \n\t" |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1633 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1634 #define MMABS_SUM_8x8_NOSPILL\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1635 MMABS(%%xmm0, %%xmm8)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1636 MMABS(%%xmm1, %%xmm9)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1637 MMABS_SUM(%%xmm2, %%xmm8, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1638 MMABS_SUM(%%xmm3, %%xmm9, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1639 MMABS_SUM(%%xmm4, %%xmm8, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1640 MMABS_SUM(%%xmm5, %%xmm9, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1641 MMABS_SUM(%%xmm6, %%xmm8, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1642 MMABS_SUM(%%xmm7, %%xmm9, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1643 "paddusw %%xmm1, %%xmm0 \n\t" |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1644 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1645 #ifdef ARCH_X86_64 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1646 #define MMABS_SUM_8x8_SSE2 MMABS_SUM_8x8_NOSPILL |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1647 #else |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1648 #define MMABS_SUM_8x8_SSE2\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1649 "movdqa %%xmm7, (%1) \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1650 MMABS(%%xmm0, %%xmm7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1651 MMABS(%%xmm1, %%xmm7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1652 MMABS_SUM(%%xmm2, %%xmm7, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1653 MMABS_SUM(%%xmm3, %%xmm7, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1654 MMABS_SUM(%%xmm4, %%xmm7, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1655 MMABS_SUM(%%xmm5, %%xmm7, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1656 MMABS_SUM(%%xmm6, %%xmm7, %%xmm0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1657 "movdqa (%1), %%xmm2 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1658 MMABS_SUM(%%xmm2, %%xmm7, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1659 "paddusw %%xmm1, %%xmm0 \n\t" |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1660 #endif |
2967 | 1661 |
936 | 1662 #define LOAD4(o, a, b, c, d)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1663 "movq "#o"(%1), "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1664 "movq "#o"+8(%1), "#b" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1665 "movq "#o"+16(%1), "#c" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1666 "movq "#o"+24(%1), "#d" \n\t"\ |
936 | 1667 |
1668 #define STORE4(o, a, b, c, d)\ | |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1669 "movq "#a", "#o"(%1) \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1670 "movq "#b", "#o"+8(%1) \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1671 "movq "#c", "#o"+16(%1) \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1672 "movq "#d", "#o"+24(%1) \n\t"\ |
936 | 1673 |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1674 /* FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get up to |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1675 * about 100k on extreme inputs. But that's very unlikely to occur in natural video, |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1676 * and it's even more unlikely to not have any alternative mvs/modes with lower cost. */ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1677 #define HSUM_MMX(a, t, dst)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1678 "movq "#a", "#t" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1679 "psrlq $32, "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1680 "paddusw "#t", "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1681 "movq "#a", "#t" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1682 "psrlq $16, "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1683 "paddusw "#t", "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1684 "movd "#a", "#dst" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1685 |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1686 #define HSUM_MMX2(a, t, dst)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1687 "pshufw $0x0E, "#a", "#t" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1688 "paddusw "#t", "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1689 "pshufw $0x01, "#a", "#t" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1690 "paddusw "#t", "#a" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1691 "movd "#a", "#dst" \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1692 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1693 #define HSUM_SSE2(a, t, dst)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1694 "movhlps "#a", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1695 "paddusw "#t", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1696 "pshuflw $0x0E, "#a", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1697 "paddusw "#t", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1698 "pshuflw $0x01, "#a", "#t" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1699 "paddusw "#t", "#a" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1700 "movd "#a", "#dst" \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1701 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1702 #define HADAMARD8_DIFF_MMX(cpu) \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1703 static int hadamard8_diff_##cpu(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1704 DECLARE_ALIGNED_8(uint64_t, temp[13]);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1705 int sum;\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1706 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1707 assert(h==8);\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1708 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1709 DIFF_PIXELS_4x8(src1, src2, stride, temp[0]);\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1710 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1711 asm volatile(\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1712 HADAMARD48\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1713 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1714 "movq %%mm7, 96(%1) \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1715 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1716 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1717 STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1718 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1719 "movq 96(%1), %%mm7 \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1720 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1721 STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1722 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1723 : "=r" (sum)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1724 : "r"(temp)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1725 );\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1726 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1727 DIFF_PIXELS_4x8(src1+4, src2+4, stride, temp[4]);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1728 \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1729 asm volatile(\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1730 HADAMARD48\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1731 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1732 "movq %%mm7, 96(%1) \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1733 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1734 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1735 STORE4(32, %%mm0, %%mm3, %%mm7, %%mm2)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1736 \ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1737 "movq 96(%1), %%mm7 \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1738 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1739 "movq %%mm7, %%mm5 \n\t"/*FIXME remove*/\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1740 "movq %%mm6, %%mm7 \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1741 "movq %%mm0, %%mm6 \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1742 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1743 LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1744 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1745 HADAMARD48\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1746 "movq %%mm7, 64(%1) \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1747 MMABS(%%mm0, %%mm7)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1748 MMABS(%%mm1, %%mm7)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1749 MMABS_SUM(%%mm2, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1750 MMABS_SUM(%%mm3, %%mm7, %%mm1)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1751 MMABS_SUM(%%mm4, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1752 MMABS_SUM(%%mm5, %%mm7, %%mm1)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1753 MMABS_SUM(%%mm6, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1754 "movq 64(%1), %%mm2 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1755 MMABS_SUM(%%mm2, %%mm7, %%mm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1756 "paddusw %%mm1, %%mm0 \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1757 "movq %%mm0, 64(%1) \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1758 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1759 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1760 LOAD4(32, %%mm4, %%mm5, %%mm6, %%mm7)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1761 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1762 HADAMARD48\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1763 "movq %%mm7, (%1) \n\t"\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1764 MMABS(%%mm0, %%mm7)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1765 MMABS(%%mm1, %%mm7)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1766 MMABS_SUM(%%mm2, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1767 MMABS_SUM(%%mm3, %%mm7, %%mm1)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1768 MMABS_SUM(%%mm4, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1769 MMABS_SUM(%%mm5, %%mm7, %%mm1)\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1770 MMABS_SUM(%%mm6, %%mm7, %%mm0)\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1771 "movq (%1), %%mm2 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1772 MMABS_SUM(%%mm2, %%mm7, %%mm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1773 "paddusw 64(%1), %%mm0 \n\t"\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1774 "paddusw %%mm1, %%mm0 \n\t"\ |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1775 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1776 HSUM(%%mm0, %%mm1, %0)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1777 \ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1778 : "=r" (sum)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1779 : "r"(temp)\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1780 );\ |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1781 return sum&0xFFFF;\ |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1782 }\ |
6056
558c1fd0ee72
Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents:
6030
diff
changeset
|
1783 WRAPPER8_16_SQ(hadamard8_diff_##cpu, hadamard8_diff16_##cpu) |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1784 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1785 #define HADAMARD8_DIFF_SSE2(cpu) \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1786 static int hadamard8_diff_##cpu(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1787 DECLARE_ALIGNED_16(uint64_t, temp[4]);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1788 int sum;\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1789 \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1790 assert(h==8);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1791 \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1792 DIFF_PIXELS_8x8(src1, src2, stride, temp[0]);\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1793 \ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1794 asm volatile(\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1795 HADAMARD8(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm6, %%xmm7)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1796 TRANSPOSE8(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm6, %%xmm7, (%1))\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1797 HADAMARD8(%%xmm0, %%xmm5, %%xmm7, %%xmm3, %%xmm6, %%xmm4, %%xmm2, %%xmm1)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1798 MMABS_SUM_8x8\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1799 HSUM_SSE2(%%xmm0, %%xmm1, %0)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1800 : "=r" (sum)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1801 : "r"(temp)\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1802 );\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1803 return sum&0xFFFF;\ |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1804 }\ |
6056
558c1fd0ee72
Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents:
6030
diff
changeset
|
1805 WRAPPER8_16_SQ(hadamard8_diff_##cpu, hadamard8_diff16_##cpu) |
936 | 1806 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1807 #define MMABS(a,z) MMABS_MMX(a,z) |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1808 #define HSUM(a,t,dst) HSUM_MMX(a,t,dst) |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1809 HADAMARD8_DIFF_MMX(mmx) |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1810 #undef MMABS |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1811 #undef HSUM |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1812 |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1813 #define MMABS(a,z) MMABS_MMX2(a,z) |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1814 #define MMABS_SUM_8x8 MMABS_SUM_8x8_SSE2 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1815 #define HSUM(a,t,dst) HSUM_MMX2(a,t,dst) |
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1816 HADAMARD8_DIFF_MMX(mmx2) |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1817 HADAMARD8_DIFF_SSE2(sse2) |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1818 #undef MMABS |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1819 #undef MMABS_SUM_8x8 |
4946
c1fb4544bd59
cosmetics: remove code duplication in hadamard8_diff_mmx
lorenm
parents:
4939
diff
changeset
|
1820 #undef HSUM |
1153 | 1821 |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1822 #ifdef HAVE_SSSE3 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1823 #define MMABS(a,z) MMABS_SSSE3(a,z) |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1824 #define MMABS_SUM_8x8 MMABS_SUM_8x8_NOSPILL |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1825 HADAMARD8_DIFF_SSE2(ssse3) |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1826 #undef MMABS |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1827 #undef MMABS_SUM_8x8 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
1828 #endif |
4749 | 1829 |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1830 #define DCT_SAD4(m,mm,o)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1831 "mov"#m" "#o"+ 0(%1), "#mm"2 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1832 "mov"#m" "#o"+16(%1), "#mm"3 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1833 "mov"#m" "#o"+32(%1), "#mm"4 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1834 "mov"#m" "#o"+48(%1), "#mm"5 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1835 MMABS_SUM(mm##2, mm##6, mm##0)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1836 MMABS_SUM(mm##3, mm##7, mm##1)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1837 MMABS_SUM(mm##4, mm##6, mm##0)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1838 MMABS_SUM(mm##5, mm##7, mm##1)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1839 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1840 #define DCT_SAD_MMX\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1841 "pxor %%mm0, %%mm0 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1842 "pxor %%mm1, %%mm1 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1843 DCT_SAD4(q, %%mm, 0)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1844 DCT_SAD4(q, %%mm, 8)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1845 DCT_SAD4(q, %%mm, 64)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1846 DCT_SAD4(q, %%mm, 72)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1847 "paddusw %%mm1, %%mm0 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1848 HSUM(%%mm0, %%mm1, %0) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1849 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1850 #define DCT_SAD_SSE2\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1851 "pxor %%xmm0, %%xmm0 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1852 "pxor %%xmm1, %%xmm1 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1853 DCT_SAD4(dqa, %%xmm, 0)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1854 DCT_SAD4(dqa, %%xmm, 64)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1855 "paddusw %%xmm1, %%xmm0 \n\t"\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1856 HSUM(%%xmm0, %%xmm1, %0) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1857 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1858 #define DCT_SAD_FUNC(cpu) \ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1859 static int sum_abs_dctelem_##cpu(DCTELEM *block){\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1860 int sum;\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1861 asm volatile(\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1862 DCT_SAD\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1863 :"=r"(sum)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1864 :"r"(block)\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1865 );\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1866 return sum&0xFFFF;\ |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1867 } |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1868 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1869 #define DCT_SAD DCT_SAD_MMX |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1870 #define HSUM(a,t,dst) HSUM_MMX(a,t,dst) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1871 #define MMABS(a,z) MMABS_MMX(a,z) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1872 DCT_SAD_FUNC(mmx) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1873 #undef MMABS |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1874 #undef HSUM |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1875 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1876 #define HSUM(a,t,dst) HSUM_MMX2(a,t,dst) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1877 #define MMABS(a,z) MMABS_MMX2(a,z) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1878 DCT_SAD_FUNC(mmx2) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1879 #undef HSUM |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1880 #undef DCT_SAD |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1881 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1882 #define DCT_SAD DCT_SAD_SSE2 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1883 #define HSUM(a,t,dst) HSUM_SSE2(a,t,dst) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1884 DCT_SAD_FUNC(sse2) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1885 #undef MMABS |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1886 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1887 #ifdef HAVE_SSSE3 |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1888 #define MMABS(a,z) MMABS_SSSE3(a,z) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1889 DCT_SAD_FUNC(ssse3) |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1890 #undef MMABS |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1891 #endif |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1892 #undef HSUM |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1893 #undef DCT_SAD |
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
1894 |
5255 | 1895 static int ssd_int8_vs_int16_mmx(const int8_t *pix1, const int16_t *pix2, int size){ |
4749 | 1896 int sum; |
1897 long i=size; | |
1898 asm volatile( | |
1899 "pxor %%mm4, %%mm4 \n" | |
1900 "1: \n" | |
1901 "sub $8, %0 \n" | |
1902 "movq (%2,%0), %%mm2 \n" | |
1903 "movq (%3,%0,2), %%mm0 \n" | |
1904 "movq 8(%3,%0,2), %%mm1 \n" | |
1905 "punpckhbw %%mm2, %%mm3 \n" | |
1906 "punpcklbw %%mm2, %%mm2 \n" | |
1907 "psraw $8, %%mm3 \n" | |
1908 "psraw $8, %%mm2 \n" | |
1909 "psubw %%mm3, %%mm1 \n" | |
1910 "psubw %%mm2, %%mm0 \n" | |
1911 "pmaddwd %%mm1, %%mm1 \n" | |
1912 "pmaddwd %%mm0, %%mm0 \n" | |
1913 "paddd %%mm1, %%mm4 \n" | |
1914 "paddd %%mm0, %%mm4 \n" | |
1915 "jg 1b \n" | |
1916 "movq %%mm4, %%mm3 \n" | |
1917 "psrlq $32, %%mm3 \n" | |
1918 "paddd %%mm3, %%mm4 \n" | |
1919 "movd %%mm4, %1 \n" | |
1920 :"+r"(i), "=r"(sum) | |
1921 :"r"(pix1), "r"(pix2) | |
1922 ); | |
1923 return sum; | |
1924 } | |
1925 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
1926 #endif //CONFIG_ENCODERS |
866 | 1927 |
954 | 1928 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\ |
2979 | 1929 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\ |
1930 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\ | |
1931 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\ | |
1932 "movq "#in7", " #m3 " \n\t" /* d */\ | |
1933 "movq "#in0", %%mm5 \n\t" /* D */\ | |
1934 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\ | |
1935 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\ | |
1936 "movq "#in1", %%mm5 \n\t" /* C */\ | |
1937 "movq "#in2", %%mm6 \n\t" /* B */\ | |
1938 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\ | |
1939 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\ | |
1940 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\ | |
1941 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\ | |
1942 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\ | |
1943 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\ | |
1944 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\ | |
1945 "psraw $5, %%mm5 \n\t"\ | |
1946 "packuswb %%mm5, %%mm5 \n\t"\ | |
954 | 1947 OP(%%mm5, out, %%mm7, d) |
1948 | |
959 | 1949 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\ |
1057 | 1950 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
954 | 1951 uint64_t temp;\ |
1952 \ | |
1953 asm volatile(\ | |
2979 | 1954 "pxor %%mm7, %%mm7 \n\t"\ |
1955 "1: \n\t"\ | |
1956 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
1957 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
1958 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
1959 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
1960 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
1961 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
1962 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
1963 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
1964 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
1965 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
1966 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
1967 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
1968 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
1969 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
1970 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
1971 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
1972 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
1973 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1974 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
1975 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
1976 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
1977 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
1978 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
1979 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
1980 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
1981 "paddw %6, %%mm6 \n\t"\ | |
1982 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1983 "psraw $5, %%mm0 \n\t"\ | |
1984 "movq %%mm0, %5 \n\t"\ | |
954 | 1985 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ |
1986 \ | |
2979 | 1987 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\ |
1988 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\ | |
1989 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\ | |
1990 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\ | |
1991 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\ | |
1992 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\ | |
1993 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\ | |
1994 "paddw %%mm0, %%mm2 \n\t" /* b */\ | |
1995 "paddw %%mm5, %%mm3 \n\t" /* c */\ | |
1996 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
1997 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
1998 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\ | |
1999 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\ | |
2000 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\ | |
2001 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\ | |
2002 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
2003 "paddw %%mm2, %%mm1 \n\t" /* a */\ | |
2004 "paddw %%mm6, %%mm4 \n\t" /* d */\ | |
2005 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
2006 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\ | |
2007 "paddw %6, %%mm1 \n\t"\ | |
2008 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\ | |
2009 "psraw $5, %%mm3 \n\t"\ | |
2010 "movq %5, %%mm1 \n\t"\ | |
2011 "packuswb %%mm3, %%mm1 \n\t"\ | |
959 | 2012 OP_MMX2(%%mm1, (%1),%%mm4, q)\ |
954 | 2013 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\ |
2014 \ | |
2979 | 2015 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\ |
2016 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\ | |
2017 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\ | |
2018 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\ | |
2019 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\ | |
2020 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\ | |
2021 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\ | |
2022 "paddw %%mm1, %%mm5 \n\t" /* b */\ | |
2023 "paddw %%mm4, %%mm0 \n\t" /* c */\ | |
2024 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
2025 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\ | |
2026 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\ | |
2027 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\ | |
2028 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\ | |
2029 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\ | |
2030 "paddw %%mm3, %%mm2 \n\t" /* d */\ | |
2031 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\ | |
2032 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\ | |
2033 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\ | |
2034 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\ | |
2035 "paddw %%mm2, %%mm6 \n\t" /* a */\ | |
2036 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\ | |
2037 "paddw %6, %%mm0 \n\t"\ | |
2038 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
2039 "psraw $5, %%mm0 \n\t"\ | |
954 | 2040 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\ |
2041 \ | |
2979 | 2042 "paddw %%mm5, %%mm3 \n\t" /* a */\ |
2043 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\ | |
2044 "paddw %%mm4, %%mm6 \n\t" /* b */\ | |
2045 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\ | |
2046 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\ | |
2047 "paddw %%mm1, %%mm4 \n\t" /* c */\ | |
2048 "paddw %%mm2, %%mm5 \n\t" /* d */\ | |
2049 "paddw %%mm6, %%mm6 \n\t" /* 2b */\ | |
2050 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\ | |
2051 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\ | |
2052 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\ | |
2053 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\ | |
2054 "paddw %6, %%mm4 \n\t"\ | |
2055 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\ | |
2056 "psraw $5, %%mm4 \n\t"\ | |
2057 "packuswb %%mm4, %%mm0 \n\t"\ | |
959 | 2058 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\ |
954 | 2059 \ |
2979 | 2060 "add %3, %0 \n\t"\ |
2061 "add %4, %1 \n\t"\ | |
2062 "decl %2 \n\t"\ | |
2063 " 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
|
2064 : "+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
|
2065 : "d"((long)srcStride), "S"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
966 | 2066 : "memory"\ |
954 | 2067 );\ |
2068 }\ | |
2069 \ | |
2070 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
2071 int i;\ | |
2072 int16_t temp[16];\ | |
2073 /* quick HACK, XXX FIXME MUST be optimized */\ | |
2074 for(i=0; i<h; i++)\ | |
2075 {\ | |
2076 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
2077 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
2078 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
2079 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
2080 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
2081 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\ | |
2082 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\ | |
2083 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\ | |
2084 temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\ | |
2085 temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\ | |
2086 temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\ | |
2087 temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\ | |
2088 temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\ | |
2089 temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\ | |
2090 temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\ | |
2091 temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\ | |
2092 asm volatile(\ | |
2979 | 2093 "movq (%0), %%mm0 \n\t"\ |
2094 "movq 8(%0), %%mm1 \n\t"\ | |
2095 "paddw %2, %%mm0 \n\t"\ | |
2096 "paddw %2, %%mm1 \n\t"\ | |
2097 "psraw $5, %%mm0 \n\t"\ | |
2098 "psraw $5, %%mm1 \n\t"\ | |
2099 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 2100 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ |
2979 | 2101 "movq 16(%0), %%mm0 \n\t"\ |
2102 "movq 24(%0), %%mm1 \n\t"\ | |
2103 "paddw %2, %%mm0 \n\t"\ | |
2104 "paddw %2, %%mm1 \n\t"\ | |
2105 "psraw $5, %%mm0 \n\t"\ | |
2106 "psraw $5, %%mm1 \n\t"\ | |
2107 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 2108 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\ |
954 | 2109 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ |
966 | 2110 : "memory"\ |
954 | 2111 );\ |
2112 dst+=dstStride;\ | |
2113 src+=srcStride;\ | |
2114 }\ | |
2115 }\ | |
2116 \ | |
1057 | 2117 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
959 | 2118 uint64_t temp;\ |
2119 \ | |
2120 asm volatile(\ | |
2979 | 2121 "pxor %%mm7, %%mm7 \n\t"\ |
2122 "1: \n\t"\ | |
2123 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
2124 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
2125 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
2126 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
2127 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
2128 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
2129 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
2130 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
2131 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
2132 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
2133 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
2134 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
2135 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
2136 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
2137 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
2138 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
2139 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
2140 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
2141 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
2142 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
2143 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
2144 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
2145 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
2146 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
2147 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
2148 "paddw %6, %%mm6 \n\t"\ | |
2149 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
2150 "psraw $5, %%mm0 \n\t"\ | |
959 | 2151 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ |
2152 \ | |
2979 | 2153 "movd 5(%0), %%mm5 \n\t" /* FGHI */\ |
2154 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\ | |
2155 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\ | |
2156 "paddw %%mm5, %%mm1 \n\t" /* a */\ | |
2157 "paddw %%mm6, %%mm2 \n\t" /* b */\ | |
2158 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\ | |
2159 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\ | |
2160 "paddw %%mm6, %%mm3 \n\t" /* c */\ | |
2161 "paddw %%mm5, %%mm4 \n\t" /* d */\ | |
2162 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
2163 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
2164 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
2165 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
2166 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\ | |
2167 "paddw %6, %%mm1 \n\t"\ | |
2168 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\ | |
2169 "psraw $5, %%mm3 \n\t"\ | |
2170 "packuswb %%mm3, %%mm0 \n\t"\ | |
959 | 2171 OP_MMX2(%%mm0, (%1), %%mm4, q)\ |
2172 \ | |
2979 | 2173 "add %3, %0 \n\t"\ |
2174 "add %4, %1 \n\t"\ | |
2175 "decl %2 \n\t"\ | |
2176 " 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
|
2177 : "+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
|
2178 : "S"((long)srcStride), "D"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
966 | 2179 : "memory"\ |
959 | 2180 );\ |
2181 }\ | |
2182 \ | |
2183 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
2184 int i;\ | |
2185 int16_t temp[8];\ | |
2186 /* quick HACK, XXX FIXME MUST be optimized */\ | |
2187 for(i=0; i<h; i++)\ | |
2188 {\ | |
2189 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
2190 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
2191 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
2192 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
2193 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
2194 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\ | |
2195 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\ | |
2196 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\ | |
2197 asm volatile(\ | |
2979 | 2198 "movq (%0), %%mm0 \n\t"\ |
2199 "movq 8(%0), %%mm1 \n\t"\ | |
2200 "paddw %2, %%mm0 \n\t"\ | |
2201 "paddw %2, %%mm1 \n\t"\ | |
2202 "psraw $5, %%mm0 \n\t"\ | |
2203 "psraw $5, %%mm1 \n\t"\ | |
2204 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 2205 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ |
2206 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ | |
966 | 2207 :"memory"\ |
959 | 2208 );\ |
2209 dst+=dstStride;\ | |
2210 src+=srcStride;\ | |
2211 }\ | |
2212 } | |
2213 | |
2214 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\ | |
2215 \ | |
2216 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
954 | 2217 uint64_t temp[17*4];\ |
2218 uint64_t *temp_ptr= temp;\ | |
2219 int count= 17;\ | |
2220 \ | |
2221 /*FIXME unroll */\ | |
2222 asm volatile(\ | |
2979 | 2223 "pxor %%mm7, %%mm7 \n\t"\ |
2224 "1: \n\t"\ | |
2225 "movq (%0), %%mm0 \n\t"\ | |
2226 "movq (%0), %%mm1 \n\t"\ | |
2227 "movq 8(%0), %%mm2 \n\t"\ | |
2228 "movq 8(%0), %%mm3 \n\t"\ | |
2229 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2230 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
2231 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
2232 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
2233 "movq %%mm0, (%1) \n\t"\ | |
2234 "movq %%mm1, 17*8(%1) \n\t"\ | |
2235 "movq %%mm2, 2*17*8(%1) \n\t"\ | |
2236 "movq %%mm3, 3*17*8(%1) \n\t"\ | |
2237 "add $8, %1 \n\t"\ | |
2238 "add %3, %0 \n\t"\ | |
2239 "decl %2 \n\t"\ | |
2240 " jnz 1b \n\t"\ | |
954 | 2241 : "+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
|
2242 : "r" ((long)srcStride)\ |
966 | 2243 : "memory"\ |
954 | 2244 );\ |
2245 \ | |
2246 temp_ptr= temp;\ | |
2247 count=4;\ | |
2248 \ | |
2249 /*FIXME reorder for speed */\ | |
2250 asm volatile(\ | |
2979 | 2251 /*"pxor %%mm7, %%mm7 \n\t"*/\ |
2252 "1: \n\t"\ | |
2253 "movq (%0), %%mm0 \n\t"\ | |
2254 "movq 8(%0), %%mm1 \n\t"\ | |
2255 "movq 16(%0), %%mm2 \n\t"\ | |
2256 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 2257 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
2258 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
2979 | 2259 "add %4, %1 \n\t"\ |
961 | 2260 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 2261 \ |
961 | 2262 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
2979 | 2263 "add %4, %1 \n\t"\ |
961 | 2264 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
2265 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\ | |
2979 | 2266 "add %4, %1 \n\t"\ |
961 | 2267 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\ |
2268 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\ | |
2979 | 2269 "add %4, %1 \n\t"\ |
961 | 2270 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\ |
2271 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\ | |
2979 | 2272 "add %4, %1 \n\t"\ |
961 | 2273 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\ |
2274 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\ | |
2979 | 2275 "add %4, %1 \n\t"\ |
961 | 2276 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\ |
954 | 2277 \ |
961 | 2278 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\ |
2979 | 2279 "add %4, %1 \n\t" \ |
961 | 2280 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\ |
2281 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\ | |
954 | 2282 \ |
2979 | 2283 "add $136, %0 \n\t"\ |
2284 "add %6, %1 \n\t"\ | |
2285 "decl %2 \n\t"\ | |
2286 " 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
|
2287 \ |
967 | 2288 : "+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
|
2289 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-14*(long)dstStride)\ |
966 | 2290 :"memory"\ |
954 | 2291 );\ |
2292 }\ | |
2293 \ | |
1057 | 2294 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2209 | 2295 uint64_t temp[9*2];\ |
954 | 2296 uint64_t *temp_ptr= temp;\ |
2297 int count= 9;\ | |
2298 \ | |
2299 /*FIXME unroll */\ | |
2300 asm volatile(\ | |
2979 | 2301 "pxor %%mm7, %%mm7 \n\t"\ |
2302 "1: \n\t"\ | |
2303 "movq (%0), %%mm0 \n\t"\ | |
2304 "movq (%0), %%mm1 \n\t"\ | |
2305 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2306 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
2307 "movq %%mm0, (%1) \n\t"\ | |
2308 "movq %%mm1, 9*8(%1) \n\t"\ | |
2309 "add $8, %1 \n\t"\ | |
2310 "add %3, %0 \n\t"\ | |
2311 "decl %2 \n\t"\ | |
2312 " jnz 1b \n\t"\ | |
954 | 2313 : "+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
|
2314 : "r" ((long)srcStride)\ |
966 | 2315 : "memory"\ |
954 | 2316 );\ |
2317 \ | |
2318 temp_ptr= temp;\ | |
2319 count=2;\ | |
2320 \ | |
2321 /*FIXME reorder for speed */\ | |
2322 asm volatile(\ | |
2979 | 2323 /*"pxor %%mm7, %%mm7 \n\t"*/\ |
2324 "1: \n\t"\ | |
2325 "movq (%0), %%mm0 \n\t"\ | |
2326 "movq 8(%0), %%mm1 \n\t"\ | |
2327 "movq 16(%0), %%mm2 \n\t"\ | |
2328 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 2329 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
2330 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
2979 | 2331 "add %4, %1 \n\t"\ |
961 | 2332 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 2333 \ |
961 | 2334 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
2979 | 2335 "add %4, %1 \n\t"\ |
961 | 2336 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
954 | 2337 \ |
961 | 2338 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\ |
2979 | 2339 "add %4, %1 \n\t"\ |
961 | 2340 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\ |
2341 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\ | |
954 | 2342 \ |
2979 | 2343 "add $72, %0 \n\t"\ |
2344 "add %6, %1 \n\t"\ | |
2345 "decl %2 \n\t"\ | |
2346 " jnz 1b \n\t"\ | |
954 | 2347 \ |
961 | 2348 : "+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
|
2349 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-6*(long)dstStride)\ |
966 | 2350 : "memory"\ |
2351 );\ | |
959 | 2352 }\ |
954 | 2353 \ |
1064 | 2354 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
6321 | 2355 OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);\ |
954 | 2356 }\ |
2357 \ | |
1064 | 2358 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2359 uint64_t temp[8];\ |
954 | 2360 uint8_t * const half= (uint8_t*)temp;\ |
2361 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
|
2362 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ |
954 | 2363 }\ |
2364 \ | |
1064 | 2365 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2366 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\ |
2367 }\ | |
2368 \ | |
1064 | 2369 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2370 uint64_t temp[8];\ |
954 | 2371 uint8_t * const half= (uint8_t*)temp;\ |
2372 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
|
2373 OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\ |
954 | 2374 }\ |
2375 \ | |
1064 | 2376 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2377 uint64_t temp[8];\ |
954 | 2378 uint8_t * const half= (uint8_t*)temp;\ |
959 | 2379 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
|
2380 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ |
954 | 2381 }\ |
2382 \ | |
1064 | 2383 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2384 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 2385 }\ |
2386 \ | |
1064 | 2387 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2388 uint64_t temp[8];\ |
954 | 2389 uint8_t * const half= (uint8_t*)temp;\ |
959 | 2390 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
|
2391 OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\ |
954 | 2392 }\ |
1064 | 2393 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2394 uint64_t half[8 + 9];\ |
2395 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2396 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2397 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2398 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
959 | 2399 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
|
2400 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 2401 }\ |
1064 | 2402 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2403 uint64_t half[8 + 9];\ |
2404 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2405 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2406 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
|
2407 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
959 | 2408 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
|
2409 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 2410 }\ |
1064 | 2411 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2412 uint64_t half[8 + 9];\ |
2413 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2414 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2415 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
|
2416 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
959 | 2417 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
|
2418 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 2419 }\ |
1064 | 2420 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2421 uint64_t half[8 + 9];\ |
2422 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2423 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2424 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
|
2425 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
959 | 2426 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
|
2427 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 2428 }\ |
1064 | 2429 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2430 uint64_t half[8 + 9];\ |
954 | 2431 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
2432 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2433 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 2434 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
|
2435 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 2436 }\ |
1064 | 2437 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2438 uint64_t half[8 + 9];\ |
954 | 2439 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
2440 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2441 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 2442 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
|
2443 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 2444 }\ |
1064 | 2445 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2446 uint64_t half[8 + 9];\ |
2447 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2448 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
|
2449 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
984 | 2450 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 2451 }\ |
1064 | 2452 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2453 uint64_t half[8 + 9];\ |
2454 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2455 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
|
2456 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
984 | 2457 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 2458 }\ |
1064 | 2459 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2460 uint64_t half[9];\ |
954 | 2461 uint8_t * const halfH= ((uint8_t*)half);\ |
2462 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 2463 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 2464 }\ |
1064 | 2465 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
6321 | 2466 OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);\ |
954 | 2467 }\ |
2468 \ | |
1064 | 2469 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2470 uint64_t temp[32];\ |
2471 uint8_t * const half= (uint8_t*)temp;\ | |
2472 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
|
2473 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ |
954 | 2474 }\ |
2475 \ | |
1064 | 2476 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2477 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\ |
2478 }\ | |
2479 \ | |
1064 | 2480 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2481 uint64_t temp[32];\ |
2482 uint8_t * const half= (uint8_t*)temp;\ | |
2483 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
|
2484 OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\ |
954 | 2485 }\ |
2486 \ | |
1064 | 2487 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2488 uint64_t temp[32];\ |
2489 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 2490 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
|
2491 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ |
954 | 2492 }\ |
2493 \ | |
1064 | 2494 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2495 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 2496 }\ |
2497 \ | |
1064 | 2498 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2499 uint64_t temp[32];\ |
2500 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 2501 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
|
2502 OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\ |
954 | 2503 }\ |
1064 | 2504 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2505 uint64_t half[16*2 + 17*2];\ |
2506 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2507 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2508 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2067
diff
changeset
|
2509 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
959 | 2510 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
|
2511 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 2512 }\ |
1064 | 2513 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2514 uint64_t half[16*2 + 17*2];\ |
2515 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2516 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2517 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
|
2518 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
959 | 2519 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
|
2520 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 2521 }\ |
1064 | 2522 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2523 uint64_t half[16*2 + 17*2];\ |
2524 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2525 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2526 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
|
2527 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
959 | 2528 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
|
2529 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 2530 }\ |
1064 | 2531 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2532 uint64_t half[16*2 + 17*2];\ |
2533 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2534 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2535 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
|
2536 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
959 | 2537 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
|
2538 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 2539 }\ |
1064 | 2540 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2541 uint64_t half[16*2 + 17*2];\ |
2542 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2543 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2544 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 2545 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
|
2546 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 2547 }\ |
1064 | 2548 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2549 uint64_t half[16*2 + 17*2];\ |
2550 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2551 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2552 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 2553 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
|
2554 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 2555 }\ |
1064 | 2556 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2557 uint64_t half[17*2];\ |
2558 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2559 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
|
2560 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
984 | 2561 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 2562 }\ |
1064 | 2563 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2564 uint64_t half[17*2];\ |
2565 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2566 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
|
2567 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
984 | 2568 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 2569 }\ |
1064 | 2570 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2571 uint64_t half[17*2];\ |
2572 uint8_t * const halfH= ((uint8_t*)half);\ | |
2573 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 2574 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 2575 } |
2576 | |
2979 | 2577 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t" |
959 | 2578 #define AVG_3DNOW_OP(a,b,temp, size) \ |
2979 | 2579 "mov" #size " " #b ", " #temp " \n\t"\ |
2580 "pavgusb " #temp ", " #a " \n\t"\ | |
2581 "mov" #size " " #a ", " #b " \n\t" | |
959 | 2582 #define AVG_MMX2_OP(a,b,temp, size) \ |
2979 | 2583 "mov" #size " " #b ", " #temp " \n\t"\ |
2584 "pavgb " #temp ", " #a " \n\t"\ | |
2585 "mov" #size " " #a ", " #b " \n\t" | |
959 | 2586 |
2587 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP) | |
2588 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP) | |
2589 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP) | |
2590 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow) | |
2591 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow) | |
2592 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow) | |
954 | 2593 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2) |
959 | 2594 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2) |
954 | 2595 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2) |
2596 | |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2597 /***********************************/ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2598 /* 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
|
2599 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2600 #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
|
2601 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
|
2602 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
|
2603 } |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2604 #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
|
2605 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
|
2606 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
|
2607 } |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2608 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2609 #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
|
2610 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
|
2611 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
|
2612 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
|
2613 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
|
2614 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
|
2615 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
|
2616 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
|
2617 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
|
2618 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
|
2619 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
|
2620 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
|
2621 }\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2622 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
|
2623 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
|
2624 }\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2625 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
|
2626 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
|
2627 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
|
2628 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
|
2629 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
|
2630 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
|
2631 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
|
2632 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
|
2633 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2634 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
|
2635 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
|
2636 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
|
2637 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
|
2638 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
|
2639 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
|
2640 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
|
2641 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
|
2642 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3777
diff
changeset
|
2643 |
393 | 2644 #if 0 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2645 static void just_return() { return; } |
393 | 2646 #endif |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2647 |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2648 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
|
2649 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
|
2650 const int w = 8; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2651 const int ix = ox>>(16+shift); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2652 const int iy = oy>>(16+shift); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2653 const int oxs = ox>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2654 const int oys = oy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2655 const int dxxs = dxx>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2656 const int dxys = dxy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2657 const int dyxs = dyx>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2658 const int dyys = dyy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2659 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
|
2660 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
|
2661 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
|
2662 const uint64_t shift2 = 2*shift; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2663 uint8_t edge_buf[(h+1)*stride]; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2664 int x, y; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2665 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2666 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
|
2667 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
|
2668 const int dxh = dxy*(h-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2669 const int dyw = dyx*(w-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2670 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
|
2671 ((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
|
2672 (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
|
2673 // 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
|
2674 || (dxx|dxy|dyx|dyy)&15 ) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2675 { |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2676 //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
|
2677 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
|
2678 return; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2679 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2680 |
3250 | 2681 src += ix + iy*stride; |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2682 if( (unsigned)ix >= width-w || |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2683 (unsigned)iy >= height-h ) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2684 { |
3250 | 2685 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
|
2686 src = edge_buf; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2687 } |
3250 | 2688 |
2689 asm volatile( | |
2690 "movd %0, %%mm6 \n\t" | |
2691 "pxor %%mm7, %%mm7 \n\t" | |
2692 "punpcklwd %%mm6, %%mm6 \n\t" | |
2693 "punpcklwd %%mm6, %%mm6 \n\t" | |
2694 :: "r"(1<<shift) | |
2695 ); | |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2696 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2697 for(x=0; x<w; x+=4){ |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2698 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
|
2699 oxs - dxys + dxxs*(x+1), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2700 oxs - dxys + dxxs*(x+2), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2701 oxs - dxys + dxxs*(x+3) }; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2702 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
|
2703 oys - dyys + dyxs*(x+1), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2704 oys - dyys + dyxs*(x+2), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2705 oys - dyys + dyxs*(x+3) }; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2706 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2707 for(y=0; y<h; y++){ |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2708 asm volatile( |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2709 "movq %0, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2710 "movq %1, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2711 "paddw %2, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2712 "paddw %3, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2713 "movq %%mm4, %0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2714 "movq %%mm5, %1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2715 "psrlw $12, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2716 "psrlw $12, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2717 : "+m"(*dx4), "+m"(*dy4) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2718 : "m"(*dxy4), "m"(*dyy4) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2719 ); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2720 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2721 asm volatile( |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2722 "movq %%mm6, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2723 "movq %%mm6, %%mm1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2724 "psubw %%mm4, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2725 "psubw %%mm5, %%mm1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2726 "movq %%mm2, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2727 "movq %%mm4, %%mm3 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2728 "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
|
2729 "pmullw %%mm5, %%mm3 \n\t" // dx*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2730 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2731 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2732 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2733 "movd %4, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2734 "movd %3, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2735 "punpcklbw %%mm7, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2736 "punpcklbw %%mm7, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2737 "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
|
2738 "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
|
2739 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2740 "movd %2, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2741 "movd %1, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2742 "punpcklbw %%mm7, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2743 "punpcklbw %%mm7, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2744 "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
|
2745 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy) |
3250 | 2746 "paddw %5, %%mm1 \n\t" |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2747 "paddw %%mm3, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2748 "paddw %%mm1, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2749 "paddw %%mm2, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2750 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2751 "psrlw %6, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2752 "packuswb %%mm0, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2753 "movd %%mm0, %0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2754 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2755 : "=m"(dst[x+y*stride]) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2756 : "m"(src[0]), "m"(src[1]), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2757 "m"(src[stride]), "m"(src[stride+1]), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2758 "m"(*r4), "m"(shift2) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2759 ); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2760 src += stride; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2761 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2762 src += 4-h*stride; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2763 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2764 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2765 |
3777 | 2766 #ifdef CONFIG_ENCODERS |
5024
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2767 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2768 #define PHADDD(a, t)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2769 "movq "#a", "#t" \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2770 "psrlq $32, "#a" \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2771 "paddd "#t", "#a" \n\t" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2772 /* |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2773 pmulhw: dst[0-15]=(src[0-15]*dst[0-15])[16-31] |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2774 pmulhrw: dst[0-15]=(src[0-15]*dst[0-15] + 0x8000)[16-31] |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2775 pmulhrsw: dst[0-15]=(src[0-15]*dst[0-15] + 0x4000)[15-30] |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2776 */ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2777 #define PMULHRW(x, y, s, o)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2778 "pmulhw " #s ", "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2779 "pmulhw " #s ", "#y " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2780 "paddw " #o ", "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2781 "paddw " #o ", "#y " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2782 "psraw $1, "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2783 "psraw $1, "#y " \n\t" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2784 #define DEF(x) x ## _mmx |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2785 #define SET_RND MOVQ_WONE |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2786 #define SCALE_OFFSET 1 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2787 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2788 #include "dsputil_mmx_qns.h" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2789 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2790 #undef DEF |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2791 #undef SET_RND |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2792 #undef SCALE_OFFSET |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2793 #undef PMULHRW |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2794 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2795 #define DEF(x) x ## _3dnow |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2796 #define SET_RND(x) |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2797 #define SCALE_OFFSET 0 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2798 #define PMULHRW(x, y, s, o)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2799 "pmulhrw " #s ", "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2800 "pmulhrw " #s ", "#y " \n\t" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2801 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2802 #include "dsputil_mmx_qns.h" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2803 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2804 #undef DEF |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2805 #undef SET_RND |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2806 #undef SCALE_OFFSET |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2807 #undef PMULHRW |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2808 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2809 #ifdef HAVE_SSSE3 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2810 #undef PHADDD |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2811 #define DEF(x) x ## _ssse3 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2812 #define SET_RND(x) |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2813 #define SCALE_OFFSET -1 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2814 #define PHADDD(a, t)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2815 "pshufw $0x0E, "#a", "#t" \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2816 "paddd "#t", "#a" \n\t" /* faster than phaddd on core2 */ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2817 #define PMULHRW(x, y, s, o)\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2818 "pmulhrsw " #s ", "#x " \n\t"\ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2819 "pmulhrsw " #s ", "#y " \n\t" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2820 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2821 #include "dsputil_mmx_qns.h" |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2822 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2823 #undef DEF |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2824 #undef SET_RND |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2825 #undef SCALE_OFFSET |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2826 #undef PMULHRW |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2827 #undef PHADDD |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2828 #endif //HAVE_SSSE3 |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
2829 |
3777 | 2830 #endif /* CONFIG_ENCODERS */ |
2754 | 2831 |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2832 #define PREFETCH(name, op) \ |
4172 | 2833 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
|
2834 const uint8_t *p= mem;\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2835 do{\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2836 asm volatile(#op" %0" :: "m"(*p));\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2837 p+= stride;\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2838 }while(--h);\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2839 } |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2840 PREFETCH(prefetch_mmx2, prefetcht0) |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2841 PREFETCH(prefetch_3dnow, prefetch) |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2842 #undef PREFETCH |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2843 |
2754 | 2844 #include "h264dsp_mmx.c" |
2967 | 2845 |
6009 | 2846 /* CAVS specific */ |
3524 | 2847 void ff_cavsdsp_init_mmx2(DSPContext* c, AVCodecContext *avctx); |
2848 | |
2849 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2850 put_pixels8_mmx(dst, src, stride, 8); | |
2851 } | |
2852 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2853 avg_pixels8_mmx(dst, src, stride, 8); | |
2854 } | |
2855 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2856 put_pixels16_mmx(dst, src, stride, 16); | |
2857 } | |
2858 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2859 avg_pixels16_mmx(dst, src, stride, 16); | |
2860 } | |
2861 | |
6030 | 2862 /* FLAC specific */ |
2863 void ff_flac_compute_autocorr_sse2(const int32_t *data, int len, int lag, | |
2864 double *autoc); | |
2865 | |
5948 | 2866 /* VC1 specific */ |
2867 void ff_vc1dsp_init_mmx(DSPContext* dsp, AVCodecContext *avctx); | |
2868 | |
2869 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) { | |
2870 put_pixels8_mmx(dst, src, stride, 8); | |
2871 } | |
2872 | |
1092 | 2873 /* external functions, from idct_mmx.c */ |
2874 void ff_mmx_idct(DCTELEM *block); | |
2875 void ff_mmxext_idct(DCTELEM *block); | |
2876 | |
2877 /* XXX: those functions should be suppressed ASAP when all IDCTs are | |
2878 converted */ | |
4020
723818b5de0f
Put libmpeg2 IDCT functions under CONFIG_GPL, fixes link failure
diego
parents:
4001
diff
changeset
|
2879 #ifdef CONFIG_GPL |
1092 | 2880 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block) |
2881 { | |
2882 ff_mmx_idct (block); | |
2883 put_pixels_clamped_mmx(block, dest, line_size); | |
2884 } | |
2885 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2886 { | |
2887 ff_mmx_idct (block); | |
2888 add_pixels_clamped_mmx(block, dest, line_size); | |
2889 } | |
2890 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block) | |
2891 { | |
2892 ff_mmxext_idct (block); | |
2893 put_pixels_clamped_mmx(block, dest, line_size); | |
2894 } | |
2895 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2896 { | |
2897 ff_mmxext_idct (block); | |
2898 add_pixels_clamped_mmx(block, dest, line_size); | |
2899 } | |
4020
723818b5de0f
Put libmpeg2 IDCT functions under CONFIG_GPL, fixes link failure
diego
parents:
4001
diff
changeset
|
2900 #endif |
2868 | 2901 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block) |
2902 { | |
2903 ff_idct_xvid_mmx (block); | |
2904 put_pixels_clamped_mmx(block, dest, line_size); | |
2905 } | |
2906 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2907 { | |
2908 ff_idct_xvid_mmx (block); | |
2909 add_pixels_clamped_mmx(block, dest, line_size); | |
2910 } | |
2911 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block) | |
2912 { | |
2913 ff_idct_xvid_mmx2 (block); | |
2914 put_pixels_clamped_mmx(block, dest, line_size); | |
2915 } | |
2916 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2917 { | |
2918 ff_idct_xvid_mmx2 (block); | |
2919 add_pixels_clamped_mmx(block, dest, line_size); | |
2920 } | |
2967 | 2921 |
3541
3fbddeb13686
10l, vorbis_inverse_coupling_sse() was really 3dnow
lorenm
parents:
3536
diff
changeset
|
2922 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
|
2923 { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2924 int i; |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2925 asm volatile("pxor %%mm7, %%mm7":); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2926 for(i=0; i<blocksize; i+=2) { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2927 asm volatile( |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2928 "movq %0, %%mm0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2929 "movq %1, %%mm1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2930 "movq %%mm0, %%mm2 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2931 "movq %%mm1, %%mm3 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2932 "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2933 "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2934 "pslld $31, %%mm2 \n\t" // keep only the sign bit |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2935 "pxor %%mm2, %%mm1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2936 "movq %%mm3, %%mm4 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2937 "pand %%mm1, %%mm3 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2938 "pandn %%mm1, %%mm4 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2939 "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
|
2940 "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
|
2941 "movq %%mm3, %1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2942 "movq %%mm0, %0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2943 :"+m"(mag[i]), "+m"(ang[i]) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2944 ::"memory" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2945 ); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2946 } |
3561 | 2947 asm volatile("femms"); |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2948 } |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2949 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
|
2950 { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2951 int i; |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2952 |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2953 asm volatile( |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2954 "movaps %0, %%xmm5 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2955 ::"m"(ff_pdw_80000000[0]) |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2956 ); |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2957 for(i=0; i<blocksize; i+=4) { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2958 asm volatile( |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2959 "movaps %0, %%xmm0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2960 "movaps %1, %%xmm1 \n\t" |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2961 "xorps %%xmm2, %%xmm2 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2962 "xorps %%xmm3, %%xmm3 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2963 "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2964 "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
|
2965 "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
|
2966 "xorps %%xmm2, %%xmm1 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2967 "movaps %%xmm3, %%xmm4 \n\t" |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2968 "andps %%xmm1, %%xmm3 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2969 "andnps %%xmm1, %%xmm4 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2970 "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
|
2971 "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
|
2972 "movaps %%xmm3, %1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2973 "movaps %%xmm0, %0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2974 :"+m"(mag[i]), "+m"(ang[i]) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2975 ::"memory" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2976 ); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2977 } |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2978 } |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2979 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2980 static void vector_fmul_3dnow(float *dst, const float *src, int len){ |
3574 | 2981 long i = (len-4)*4; |
2982 asm volatile( | |
2983 "1: \n\t" | |
2984 "movq (%1,%0), %%mm0 \n\t" | |
2985 "movq 8(%1,%0), %%mm1 \n\t" | |
2986 "pfmul (%2,%0), %%mm0 \n\t" | |
2987 "pfmul 8(%2,%0), %%mm1 \n\t" | |
2988 "movq %%mm0, (%1,%0) \n\t" | |
2989 "movq %%mm1, 8(%1,%0) \n\t" | |
2990 "sub $16, %0 \n\t" | |
2991 "jge 1b \n\t" | |
2992 "femms \n\t" | |
2993 :"+r"(i) | |
2994 :"r"(dst), "r"(src) | |
2995 :"memory" | |
2996 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2997 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2998 static void vector_fmul_sse(float *dst, const float *src, int len){ |
3574 | 2999 long i = (len-8)*4; |
3000 asm volatile( | |
3001 "1: \n\t" | |
3002 "movaps (%1,%0), %%xmm0 \n\t" | |
3003 "movaps 16(%1,%0), %%xmm1 \n\t" | |
3004 "mulps (%2,%0), %%xmm0 \n\t" | |
3005 "mulps 16(%2,%0), %%xmm1 \n\t" | |
3006 "movaps %%xmm0, (%1,%0) \n\t" | |
3007 "movaps %%xmm1, 16(%1,%0) \n\t" | |
3008 "sub $32, %0 \n\t" | |
3009 "jge 1b \n\t" | |
3010 :"+r"(i) | |
3011 :"r"(dst), "r"(src) | |
3012 :"memory" | |
3013 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3014 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3015 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3016 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
|
3017 long i = len*4-16; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3018 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3019 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3020 "pswapd 8(%1), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3021 "pswapd (%1), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3022 "pfmul (%3,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3023 "pfmul 8(%3,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3024 "movq %%mm0, (%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3025 "movq %%mm1, 8(%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3026 "add $16, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3027 "sub $16, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3028 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3029 :"+r"(i), "+r"(src1) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3030 :"r"(dst), "r"(src0) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3031 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3032 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3033 } |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3034 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
|
3035 long i = len*4-32; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3036 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3037 "1: \n\t" |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3038 "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
|
3039 "movaps (%1), %%xmm1 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3040 "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
|
3041 "shufps $0x1b, %%xmm1, %%xmm1 \n\t" |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3042 "mulps (%3,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3043 "mulps 16(%3,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3044 "movaps %%xmm0, (%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3045 "movaps %%xmm1, 16(%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3046 "add $32, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3047 "sub $32, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3048 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3049 :"+r"(i), "+r"(src1) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3050 :"r"(dst), "r"(src0) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3051 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3052 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3053 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3054 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
|
3055 const float *src2, int src3, int len, int step){ |
3574 | 3056 long i = (len-4)*4; |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3057 if(step == 2 && src3 == 0){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3058 dst += (len-4)*2; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3059 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3060 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3061 "movq (%2,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3062 "movq 8(%2,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3063 "pfmul (%3,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3064 "pfmul 8(%3,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3065 "pfadd (%4,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3066 "pfadd 8(%4,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3067 "movd %%mm0, (%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3068 "movd %%mm1, 16(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3069 "psrlq $32, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3070 "psrlq $32, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3071 "movd %%mm0, 8(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3072 "movd %%mm1, 24(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3073 "sub $32, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3074 "sub $16, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3075 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3076 :"+r"(i), "+r"(dst) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3077 :"r"(src0), "r"(src1), "r"(src2) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3078 :"memory" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3079 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3080 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3081 else if(step == 1 && src3 == 0){ |
3574 | 3082 asm volatile( |
3083 "1: \n\t" | |
3084 "movq (%2,%0), %%mm0 \n\t" | |
3085 "movq 8(%2,%0), %%mm1 \n\t" | |
3086 "pfmul (%3,%0), %%mm0 \n\t" | |
3087 "pfmul 8(%3,%0), %%mm1 \n\t" | |
3088 "pfadd (%4,%0), %%mm0 \n\t" | |
3089 "pfadd 8(%4,%0), %%mm1 \n\t" | |
3090 "movq %%mm0, (%1,%0) \n\t" | |
3091 "movq %%mm1, 8(%1,%0) \n\t" | |
3092 "sub $16, %0 \n\t" | |
3093 "jge 1b \n\t" | |
3094 :"+r"(i) | |
3095 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
3096 :"memory" | |
3097 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3098 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3099 else |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3100 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
|
3101 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3102 } |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3103 static void vector_fmul_add_add_sse(float *dst, const float *src0, const float *src1, |
3574 | 3104 const float *src2, int src3, int len, int step){ |
3105 long i = (len-8)*4; | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3106 if(step == 2 && src3 == 0){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3107 dst += (len-8)*2; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3108 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3109 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3110 "movaps (%2,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3111 "movaps 16(%2,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3112 "mulps (%3,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3113 "mulps 16(%3,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3114 "addps (%4,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3115 "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
|
3116 "movss %%xmm0, (%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3117 "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
|
3118 "movhlps %%xmm0, %%xmm2 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3119 "movhlps %%xmm1, %%xmm3 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3120 "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
|
3121 "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
|
3122 "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
|
3123 "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
|
3124 "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
|
3125 "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
|
3126 "movhlps %%xmm0, %%xmm2 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3127 "movhlps %%xmm1, %%xmm3 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
3128 "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
|
3129 "movss %%xmm3, 56(%1) \n\t" |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3130 "sub $64, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3131 "sub $32, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3132 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3133 :"+r"(i), "+r"(dst) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3134 :"r"(src0), "r"(src1), "r"(src2) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3135 :"memory" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3136 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3137 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3138 else if(step == 1 && src3 == 0){ |
3574 | 3139 asm volatile( |
3140 "1: \n\t" | |
3141 "movaps (%2,%0), %%xmm0 \n\t" | |
3142 "movaps 16(%2,%0), %%xmm1 \n\t" | |
3143 "mulps (%3,%0), %%xmm0 \n\t" | |
3144 "mulps 16(%3,%0), %%xmm1 \n\t" | |
3145 "addps (%4,%0), %%xmm0 \n\t" | |
3146 "addps 16(%4,%0), %%xmm1 \n\t" | |
3147 "movaps %%xmm0, (%1,%0) \n\t" | |
3148 "movaps %%xmm1, 16(%1,%0) \n\t" | |
3149 "sub $32, %0 \n\t" | |
3150 "jge 1b \n\t" | |
3151 :"+r"(i) | |
3152 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
3153 :"memory" | |
3154 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3155 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3156 else |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3157 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
|
3158 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3159 |
4172 | 3160 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
|
3161 // 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
|
3162 int i; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3163 for(i=0; i<len; i+=4) { |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3164 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3165 "pf2id %1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3166 "pf2id %2, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3167 "packssdw %%mm1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3168 "movq %%mm0, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3169 :"=m"(dst[i]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3170 :"m"(src[i]), "m"(src[i+2]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3171 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3172 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3173 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3174 } |
4172 | 3175 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
|
3176 int i; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3177 for(i=0; i<len; i+=4) { |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3178 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3179 "cvtps2pi %1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3180 "cvtps2pi %2, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3181 "packssdw %%mm1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3182 "movq %%mm0, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3183 :"=m"(dst[i]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3184 :"m"(src[i]), "m"(src[i+2]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3185 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3186 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3187 asm volatile("emms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3188 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3189 |
6195
5f704e9cb518
fix prototypes, remove warning: i386/dsputil_mmx.c:3594: warning: assignment from incompatible pointer type
bcoudurier
parents:
6181
diff
changeset
|
3190 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
|
3191 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
|
3192 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
|
3193 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
|
3194 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
|
3195 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
|
3196 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
|
3197 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
|
3198 |
1092 | 3199 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx) |
0 | 3200 { |
4197 | 3201 mm_flags = mm_support(); |
1115 | 3202 |
1122 | 3203 if (avctx->dsp_mask) { |
2979 | 3204 if (avctx->dsp_mask & FF_MM_FORCE) |
4197 | 3205 mm_flags |= (avctx->dsp_mask & 0xffff); |
2979 | 3206 else |
4197 | 3207 mm_flags &= ~(avctx->dsp_mask & 0xffff); |
1122 | 3208 } |
1115 | 3209 |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
3210 #if 0 |
1868 | 3211 av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:"); |
4197 | 3212 if (mm_flags & MM_MMX) |
1868 | 3213 av_log(avctx, AV_LOG_INFO, " mmx"); |
4197 | 3214 if (mm_flags & MM_MMXEXT) |
1868 | 3215 av_log(avctx, AV_LOG_INFO, " mmxext"); |
4197 | 3216 if (mm_flags & MM_3DNOW) |
1868 | 3217 av_log(avctx, AV_LOG_INFO, " 3dnow"); |
4197 | 3218 if (mm_flags & MM_SSE) |
1868 | 3219 av_log(avctx, AV_LOG_INFO, " sse"); |
4197 | 3220 if (mm_flags & MM_SSE2) |
1868 | 3221 av_log(avctx, AV_LOG_INFO, " sse2"); |
3222 av_log(avctx, AV_LOG_INFO, "\n"); | |
0 | 3223 #endif |
3224 | |
4197 | 3225 if (mm_flags & MM_MMX) { |
1092 | 3226 const int idct_algo= avctx->idct_algo; |
3227 | |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1186
diff
changeset
|
3228 #ifdef CONFIG_ENCODERS |
2024
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1985
diff
changeset
|
3229 const int dct_algo = avctx->dct_algo; |
1565 | 3230 if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){ |
4197 | 3231 if(mm_flags & MM_SSE2){ |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
3232 c->fdct = ff_fdct_sse2; |
4197 | 3233 }else if(mm_flags & MM_MMXEXT){ |
1565 | 3234 c->fdct = ff_fdct_mmx2; |
3235 }else{ | |
3236 c->fdct = ff_fdct_mmx; | |
3237 } | |
3238 } | |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1186
diff
changeset
|
3239 #endif //CONFIG_ENCODERS |
2256 | 3240 if(avctx->lowres==0){ |
3241 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){ | |
3242 c->idct_put= ff_simple_idct_put_mmx; | |
3243 c->idct_add= ff_simple_idct_add_mmx; | |
3244 c->idct = ff_simple_idct_mmx; | |
3245 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
|
3246 #ifdef CONFIG_GPL |
2256 | 3247 }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){ |
4197 | 3248 if(mm_flags & MM_MMXEXT){ |
2256 | 3249 c->idct_put= ff_libmpeg2mmx2_idct_put; |
3250 c->idct_add= ff_libmpeg2mmx2_idct_add; | |
3251 c->idct = ff_mmxext_idct; | |
3252 }else{ | |
3253 c->idct_put= ff_libmpeg2mmx_idct_put; | |
3254 c->idct_add= ff_libmpeg2mmx_idct_add; | |
3255 c->idct = ff_mmx_idct; | |
3256 } | |
3257 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
|
3258 #endif |
5007 | 3259 }else if((ENABLE_VP3_DECODER || ENABLE_VP5_DECODER || ENABLE_VP6_DECODER) && |
3260 idct_algo==FF_IDCT_VP3 && | |
3721
2000e401593d
disable vp3 mmx idct for theora files to avoid artifacts
aurel
parents:
3717
diff
changeset
|
3261 avctx->codec->id!=CODEC_ID_THEORA && |
3712
f7f75f718efb
Enables back the mmx/sse optimized version of the vp3 idct.
aurel
parents:
3666
diff
changeset
|
3262 !(avctx->flags & CODEC_FLAG_BITEXACT)){ |
4197 | 3263 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
|
3264 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
|
3265 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
|
3266 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
|
3267 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
|
3268 }else{ |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3269 ff_vp3_dsp_init_mmx(); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3270 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
|
3271 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
|
3272 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
|
3273 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
|
3274 } |
3524 | 3275 }else if(idct_algo==FF_IDCT_CAVS){ |
3276 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM; | |
2868 | 3277 }else if(idct_algo==FF_IDCT_XVIDMMX){ |
4197 | 3278 if(mm_flags & MM_MMXEXT){ |
2868 | 3279 c->idct_put= ff_idct_xvid_mmx2_put; |
3280 c->idct_add= ff_idct_xvid_mmx2_add; | |
3281 c->idct = ff_idct_xvid_mmx2; | |
3282 }else{ | |
3283 c->idct_put= ff_idct_xvid_mmx_put; | |
3284 c->idct_add= ff_idct_xvid_mmx_add; | |
3285 c->idct = ff_idct_xvid_mmx; | |
3286 } | |
1092 | 3287 } |
3288 } | |
1868 | 3289 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3290 #ifdef CONFIG_ENCODERS |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3291 c->get_pixels = get_pixels_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3292 c->diff_pixels = diff_pixels_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3293 #endif //CONFIG_ENCODERS |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3294 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
|
3295 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
|
3296 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
|
3297 c->clear_blocks = clear_blocks_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3298 #ifdef CONFIG_ENCODERS |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3299 c->pix_sum = pix_sum16_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3300 #endif //CONFIG_ENCODERS |
415 | 3301 |
6327 | 3302 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU) \ |
3303 c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## SIZE ## _ ## CPU; \ | |
3304 c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## SIZE ## _x2_ ## CPU; \ | |
3305 c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## SIZE ## _y2_ ## CPU; \ | |
3306 c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU | |
3307 | |
3308 SET_HPEL_FUNCS(put, 0, 16, mmx); | |
3309 SET_HPEL_FUNCS(put_no_rnd, 0, 16, mmx); | |
3310 SET_HPEL_FUNCS(avg, 0, 16, mmx); | |
3311 SET_HPEL_FUNCS(avg_no_rnd, 0, 16, mmx); | |
3312 SET_HPEL_FUNCS(put, 1, 8, mmx); | |
3313 SET_HPEL_FUNCS(put_no_rnd, 1, 8, mmx); | |
3314 SET_HPEL_FUNCS(avg, 1, 8, mmx); | |
3315 SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx); | |
2967 | 3316 |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
3317 c->gmc= gmc_mmx; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
3318 |
866 | 3319 c->add_bytes= add_bytes_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3320 #ifdef CONFIG_ENCODERS |
866 | 3321 c->diff_bytes= diff_bytes_mmx; |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
3322 c->sum_abs_dctelem= sum_abs_dctelem_mmx; |
2967 | 3323 |
936 | 3324 c->hadamard8_diff[0]= hadamard8_diff16_mmx; |
3325 c->hadamard8_diff[1]= hadamard8_diff_mmx; | |
2967 | 3326 |
2979 | 3327 c->pix_norm1 = pix_norm1_mmx; |
4197 | 3328 c->sse[0] = (mm_flags & MM_SSE2) ? sse16_sse2 : sse16_mmx; |
2979 | 3329 c->sse[1] = sse8_mmx; |
1729 | 3330 c->vsad[4]= vsad_intra16_mmx; |
3331 | |
2979 | 3332 c->nsse[0] = nsse16_mmx; |
3333 c->nsse[1] = nsse8_mmx; | |
1729 | 3334 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
3335 c->vsad[0] = vsad16_mmx; | |
3336 } | |
2967 | 3337 |
1784 | 3338 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
3339 c->try_8x8basis= try_8x8basis_mmx; | |
3340 } | |
3341 c->add_8x8basis= add_8x8basis_mmx; | |
2967 | 3342 |
4749 | 3343 c->ssd_int8_vs_int16 = ssd_int8_vs_int16_mmx; |
3344 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3345 #endif //CONFIG_ENCODERS |
1647 | 3346 |
5277
7b3fcb7c61ce
Avoid linking with h263.c functions when the relevant codecs
aurel
parents:
5255
diff
changeset
|
3347 if (ENABLE_ANY_H263) { |
5278 | 3348 c->h263_v_loop_filter= h263_v_loop_filter_mmx; |
3349 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
|
3350 } |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
3351 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
|
3352 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
|
3353 c->put_no_rnd_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_mmx_nornd; |
2967 | 3354 |
3173 | 3355 c->h264_idct_dc_add= |
3356 c->h264_idct_add= ff_h264_idct_add_mmx; | |
3174 | 3357 c->h264_idct8_dc_add= |
3358 c->h264_idct8_add= ff_h264_idct8_add_mmx; | |
6320 | 3359 if (mm_flags & MM_SSE2) |
3360 c->h264_idct8_add= ff_h264_idct8_add_sse2; | |
3173 | 3361 |
4197 | 3362 if (mm_flags & MM_MMXEXT) { |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3363 c->prefetch = prefetch_mmx2; |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3364 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3365 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
|
3366 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2; |
651 | 3367 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3368 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
|
3369 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
|
3370 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2; |
415 | 3371 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3372 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
|
3373 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2; |
651 | 3374 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3375 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
|
3376 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
|
3377 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2; |
1092 | 3378 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3379 #ifdef CONFIG_ENCODERS |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
3380 c->sum_abs_dctelem= sum_abs_dctelem_mmx2; |
1153 | 3381 c->hadamard8_diff[0]= hadamard8_diff16_mmx2; |
3382 c->hadamard8_diff[1]= hadamard8_diff_mmx2; | |
1729 | 3383 c->vsad[4]= vsad_intra16_mmx2; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3384 #endif //CONFIG_ENCODERS |
1153 | 3385 |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3089
diff
changeset
|
3386 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
|
3387 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2; |
2745 | 3388 |
1092 | 3389 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
3390 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2; | |
3391 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2; | |
3392 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2; | |
3393 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2; | |
3394 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2; | |
3395 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2; | |
1772
8cd5257195c9
vsad16_mmx2 only applies if encoders are turned on
melanson
parents:
1765
diff
changeset
|
3396 #ifdef CONFIG_ENCODERS |
1729 | 3397 c->vsad[0] = vsad16_mmx2; |
1772
8cd5257195c9
vsad16_mmx2 only applies if encoders are turned on
melanson
parents:
1765
diff
changeset
|
3398 #endif //CONFIG_ENCODERS |
1092 | 3399 } |
959 | 3400 |
6327 | 3401 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU) \ |
3402 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## SIZE ## _mc00_ ## CPU; \ | |
3403 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## SIZE ## _mc10_ ## CPU; \ | |
3404 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## SIZE ## _mc20_ ## CPU; \ | |
3405 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## SIZE ## _mc30_ ## CPU; \ | |
3406 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## SIZE ## _mc01_ ## CPU; \ | |
3407 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## SIZE ## _mc11_ ## CPU; \ | |
3408 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## SIZE ## _mc21_ ## CPU; \ | |
3409 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## SIZE ## _mc31_ ## CPU; \ | |
3410 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## SIZE ## _mc02_ ## CPU; \ | |
3411 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## SIZE ## _mc12_ ## CPU; \ | |
3412 c->PFX ## _pixels_tab[IDX][10] = PFX ## SIZE ## _mc22_ ## CPU; \ | |
3413 c->PFX ## _pixels_tab[IDX][11] = PFX ## SIZE ## _mc32_ ## CPU; \ | |
3414 c->PFX ## _pixels_tab[IDX][12] = PFX ## SIZE ## _mc03_ ## CPU; \ | |
3415 c->PFX ## _pixels_tab[IDX][13] = PFX ## SIZE ## _mc13_ ## CPU; \ | |
3416 c->PFX ## _pixels_tab[IDX][14] = PFX ## SIZE ## _mc23_ ## CPU; \ | |
3417 c->PFX ## _pixels_tab[IDX][15] = PFX ## SIZE ## _mc33_ ## CPU | |
3418 | |
3419 SET_QPEL_FUNCS(put_qpel, 0, 16, mmx2); | |
3420 SET_QPEL_FUNCS(put_qpel, 1, 8, mmx2); | |
3421 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmx2); | |
3422 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, mmx2); | |
3423 SET_QPEL_FUNCS(avg_qpel, 0, 16, mmx2); | |
3424 SET_QPEL_FUNCS(avg_qpel, 1, 8, mmx2); | |
3425 | |
3426 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, mmx2); | |
3427 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, mmx2); | |
3428 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, mmx2); | |
3429 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, mmx2); | |
3430 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, mmx2); | |
3431 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, mmx2); | |
3432 | |
3433 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, mmx2); | |
3434 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, mmx2); | |
3435 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, mmx2); | |
3436 SET_QPEL_FUNCS(avg_2tap_qpel, 1, 8, mmx2); | |
2209 | 3437 |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
3438 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_mmx2_rnd; |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
3439 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_mmx2; |
3213 | 3440 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_mmx2; |
3441 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_mmx2; | |
2633 | 3442 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_mmx2; |
3443 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_mmx2; | |
3444 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_mmx2; | |
3445 c->h264_h_loop_filter_chroma= h264_h_loop_filter_chroma_mmx2; | |
2707
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
3446 c->h264_v_loop_filter_chroma_intra= h264_v_loop_filter_chroma_intra_mmx2; |
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
3447 c->h264_h_loop_filter_chroma_intra= h264_h_loop_filter_chroma_intra_mmx2; |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3576
diff
changeset
|
3448 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2; |
2633 | 3449 |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3450 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3451 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3452 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3453 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3454 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3455 c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3456 c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3457 c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3458 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3459 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3460 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3461 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3462 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3463 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3464 c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3465 c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3466 c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3467 |
5949 | 3468 if (ENABLE_CAVS_DECODER) |
5950 | 3469 ff_cavsdsp_init_mmx2(c, avctx); |
5949 | 3470 |
3471 if (ENABLE_VC1_DECODER || ENABLE_WMV3_DECODER) | |
5950 | 3472 ff_vc1dsp_init_mmx(c, avctx); |
5933 | 3473 |
1686
68abbec33289
Here are juste two added #ifdef CONFIG_ENCODERS to allow
michael
parents:
1648
diff
changeset
|
3474 #ifdef CONFIG_ENCODERS |
1527 | 3475 c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_mmx2; |
1686
68abbec33289
Here are juste two added #ifdef CONFIG_ENCODERS to allow
michael
parents:
1648
diff
changeset
|
3476 #endif //CONFIG_ENCODERS |
4197 | 3477 } else if (mm_flags & MM_3DNOW) { |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3478 c->prefetch = prefetch_3dnow; |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3479 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3480 c->put_pixels_tab[0][1] = put_pixels16_x2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3481 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow; |
393 | 3482 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3483 c->avg_pixels_tab[0][0] = avg_pixels16_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3484 c->avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3485 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow; |
651 | 3486 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3487 c->put_pixels_tab[1][1] = put_pixels8_x2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3488 c->put_pixels_tab[1][2] = put_pixels8_y2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3489 |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3490 c->avg_pixels_tab[1][0] = avg_pixels8_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3491 c->avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3492 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow; |
1092 | 3493 |
3494 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
3495 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow; | |
3496 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow; | |
3497 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow; | |
3498 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow; | |
3499 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow; | |
3500 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow; | |
3501 } | |
984 | 3502 |
6327 | 3503 SET_QPEL_FUNCS(put_qpel, 0, 16, 3dnow); |
3504 SET_QPEL_FUNCS(put_qpel, 1, 8, 3dnow); | |
3505 SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, 3dnow); | |
3506 SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, 3dnow); | |
3507 SET_QPEL_FUNCS(avg_qpel, 0, 16, 3dnow); | |
3508 SET_QPEL_FUNCS(avg_qpel, 1, 8, 3dnow); | |
3509 | |
3510 SET_QPEL_FUNCS(put_h264_qpel, 0, 16, 3dnow); | |
3511 SET_QPEL_FUNCS(put_h264_qpel, 1, 8, 3dnow); | |
3512 SET_QPEL_FUNCS(put_h264_qpel, 2, 4, 3dnow); | |
3513 SET_QPEL_FUNCS(avg_h264_qpel, 0, 16, 3dnow); | |
3514 SET_QPEL_FUNCS(avg_h264_qpel, 1, 8, 3dnow); | |
3515 SET_QPEL_FUNCS(avg_h264_qpel, 2, 4, 3dnow); | |
3516 | |
3517 SET_QPEL_FUNCS(put_2tap_qpel, 0, 16, 3dnow); | |
3518 SET_QPEL_FUNCS(put_2tap_qpel, 1, 8, 3dnow); | |
3519 SET_QPEL_FUNCS(avg_2tap_qpel, 0, 16, 3dnow); | |
3520 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
|
3521 |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
6056
diff
changeset
|
3522 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
|
3523 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_3dnow; |
0 | 3524 } |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3525 |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
3526 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
3527 #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
|
3528 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
|
3529 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
|
3530 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
|
3531 c->avg_h264_qpel_pixels_tab[1][x+y*4] = avg_h264_qpel8_mc##x##y##_##CPU; |
6331 | 3532 if((mm_flags & MM_SSE2) && !(mm_flags & MM_3DNOW)){ |
3533 // 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
|
3534 /* FIXME works in most codecs, but crashes svq1 due to unaligned chroma |
6331 | 3535 c->put_pixels_tab[0][0] = put_pixels16_sse2; |
3536 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
|
3537 */ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
3538 H264_QPEL_FUNCS(0, 0, sse2); |
6331 | 3539 } |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
3540 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
|
3541 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
|
3542 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
|
3543 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
|
3544 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
|
3545 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
|
3546 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
|
3547 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
|
3548 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
|
3549 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
|
3550 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
|
3551 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
|
3552 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
|
3553 } |
6331 | 3554 #ifdef HAVE_SSSE3 |
3555 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
|
3556 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
|
3557 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
|
3558 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
|
3559 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
|
3560 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
|
3561 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
|
3562 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
|
3563 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
|
3564 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
|
3565 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
|
3566 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
|
3567 H264_QPEL_FUNCS(3, 3, ssse3); |
6331 | 3568 } |
3569 #endif | |
3570 | |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3571 #ifdef CONFIG_ENCODERS |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3572 if(mm_flags & MM_SSE2){ |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
3573 c->sum_abs_dctelem= sum_abs_dctelem_sse2; |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3574 c->hadamard8_diff[0]= hadamard8_diff16_sse2; |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3575 c->hadamard8_diff[1]= hadamard8_diff_sse2; |
6030 | 3576 if (ENABLE_FLAC_ENCODER) |
3577 c->flac_compute_autocorr = ff_flac_compute_autocorr_sse2; | |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3578 } |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3579 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3580 #ifdef HAVE_SSSE3 |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3581 if(mm_flags & MM_SSSE3){ |
5024
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3582 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3583 c->try_8x8basis= try_8x8basis_ssse3; |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3584 } |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3585 c->add_8x8basis= add_8x8basis_ssse3; |
4988
689490842cf5
factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents:
4987
diff
changeset
|
3586 c->sum_abs_dctelem= sum_abs_dctelem_ssse3; |
4987
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3587 c->hadamard8_diff[0]= hadamard8_diff16_ssse3; |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3588 c->hadamard8_diff[1]= hadamard8_diff_ssse3; |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3589 } |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3590 #endif |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3591 #endif |
02199b094850
sse2 & ssse3 versions of hadamard. unroll and inline diff_pixels.
lorenm
parents:
4946
diff
changeset
|
3592 |
4589
30261f4ed12d
Fix wrong conditional, Snow decoding, not encoding, was SIMD-accelerated.
diego
parents:
4436
diff
changeset
|
3593 #ifdef CONFIG_SNOW_DECODER |
5591 | 3594 if(mm_flags & MM_SSE2 & 0){ |
3210 | 3595 c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2; |
5602 | 3596 #ifdef HAVE_7REGS |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3597 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
|
3598 #endif |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
3599 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
|
3600 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3601 else{ |
5594
384629ebcb93
avoid overflow in the 3rd lifting step, this now needs mmx2 at minimum
michael
parents:
5591
diff
changeset
|
3602 if(mm_flags & MM_MMXEXT){ |
3210 | 3603 c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx; |
5602 | 3604 #ifdef HAVE_7REGS |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3605 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
|
3606 #endif |
5594
384629ebcb93
avoid overflow in the 3rd lifting step, this now needs mmx2 at minimum
michael
parents:
5591
diff
changeset
|
3607 } |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
3608 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
|
3609 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3610 #endif |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
3611 |
4197 | 3612 if(mm_flags & MM_3DNOW){ |
5024
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3613 #ifdef CONFIG_ENCODERS |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3614 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3615 c->try_8x8basis= try_8x8basis_3dnow; |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3616 } |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3617 c->add_8x8basis= add_8x8basis_3dnow; |
8a3bc96c366f
3DNow! and SSSE3 optimization to QNS DSP functions; use pmulhrw/pmulhrsw instead of pmulhw
zuxy
parents:
5014
diff
changeset
|
3618 #endif //CONFIG_ENCODERS |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3619 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
|
3620 c->vector_fmul = vector_fmul_3dnow; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3621 if(!(avctx->flags & CODEC_FLAG_BITEXACT)) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3622 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
|
3623 } |
4197 | 3624 if(mm_flags & MM_3DNOWEXT) |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3625 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2; |
4197 | 3626 if(mm_flags & MM_SSE){ |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
3627 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
|
3628 c->vector_fmul = vector_fmul_sse; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3629 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
|
3630 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
|
3631 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
|
3632 } |
4197 | 3633 if(mm_flags & MM_3DNOW) |
3574 | 3634 c->vector_fmul_add_add = vector_fmul_add_add_3dnow; // faster than sse |
0 | 3635 } |
2967 | 3636 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3637 #ifdef CONFIG_ENCODERS |
1092 | 3638 dsputil_init_pix_mmx(c, avctx); |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3639 #endif //CONFIG_ENCODERS |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3640 #if 0 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3641 // for speed testing |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3642 get_pixels = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3643 put_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3644 add_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3645 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3646 pix_abs16x16 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3647 pix_abs16x16_x2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3648 pix_abs16x16_y2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3649 pix_abs16x16_xy2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3650 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3651 put_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3652 put_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3653 put_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3654 put_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3655 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3656 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
|
3657 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
|
3658 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
|
3659 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
|
3660 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3661 avg_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3662 avg_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3663 avg_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3664 avg_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3665 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3666 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
|
3667 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
|
3668 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
|
3669 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
|
3670 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3671 //av_fdct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3672 //ff_idct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3673 #endif |
0 | 3674 } |