Mercurial > libavcodec.hg
annotate i386/dsputil_mmx.c @ 3645:47821be55b6c libavcodec
mmx implementation of deblocking strength decision.
2-3% faster h264.
author | lorenm |
---|---|
date | Mon, 28 Aug 2006 09:33:01 +0000 |
parents | f7125bf10892 |
children | faa8dc533ad3 |
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 * |
429 | 6 * This library is free software; you can redistribute it and/or |
7 * modify it under the terms of the GNU Lesser General Public | |
8 * License as published by the Free Software Foundation; either | |
9 * version 2 of the License, or (at your option) any later version. | |
0 | 10 * |
429 | 11 * This library is distributed in the hope that it will be useful, |
0 | 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
429 | 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
14 * Lesser General Public License for more details. | |
0 | 15 * |
429 | 16 * You should have received a copy of the GNU Lesser General Public |
17 * License along with this library; if not, write to the Free Software | |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
0 | 19 * |
20 * MMX optimization by Nick Kurshev <nickols_k@mail.ru> | |
21 */ | |
22 | |
23 #include "../dsputil.h" | |
1092 | 24 #include "../simple_idct.h" |
2067 | 25 #include "../mpegvideo.h" |
3398
e0927bc44a10
Move REG_* macros from libavcodec/i386/mmx.h to libavutil/x86_cpu.h
lucabe
parents:
3250
diff
changeset
|
26 #include "x86_cpu.h" |
1984
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
27 #include "mmx.h" |
0 | 28 |
1729 | 29 //#undef NDEBUG |
30 //#include <assert.h> | |
31 | |
1647 | 32 extern const uint8_t ff_h263_loop_filter_strength[32]; |
2868 | 33 extern void ff_idct_xvid_mmx(short *block); |
34 extern void ff_idct_xvid_mmx2(short *block); | |
1647 | 35 |
5 | 36 int mm_flags; /* multimedia extension flags */ |
936 | 37 |
0 | 38 /* pixel operations */ |
1845
3054613980a8
attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents:
1784
diff
changeset
|
39 static const uint64_t mm_bone attribute_used __attribute__ ((aligned(8))) = 0x0101010101010101ULL; |
3054613980a8
attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents:
1784
diff
changeset
|
40 static const uint64_t mm_wone attribute_used __attribute__ ((aligned(8))) = 0x0001000100010001ULL; |
3054613980a8
attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents:
1784
diff
changeset
|
41 static const uint64_t mm_wtwo attribute_used __attribute__ ((aligned(8))) = 0x0002000200020002ULL; |
0 | 42 |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
43 static const uint64_t ff_pdw_80000000[2] attribute_used __attribute__ ((aligned(16))) = |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
44 {0x8000000080000000ULL, 0x8000000080000000ULL}; |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
45 |
1845
3054613980a8
attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents:
1784
diff
changeset
|
46 static const uint64_t ff_pw_20 attribute_used __attribute__ ((aligned(8))) = 0x0014001400140014ULL; |
3054613980a8
attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents:
1784
diff
changeset
|
47 static const uint64_t ff_pw_3 attribute_used __attribute__ ((aligned(8))) = 0x0003000300030003ULL; |
2633 | 48 static const uint64_t ff_pw_4 attribute_used __attribute__ ((aligned(8))) = 0x0004000400040004ULL; |
2209 | 49 static const uint64_t ff_pw_5 attribute_used __attribute__ ((aligned(8))) = 0x0005000500050005ULL; |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
50 static const uint64_t ff_pw_8 attribute_used __attribute__ ((aligned(8))) = 0x0008000800080008ULL; |
1845
3054613980a8
attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents:
1784
diff
changeset
|
51 static const uint64_t ff_pw_16 attribute_used __attribute__ ((aligned(8))) = 0x0010001000100010ULL; |
2209 | 52 static const uint64_t ff_pw_32 attribute_used __attribute__ ((aligned(8))) = 0x0020002000200020ULL; |
2754 | 53 static const uint64_t ff_pw_64 attribute_used __attribute__ ((aligned(8))) = 0x0040004000400040ULL; |
1845
3054613980a8
attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents:
1784
diff
changeset
|
54 static const uint64_t ff_pw_15 attribute_used __attribute__ ((aligned(8))) = 0x000F000F000F000FULL; |
954 | 55 |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3576
diff
changeset
|
56 static const uint64_t ff_pb_1 attribute_used __attribute__ ((aligned(8))) = 0x0101010101010101ULL; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3576
diff
changeset
|
57 static const uint64_t ff_pb_3 attribute_used __attribute__ ((aligned(8))) = 0x0303030303030303ULL; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3576
diff
changeset
|
58 static const uint64_t ff_pb_7 attribute_used __attribute__ ((aligned(8))) = 0x0707070707070707ULL; |
2707
360024d31dab
H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents:
2696
diff
changeset
|
59 static const uint64_t ff_pb_3F attribute_used __attribute__ ((aligned(8))) = 0x3F3F3F3F3F3F3F3FULL; |
1845
3054613980a8
attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents:
1784
diff
changeset
|
60 static const uint64_t ff_pb_FC attribute_used __attribute__ ((aligned(8))) = 0xFCFCFCFCFCFCFCFCULL; |
1647 | 61 |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
62 #define JUMPALIGN() __asm __volatile (ASMALIGN(3)::) |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
63 #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
|
64 |
448 | 65 #define MOVQ_WONE(regd) \ |
66 __asm __volatile ( \ | |
67 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ | |
68 "psrlw $15, %%" #regd ::) | |
69 | |
70 #define MOVQ_BFE(regd) \ | |
71 __asm __volatile ( \ | |
72 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\ | |
73 "paddb %%" #regd ", %%" #regd " \n\t" ::) | |
74 | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
75 #ifndef PIC |
448 | 76 #define MOVQ_BONE(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(mm_bone)) |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
77 #define MOVQ_WTWO(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(mm_wtwo)) |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
78 #else |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
79 // 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
|
80 // pcmpeqd -> -1 |
448 | 81 #define MOVQ_BONE(regd) \ |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
82 __asm __volatile ( \ |
448 | 83 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
84 "psrlw $15, %%" #regd " \n\t" \ | |
85 "packuswb %%" #regd ", %%" #regd " \n\t" ::) | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
86 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
87 #define MOVQ_WTWO(regd) \ |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
88 __asm __volatile ( \ |
448 | 89 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
90 "psrlw $15, %%" #regd " \n\t" \ | |
91 "psllw $1, %%" #regd " \n\t"::) | |
387 | 92 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
93 #endif |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
94 |
448 | 95 // 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
|
96 // first argument is unmodifed and second is trashed |
471 | 97 // regfe is supposed to contain 0xfefefefefefefefe |
98 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \ | |
2979 | 99 "movq " #rega ", " #regr " \n\t"\ |
100 "pand " #regb ", " #regr " \n\t"\ | |
101 "pxor " #rega ", " #regb " \n\t"\ | |
102 "pand " #regfe "," #regb " \n\t"\ | |
103 "psrlq $1, " #regb " \n\t"\ | |
104 "paddb " #regb ", " #regr " \n\t" | |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
105 |
471 | 106 #define PAVGB_MMX(rega, regb, regr, regfe) \ |
2979 | 107 "movq " #rega ", " #regr " \n\t"\ |
108 "por " #regb ", " #regr " \n\t"\ | |
109 "pxor " #rega ", " #regb " \n\t"\ | |
110 "pand " #regfe "," #regb " \n\t"\ | |
111 "psrlq $1, " #regb " \n\t"\ | |
112 "psubb " #regb ", " #regr " \n\t" | |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
113 |
471 | 114 // mm6 is supposed to contain 0xfefefefefefefefe |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
115 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \ |
2979 | 116 "movq " #rega ", " #regr " \n\t"\ |
117 "movq " #regc ", " #regp " \n\t"\ | |
118 "pand " #regb ", " #regr " \n\t"\ | |
119 "pand " #regd ", " #regp " \n\t"\ | |
120 "pxor " #rega ", " #regb " \n\t"\ | |
121 "pxor " #regc ", " #regd " \n\t"\ | |
122 "pand %%mm6, " #regb " \n\t"\ | |
123 "pand %%mm6, " #regd " \n\t"\ | |
124 "psrlq $1, " #regb " \n\t"\ | |
125 "psrlq $1, " #regd " \n\t"\ | |
126 "paddb " #regb ", " #regr " \n\t"\ | |
127 "paddb " #regd ", " #regp " \n\t" | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
128 |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
129 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \ |
2979 | 130 "movq " #rega ", " #regr " \n\t"\ |
131 "movq " #regc ", " #regp " \n\t"\ | |
132 "por " #regb ", " #regr " \n\t"\ | |
133 "por " #regd ", " #regp " \n\t"\ | |
134 "pxor " #rega ", " #regb " \n\t"\ | |
135 "pxor " #regc ", " #regd " \n\t"\ | |
136 "pand %%mm6, " #regb " \n\t"\ | |
137 "pand %%mm6, " #regd " \n\t"\ | |
138 "psrlq $1, " #regd " \n\t"\ | |
139 "psrlq $1, " #regb " \n\t"\ | |
140 "psubb " #regb ", " #regr " \n\t"\ | |
141 "psubb " #regd ", " #regp " \n\t" | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
142 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
143 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
144 /* MMX no rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
145 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx |
448 | 146 #define SET_RND MOVQ_WONE |
2979 | 147 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX_NO_RND(a, b, c, d, e, f) |
148 #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
|
149 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
150 #include "dsputil_mmx_rnd.h" |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
151 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
152 #undef DEF |
448 | 153 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
154 #undef PAVGBP |
471 | 155 #undef PAVGB |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
156 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
157 /* MMX rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
158 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
159 #define DEF(x, y) x ## _ ## y ##_mmx |
448 | 160 #define SET_RND MOVQ_WTWO |
2979 | 161 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f) |
162 #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
|
163 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
164 #include "dsputil_mmx_rnd.h" |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
165 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
166 #undef DEF |
448 | 167 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
168 #undef PAVGBP |
471 | 169 #undef PAVGB |
387 | 170 |
0 | 171 /***********************************/ |
172 /* 3Dnow specific */ | |
173 | |
174 #define DEF(x) x ## _3dnow | |
175 /* for Athlons PAVGUSB is prefered */ | |
176 #define PAVGB "pavgusb" | |
177 | |
178 #include "dsputil_mmx_avg.h" | |
179 | |
180 #undef DEF | |
181 #undef PAVGB | |
182 | |
183 /***********************************/ | |
184 /* MMX2 specific */ | |
185 | |
386 | 186 #define DEF(x) x ## _mmx2 |
0 | 187 |
188 /* Introduced only in MMX2 set */ | |
189 #define PAVGB "pavgb" | |
190 | |
191 #include "dsputil_mmx_avg.h" | |
192 | |
193 #undef DEF | |
194 #undef PAVGB | |
195 | |
3416 | 196 #define SBUTTERFLY(a,b,t,n)\ |
197 "movq " #a ", " #t " \n\t" /* abcd */\ | |
198 "punpckl" #n " " #b ", " #a " \n\t" /* aebf */\ | |
199 "punpckh" #n " " #b ", " #t " \n\t" /* cgdh */\ | |
200 | |
0 | 201 /***********************************/ |
202 /* standard MMX */ | |
203 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
204 #ifdef CONFIG_ENCODERS |
1064 | 205 static void get_pixels_mmx(DCTELEM *block, const uint8_t *pixels, int line_size) |
0 | 206 { |
386 | 207 asm volatile( |
2979 | 208 "mov $-128, %%"REG_a" \n\t" |
209 "pxor %%mm7, %%mm7 \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
210 ASMALIGN(4) |
2979 | 211 "1: \n\t" |
212 "movq (%0), %%mm0 \n\t" | |
213 "movq (%0, %2), %%mm2 \n\t" | |
214 "movq %%mm0, %%mm1 \n\t" | |
215 "movq %%mm2, %%mm3 \n\t" | |
216 "punpcklbw %%mm7, %%mm0 \n\t" | |
217 "punpckhbw %%mm7, %%mm1 \n\t" | |
218 "punpcklbw %%mm7, %%mm2 \n\t" | |
219 "punpckhbw %%mm7, %%mm3 \n\t" | |
220 "movq %%mm0, (%1, %%"REG_a") \n\t" | |
221 "movq %%mm1, 8(%1, %%"REG_a") \n\t" | |
222 "movq %%mm2, 16(%1, %%"REG_a") \n\t" | |
223 "movq %%mm3, 24(%1, %%"REG_a") \n\t" | |
224 "add %3, %0 \n\t" | |
225 "add $32, %%"REG_a" \n\t" | |
226 "js 1b \n\t" | |
386 | 227 : "+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
|
228 : "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
|
229 : "%"REG_a |
386 | 230 ); |
0 | 231 } |
232 | |
1064 | 233 static inline void diff_pixels_mmx(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride) |
324 | 234 { |
235 asm volatile( | |
2979 | 236 "pxor %%mm7, %%mm7 \n\t" |
237 "mov $-128, %%"REG_a" \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
238 ASMALIGN(4) |
2979 | 239 "1: \n\t" |
240 "movq (%0), %%mm0 \n\t" | |
241 "movq (%1), %%mm2 \n\t" | |
242 "movq %%mm0, %%mm1 \n\t" | |
243 "movq %%mm2, %%mm3 \n\t" | |
244 "punpcklbw %%mm7, %%mm0 \n\t" | |
245 "punpckhbw %%mm7, %%mm1 \n\t" | |
246 "punpcklbw %%mm7, %%mm2 \n\t" | |
247 "punpckhbw %%mm7, %%mm3 \n\t" | |
248 "psubw %%mm2, %%mm0 \n\t" | |
249 "psubw %%mm3, %%mm1 \n\t" | |
250 "movq %%mm0, (%2, %%"REG_a") \n\t" | |
251 "movq %%mm1, 8(%2, %%"REG_a") \n\t" | |
252 "add %3, %0 \n\t" | |
253 "add %3, %1 \n\t" | |
254 "add $16, %%"REG_a" \n\t" | |
255 "jnz 1b \n\t" | |
324 | 256 : "+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
|
257 : "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
|
258 : "%"REG_a |
324 | 259 ); |
260 } | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
261 #endif //CONFIG_ENCODERS |
324 | 262 |
1064 | 263 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 264 { |
265 const DCTELEM *p; | |
1064 | 266 uint8_t *pix; |
0 | 267 |
268 /* read the pixels */ | |
269 p = block; | |
270 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
271 /* unrolled loop */ |
2979 | 272 __asm __volatile( |
273 "movq %3, %%mm0 \n\t" | |
274 "movq 8%3, %%mm1 \n\t" | |
275 "movq 16%3, %%mm2 \n\t" | |
276 "movq 24%3, %%mm3 \n\t" | |
277 "movq 32%3, %%mm4 \n\t" | |
278 "movq 40%3, %%mm5 \n\t" | |
279 "movq 48%3, %%mm6 \n\t" | |
280 "movq 56%3, %%mm7 \n\t" | |
281 "packuswb %%mm1, %%mm0 \n\t" | |
282 "packuswb %%mm3, %%mm2 \n\t" | |
283 "packuswb %%mm5, %%mm4 \n\t" | |
284 "packuswb %%mm7, %%mm6 \n\t" | |
285 "movq %%mm0, (%0) \n\t" | |
286 "movq %%mm2, (%0, %1) \n\t" | |
287 "movq %%mm4, (%0, %1, 2) \n\t" | |
288 "movq %%mm6, (%0, %2) \n\t" | |
289 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "m"(*p) | |
290 :"memory"); | |
0 | 291 pix += line_size*4; |
292 p += 32; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
293 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
294 // 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
|
295 // compiler would generate some very strange code |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
296 // thus using "r" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
297 __asm __volatile( |
2979 | 298 "movq (%3), %%mm0 \n\t" |
299 "movq 8(%3), %%mm1 \n\t" | |
300 "movq 16(%3), %%mm2 \n\t" | |
301 "movq 24(%3), %%mm3 \n\t" | |
302 "movq 32(%3), %%mm4 \n\t" | |
303 "movq 40(%3), %%mm5 \n\t" | |
304 "movq 48(%3), %%mm6 \n\t" | |
305 "movq 56(%3), %%mm7 \n\t" | |
306 "packuswb %%mm1, %%mm0 \n\t" | |
307 "packuswb %%mm3, %%mm2 \n\t" | |
308 "packuswb %%mm5, %%mm4 \n\t" | |
309 "packuswb %%mm7, %%mm6 \n\t" | |
310 "movq %%mm0, (%0) \n\t" | |
311 "movq %%mm2, (%0, %1) \n\t" | |
312 "movq %%mm4, (%0, %1, 2) \n\t" | |
313 "movq %%mm6, (%0, %2) \n\t" | |
314 ::"r" (pix), "r" ((long)line_size), "r" ((long)line_size*3), "r"(p) | |
315 :"memory"); | |
0 | 316 } |
317 | |
3089 | 318 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
|
319 { 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
|
320 |
1984
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
321 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
|
322 { |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
323 int i; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
324 |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
325 movq_m2r(*vector128, mm1); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
326 for (i = 0; i < 8; i++) { |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
327 movq_m2r(*(block), mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
328 packsswb_m2r(*(block + 4), mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
329 block += 8; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
330 paddb_r2r(mm1, mm0); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
331 movq_r2m(mm0, *pixels); |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
332 pixels += line_size; |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
333 } |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
334 } |
ef919e9ef73e
separate out put_signed_pixels_clamped() into its own function and
melanson
parents:
1977
diff
changeset
|
335 |
1064 | 336 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 337 { |
338 const DCTELEM *p; | |
1064 | 339 uint8_t *pix; |
0 | 340 int i; |
341 | |
342 /* read the pixels */ | |
343 p = block; | |
344 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
345 MOVQ_ZERO(mm7); |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
346 i = 4; |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
347 do { |
2979 | 348 __asm __volatile( |
349 "movq (%2), %%mm0 \n\t" | |
350 "movq 8(%2), %%mm1 \n\t" | |
351 "movq 16(%2), %%mm2 \n\t" | |
352 "movq 24(%2), %%mm3 \n\t" | |
353 "movq %0, %%mm4 \n\t" | |
354 "movq %1, %%mm6 \n\t" | |
355 "movq %%mm4, %%mm5 \n\t" | |
356 "punpcklbw %%mm7, %%mm4 \n\t" | |
357 "punpckhbw %%mm7, %%mm5 \n\t" | |
358 "paddsw %%mm4, %%mm0 \n\t" | |
359 "paddsw %%mm5, %%mm1 \n\t" | |
360 "movq %%mm6, %%mm5 \n\t" | |
361 "punpcklbw %%mm7, %%mm6 \n\t" | |
362 "punpckhbw %%mm7, %%mm5 \n\t" | |
363 "paddsw %%mm6, %%mm2 \n\t" | |
364 "paddsw %%mm5, %%mm3 \n\t" | |
365 "packuswb %%mm1, %%mm0 \n\t" | |
366 "packuswb %%mm3, %%mm2 \n\t" | |
367 "movq %%mm0, %0 \n\t" | |
368 "movq %%mm2, %1 \n\t" | |
369 :"+m"(*pix), "+m"(*(pix+line_size)) | |
370 :"r"(p) | |
371 :"memory"); | |
0 | 372 pix += line_size*2; |
373 p += 16; | |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
374 } while (--i); |
0 | 375 } |
376 | |
2209 | 377 static void put_pixels4_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
378 { | |
379 __asm __volatile( | |
2979 | 380 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
381 ASMALIGN(3) |
2979 | 382 "1: \n\t" |
383 "movd (%1), %%mm0 \n\t" | |
384 "movd (%1, %3), %%mm1 \n\t" | |
385 "movd %%mm0, (%2) \n\t" | |
386 "movd %%mm1, (%2, %3) \n\t" | |
387 "add %%"REG_a", %1 \n\t" | |
388 "add %%"REG_a", %2 \n\t" | |
389 "movd (%1), %%mm0 \n\t" | |
390 "movd (%1, %3), %%mm1 \n\t" | |
391 "movd %%mm0, (%2) \n\t" | |
392 "movd %%mm1, (%2, %3) \n\t" | |
393 "add %%"REG_a", %1 \n\t" | |
394 "add %%"REG_a", %2 \n\t" | |
395 "subl $4, %0 \n\t" | |
396 "jnz 1b \n\t" | |
397 : "+g"(h), "+r" (pixels), "+r" (block) | |
398 : "r"((long)line_size) | |
399 : "%"REG_a, "memory" | |
400 ); | |
2209 | 401 } |
402 | |
1064 | 403 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 404 { |
471 | 405 __asm __volatile( |
2979 | 406 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
407 ASMALIGN(3) |
2979 | 408 "1: \n\t" |
409 "movq (%1), %%mm0 \n\t" | |
410 "movq (%1, %3), %%mm1 \n\t" | |
411 "movq %%mm0, (%2) \n\t" | |
412 "movq %%mm1, (%2, %3) \n\t" | |
413 "add %%"REG_a", %1 \n\t" | |
414 "add %%"REG_a", %2 \n\t" | |
415 "movq (%1), %%mm0 \n\t" | |
416 "movq (%1, %3), %%mm1 \n\t" | |
417 "movq %%mm0, (%2) \n\t" | |
418 "movq %%mm1, (%2, %3) \n\t" | |
419 "add %%"REG_a", %1 \n\t" | |
420 "add %%"REG_a", %2 \n\t" | |
421 "subl $4, %0 \n\t" | |
422 "jnz 1b \n\t" | |
423 : "+g"(h), "+r" (pixels), "+r" (block) | |
424 : "r"((long)line_size) | |
425 : "%"REG_a, "memory" | |
426 ); | |
0 | 427 } |
428 | |
1064 | 429 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
651 | 430 { |
431 __asm __volatile( | |
2979 | 432 "lea (%3, %3), %%"REG_a" \n\t" |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3574
diff
changeset
|
433 ASMALIGN(3) |
2979 | 434 "1: \n\t" |
435 "movq (%1), %%mm0 \n\t" | |
436 "movq 8(%1), %%mm4 \n\t" | |
437 "movq (%1, %3), %%mm1 \n\t" | |
438 "movq 8(%1, %3), %%mm5 \n\t" | |
439 "movq %%mm0, (%2) \n\t" | |
440 "movq %%mm4, 8(%2) \n\t" | |
441 "movq %%mm1, (%2, %3) \n\t" | |
442 "movq %%mm5, 8(%2, %3) \n\t" | |
443 "add %%"REG_a", %1 \n\t" | |
444 "add %%"REG_a", %2 \n\t" | |
445 "movq (%1), %%mm0 \n\t" | |
446 "movq 8(%1), %%mm4 \n\t" | |
447 "movq (%1, %3), %%mm1 \n\t" | |
448 "movq 8(%1, %3), %%mm5 \n\t" | |
449 "movq %%mm0, (%2) \n\t" | |
450 "movq %%mm4, 8(%2) \n\t" | |
451 "movq %%mm1, (%2, %3) \n\t" | |
452 "movq %%mm5, 8(%2, %3) \n\t" | |
453 "add %%"REG_a", %1 \n\t" | |
454 "add %%"REG_a", %2 \n\t" | |
455 "subl $4, %0 \n\t" | |
456 "jnz 1b \n\t" | |
457 : "+g"(h), "+r" (pixels), "+r" (block) | |
458 : "r"((long)line_size) | |
459 : "%"REG_a, "memory" | |
460 ); | |
651 | 461 } |
462 | |
296 | 463 static void clear_blocks_mmx(DCTELEM *blocks) |
464 { | |
471 | 465 __asm __volatile( |
2979 | 466 "pxor %%mm7, %%mm7 \n\t" |
467 "mov $-128*6, %%"REG_a" \n\t" | |
468 "1: \n\t" | |
469 "movq %%mm7, (%0, %%"REG_a") \n\t" | |
470 "movq %%mm7, 8(%0, %%"REG_a") \n\t" | |
471 "movq %%mm7, 16(%0, %%"REG_a") \n\t" | |
472 "movq %%mm7, 24(%0, %%"REG_a") \n\t" | |
473 "add $32, %%"REG_a" \n\t" | |
474 " 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
|
475 : : "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
|
476 : "%"REG_a |
296 | 477 ); |
478 } | |
479 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
480 #ifdef CONFIG_ENCODERS |
1064 | 481 static int pix_sum16_mmx(uint8_t * pix, int line_size){ |
688 | 482 const int h=16; |
483 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
|
484 long index= -line_size*h; |
688 | 485 |
486 __asm __volatile( | |
2979 | 487 "pxor %%mm7, %%mm7 \n\t" |
488 "pxor %%mm6, %%mm6 \n\t" | |
489 "1: \n\t" | |
490 "movq (%2, %1), %%mm0 \n\t" | |
491 "movq (%2, %1), %%mm1 \n\t" | |
492 "movq 8(%2, %1), %%mm2 \n\t" | |
493 "movq 8(%2, %1), %%mm3 \n\t" | |
494 "punpcklbw %%mm7, %%mm0 \n\t" | |
495 "punpckhbw %%mm7, %%mm1 \n\t" | |
496 "punpcklbw %%mm7, %%mm2 \n\t" | |
497 "punpckhbw %%mm7, %%mm3 \n\t" | |
498 "paddw %%mm0, %%mm1 \n\t" | |
499 "paddw %%mm2, %%mm3 \n\t" | |
500 "paddw %%mm1, %%mm3 \n\t" | |
501 "paddw %%mm3, %%mm6 \n\t" | |
502 "add %3, %1 \n\t" | |
503 " js 1b \n\t" | |
504 "movq %%mm6, %%mm5 \n\t" | |
505 "psrlq $32, %%mm6 \n\t" | |
506 "paddw %%mm5, %%mm6 \n\t" | |
507 "movq %%mm6, %%mm5 \n\t" | |
508 "psrlq $16, %%mm6 \n\t" | |
509 "paddw %%mm5, %%mm6 \n\t" | |
510 "movd %%mm6, %0 \n\t" | |
511 "andl $0xFFFF, %0 \n\t" | |
688 | 512 : "=&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
|
513 : "r" (pix - index), "r" ((long)line_size) |
688 | 514 ); |
515 | |
516 return sum; | |
517 } | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
518 #endif //CONFIG_ENCODERS |
688 | 519 |
866 | 520 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
|
521 long i=0; |
866 | 522 asm volatile( |
2979 | 523 "1: \n\t" |
524 "movq (%1, %0), %%mm0 \n\t" | |
525 "movq (%2, %0), %%mm1 \n\t" | |
526 "paddb %%mm0, %%mm1 \n\t" | |
527 "movq %%mm1, (%2, %0) \n\t" | |
528 "movq 8(%1, %0), %%mm0 \n\t" | |
529 "movq 8(%2, %0), %%mm1 \n\t" | |
530 "paddb %%mm0, %%mm1 \n\t" | |
531 "movq %%mm1, 8(%2, %0) \n\t" | |
532 "add $16, %0 \n\t" | |
533 "cmp %3, %0 \n\t" | |
534 " jb 1b \n\t" | |
866 | 535 : "+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
|
536 : "r"(src), "r"(dst), "r"((long)w-15) |
866 | 537 ); |
538 for(; i<w; i++) | |
539 dst[i+0] += src[i+0]; | |
540 } | |
541 | |
1648 | 542 #define H263_LOOP_FILTER \ |
2979 | 543 "pxor %%mm7, %%mm7 \n\t"\ |
544 "movq %0, %%mm0 \n\t"\ | |
545 "movq %0, %%mm1 \n\t"\ | |
546 "movq %3, %%mm2 \n\t"\ | |
547 "movq %3, %%mm3 \n\t"\ | |
548 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
549 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
550 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
551 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
552 "psubw %%mm2, %%mm0 \n\t"\ | |
553 "psubw %%mm3, %%mm1 \n\t"\ | |
554 "movq %1, %%mm2 \n\t"\ | |
555 "movq %1, %%mm3 \n\t"\ | |
556 "movq %2, %%mm4 \n\t"\ | |
557 "movq %2, %%mm5 \n\t"\ | |
558 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
559 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
560 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
561 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
562 "psubw %%mm2, %%mm4 \n\t"\ | |
563 "psubw %%mm3, %%mm5 \n\t"\ | |
564 "psllw $2, %%mm4 \n\t"\ | |
565 "psllw $2, %%mm5 \n\t"\ | |
566 "paddw %%mm0, %%mm4 \n\t"\ | |
567 "paddw %%mm1, %%mm5 \n\t"\ | |
568 "pxor %%mm6, %%mm6 \n\t"\ | |
569 "pcmpgtw %%mm4, %%mm6 \n\t"\ | |
570 "pcmpgtw %%mm5, %%mm7 \n\t"\ | |
571 "pxor %%mm6, %%mm4 \n\t"\ | |
572 "pxor %%mm7, %%mm5 \n\t"\ | |
573 "psubw %%mm6, %%mm4 \n\t"\ | |
574 "psubw %%mm7, %%mm5 \n\t"\ | |
575 "psrlw $3, %%mm4 \n\t"\ | |
576 "psrlw $3, %%mm5 \n\t"\ | |
577 "packuswb %%mm5, %%mm4 \n\t"\ | |
578 "packsswb %%mm7, %%mm6 \n\t"\ | |
579 "pxor %%mm7, %%mm7 \n\t"\ | |
580 "movd %4, %%mm2 \n\t"\ | |
581 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
582 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
583 "punpcklbw %%mm2, %%mm2 \n\t"\ | |
584 "psubusb %%mm4, %%mm2 \n\t"\ | |
585 "movq %%mm2, %%mm3 \n\t"\ | |
586 "psubusb %%mm4, %%mm3 \n\t"\ | |
587 "psubb %%mm3, %%mm2 \n\t"\ | |
588 "movq %1, %%mm3 \n\t"\ | |
589 "movq %2, %%mm4 \n\t"\ | |
590 "pxor %%mm6, %%mm3 \n\t"\ | |
591 "pxor %%mm6, %%mm4 \n\t"\ | |
592 "paddusb %%mm2, %%mm3 \n\t"\ | |
593 "psubusb %%mm2, %%mm4 \n\t"\ | |
594 "pxor %%mm6, %%mm3 \n\t"\ | |
595 "pxor %%mm6, %%mm4 \n\t"\ | |
596 "paddusb %%mm2, %%mm2 \n\t"\ | |
597 "packsswb %%mm1, %%mm0 \n\t"\ | |
598 "pcmpgtb %%mm0, %%mm7 \n\t"\ | |
599 "pxor %%mm7, %%mm0 \n\t"\ | |
600 "psubb %%mm7, %%mm0 \n\t"\ | |
601 "movq %%mm0, %%mm1 \n\t"\ | |
602 "psubusb %%mm2, %%mm0 \n\t"\ | |
603 "psubb %%mm0, %%mm1 \n\t"\ | |
604 "pand %5, %%mm1 \n\t"\ | |
605 "psrlw $2, %%mm1 \n\t"\ | |
606 "pxor %%mm7, %%mm1 \n\t"\ | |
607 "psubb %%mm7, %%mm1 \n\t"\ | |
608 "movq %0, %%mm5 \n\t"\ | |
609 "movq %3, %%mm6 \n\t"\ | |
610 "psubb %%mm1, %%mm5 \n\t"\ | |
611 "paddb %%mm1, %%mm6 \n\t" | |
1648 | 612 |
1647 | 613 static void h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale){ |
614 const int strength= ff_h263_loop_filter_strength[qscale]; | |
615 | |
616 asm volatile( | |
2967 | 617 |
1648 | 618 H263_LOOP_FILTER |
2967 | 619 |
2979 | 620 "movq %%mm3, %1 \n\t" |
621 "movq %%mm4, %2 \n\t" | |
622 "movq %%mm5, %0 \n\t" | |
623 "movq %%mm6, %3 \n\t" | |
1647 | 624 : "+m" (*(uint64_t*)(src - 2*stride)), |
625 "+m" (*(uint64_t*)(src - 1*stride)), | |
626 "+m" (*(uint64_t*)(src + 0*stride)), | |
627 "+m" (*(uint64_t*)(src + 1*stride)) | |
628 : "g" (2*strength), "m"(ff_pb_FC) | |
629 ); | |
630 } | |
631 | |
1648 | 632 static inline void transpose4x4(uint8_t *dst, uint8_t *src, int dst_stride, int src_stride){ |
633 asm volatile( //FIXME could save 1 instruction if done as 8x4 ... | |
2979 | 634 "movd %4, %%mm0 \n\t" |
635 "movd %5, %%mm1 \n\t" | |
636 "movd %6, %%mm2 \n\t" | |
637 "movd %7, %%mm3 \n\t" | |
638 "punpcklbw %%mm1, %%mm0 \n\t" | |
639 "punpcklbw %%mm3, %%mm2 \n\t" | |
640 "movq %%mm0, %%mm1 \n\t" | |
641 "punpcklwd %%mm2, %%mm0 \n\t" | |
642 "punpckhwd %%mm2, %%mm1 \n\t" | |
643 "movd %%mm0, %0 \n\t" | |
644 "punpckhdq %%mm0, %%mm0 \n\t" | |
645 "movd %%mm0, %1 \n\t" | |
646 "movd %%mm1, %2 \n\t" | |
647 "punpckhdq %%mm1, %%mm1 \n\t" | |
648 "movd %%mm1, %3 \n\t" | |
2967 | 649 |
1648 | 650 : "=m" (*(uint32_t*)(dst + 0*dst_stride)), |
651 "=m" (*(uint32_t*)(dst + 1*dst_stride)), | |
652 "=m" (*(uint32_t*)(dst + 2*dst_stride)), | |
653 "=m" (*(uint32_t*)(dst + 3*dst_stride)) | |
654 : "m" (*(uint32_t*)(src + 0*src_stride)), | |
655 "m" (*(uint32_t*)(src + 1*src_stride)), | |
656 "m" (*(uint32_t*)(src + 2*src_stride)), | |
657 "m" (*(uint32_t*)(src + 3*src_stride)) | |
658 ); | |
659 } | |
660 | |
661 static void h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale){ | |
662 const int strength= ff_h263_loop_filter_strength[qscale]; | |
663 uint64_t temp[4] __attribute__ ((aligned(8))); | |
664 uint8_t *btemp= (uint8_t*)temp; | |
2967 | 665 |
1648 | 666 src -= 2; |
667 | |
668 transpose4x4(btemp , src , 8, stride); | |
669 transpose4x4(btemp+4, src + 4*stride, 8, stride); | |
670 asm volatile( | |
671 H263_LOOP_FILTER // 5 3 4 6 | |
2967 | 672 |
1648 | 673 : "+m" (temp[0]), |
674 "+m" (temp[1]), | |
675 "+m" (temp[2]), | |
676 "+m" (temp[3]) | |
677 : "g" (2*strength), "m"(ff_pb_FC) | |
678 ); | |
679 | |
680 asm volatile( | |
2979 | 681 "movq %%mm5, %%mm1 \n\t" |
682 "movq %%mm4, %%mm0 \n\t" | |
683 "punpcklbw %%mm3, %%mm5 \n\t" | |
684 "punpcklbw %%mm6, %%mm4 \n\t" | |
685 "punpckhbw %%mm3, %%mm1 \n\t" | |
686 "punpckhbw %%mm6, %%mm0 \n\t" | |
687 "movq %%mm5, %%mm3 \n\t" | |
688 "movq %%mm1, %%mm6 \n\t" | |
689 "punpcklwd %%mm4, %%mm5 \n\t" | |
690 "punpcklwd %%mm0, %%mm1 \n\t" | |
691 "punpckhwd %%mm4, %%mm3 \n\t" | |
692 "punpckhwd %%mm0, %%mm6 \n\t" | |
693 "movd %%mm5, (%0) \n\t" | |
694 "punpckhdq %%mm5, %%mm5 \n\t" | |
695 "movd %%mm5, (%0,%2) \n\t" | |
696 "movd %%mm3, (%0,%2,2) \n\t" | |
697 "punpckhdq %%mm3, %%mm3 \n\t" | |
698 "movd %%mm3, (%0,%3) \n\t" | |
699 "movd %%mm1, (%1) \n\t" | |
700 "punpckhdq %%mm1, %%mm1 \n\t" | |
701 "movd %%mm1, (%1,%2) \n\t" | |
702 "movd %%mm6, (%1,%2,2) \n\t" | |
703 "punpckhdq %%mm6, %%mm6 \n\t" | |
704 "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
|
705 :: "r" (src), |
86e2b1424801
optimization and gcc 4.0 bug workaround patch by (Martin Drab >drab kepler.fjfi.cvut cz<)
michael
parents:
2293
diff
changeset
|
706 "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
|
707 "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
|
708 "r" ((long)(3*stride)) |
1648 | 709 ); |
710 } | |
711 | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
712 #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
|
713 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
|
714 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
|
715 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
|
716 "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
|
717 "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
|
718 "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
|
719 "1:\n" |
2979 | 720 "movq (%0),%%mm2\n" /* mm2 = pix[0-7] */ |
721 "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
|
722 |
2979 | 723 "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
|
724 |
2979 | 725 "punpckhbw %%mm0,%%mm1\n" /* mm1 = [pix4-7] */ |
726 "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
|
727 |
2979 | 728 "movq %%mm3,%%mm4\n" /* mm4 = mm3 = pix[8-15] */ |
729 "punpckhbw %%mm0,%%mm3\n" /* mm3 = [pix12-15] */ | |
730 "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
|
731 |
2979 | 732 "pmaddwd %%mm1,%%mm1\n" /* mm1 = (pix0^2+pix1^2,pix2^2+pix3^2) */ |
733 "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
|
734 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
735 "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
|
736 "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
|
737 |
2979 | 738 "paddd %%mm1,%%mm2\n" /* mm2 = (pix0^2+pix1^2+pix4^2+pix5^2, |
739 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
|
740 "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
|
741 "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
|
742 |
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
|
743 "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
|
744 "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
|
745 "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
|
746 "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
|
747 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
748 "movq %%mm7,%%mm1\n" |
2979 | 749 "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
|
750 "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
|
751 "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
|
752 : "+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
|
753 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
|
754 } |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
755 |
2067 | 756 static int sse8_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { |
757 int tmp; | |
758 asm volatile ( | |
759 "movl %4,%%ecx\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
760 "shr $1,%%ecx\n" |
2979 | 761 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */ |
762 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */ | |
2067 | 763 "1:\n" |
2979 | 764 "movq (%0),%%mm1\n" /* mm1 = pix1[0][0-7] */ |
765 "movq (%1),%%mm2\n" /* mm2 = pix2[0][0-7] */ | |
766 "movq (%0,%3),%%mm3\n" /* mm3 = pix1[1][0-7] */ | |
767 "movq (%1,%3),%%mm4\n" /* mm4 = pix2[1][0-7] */ | |
2067 | 768 |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
769 /* todo: mm1-mm2, mm3-mm4 */ |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
770 /* algo: substract mm1 from mm2 with saturation and vice versa */ |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
771 /* OR the results to get absolute difference */ |
2067 | 772 "movq %%mm1,%%mm5\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
773 "movq %%mm3,%%mm6\n" |
2067 | 774 "psubusb %%mm2,%%mm1\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
775 "psubusb %%mm4,%%mm3\n" |
2067 | 776 "psubusb %%mm5,%%mm2\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
777 "psubusb %%mm6,%%mm4\n" |
2067 | 778 |
779 "por %%mm1,%%mm2\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
780 "por %%mm3,%%mm4\n" |
2067 | 781 |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
782 /* now convert to 16-bit vectors so we can square them */ |
2067 | 783 "movq %%mm2,%%mm1\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
784 "movq %%mm4,%%mm3\n" |
2067 | 785 |
786 "punpckhbw %%mm0,%%mm2\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
787 "punpckhbw %%mm0,%%mm4\n" |
2979 | 788 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */ |
789 "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (mm3,mm4) */ | |
2067 | 790 |
791 "pmaddwd %%mm2,%%mm2\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
792 "pmaddwd %%mm4,%%mm4\n" |
2067 | 793 "pmaddwd %%mm1,%%mm1\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
794 "pmaddwd %%mm3,%%mm3\n" |
2067 | 795 |
2979 | 796 "lea (%0,%3,2), %0\n" /* pix1 += 2*line_size */ |
797 "lea (%1,%3,2), %1\n" /* pix2 += 2*line_size */ | |
2067 | 798 |
799 "paddd %%mm2,%%mm1\n" | |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
800 "paddd %%mm4,%%mm3\n" |
2067 | 801 "paddd %%mm1,%%mm7\n" |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
802 "paddd %%mm3,%%mm7\n" |
2067 | 803 |
804 "decl %%ecx\n" | |
805 "jnz 1b\n" | |
806 | |
807 "movq %%mm7,%%mm1\n" | |
2979 | 808 "psrlq $32, %%mm7\n" /* shift hi dword to lo */ |
2067 | 809 "paddd %%mm7,%%mm1\n" |
810 "movd %%mm1,%2\n" | |
2967 | 811 : "+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
|
812 : "r" ((long)line_size) , "m" (h) |
2067 | 813 : "%ecx"); |
814 return tmp; | |
815 } | |
816 | |
1708 | 817 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
|
818 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
|
819 asm volatile ( |
1708 | 820 "movl %4,%%ecx\n" |
2979 | 821 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */ |
822 "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
|
823 "1:\n" |
2979 | 824 "movq (%0),%%mm1\n" /* mm1 = pix1[0-7] */ |
825 "movq (%1),%%mm2\n" /* mm2 = pix2[0-7] */ | |
826 "movq 8(%0),%%mm3\n" /* mm3 = pix1[8-15] */ | |
827 "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
|
828 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
829 /* todo: mm1-mm2, mm3-mm4 */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
830 /* algo: substract mm1 from mm2 with saturation and vice versa */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
831 /* 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
|
832 "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
|
833 "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
|
834 "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
|
835 "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
|
836 "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
|
837 "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
|
838 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
839 "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
|
840 "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
|
841 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
842 /* 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
|
843 "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
|
844 "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
|
845 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
846 "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
|
847 "punpckhbw %%mm0,%%mm4\n" |
2979 | 848 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */ |
849 "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
|
850 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
851 "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
|
852 "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
|
853 "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
|
854 "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
|
855 |
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
|
856 "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
|
857 "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
|
858 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
859 "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
|
860 "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
|
861 "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
|
862 "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
|
863 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
864 "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
|
865 "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
|
866 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
867 "movq %%mm7,%%mm1\n" |
2979 | 868 "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
|
869 "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
|
870 "movd %%mm1,%2\n" |
2967 | 871 : "+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
|
872 : "r" ((long)line_size) , "m" (h) |
1708 | 873 : "%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
|
874 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
|
875 } |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
876 |
2899
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
877 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
|
878 int tmp; |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
879 asm volatile ( |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
880 "shr $1,%2\n" |
2979 | 881 "pxor %%xmm0,%%xmm0\n" /* mm0 = 0 */ |
882 "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
|
883 "1:\n" |
2979 | 884 "movdqu (%0),%%xmm1\n" /* mm1 = pix1[0][0-15] */ |
885 "movdqu (%1),%%xmm2\n" /* mm2 = pix2[0][0-15] */ | |
886 "movdqu (%0,%4),%%xmm3\n" /* mm3 = pix1[1][0-15] */ | |
887 "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
|
888 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
889 /* todo: mm1-mm2, mm3-mm4 */ |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
890 /* algo: substract mm1 from mm2 with saturation and vice versa */ |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
891 /* OR the results to get absolute difference */ |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
892 "movdqa %%xmm1,%%xmm5\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
893 "movdqa %%xmm3,%%xmm6\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
894 "psubusb %%xmm2,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
895 "psubusb %%xmm4,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
896 "psubusb %%xmm5,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
897 "psubusb %%xmm6,%%xmm4\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
898 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
899 "por %%xmm1,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
900 "por %%xmm3,%%xmm4\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
901 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
902 /* 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
|
903 "movdqa %%xmm2,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
904 "movdqa %%xmm4,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
905 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
906 "punpckhbw %%xmm0,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
907 "punpckhbw %%xmm0,%%xmm4\n" |
2979 | 908 "punpcklbw %%xmm0,%%xmm1\n" /* mm1 now spread over (mm1,mm2) */ |
909 "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
|
910 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
911 "pmaddwd %%xmm2,%%xmm2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
912 "pmaddwd %%xmm4,%%xmm4\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
913 "pmaddwd %%xmm1,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
914 "pmaddwd %%xmm3,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
915 |
2979 | 916 "lea (%0,%4,2), %0\n" /* pix1 += 2*line_size */ |
917 "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
|
918 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
919 "paddd %%xmm2,%%xmm1\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
920 "paddd %%xmm4,%%xmm3\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
921 "paddd %%xmm1,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
922 "paddd %%xmm3,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
923 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
924 "decl %2\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
925 "jnz 1b\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
926 |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
927 "movdqa %%xmm7,%%xmm1\n" |
2979 | 928 "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
|
929 "paddd %%xmm1,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
930 "movdqa %%xmm7,%%xmm1\n" |
2979 | 931 "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
|
932 "paddd %%xmm1,%%xmm7\n" |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
933 "movd %%xmm7,%3\n" |
2967 | 934 : "+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
|
935 : "r" ((long)line_size)); |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
936 return tmp; |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
937 } |
d3a726717baf
sse2 16x16 sum squared diff (306=>268 cycles on a K8)
lorenm
parents:
2892
diff
changeset
|
938 |
2067 | 939 static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) { |
940 int tmp; | |
941 asm volatile ( | |
942 "movl %3,%%ecx\n" | |
943 "pxor %%mm7,%%mm7\n" | |
944 "pxor %%mm6,%%mm6\n" | |
2967 | 945 |
2067 | 946 "movq (%0),%%mm0\n" |
947 "movq %%mm0, %%mm1\n" | |
948 "psllq $8, %%mm0\n" | |
949 "psrlq $8, %%mm1\n" | |
950 "psrlq $8, %%mm0\n" | |
951 "movq %%mm0, %%mm2\n" | |
952 "movq %%mm1, %%mm3\n" | |
953 "punpcklbw %%mm7,%%mm0\n" | |
954 "punpcklbw %%mm7,%%mm1\n" | |
955 "punpckhbw %%mm7,%%mm2\n" | |
956 "punpckhbw %%mm7,%%mm3\n" | |
957 "psubw %%mm1, %%mm0\n" | |
958 "psubw %%mm3, %%mm2\n" | |
2967 | 959 |
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
|
960 "add %2,%0\n" |
2967 | 961 |
2067 | 962 "movq (%0),%%mm4\n" |
963 "movq %%mm4, %%mm1\n" | |
964 "psllq $8, %%mm4\n" | |
965 "psrlq $8, %%mm1\n" | |
966 "psrlq $8, %%mm4\n" | |
967 "movq %%mm4, %%mm5\n" | |
968 "movq %%mm1, %%mm3\n" | |
969 "punpcklbw %%mm7,%%mm4\n" | |
970 "punpcklbw %%mm7,%%mm1\n" | |
971 "punpckhbw %%mm7,%%mm5\n" | |
972 "punpckhbw %%mm7,%%mm3\n" | |
973 "psubw %%mm1, %%mm4\n" | |
974 "psubw %%mm3, %%mm5\n" | |
975 "psubw %%mm4, %%mm0\n" | |
976 "psubw %%mm5, %%mm2\n" | |
977 "pxor %%mm3, %%mm3\n" | |
978 "pxor %%mm1, %%mm1\n" | |
979 "pcmpgtw %%mm0, %%mm3\n\t" | |
980 "pcmpgtw %%mm2, %%mm1\n\t" | |
981 "pxor %%mm3, %%mm0\n" | |
982 "pxor %%mm1, %%mm2\n" | |
2967 | 983 "psubw %%mm3, %%mm0\n" |
2067 | 984 "psubw %%mm1, %%mm2\n" |
985 "paddw %%mm0, %%mm2\n" | |
986 "paddw %%mm2, %%mm6\n" | |
987 | |
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
|
988 "add %2,%0\n" |
2067 | 989 "1:\n" |
2967 | 990 |
2067 | 991 "movq (%0),%%mm0\n" |
992 "movq %%mm0, %%mm1\n" | |
993 "psllq $8, %%mm0\n" | |
994 "psrlq $8, %%mm1\n" | |
995 "psrlq $8, %%mm0\n" | |
996 "movq %%mm0, %%mm2\n" | |
997 "movq %%mm1, %%mm3\n" | |
998 "punpcklbw %%mm7,%%mm0\n" | |
999 "punpcklbw %%mm7,%%mm1\n" | |
1000 "punpckhbw %%mm7,%%mm2\n" | |
1001 "punpckhbw %%mm7,%%mm3\n" | |
1002 "psubw %%mm1, %%mm0\n" | |
1003 "psubw %%mm3, %%mm2\n" | |
1004 "psubw %%mm0, %%mm4\n" | |
1005 "psubw %%mm2, %%mm5\n" | |
1006 "pxor %%mm3, %%mm3\n" | |
1007 "pxor %%mm1, %%mm1\n" | |
1008 "pcmpgtw %%mm4, %%mm3\n\t" | |
1009 "pcmpgtw %%mm5, %%mm1\n\t" | |
1010 "pxor %%mm3, %%mm4\n" | |
1011 "pxor %%mm1, %%mm5\n" | |
2967 | 1012 "psubw %%mm3, %%mm4\n" |
2067 | 1013 "psubw %%mm1, %%mm5\n" |
1014 "paddw %%mm4, %%mm5\n" | |
1015 "paddw %%mm5, %%mm6\n" | |
2967 | 1016 |
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
|
1017 "add %2,%0\n" |
2967 | 1018 |
2067 | 1019 "movq (%0),%%mm4\n" |
1020 "movq %%mm4, %%mm1\n" | |
1021 "psllq $8, %%mm4\n" | |
1022 "psrlq $8, %%mm1\n" | |
1023 "psrlq $8, %%mm4\n" | |
1024 "movq %%mm4, %%mm5\n" | |
1025 "movq %%mm1, %%mm3\n" | |
1026 "punpcklbw %%mm7,%%mm4\n" | |
1027 "punpcklbw %%mm7,%%mm1\n" | |
1028 "punpckhbw %%mm7,%%mm5\n" | |
1029 "punpckhbw %%mm7,%%mm3\n" | |
1030 "psubw %%mm1, %%mm4\n" | |
1031 "psubw %%mm3, %%mm5\n" | |
1032 "psubw %%mm4, %%mm0\n" | |
1033 "psubw %%mm5, %%mm2\n" | |
1034 "pxor %%mm3, %%mm3\n" | |
1035 "pxor %%mm1, %%mm1\n" | |
1036 "pcmpgtw %%mm0, %%mm3\n\t" | |
1037 "pcmpgtw %%mm2, %%mm1\n\t" | |
1038 "pxor %%mm3, %%mm0\n" | |
1039 "pxor %%mm1, %%mm2\n" | |
2967 | 1040 "psubw %%mm3, %%mm0\n" |
2067 | 1041 "psubw %%mm1, %%mm2\n" |
1042 "paddw %%mm0, %%mm2\n" | |
1043 "paddw %%mm2, %%mm6\n" | |
1044 | |
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
|
1045 "add %2,%0\n" |
2067 | 1046 "subl $2, %%ecx\n" |
1047 " jnz 1b\n" | |
1048 | |
1049 "movq %%mm6, %%mm0\n" | |
1050 "punpcklwd %%mm7,%%mm0\n" | |
1051 "punpckhwd %%mm7,%%mm6\n" | |
1052 "paddd %%mm0, %%mm6\n" | |
2967 | 1053 |
2067 | 1054 "movq %%mm6,%%mm0\n" |
1055 "psrlq $32, %%mm6\n" | |
1056 "paddd %%mm6,%%mm0\n" | |
1057 "movd %%mm0,%1\n" | |
2967 | 1058 : "+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
|
1059 : "r" ((long)line_size) , "g" (h-2) |
2067 | 1060 : "%ecx"); |
1061 return tmp; | |
1062 } | |
1063 | |
1064 static int hf_noise16_mmx(uint8_t * pix1, int line_size, int h) { | |
1065 int tmp; | |
1066 uint8_t * pix= pix1; | |
1067 asm volatile ( | |
1068 "movl %3,%%ecx\n" | |
1069 "pxor %%mm7,%%mm7\n" | |
1070 "pxor %%mm6,%%mm6\n" | |
2967 | 1071 |
2067 | 1072 "movq (%0),%%mm0\n" |
1073 "movq 1(%0),%%mm1\n" | |
1074 "movq %%mm0, %%mm2\n" | |
1075 "movq %%mm1, %%mm3\n" | |
1076 "punpcklbw %%mm7,%%mm0\n" | |
1077 "punpcklbw %%mm7,%%mm1\n" | |
1078 "punpckhbw %%mm7,%%mm2\n" | |
1079 "punpckhbw %%mm7,%%mm3\n" | |
1080 "psubw %%mm1, %%mm0\n" | |
1081 "psubw %%mm3, %%mm2\n" | |
2967 | 1082 |
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
|
1083 "add %2,%0\n" |
2967 | 1084 |
2067 | 1085 "movq (%0),%%mm4\n" |
1086 "movq 1(%0),%%mm1\n" | |
1087 "movq %%mm4, %%mm5\n" | |
1088 "movq %%mm1, %%mm3\n" | |
1089 "punpcklbw %%mm7,%%mm4\n" | |
1090 "punpcklbw %%mm7,%%mm1\n" | |
1091 "punpckhbw %%mm7,%%mm5\n" | |
1092 "punpckhbw %%mm7,%%mm3\n" | |
1093 "psubw %%mm1, %%mm4\n" | |
1094 "psubw %%mm3, %%mm5\n" | |
1095 "psubw %%mm4, %%mm0\n" | |
1096 "psubw %%mm5, %%mm2\n" | |
1097 "pxor %%mm3, %%mm3\n" | |
1098 "pxor %%mm1, %%mm1\n" | |
1099 "pcmpgtw %%mm0, %%mm3\n\t" | |
1100 "pcmpgtw %%mm2, %%mm1\n\t" | |
1101 "pxor %%mm3, %%mm0\n" | |
1102 "pxor %%mm1, %%mm2\n" | |
2967 | 1103 "psubw %%mm3, %%mm0\n" |
2067 | 1104 "psubw %%mm1, %%mm2\n" |
1105 "paddw %%mm0, %%mm2\n" | |
1106 "paddw %%mm2, %%mm6\n" | |
1107 | |
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
|
1108 "add %2,%0\n" |
2067 | 1109 "1:\n" |
2967 | 1110 |
2067 | 1111 "movq (%0),%%mm0\n" |
1112 "movq 1(%0),%%mm1\n" | |
1113 "movq %%mm0, %%mm2\n" | |
1114 "movq %%mm1, %%mm3\n" | |
1115 "punpcklbw %%mm7,%%mm0\n" | |
1116 "punpcklbw %%mm7,%%mm1\n" | |
1117 "punpckhbw %%mm7,%%mm2\n" | |
1118 "punpckhbw %%mm7,%%mm3\n" | |
1119 "psubw %%mm1, %%mm0\n" | |
1120 "psubw %%mm3, %%mm2\n" | |
1121 "psubw %%mm0, %%mm4\n" | |
1122 "psubw %%mm2, %%mm5\n" | |
1123 "pxor %%mm3, %%mm3\n" | |
1124 "pxor %%mm1, %%mm1\n" | |
1125 "pcmpgtw %%mm4, %%mm3\n\t" | |
1126 "pcmpgtw %%mm5, %%mm1\n\t" | |
1127 "pxor %%mm3, %%mm4\n" | |
1128 "pxor %%mm1, %%mm5\n" | |
1129 "psubw %%mm3, %%mm4\n" | |
1130 "psubw %%mm1, %%mm5\n" | |
1131 "paddw %%mm4, %%mm5\n" | |
1132 "paddw %%mm5, %%mm6\n" | |
2967 | 1133 |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1134 "add %2,%0\n" |
2967 | 1135 |
2067 | 1136 "movq (%0),%%mm4\n" |
1137 "movq 1(%0),%%mm1\n" | |
1138 "movq %%mm4, %%mm5\n" | |
1139 "movq %%mm1, %%mm3\n" | |
1140 "punpcklbw %%mm7,%%mm4\n" | |
1141 "punpcklbw %%mm7,%%mm1\n" | |
1142 "punpckhbw %%mm7,%%mm5\n" | |
1143 "punpckhbw %%mm7,%%mm3\n" | |
1144 "psubw %%mm1, %%mm4\n" | |
1145 "psubw %%mm3, %%mm5\n" | |
1146 "psubw %%mm4, %%mm0\n" | |
1147 "psubw %%mm5, %%mm2\n" | |
1148 "pxor %%mm3, %%mm3\n" | |
1149 "pxor %%mm1, %%mm1\n" | |
1150 "pcmpgtw %%mm0, %%mm3\n\t" | |
1151 "pcmpgtw %%mm2, %%mm1\n\t" | |
1152 "pxor %%mm3, %%mm0\n" | |
1153 "pxor %%mm1, %%mm2\n" | |
2967 | 1154 "psubw %%mm3, %%mm0\n" |
2067 | 1155 "psubw %%mm1, %%mm2\n" |
1156 "paddw %%mm0, %%mm2\n" | |
1157 "paddw %%mm2, %%mm6\n" | |
1158 | |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1159 "add %2,%0\n" |
2067 | 1160 "subl $2, %%ecx\n" |
1161 " jnz 1b\n" | |
1162 | |
1163 "movq %%mm6, %%mm0\n" | |
1164 "punpcklwd %%mm7,%%mm0\n" | |
1165 "punpckhwd %%mm7,%%mm6\n" | |
1166 "paddd %%mm0, %%mm6\n" | |
2967 | 1167 |
2067 | 1168 "movq %%mm6,%%mm0\n" |
1169 "psrlq $32, %%mm6\n" | |
1170 "paddd %%mm6,%%mm0\n" | |
1171 "movd %%mm0,%1\n" | |
2967 | 1172 : "+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
|
1173 : "r" ((long)line_size) , "g" (h-2) |
2067 | 1174 : "%ecx"); |
1175 return tmp + hf_noise8_mmx(pix+8, line_size, h); | |
1176 } | |
1177 | |
2864
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1178 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
|
1179 MpegEncContext *c = p; |
2940 | 1180 int score1, score2; |
1181 | |
1182 if(c) score1 = c->dsp.sse[0](c, pix1, pix2, line_size, h); | |
1183 else score1 = sse16_mmx(c, pix1, pix2, line_size, h); | |
1184 score2= hf_noise16_mmx(pix1, line_size, h) - hf_noise16_mmx(pix2, line_size, h); | |
2067 | 1185 |
1186 if(c) return score1 + ABS(score2)*c->avctx->nsse_weight; | |
1187 else return score1 + ABS(score2)*8; | |
1188 } | |
1189 | |
2864
95bac7109ff0
Kill some compiler warnings. Compiled code verified identical after changes.
mru
parents:
2754
diff
changeset
|
1190 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
|
1191 MpegEncContext *c = p; |
2067 | 1192 int score1= sse8_mmx(c, pix1, pix2, line_size, h); |
1193 int score2= hf_noise8_mmx(pix1, line_size, h) - hf_noise8_mmx(pix2, line_size, h); | |
1194 | |
1195 if(c) return score1 + ABS(score2)*c->avctx->nsse_weight; | |
1196 else return score1 + ABS(score2)*8; | |
1197 } | |
1198 | |
1729 | 1199 static int vsad_intra16_mmx(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) { |
1200 int tmp; | |
2967 | 1201 |
1729 | 1202 assert( (((int)pix) & 7) == 0); |
1203 assert((line_size &7) ==0); | |
2967 | 1204 |
1729 | 1205 #define SUM(in0, in1, out0, out1) \ |
1206 "movq (%0), %%mm2\n"\ | |
1207 "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
|
1208 "add %2,%0\n"\ |
1729 | 1209 "movq %%mm2, " #out0 "\n"\ |
1210 "movq %%mm3, " #out1 "\n"\ | |
1211 "psubusb " #in0 ", %%mm2\n"\ | |
1212 "psubusb " #in1 ", %%mm3\n"\ | |
1213 "psubusb " #out0 ", " #in0 "\n"\ | |
1214 "psubusb " #out1 ", " #in1 "\n"\ | |
1215 "por %%mm2, " #in0 "\n"\ | |
1216 "por %%mm3, " #in1 "\n"\ | |
1217 "movq " #in0 ", %%mm2\n"\ | |
1218 "movq " #in1 ", %%mm3\n"\ | |
1219 "punpcklbw %%mm7, " #in0 "\n"\ | |
1220 "punpcklbw %%mm7, " #in1 "\n"\ | |
1221 "punpckhbw %%mm7, %%mm2\n"\ | |
1222 "punpckhbw %%mm7, %%mm3\n"\ | |
1223 "paddw " #in1 ", " #in0 "\n"\ | |
1224 "paddw %%mm3, %%mm2\n"\ | |
1225 "paddw %%mm2, " #in0 "\n"\ | |
1226 "paddw " #in0 ", %%mm6\n" | |
1227 | |
2967 | 1228 |
1729 | 1229 asm volatile ( |
1230 "movl %3,%%ecx\n" | |
1231 "pxor %%mm6,%%mm6\n" | |
1232 "pxor %%mm7,%%mm7\n" | |
1233 "movq (%0),%%mm0\n" | |
1234 "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
|
1235 "add %2,%0\n" |
1729 | 1236 "subl $2, %%ecx\n" |
1237 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1238 "1:\n" | |
2967 | 1239 |
1729 | 1240 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1241 |
1729 | 1242 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1243 |
1729 | 1244 "subl $2, %%ecx\n" |
1245 "jnz 1b\n" | |
1246 | |
1247 "movq %%mm6,%%mm0\n" | |
1248 "psrlq $32, %%mm6\n" | |
1249 "paddw %%mm6,%%mm0\n" | |
1250 "movq %%mm0,%%mm6\n" | |
1251 "psrlq $16, %%mm0\n" | |
1252 "paddw %%mm6,%%mm0\n" | |
1253 "movd %%mm0,%1\n" | |
2967 | 1254 : "+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
|
1255 : "r" ((long)line_size) , "m" (h) |
1729 | 1256 : "%ecx"); |
1257 return tmp & 0xFFFF; | |
1258 } | |
1259 #undef SUM | |
1260 | |
1261 static int vsad_intra16_mmx2(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) { | |
1262 int tmp; | |
2967 | 1263 |
1729 | 1264 assert( (((int)pix) & 7) == 0); |
1265 assert((line_size &7) ==0); | |
2967 | 1266 |
1729 | 1267 #define SUM(in0, in1, out0, out1) \ |
1268 "movq (%0), " #out0 "\n"\ | |
1269 "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
|
1270 "add %2,%0\n"\ |
1729 | 1271 "psadbw " #out0 ", " #in0 "\n"\ |
1272 "psadbw " #out1 ", " #in1 "\n"\ | |
1273 "paddw " #in1 ", " #in0 "\n"\ | |
1274 "paddw " #in0 ", %%mm6\n" | |
1275 | |
1276 asm volatile ( | |
1277 "movl %3,%%ecx\n" | |
1278 "pxor %%mm6,%%mm6\n" | |
1279 "pxor %%mm7,%%mm7\n" | |
1280 "movq (%0),%%mm0\n" | |
1281 "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
|
1282 "add %2,%0\n" |
1729 | 1283 "subl $2, %%ecx\n" |
1284 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1285 "1:\n" | |
2967 | 1286 |
1729 | 1287 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1288 |
1729 | 1289 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1290 |
1729 | 1291 "subl $2, %%ecx\n" |
1292 "jnz 1b\n" | |
1293 | |
1294 "movd %%mm6,%1\n" | |
2967 | 1295 : "+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
|
1296 : "r" ((long)line_size) , "m" (h) |
1729 | 1297 : "%ecx"); |
1298 return tmp; | |
1299 } | |
1300 #undef SUM | |
1301 | |
1302 static int vsad16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { | |
1303 int tmp; | |
2967 | 1304 |
1729 | 1305 assert( (((int)pix1) & 7) == 0); |
1306 assert( (((int)pix2) & 7) == 0); | |
1307 assert((line_size &7) ==0); | |
2967 | 1308 |
1729 | 1309 #define SUM(in0, in1, out0, out1) \ |
1310 "movq (%0),%%mm2\n"\ | |
1311 "movq (%1)," #out0 "\n"\ | |
1312 "movq 8(%0),%%mm3\n"\ | |
1313 "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
|
1314 "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
|
1315 "add %3,%1\n"\ |
1729 | 1316 "psubb " #out0 ", %%mm2\n"\ |
1317 "psubb " #out1 ", %%mm3\n"\ | |
1318 "pxor %%mm7, %%mm2\n"\ | |
1319 "pxor %%mm7, %%mm3\n"\ | |
1320 "movq %%mm2, " #out0 "\n"\ | |
1321 "movq %%mm3, " #out1 "\n"\ | |
1322 "psubusb " #in0 ", %%mm2\n"\ | |
1323 "psubusb " #in1 ", %%mm3\n"\ | |
1324 "psubusb " #out0 ", " #in0 "\n"\ | |
1325 "psubusb " #out1 ", " #in1 "\n"\ | |
1326 "por %%mm2, " #in0 "\n"\ | |
1327 "por %%mm3, " #in1 "\n"\ | |
1328 "movq " #in0 ", %%mm2\n"\ | |
1329 "movq " #in1 ", %%mm3\n"\ | |
1330 "punpcklbw %%mm7, " #in0 "\n"\ | |
1331 "punpcklbw %%mm7, " #in1 "\n"\ | |
1332 "punpckhbw %%mm7, %%mm2\n"\ | |
1333 "punpckhbw %%mm7, %%mm3\n"\ | |
1334 "paddw " #in1 ", " #in0 "\n"\ | |
1335 "paddw %%mm3, %%mm2\n"\ | |
1336 "paddw %%mm2, " #in0 "\n"\ | |
1337 "paddw " #in0 ", %%mm6\n" | |
1338 | |
2967 | 1339 |
1729 | 1340 asm volatile ( |
1341 "movl %4,%%ecx\n" | |
1342 "pxor %%mm6,%%mm6\n" | |
1343 "pcmpeqw %%mm7,%%mm7\n" | |
1344 "psllw $15, %%mm7\n" | |
1345 "packsswb %%mm7, %%mm7\n" | |
1346 "movq (%0),%%mm0\n" | |
1347 "movq (%1),%%mm2\n" | |
1348 "movq 8(%0),%%mm1\n" | |
1349 "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
|
1350 "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
|
1351 "add %3,%1\n" |
1729 | 1352 "subl $2, %%ecx\n" |
1353 "psubb %%mm2, %%mm0\n" | |
1354 "psubb %%mm3, %%mm1\n" | |
1355 "pxor %%mm7, %%mm0\n" | |
1356 "pxor %%mm7, %%mm1\n" | |
1357 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1358 "1:\n" | |
2967 | 1359 |
1729 | 1360 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1361 |
1729 | 1362 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1363 |
1729 | 1364 "subl $2, %%ecx\n" |
1365 "jnz 1b\n" | |
1366 | |
1367 "movq %%mm6,%%mm0\n" | |
1368 "psrlq $32, %%mm6\n" | |
1369 "paddw %%mm6,%%mm0\n" | |
1370 "movq %%mm0,%%mm6\n" | |
1371 "psrlq $16, %%mm0\n" | |
1372 "paddw %%mm6,%%mm0\n" | |
1373 "movd %%mm0,%2\n" | |
2967 | 1374 : "+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
|
1375 : "r" ((long)line_size) , "m" (h) |
1729 | 1376 : "%ecx"); |
1377 return tmp & 0x7FFF; | |
1378 } | |
1379 #undef SUM | |
1380 | |
1381 static int vsad16_mmx2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) { | |
1382 int tmp; | |
2967 | 1383 |
1729 | 1384 assert( (((int)pix1) & 7) == 0); |
1385 assert( (((int)pix2) & 7) == 0); | |
1386 assert((line_size &7) ==0); | |
2967 | 1387 |
1729 | 1388 #define SUM(in0, in1, out0, out1) \ |
1389 "movq (%0)," #out0 "\n"\ | |
1390 "movq (%1),%%mm2\n"\ | |
1391 "movq 8(%0)," #out1 "\n"\ | |
1392 "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
|
1393 "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
|
1394 "add %3,%1\n"\ |
1729 | 1395 "psubb %%mm2, " #out0 "\n"\ |
1396 "psubb %%mm3, " #out1 "\n"\ | |
1397 "pxor %%mm7, " #out0 "\n"\ | |
1398 "pxor %%mm7, " #out1 "\n"\ | |
1399 "psadbw " #out0 ", " #in0 "\n"\ | |
1400 "psadbw " #out1 ", " #in1 "\n"\ | |
1401 "paddw " #in1 ", " #in0 "\n"\ | |
1402 "paddw " #in0 ", %%mm6\n" | |
1403 | |
1404 asm volatile ( | |
1405 "movl %4,%%ecx\n" | |
1406 "pxor %%mm6,%%mm6\n" | |
1407 "pcmpeqw %%mm7,%%mm7\n" | |
1408 "psllw $15, %%mm7\n" | |
1409 "packsswb %%mm7, %%mm7\n" | |
1410 "movq (%0),%%mm0\n" | |
1411 "movq (%1),%%mm2\n" | |
1412 "movq 8(%0),%%mm1\n" | |
1413 "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
|
1414 "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
|
1415 "add %3,%1\n" |
1729 | 1416 "subl $2, %%ecx\n" |
1417 "psubb %%mm2, %%mm0\n" | |
1418 "psubb %%mm3, %%mm1\n" | |
1419 "pxor %%mm7, %%mm0\n" | |
1420 "pxor %%mm7, %%mm1\n" | |
1421 SUM(%%mm0, %%mm1, %%mm4, %%mm5) | |
1422 "1:\n" | |
2967 | 1423 |
1729 | 1424 SUM(%%mm4, %%mm5, %%mm0, %%mm1) |
2967 | 1425 |
1729 | 1426 SUM(%%mm0, %%mm1, %%mm4, %%mm5) |
2967 | 1427 |
1729 | 1428 "subl $2, %%ecx\n" |
1429 "jnz 1b\n" | |
1430 | |
1431 "movd %%mm6,%2\n" | |
2967 | 1432 : "+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
|
1433 : "r" ((long)line_size) , "m" (h) |
1729 | 1434 : "%ecx"); |
1435 return tmp; | |
1436 } | |
1437 #undef SUM | |
1438 | |
866 | 1439 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
|
1440 long i=0; |
866 | 1441 asm volatile( |
2979 | 1442 "1: \n\t" |
1443 "movq (%2, %0), %%mm0 \n\t" | |
1444 "movq (%1, %0), %%mm1 \n\t" | |
1445 "psubb %%mm0, %%mm1 \n\t" | |
1446 "movq %%mm1, (%3, %0) \n\t" | |
1447 "movq 8(%2, %0), %%mm0 \n\t" | |
1448 "movq 8(%1, %0), %%mm1 \n\t" | |
1449 "psubb %%mm0, %%mm1 \n\t" | |
1450 "movq %%mm1, 8(%3, %0) \n\t" | |
1451 "add $16, %0 \n\t" | |
1452 "cmp %4, %0 \n\t" | |
1453 " jb 1b \n\t" | |
866 | 1454 : "+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
|
1455 : "r"(src1), "r"(src2), "r"(dst), "r"((long)w-15) |
866 | 1456 ); |
1457 for(; i<w; i++) | |
1458 dst[i+0] = src1[i+0]-src2[i+0]; | |
1459 } | |
1527 | 1460 |
1461 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
|
1462 long i=0; |
1527 | 1463 uint8_t l, lt; |
2967 | 1464 |
1527 | 1465 asm volatile( |
2979 | 1466 "1: \n\t" |
1467 "movq -1(%1, %0), %%mm0 \n\t" // LT | |
1468 "movq (%1, %0), %%mm1 \n\t" // T | |
1469 "movq -1(%2, %0), %%mm2 \n\t" // L | |
1470 "movq (%2, %0), %%mm3 \n\t" // X | |
1471 "movq %%mm2, %%mm4 \n\t" // L | |
1472 "psubb %%mm0, %%mm2 \n\t" | |
1473 "paddb %%mm1, %%mm2 \n\t" // L + T - LT | |
1474 "movq %%mm4, %%mm5 \n\t" // L | |
1475 "pmaxub %%mm1, %%mm4 \n\t" // max(T, L) | |
1476 "pminub %%mm5, %%mm1 \n\t" // min(T, L) | |
1477 "pminub %%mm2, %%mm4 \n\t" | |
1478 "pmaxub %%mm1, %%mm4 \n\t" | |
1479 "psubb %%mm4, %%mm3 \n\t" // dst - pred | |
1480 "movq %%mm3, (%3, %0) \n\t" | |
1481 "add $8, %0 \n\t" | |
1482 "cmp %4, %0 \n\t" | |
1483 " jb 1b \n\t" | |
1527 | 1484 : "+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
|
1485 : "r"(src1), "r"(src2), "r"(dst), "r"((long)w) |
1527 | 1486 ); |
1487 | |
1488 l= *left; | |
1489 lt= *left_top; | |
2967 | 1490 |
1527 | 1491 dst[0]= src2[0] - mid_pred(l, src1[0], (l + src1[0] - lt)&0xFF); |
2967 | 1492 |
1527 | 1493 *left_top= src1[w-1]; |
1494 *left = src2[w-1]; | |
1495 } | |
1496 | |
1153 | 1497 #define LBUTTERFLY2(a1,b1,a2,b2)\ |
2979 | 1498 "paddw " #b1 ", " #a1 " \n\t"\ |
1499 "paddw " #b2 ", " #a2 " \n\t"\ | |
1500 "paddw " #b1 ", " #b1 " \n\t"\ | |
1501 "paddw " #b2 ", " #b2 " \n\t"\ | |
1502 "psubw " #a1 ", " #b1 " \n\t"\ | |
1503 "psubw " #a2 ", " #b2 " \n\t" | |
866 | 1504 |
936 | 1505 #define HADAMARD48\ |
1153 | 1506 LBUTTERFLY2(%%mm0, %%mm1, %%mm2, %%mm3)\ |
1507 LBUTTERFLY2(%%mm4, %%mm5, %%mm6, %%mm7)\ | |
1508 LBUTTERFLY2(%%mm0, %%mm2, %%mm1, %%mm3)\ | |
1509 LBUTTERFLY2(%%mm4, %%mm6, %%mm5, %%mm7)\ | |
1510 LBUTTERFLY2(%%mm0, %%mm4, %%mm1, %%mm5)\ | |
1511 LBUTTERFLY2(%%mm2, %%mm6, %%mm3, %%mm7)\ | |
936 | 1512 |
1513 #define MMABS(a,z)\ | |
2979 | 1514 "pxor " #z ", " #z " \n\t"\ |
1515 "pcmpgtw " #a ", " #z " \n\t"\ | |
1516 "pxor " #z ", " #a " \n\t"\ | |
1517 "psubw " #z ", " #a " \n\t" | |
936 | 1518 |
1519 #define MMABS_SUM(a,z, sum)\ | |
2979 | 1520 "pxor " #z ", " #z " \n\t"\ |
1521 "pcmpgtw " #a ", " #z " \n\t"\ | |
1522 "pxor " #z ", " #a " \n\t"\ | |
1523 "psubw " #z ", " #a " \n\t"\ | |
1524 "paddusw " #a ", " #sum " \n\t" | |
936 | 1525 |
1153 | 1526 #define MMABS_MMX2(a,z)\ |
2979 | 1527 "pxor " #z ", " #z " \n\t"\ |
1528 "psubw " #a ", " #z " \n\t"\ | |
1529 "pmaxsw " #z ", " #a " \n\t" | |
1153 | 1530 |
1531 #define MMABS_SUM_MMX2(a,z, sum)\ | |
2979 | 1532 "pxor " #z ", " #z " \n\t"\ |
1533 "psubw " #a ", " #z " \n\t"\ | |
1534 "pmaxsw " #z ", " #a " \n\t"\ | |
1535 "paddusw " #a ", " #sum " \n\t" | |
2967 | 1536 |
936 | 1537 #define TRANSPOSE4(a,b,c,d,t)\ |
1538 SBUTTERFLY(a,b,t,wd) /* a=aebf t=cgdh */\ | |
1539 SBUTTERFLY(c,d,b,wd) /* c=imjn b=kolp */\ | |
1540 SBUTTERFLY(a,c,d,dq) /* a=aeim d=bfjn */\ | |
1541 SBUTTERFLY(t,b,c,dq) /* t=cgko c=dhlp */ | |
1542 | |
1543 #define LOAD4(o, a, b, c, d)\ | |
2979 | 1544 "movq "#o"(%1), " #a " \n\t"\ |
1545 "movq "#o"+16(%1), " #b " \n\t"\ | |
1546 "movq "#o"+32(%1), " #c " \n\t"\ | |
1547 "movq "#o"+48(%1), " #d " \n\t" | |
936 | 1548 |
1549 #define STORE4(o, a, b, c, d)\ | |
2979 | 1550 "movq "#a", "#o"(%1) \n\t"\ |
1551 "movq "#b", "#o"+16(%1) \n\t"\ | |
1552 "movq "#c", "#o"+32(%1) \n\t"\ | |
1553 "movq "#d", "#o"+48(%1) \n\t"\ | |
936 | 1554 |
1708 | 1555 static int hadamard8_diff_mmx(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){ |
3089 | 1556 DECLARE_ALIGNED_8(uint64_t, temp[16]); |
936 | 1557 int sum=0; |
2967 | 1558 |
1708 | 1559 assert(h==8); |
936 | 1560 |
1561 diff_pixels_mmx((DCTELEM*)temp, src1, src2, stride); | |
1562 | |
1563 asm volatile( | |
1564 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3) | |
1565 LOAD4(64, %%mm4, %%mm5, %%mm6, %%mm7) | |
2967 | 1566 |
936 | 1567 HADAMARD48 |
2967 | 1568 |
2979 | 1569 "movq %%mm7, 112(%1) \n\t" |
2967 | 1570 |
936 | 1571 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7) |
1572 STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2) | |
2967 | 1573 |
2979 | 1574 "movq 112(%1), %%mm7 \n\t" |
936 | 1575 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0) |
1576 STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6) | |
1577 | |
1578 LOAD4(8 , %%mm0, %%mm1, %%mm2, %%mm3) | |
1579 LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7) | |
2967 | 1580 |
936 | 1581 HADAMARD48 |
2967 | 1582 |
2979 | 1583 "movq %%mm7, 120(%1) \n\t" |
2967 | 1584 |
936 | 1585 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7) |
1586 STORE4(8 , %%mm0, %%mm3, %%mm7, %%mm2) | |
2967 | 1587 |
2979 | 1588 "movq 120(%1), %%mm7 \n\t" |
936 | 1589 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0) |
2979 | 1590 "movq %%mm7, %%mm5 \n\t"//FIXME remove |
1591 "movq %%mm6, %%mm7 \n\t" | |
1592 "movq %%mm0, %%mm6 \n\t" | |
936 | 1593 // STORE4(72, %%mm4, %%mm7, %%mm0, %%mm6) //FIXME remove |
2967 | 1594 |
936 | 1595 LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3) |
1596 // LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7) | |
2967 | 1597 |
936 | 1598 HADAMARD48 |
2979 | 1599 "movq %%mm7, 64(%1) \n\t" |
936 | 1600 MMABS(%%mm0, %%mm7) |
1601 MMABS_SUM(%%mm1, %%mm7, %%mm0) | |
1602 MMABS_SUM(%%mm2, %%mm7, %%mm0) | |
1603 MMABS_SUM(%%mm3, %%mm7, %%mm0) | |
1604 MMABS_SUM(%%mm4, %%mm7, %%mm0) | |
1605 MMABS_SUM(%%mm5, %%mm7, %%mm0) | |
1606 MMABS_SUM(%%mm6, %%mm7, %%mm0) | |
2979 | 1607 "movq 64(%1), %%mm1 \n\t" |
936 | 1608 MMABS_SUM(%%mm1, %%mm7, %%mm0) |
2979 | 1609 "movq %%mm0, 64(%1) \n\t" |
2967 | 1610 |
936 | 1611 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3) |
1612 LOAD4(8 , %%mm4, %%mm5, %%mm6, %%mm7) | |
2967 | 1613 |
936 | 1614 HADAMARD48 |
2979 | 1615 "movq %%mm7, (%1) \n\t" |
936 | 1616 MMABS(%%mm0, %%mm7) |
1617 MMABS_SUM(%%mm1, %%mm7, %%mm0) | |
1618 MMABS_SUM(%%mm2, %%mm7, %%mm0) | |
1619 MMABS_SUM(%%mm3, %%mm7, %%mm0) | |
1620 MMABS_SUM(%%mm4, %%mm7, %%mm0) | |
1621 MMABS_SUM(%%mm5, %%mm7, %%mm0) | |
1622 MMABS_SUM(%%mm6, %%mm7, %%mm0) | |
2979 | 1623 "movq (%1), %%mm1 \n\t" |
936 | 1624 MMABS_SUM(%%mm1, %%mm7, %%mm0) |
2979 | 1625 "movq 64(%1), %%mm1 \n\t" |
936 | 1626 MMABS_SUM(%%mm1, %%mm7, %%mm0) |
2967 | 1627 |
2979 | 1628 "movq %%mm0, %%mm1 \n\t" |
1629 "psrlq $32, %%mm0 \n\t" | |
1630 "paddusw %%mm1, %%mm0 \n\t" | |
1631 "movq %%mm0, %%mm1 \n\t" | |
1632 "psrlq $16, %%mm0 \n\t" | |
1633 "paddusw %%mm1, %%mm0 \n\t" | |
1634 "movd %%mm0, %0 \n\t" | |
2967 | 1635 |
936 | 1636 : "=r" (sum) |
1637 : "r"(temp) | |
1638 ); | |
1639 return sum&0xFFFF; | |
1640 } | |
1641 | |
1708 | 1642 static int hadamard8_diff_mmx2(void *s, uint8_t *src1, uint8_t *src2, int stride, int h){ |
3089 | 1643 DECLARE_ALIGNED_8(uint64_t, temp[16]); |
1153 | 1644 int sum=0; |
2967 | 1645 |
1708 | 1646 assert(h==8); |
1153 | 1647 |
1648 diff_pixels_mmx((DCTELEM*)temp, src1, src2, stride); | |
1649 | |
1650 asm volatile( | |
1651 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3) | |
1652 LOAD4(64, %%mm4, %%mm5, %%mm6, %%mm7) | |
2967 | 1653 |
1153 | 1654 HADAMARD48 |
2967 | 1655 |
2979 | 1656 "movq %%mm7, 112(%1) \n\t" |
2967 | 1657 |
1153 | 1658 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7) |
1659 STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2) | |
2967 | 1660 |
2979 | 1661 "movq 112(%1), %%mm7 \n\t" |
1153 | 1662 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0) |
1663 STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6) | |
1664 | |
1665 LOAD4(8 , %%mm0, %%mm1, %%mm2, %%mm3) | |
1666 LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7) | |
2967 | 1667 |
1153 | 1668 HADAMARD48 |
2967 | 1669 |
2979 | 1670 "movq %%mm7, 120(%1) \n\t" |
2967 | 1671 |
1153 | 1672 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7) |
1673 STORE4(8 , %%mm0, %%mm3, %%mm7, %%mm2) | |
2967 | 1674 |
2979 | 1675 "movq 120(%1), %%mm7 \n\t" |
1153 | 1676 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0) |
2979 | 1677 "movq %%mm7, %%mm5 \n\t"//FIXME remove |
1678 "movq %%mm6, %%mm7 \n\t" | |
1679 "movq %%mm0, %%mm6 \n\t" | |
1153 | 1680 // STORE4(72, %%mm4, %%mm7, %%mm0, %%mm6) //FIXME remove |
2967 | 1681 |
1153 | 1682 LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3) |
1683 // LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7) | |
2967 | 1684 |
1153 | 1685 HADAMARD48 |
2979 | 1686 "movq %%mm7, 64(%1) \n\t" |
1153 | 1687 MMABS_MMX2(%%mm0, %%mm7) |
1688 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0) | |
1689 MMABS_SUM_MMX2(%%mm2, %%mm7, %%mm0) | |
1690 MMABS_SUM_MMX2(%%mm3, %%mm7, %%mm0) | |
1691 MMABS_SUM_MMX2(%%mm4, %%mm7, %%mm0) | |
1692 MMABS_SUM_MMX2(%%mm5, %%mm7, %%mm0) | |
1693 MMABS_SUM_MMX2(%%mm6, %%mm7, %%mm0) | |
2979 | 1694 "movq 64(%1), %%mm1 \n\t" |
1153 | 1695 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0) |
2979 | 1696 "movq %%mm0, 64(%1) \n\t" |
2967 | 1697 |
1153 | 1698 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3) |
1699 LOAD4(8 , %%mm4, %%mm5, %%mm6, %%mm7) | |
2967 | 1700 |
1153 | 1701 HADAMARD48 |
2979 | 1702 "movq %%mm7, (%1) \n\t" |
1153 | 1703 MMABS_MMX2(%%mm0, %%mm7) |
1704 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0) | |
1705 MMABS_SUM_MMX2(%%mm2, %%mm7, %%mm0) | |
1706 MMABS_SUM_MMX2(%%mm3, %%mm7, %%mm0) | |
1707 MMABS_SUM_MMX2(%%mm4, %%mm7, %%mm0) | |
1708 MMABS_SUM_MMX2(%%mm5, %%mm7, %%mm0) | |
1709 MMABS_SUM_MMX2(%%mm6, %%mm7, %%mm0) | |
2979 | 1710 "movq (%1), %%mm1 \n\t" |
1153 | 1711 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0) |
2979 | 1712 "movq 64(%1), %%mm1 \n\t" |
1153 | 1713 MMABS_SUM_MMX2(%%mm1, %%mm7, %%mm0) |
2967 | 1714 |
2892
41315d0120b3
replace a few mov + psrlq with pshufw, there are more cases which could benefit from this but they would require us to duplicate some functions ...
michael
parents:
2871
diff
changeset
|
1715 "pshufw $0x0E, %%mm0, %%mm1 \n\t" |
2979 | 1716 "paddusw %%mm1, %%mm0 \n\t" |
2892
41315d0120b3
replace a few mov + psrlq with pshufw, there are more cases which could benefit from this but they would require us to duplicate some functions ...
michael
parents:
2871
diff
changeset
|
1717 "pshufw $0x01, %%mm0, %%mm1 \n\t" |
2979 | 1718 "paddusw %%mm1, %%mm0 \n\t" |
1719 "movd %%mm0, %0 \n\t" | |
2967 | 1720 |
1153 | 1721 : "=r" (sum) |
1722 : "r"(temp) | |
1723 ); | |
1724 return sum&0xFFFF; | |
1725 } | |
1726 | |
1727 | |
1708 | 1728 WARPER8_16_SQ(hadamard8_diff_mmx, hadamard8_diff16_mmx) |
1729 WARPER8_16_SQ(hadamard8_diff_mmx2, hadamard8_diff16_mmx2) | |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
1730 #endif //CONFIG_ENCODERS |
866 | 1731 |
959 | 1732 #define put_no_rnd_pixels8_mmx(a,b,c,d) put_pixels8_mmx(a,b,c,d) |
1733 #define put_no_rnd_pixels16_mmx(a,b,c,d) put_pixels16_mmx(a,b,c,d) | |
1734 | |
954 | 1735 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\ |
2979 | 1736 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\ |
1737 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\ | |
1738 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\ | |
1739 "movq "#in7", " #m3 " \n\t" /* d */\ | |
1740 "movq "#in0", %%mm5 \n\t" /* D */\ | |
1741 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\ | |
1742 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\ | |
1743 "movq "#in1", %%mm5 \n\t" /* C */\ | |
1744 "movq "#in2", %%mm6 \n\t" /* B */\ | |
1745 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\ | |
1746 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\ | |
1747 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\ | |
1748 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\ | |
1749 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\ | |
1750 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\ | |
1751 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\ | |
1752 "psraw $5, %%mm5 \n\t"\ | |
1753 "packuswb %%mm5, %%mm5 \n\t"\ | |
954 | 1754 OP(%%mm5, out, %%mm7, d) |
1755 | |
959 | 1756 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\ |
1057 | 1757 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
954 | 1758 uint64_t temp;\ |
1759 \ | |
1760 asm volatile(\ | |
2979 | 1761 "pxor %%mm7, %%mm7 \n\t"\ |
1762 "1: \n\t"\ | |
1763 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
1764 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
1765 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
1766 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
1767 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
1768 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
1769 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
1770 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
1771 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
1772 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
1773 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
1774 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
1775 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
1776 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
1777 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
1778 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
1779 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
1780 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1781 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
1782 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
1783 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
1784 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
1785 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
1786 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
1787 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
1788 "paddw %6, %%mm6 \n\t"\ | |
1789 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1790 "psraw $5, %%mm0 \n\t"\ | |
1791 "movq %%mm0, %5 \n\t"\ | |
954 | 1792 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ |
1793 \ | |
2979 | 1794 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\ |
1795 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\ | |
1796 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\ | |
1797 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\ | |
1798 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\ | |
1799 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\ | |
1800 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\ | |
1801 "paddw %%mm0, %%mm2 \n\t" /* b */\ | |
1802 "paddw %%mm5, %%mm3 \n\t" /* c */\ | |
1803 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
1804 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
1805 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\ | |
1806 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\ | |
1807 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\ | |
1808 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\ | |
1809 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
1810 "paddw %%mm2, %%mm1 \n\t" /* a */\ | |
1811 "paddw %%mm6, %%mm4 \n\t" /* d */\ | |
1812 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
1813 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\ | |
1814 "paddw %6, %%mm1 \n\t"\ | |
1815 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\ | |
1816 "psraw $5, %%mm3 \n\t"\ | |
1817 "movq %5, %%mm1 \n\t"\ | |
1818 "packuswb %%mm3, %%mm1 \n\t"\ | |
959 | 1819 OP_MMX2(%%mm1, (%1),%%mm4, q)\ |
954 | 1820 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\ |
1821 \ | |
2979 | 1822 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\ |
1823 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\ | |
1824 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\ | |
1825 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\ | |
1826 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\ | |
1827 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\ | |
1828 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\ | |
1829 "paddw %%mm1, %%mm5 \n\t" /* b */\ | |
1830 "paddw %%mm4, %%mm0 \n\t" /* c */\ | |
1831 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1832 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\ | |
1833 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\ | |
1834 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\ | |
1835 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\ | |
1836 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\ | |
1837 "paddw %%mm3, %%mm2 \n\t" /* d */\ | |
1838 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\ | |
1839 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\ | |
1840 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\ | |
1841 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\ | |
1842 "paddw %%mm2, %%mm6 \n\t" /* a */\ | |
1843 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\ | |
1844 "paddw %6, %%mm0 \n\t"\ | |
1845 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1846 "psraw $5, %%mm0 \n\t"\ | |
954 | 1847 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\ |
1848 \ | |
2979 | 1849 "paddw %%mm5, %%mm3 \n\t" /* a */\ |
1850 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\ | |
1851 "paddw %%mm4, %%mm6 \n\t" /* b */\ | |
1852 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\ | |
1853 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\ | |
1854 "paddw %%mm1, %%mm4 \n\t" /* c */\ | |
1855 "paddw %%mm2, %%mm5 \n\t" /* d */\ | |
1856 "paddw %%mm6, %%mm6 \n\t" /* 2b */\ | |
1857 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\ | |
1858 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\ | |
1859 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\ | |
1860 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\ | |
1861 "paddw %6, %%mm4 \n\t"\ | |
1862 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\ | |
1863 "psraw $5, %%mm4 \n\t"\ | |
1864 "packuswb %%mm4, %%mm0 \n\t"\ | |
959 | 1865 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\ |
954 | 1866 \ |
2979 | 1867 "add %3, %0 \n\t"\ |
1868 "add %4, %1 \n\t"\ | |
1869 "decl %2 \n\t"\ | |
1870 " jnz 1b \n\t"\ | |
967 | 1871 : "+a"(src), "+c"(dst), "+m"(h)\ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1872 : "d"((long)srcStride), "S"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
966 | 1873 : "memory"\ |
954 | 1874 );\ |
1875 }\ | |
1876 \ | |
1877 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1878 int i;\ | |
1879 int16_t temp[16];\ | |
1880 /* quick HACK, XXX FIXME MUST be optimized */\ | |
1881 for(i=0; i<h; i++)\ | |
1882 {\ | |
1883 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
1884 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
1885 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
1886 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
1887 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
1888 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\ | |
1889 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\ | |
1890 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\ | |
1891 temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\ | |
1892 temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\ | |
1893 temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\ | |
1894 temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\ | |
1895 temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\ | |
1896 temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\ | |
1897 temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\ | |
1898 temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\ | |
1899 asm volatile(\ | |
2979 | 1900 "movq (%0), %%mm0 \n\t"\ |
1901 "movq 8(%0), %%mm1 \n\t"\ | |
1902 "paddw %2, %%mm0 \n\t"\ | |
1903 "paddw %2, %%mm1 \n\t"\ | |
1904 "psraw $5, %%mm0 \n\t"\ | |
1905 "psraw $5, %%mm1 \n\t"\ | |
1906 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 1907 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ |
2979 | 1908 "movq 16(%0), %%mm0 \n\t"\ |
1909 "movq 24(%0), %%mm1 \n\t"\ | |
1910 "paddw %2, %%mm0 \n\t"\ | |
1911 "paddw %2, %%mm1 \n\t"\ | |
1912 "psraw $5, %%mm0 \n\t"\ | |
1913 "psraw $5, %%mm1 \n\t"\ | |
1914 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 1915 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\ |
954 | 1916 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ |
966 | 1917 : "memory"\ |
954 | 1918 );\ |
1919 dst+=dstStride;\ | |
1920 src+=srcStride;\ | |
1921 }\ | |
1922 }\ | |
1923 \ | |
1057 | 1924 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
959 | 1925 uint64_t temp;\ |
1926 \ | |
1927 asm volatile(\ | |
2979 | 1928 "pxor %%mm7, %%mm7 \n\t"\ |
1929 "1: \n\t"\ | |
1930 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
1931 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
1932 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
1933 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
1934 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
1935 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
1936 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
1937 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
1938 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
1939 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
1940 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
1941 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
1942 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
1943 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
1944 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
1945 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
1946 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
1947 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
1948 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
1949 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
1950 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ | |
1951 "paddw %%mm4, %%mm0 \n\t" /* a */\ | |
1952 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
1953 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ | |
1954 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ | |
1955 "paddw %6, %%mm6 \n\t"\ | |
1956 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ | |
1957 "psraw $5, %%mm0 \n\t"\ | |
959 | 1958 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ |
1959 \ | |
2979 | 1960 "movd 5(%0), %%mm5 \n\t" /* FGHI */\ |
1961 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\ | |
1962 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\ | |
1963 "paddw %%mm5, %%mm1 \n\t" /* a */\ | |
1964 "paddw %%mm6, %%mm2 \n\t" /* b */\ | |
1965 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\ | |
1966 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\ | |
1967 "paddw %%mm6, %%mm3 \n\t" /* c */\ | |
1968 "paddw %%mm5, %%mm4 \n\t" /* d */\ | |
1969 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
1970 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
1971 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ | |
1972 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
1973 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\ | |
1974 "paddw %6, %%mm1 \n\t"\ | |
1975 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\ | |
1976 "psraw $5, %%mm3 \n\t"\ | |
1977 "packuswb %%mm3, %%mm0 \n\t"\ | |
959 | 1978 OP_MMX2(%%mm0, (%1), %%mm4, q)\ |
1979 \ | |
2979 | 1980 "add %3, %0 \n\t"\ |
1981 "add %4, %1 \n\t"\ | |
1982 "decl %2 \n\t"\ | |
1983 " jnz 1b \n\t"\ | |
967 | 1984 : "+a"(src), "+c"(dst), "+m"(h)\ |
2293
15cfba1b97b5
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64 patch by (Aurelien Jacobs <aurel at gnuage dot org>)
michael
parents:
2256
diff
changeset
|
1985 : "S"((long)srcStride), "D"((long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
966 | 1986 : "memory"\ |
959 | 1987 );\ |
1988 }\ | |
1989 \ | |
1990 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
1991 int i;\ | |
1992 int16_t temp[8];\ | |
1993 /* quick HACK, XXX FIXME MUST be optimized */\ | |
1994 for(i=0; i<h; i++)\ | |
1995 {\ | |
1996 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
1997 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
1998 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
1999 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
2000 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
2001 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\ | |
2002 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\ | |
2003 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\ | |
2004 asm volatile(\ | |
2979 | 2005 "movq (%0), %%mm0 \n\t"\ |
2006 "movq 8(%0), %%mm1 \n\t"\ | |
2007 "paddw %2, %%mm0 \n\t"\ | |
2008 "paddw %2, %%mm1 \n\t"\ | |
2009 "psraw $5, %%mm0 \n\t"\ | |
2010 "psraw $5, %%mm1 \n\t"\ | |
2011 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 2012 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ |
2013 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ | |
966 | 2014 :"memory"\ |
959 | 2015 );\ |
2016 dst+=dstStride;\ | |
2017 src+=srcStride;\ | |
2018 }\ | |
2019 } | |
2020 | |
2021 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\ | |
2022 \ | |
2023 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
954 | 2024 uint64_t temp[17*4];\ |
2025 uint64_t *temp_ptr= temp;\ | |
2026 int count= 17;\ | |
2027 \ | |
2028 /*FIXME unroll */\ | |
2029 asm volatile(\ | |
2979 | 2030 "pxor %%mm7, %%mm7 \n\t"\ |
2031 "1: \n\t"\ | |
2032 "movq (%0), %%mm0 \n\t"\ | |
2033 "movq (%0), %%mm1 \n\t"\ | |
2034 "movq 8(%0), %%mm2 \n\t"\ | |
2035 "movq 8(%0), %%mm3 \n\t"\ | |
2036 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2037 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
2038 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
2039 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
2040 "movq %%mm0, (%1) \n\t"\ | |
2041 "movq %%mm1, 17*8(%1) \n\t"\ | |
2042 "movq %%mm2, 2*17*8(%1) \n\t"\ | |
2043 "movq %%mm3, 3*17*8(%1) \n\t"\ | |
2044 "add $8, %1 \n\t"\ | |
2045 "add %3, %0 \n\t"\ | |
2046 "decl %2 \n\t"\ | |
2047 " jnz 1b \n\t"\ | |
954 | 2048 : "+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
|
2049 : "r" ((long)srcStride)\ |
966 | 2050 : "memory"\ |
954 | 2051 );\ |
2052 \ | |
2053 temp_ptr= temp;\ | |
2054 count=4;\ | |
2055 \ | |
2056 /*FIXME reorder for speed */\ | |
2057 asm volatile(\ | |
2979 | 2058 /*"pxor %%mm7, %%mm7 \n\t"*/\ |
2059 "1: \n\t"\ | |
2060 "movq (%0), %%mm0 \n\t"\ | |
2061 "movq 8(%0), %%mm1 \n\t"\ | |
2062 "movq 16(%0), %%mm2 \n\t"\ | |
2063 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 2064 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
2065 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
2979 | 2066 "add %4, %1 \n\t"\ |
961 | 2067 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 2068 \ |
961 | 2069 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
2979 | 2070 "add %4, %1 \n\t"\ |
961 | 2071 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
2072 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\ | |
2979 | 2073 "add %4, %1 \n\t"\ |
961 | 2074 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\ |
2075 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\ | |
2979 | 2076 "add %4, %1 \n\t"\ |
961 | 2077 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\ |
2078 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\ | |
2979 | 2079 "add %4, %1 \n\t"\ |
961 | 2080 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\ |
2081 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\ | |
2979 | 2082 "add %4, %1 \n\t"\ |
961 | 2083 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\ |
954 | 2084 \ |
961 | 2085 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\ |
2979 | 2086 "add %4, %1 \n\t" \ |
961 | 2087 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\ |
2088 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\ | |
954 | 2089 \ |
2979 | 2090 "add $136, %0 \n\t"\ |
2091 "add %6, %1 \n\t"\ | |
2092 "decl %2 \n\t"\ | |
2093 " 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
|
2094 \ |
967 | 2095 : "+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
|
2096 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-14*(long)dstStride)\ |
966 | 2097 :"memory"\ |
954 | 2098 );\ |
2099 }\ | |
2100 \ | |
1057 | 2101 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2209 | 2102 uint64_t temp[9*2];\ |
954 | 2103 uint64_t *temp_ptr= temp;\ |
2104 int count= 9;\ | |
2105 \ | |
2106 /*FIXME unroll */\ | |
2107 asm volatile(\ | |
2979 | 2108 "pxor %%mm7, %%mm7 \n\t"\ |
2109 "1: \n\t"\ | |
2110 "movq (%0), %%mm0 \n\t"\ | |
2111 "movq (%0), %%mm1 \n\t"\ | |
2112 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2113 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
2114 "movq %%mm0, (%1) \n\t"\ | |
2115 "movq %%mm1, 9*8(%1) \n\t"\ | |
2116 "add $8, %1 \n\t"\ | |
2117 "add %3, %0 \n\t"\ | |
2118 "decl %2 \n\t"\ | |
2119 " jnz 1b \n\t"\ | |
954 | 2120 : "+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
|
2121 : "r" ((long)srcStride)\ |
966 | 2122 : "memory"\ |
954 | 2123 );\ |
2124 \ | |
2125 temp_ptr= temp;\ | |
2126 count=2;\ | |
2127 \ | |
2128 /*FIXME reorder for speed */\ | |
2129 asm volatile(\ | |
2979 | 2130 /*"pxor %%mm7, %%mm7 \n\t"*/\ |
2131 "1: \n\t"\ | |
2132 "movq (%0), %%mm0 \n\t"\ | |
2133 "movq 8(%0), %%mm1 \n\t"\ | |
2134 "movq 16(%0), %%mm2 \n\t"\ | |
2135 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 2136 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
2137 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
2979 | 2138 "add %4, %1 \n\t"\ |
961 | 2139 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 2140 \ |
961 | 2141 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
2979 | 2142 "add %4, %1 \n\t"\ |
961 | 2143 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
954 | 2144 \ |
961 | 2145 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\ |
2979 | 2146 "add %4, %1 \n\t"\ |
961 | 2147 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\ |
2148 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\ | |
954 | 2149 \ |
2979 | 2150 "add $72, %0 \n\t"\ |
2151 "add %6, %1 \n\t"\ | |
2152 "decl %2 \n\t"\ | |
2153 " jnz 1b \n\t"\ | |
954 | 2154 \ |
961 | 2155 : "+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
|
2156 : "r"((long)dstStride), "r"(2*(long)dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-6*(long)dstStride)\ |
966 | 2157 : "memory"\ |
2158 );\ | |
959 | 2159 }\ |
954 | 2160 \ |
1064 | 2161 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2162 OPNAME ## pixels8_mmx(dst, src, stride, 8);\ |
954 | 2163 }\ |
2164 \ | |
1064 | 2165 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2166 uint64_t temp[8];\ |
954 | 2167 uint8_t * const half= (uint8_t*)temp;\ |
2168 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
|
2169 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ |
954 | 2170 }\ |
2171 \ | |
1064 | 2172 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2173 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\ |
2174 }\ | |
2175 \ | |
1064 | 2176 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2177 uint64_t temp[8];\ |
954 | 2178 uint8_t * const half= (uint8_t*)temp;\ |
2179 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
|
2180 OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\ |
954 | 2181 }\ |
2182 \ | |
1064 | 2183 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2184 uint64_t temp[8];\ |
954 | 2185 uint8_t * const half= (uint8_t*)temp;\ |
959 | 2186 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
|
2187 OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\ |
954 | 2188 }\ |
2189 \ | |
1064 | 2190 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2191 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 2192 }\ |
2193 \ | |
1064 | 2194 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2195 uint64_t temp[8];\ |
954 | 2196 uint8_t * const half= (uint8_t*)temp;\ |
959 | 2197 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
|
2198 OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\ |
954 | 2199 }\ |
1064 | 2200 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2201 uint64_t half[8 + 9];\ |
2202 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2203 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2204 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
|
2205 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
959 | 2206 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
|
2207 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 2208 }\ |
1064 | 2209 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2210 uint64_t half[8 + 9];\ |
2211 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2212 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2213 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
|
2214 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
959 | 2215 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
|
2216 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 2217 }\ |
1064 | 2218 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2219 uint64_t half[8 + 9];\ |
2220 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2221 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2222 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
|
2223 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
959 | 2224 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
|
2225 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 2226 }\ |
1064 | 2227 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2228 uint64_t half[8 + 9];\ |
2229 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
2230 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2231 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
|
2232 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
959 | 2233 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
|
2234 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 2235 }\ |
1064 | 2236 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2237 uint64_t half[8 + 9];\ |
954 | 2238 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
2239 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2240 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 2241 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
|
2242 OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 2243 }\ |
1064 | 2244 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2245 uint64_t half[8 + 9];\ |
954 | 2246 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
2247 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2248 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 2249 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
|
2250 OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 2251 }\ |
1064 | 2252 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2253 uint64_t half[8 + 9];\ |
2254 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2255 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
|
2256 put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\ |
984 | 2257 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 2258 }\ |
1064 | 2259 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2260 uint64_t half[8 + 9];\ |
2261 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2262 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
|
2263 put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\ |
984 | 2264 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 2265 }\ |
1064 | 2266 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2267 uint64_t half[9];\ |
954 | 2268 uint8_t * const halfH= ((uint8_t*)half);\ |
2269 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 2270 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 2271 }\ |
1064 | 2272 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2273 OPNAME ## pixels16_mmx(dst, src, stride, 16);\ |
954 | 2274 }\ |
2275 \ | |
1064 | 2276 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2277 uint64_t temp[32];\ |
2278 uint8_t * const half= (uint8_t*)temp;\ | |
2279 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
|
2280 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ |
954 | 2281 }\ |
2282 \ | |
1064 | 2283 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2284 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\ |
2285 }\ | |
2286 \ | |
1064 | 2287 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2288 uint64_t temp[32];\ |
2289 uint8_t * const half= (uint8_t*)temp;\ | |
2290 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
|
2291 OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\ |
954 | 2292 }\ |
2293 \ | |
1064 | 2294 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2295 uint64_t temp[32];\ |
2296 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 2297 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
|
2298 OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\ |
954 | 2299 }\ |
2300 \ | |
1064 | 2301 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 2302 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 2303 }\ |
2304 \ | |
1064 | 2305 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2306 uint64_t temp[32];\ |
2307 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 2308 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
|
2309 OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\ |
954 | 2310 }\ |
1064 | 2311 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2312 uint64_t half[16*2 + 17*2];\ |
2313 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2314 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2315 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
|
2316 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
959 | 2317 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
|
2318 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 2319 }\ |
1064 | 2320 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2321 uint64_t half[16*2 + 17*2];\ |
2322 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2323 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2324 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
|
2325 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
959 | 2326 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
|
2327 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 2328 }\ |
1064 | 2329 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2330 uint64_t half[16*2 + 17*2];\ |
2331 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2332 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 2333 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
|
2334 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
959 | 2335 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
|
2336 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 2337 }\ |
1064 | 2338 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2339 uint64_t half[16*2 + 17*2];\ |
2340 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2341 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2342 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
|
2343 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
959 | 2344 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
|
2345 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 2346 }\ |
1064 | 2347 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2348 uint64_t half[16*2 + 17*2];\ |
2349 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2350 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2351 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 2352 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
|
2353 OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 2354 }\ |
1064 | 2355 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2356 uint64_t half[16*2 + 17*2];\ |
2357 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
2358 uint8_t * const halfHV= ((uint8_t*)half);\ | |
2359 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 2360 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
|
2361 OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 2362 }\ |
1064 | 2363 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2364 uint64_t half[17*2];\ |
2365 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2366 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
|
2367 put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\ |
984 | 2368 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 2369 }\ |
1064 | 2370 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 2371 uint64_t half[17*2];\ |
2372 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 2373 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
|
2374 put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\ |
984 | 2375 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 2376 }\ |
1064 | 2377 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 2378 uint64_t half[17*2];\ |
2379 uint8_t * const halfH= ((uint8_t*)half);\ | |
2380 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 2381 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 2382 } |
2383 | |
2979 | 2384 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t" |
959 | 2385 #define AVG_3DNOW_OP(a,b,temp, size) \ |
2979 | 2386 "mov" #size " " #b ", " #temp " \n\t"\ |
2387 "pavgusb " #temp ", " #a " \n\t"\ | |
2388 "mov" #size " " #a ", " #b " \n\t" | |
959 | 2389 #define AVG_MMX2_OP(a,b,temp, size) \ |
2979 | 2390 "mov" #size " " #b ", " #temp " \n\t"\ |
2391 "pavgb " #temp ", " #a " \n\t"\ | |
2392 "mov" #size " " #a ", " #b " \n\t" | |
959 | 2393 |
2394 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP) | |
2395 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP) | |
2396 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP) | |
2397 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow) | |
2398 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow) | |
2399 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow) | |
954 | 2400 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2) |
959 | 2401 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2) |
954 | 2402 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2) |
2403 | |
393 | 2404 #if 0 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2405 static void just_return() { return; } |
393 | 2406 #endif |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
2407 |
954 | 2408 #define SET_QPEL_FUNC(postfix1, postfix2) \ |
2409 c->put_ ## postfix1 = put_ ## postfix2;\ | |
2410 c->put_no_rnd_ ## postfix1 = put_no_rnd_ ## postfix2;\ | |
2411 c->avg_ ## postfix1 = avg_ ## postfix2; | |
1092 | 2412 |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2413 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
|
2414 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
|
2415 const int w = 8; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2416 const int ix = ox>>(16+shift); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2417 const int iy = oy>>(16+shift); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2418 const int oxs = ox>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2419 const int oys = oy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2420 const int dxxs = dxx>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2421 const int dxys = dxy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2422 const int dyxs = dyx>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2423 const int dyys = dyy>>4; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2424 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
|
2425 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
|
2426 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
|
2427 const uint64_t shift2 = 2*shift; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2428 uint8_t edge_buf[(h+1)*stride]; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2429 int x, y; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2430 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2431 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
|
2432 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
|
2433 const int dxh = dxy*(h-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2434 const int dyw = dyx*(w-1); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2435 if( // non-constant fullpel offset (3% of blocks) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2436 (ox^(ox+dxw) | ox^(ox+dxh) | ox^(ox+dxw+dxh) | |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2437 oy^(oy+dyw) | oy^(oy+dyh) | oy^(oy+dyw+dyh)) >> (16+shift) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2438 // 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
|
2439 || (dxx|dxy|dyx|dyy)&15 ) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2440 { |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2441 //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
|
2442 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
|
2443 return; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2444 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2445 |
3250 | 2446 src += ix + iy*stride; |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2447 if( (unsigned)ix >= width-w || |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2448 (unsigned)iy >= height-h ) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2449 { |
3250 | 2450 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
|
2451 src = edge_buf; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2452 } |
3250 | 2453 |
2454 asm volatile( | |
2455 "movd %0, %%mm6 \n\t" | |
2456 "pxor %%mm7, %%mm7 \n\t" | |
2457 "punpcklwd %%mm6, %%mm6 \n\t" | |
2458 "punpcklwd %%mm6, %%mm6 \n\t" | |
2459 :: "r"(1<<shift) | |
2460 ); | |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2461 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2462 for(x=0; x<w; x+=4){ |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2463 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
|
2464 oxs - dxys + dxxs*(x+1), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2465 oxs - dxys + dxxs*(x+2), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2466 oxs - dxys + dxxs*(x+3) }; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2467 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
|
2468 oys - dyys + dyxs*(x+1), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2469 oys - dyys + dyxs*(x+2), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2470 oys - dyys + dyxs*(x+3) }; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2471 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2472 for(y=0; y<h; y++){ |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2473 asm volatile( |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2474 "movq %0, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2475 "movq %1, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2476 "paddw %2, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2477 "paddw %3, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2478 "movq %%mm4, %0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2479 "movq %%mm5, %1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2480 "psrlw $12, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2481 "psrlw $12, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2482 : "+m"(*dx4), "+m"(*dy4) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2483 : "m"(*dxy4), "m"(*dyy4) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2484 ); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2485 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2486 asm volatile( |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2487 "movq %%mm6, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2488 "movq %%mm6, %%mm1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2489 "psubw %%mm4, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2490 "psubw %%mm5, %%mm1 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2491 "movq %%mm2, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2492 "movq %%mm4, %%mm3 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2493 "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
|
2494 "pmullw %%mm5, %%mm3 \n\t" // dx*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2495 "pmullw %%mm5, %%mm2 \n\t" // (s-dx)*dy |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2496 "pmullw %%mm4, %%mm1 \n\t" // dx*(s-dy) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2497 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2498 "movd %4, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2499 "movd %3, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2500 "punpcklbw %%mm7, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2501 "punpcklbw %%mm7, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2502 "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
|
2503 "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
|
2504 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2505 "movd %2, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2506 "movd %1, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2507 "punpcklbw %%mm7, %%mm5 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2508 "punpcklbw %%mm7, %%mm4 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2509 "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
|
2510 "pmullw %%mm4, %%mm0 \n\t" // src[0,0] * (s-dx)*(s-dy) |
3250 | 2511 "paddw %5, %%mm1 \n\t" |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2512 "paddw %%mm3, %%mm2 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2513 "paddw %%mm1, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2514 "paddw %%mm2, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2515 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2516 "psrlw %6, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2517 "packuswb %%mm0, %%mm0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2518 "movd %%mm0, %0 \n\t" |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2519 |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2520 : "=m"(dst[x+y*stride]) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2521 : "m"(src[0]), "m"(src[1]), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2522 "m"(src[stride]), "m"(src[stride+1]), |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2523 "m"(*r4), "m"(shift2) |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2524 ); |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2525 src += stride; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2526 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2527 src += 4-h*stride; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2528 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2529 } |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
2530 |
1784 | 2531 static int try_8x8basis_mmx(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale){ |
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
|
2532 long i=0; |
2967 | 2533 |
1784 | 2534 assert(ABS(scale) < 256); |
2535 scale<<= 16 + 1 - BASIS_SHIFT + RECON_SHIFT; | |
2536 | |
2537 asm volatile( | |
2979 | 2538 "pcmpeqw %%mm6, %%mm6 \n\t" // -1w |
2539 "psrlw $15, %%mm6 \n\t" // 1w | |
2540 "pxor %%mm7, %%mm7 \n\t" | |
2541 "movd %4, %%mm5 \n\t" | |
2542 "punpcklwd %%mm5, %%mm5 \n\t" | |
2543 "punpcklwd %%mm5, %%mm5 \n\t" | |
2544 "1: \n\t" | |
2545 "movq (%1, %0), %%mm0 \n\t" | |
2546 "movq 8(%1, %0), %%mm1 \n\t" | |
2547 "pmulhw %%mm5, %%mm0 \n\t" | |
2548 "pmulhw %%mm5, %%mm1 \n\t" | |
2549 "paddw %%mm6, %%mm0 \n\t" | |
2550 "paddw %%mm6, %%mm1 \n\t" | |
2551 "psraw $1, %%mm0 \n\t" | |
2552 "psraw $1, %%mm1 \n\t" | |
2553 "paddw (%2, %0), %%mm0 \n\t" | |
2554 "paddw 8(%2, %0), %%mm1 \n\t" | |
2555 "psraw $6, %%mm0 \n\t" | |
2556 "psraw $6, %%mm1 \n\t" | |
2557 "pmullw (%3, %0), %%mm0 \n\t" | |
2558 "pmullw 8(%3, %0), %%mm1 \n\t" | |
2559 "pmaddwd %%mm0, %%mm0 \n\t" | |
2560 "pmaddwd %%mm1, %%mm1 \n\t" | |
2561 "paddd %%mm1, %%mm0 \n\t" | |
2562 "psrld $4, %%mm0 \n\t" | |
2563 "paddd %%mm0, %%mm7 \n\t" | |
2564 "add $16, %0 \n\t" | |
2565 "cmp $128, %0 \n\t" //FIXME optimize & bench | |
2566 " jb 1b \n\t" | |
2567 "movq %%mm7, %%mm6 \n\t" | |
2568 "psrlq $32, %%mm7 \n\t" | |
2569 "paddd %%mm6, %%mm7 \n\t" | |
2570 "psrld $2, %%mm7 \n\t" | |
2571 "movd %%mm7, %0 \n\t" | |
2967 | 2572 |
1784 | 2573 : "+r" (i) |
2574 : "r"(basis), "r"(rem), "r"(weight), "g"(scale) | |
2575 ); | |
2576 return i; | |
2577 } | |
2578 | |
2579 static void add_8x8basis_mmx(int16_t rem[64], int16_t basis[64], int scale){ | |
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
|
2580 long i=0; |
2967 | 2581 |
1784 | 2582 if(ABS(scale) < 256){ |
2583 scale<<= 16 + 1 - BASIS_SHIFT + RECON_SHIFT; | |
2584 asm volatile( | |
2979 | 2585 "pcmpeqw %%mm6, %%mm6 \n\t" // -1w |
2586 "psrlw $15, %%mm6 \n\t" // 1w | |
2587 "movd %3, %%mm5 \n\t" | |
2588 "punpcklwd %%mm5, %%mm5 \n\t" | |
2589 "punpcklwd %%mm5, %%mm5 \n\t" | |
2590 "1: \n\t" | |
2591 "movq (%1, %0), %%mm0 \n\t" | |
2592 "movq 8(%1, %0), %%mm1 \n\t" | |
2593 "pmulhw %%mm5, %%mm0 \n\t" | |
2594 "pmulhw %%mm5, %%mm1 \n\t" | |
2595 "paddw %%mm6, %%mm0 \n\t" | |
2596 "paddw %%mm6, %%mm1 \n\t" | |
2597 "psraw $1, %%mm0 \n\t" | |
2598 "psraw $1, %%mm1 \n\t" | |
2599 "paddw (%2, %0), %%mm0 \n\t" | |
2600 "paddw 8(%2, %0), %%mm1 \n\t" | |
2601 "movq %%mm0, (%2, %0) \n\t" | |
2602 "movq %%mm1, 8(%2, %0) \n\t" | |
2603 "add $16, %0 \n\t" | |
2604 "cmp $128, %0 \n\t" //FIXME optimize & bench | |
2605 " jb 1b \n\t" | |
2967 | 2606 |
1784 | 2607 : "+r" (i) |
2608 : "r"(basis), "r"(rem), "g"(scale) | |
2609 ); | |
2610 }else{ | |
2611 for(i=0; i<8*8; i++){ | |
2612 rem[i] += (basis[i]*scale + (1<<(BASIS_SHIFT - RECON_SHIFT-1)))>>(BASIS_SHIFT - RECON_SHIFT); | |
2967 | 2613 } |
1784 | 2614 } |
2615 } | |
2754 | 2616 |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2617 #define PREFETCH(name, op) \ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2618 void name(void *mem, int stride, int h){\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2619 const uint8_t *p= mem;\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2620 do{\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2621 asm volatile(#op" %0" :: "m"(*p));\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2622 p+= stride;\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2623 }while(--h);\ |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2624 } |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2625 PREFETCH(prefetch_mmx2, prefetcht0) |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2626 PREFETCH(prefetch_3dnow, prefetch) |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2627 #undef PREFETCH |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
2628 |
2754 | 2629 #include "h264dsp_mmx.c" |
2967 | 2630 |
3524 | 2631 /* AVS specific */ |
2632 void ff_cavsdsp_init_mmx2(DSPContext* c, AVCodecContext *avctx); | |
2633 | |
2634 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2635 put_pixels8_mmx(dst, src, stride, 8); | |
2636 } | |
2637 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2638 avg_pixels8_mmx(dst, src, stride, 8); | |
2639 } | |
2640 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2641 put_pixels16_mmx(dst, src, stride, 16); | |
2642 } | |
2643 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride) { | |
2644 avg_pixels16_mmx(dst, src, stride, 16); | |
2645 } | |
2646 | |
1092 | 2647 /* external functions, from idct_mmx.c */ |
2648 void ff_mmx_idct(DCTELEM *block); | |
2649 void ff_mmxext_idct(DCTELEM *block); | |
2650 | |
2696
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2651 void ff_vp3_idct_sse2(int16_t *input_data); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2652 void ff_vp3_idct_mmx(int16_t *data); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2653 void ff_vp3_dsp_init_mmx(void); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2654 |
1092 | 2655 /* XXX: those functions should be suppressed ASAP when all IDCTs are |
2656 converted */ | |
2657 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block) | |
2658 { | |
2659 ff_mmx_idct (block); | |
2660 put_pixels_clamped_mmx(block, dest, line_size); | |
2661 } | |
2662 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2663 { | |
2664 ff_mmx_idct (block); | |
2665 add_pixels_clamped_mmx(block, dest, line_size); | |
2666 } | |
2667 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block) | |
2668 { | |
2669 ff_mmxext_idct (block); | |
2670 put_pixels_clamped_mmx(block, dest, line_size); | |
2671 } | |
2672 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2673 { | |
2674 ff_mmxext_idct (block); | |
2675 add_pixels_clamped_mmx(block, dest, line_size); | |
2676 } | |
2696
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2677 static void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block) |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2678 { |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2679 ff_vp3_idct_sse2(block); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2680 put_signed_pixels_clamped_mmx(block, dest, line_size); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2681 } |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2682 static void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block) |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2683 { |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2684 ff_vp3_idct_sse2(block); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2685 add_pixels_clamped_mmx(block, dest, line_size); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2686 } |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2687 static void ff_vp3_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block) |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2688 { |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2689 ff_vp3_idct_mmx(block); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2690 put_signed_pixels_clamped_mmx(block, dest, line_size); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2691 } |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2692 static void ff_vp3_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block) |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2693 { |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2694 ff_vp3_idct_mmx(block); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2695 add_pixels_clamped_mmx(block, dest, line_size); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
2696 } |
2868 | 2697 #ifdef CONFIG_GPL |
2698 static void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block) | |
2699 { | |
2700 ff_idct_xvid_mmx (block); | |
2701 put_pixels_clamped_mmx(block, dest, line_size); | |
2702 } | |
2703 static void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2704 { | |
2705 ff_idct_xvid_mmx (block); | |
2706 add_pixels_clamped_mmx(block, dest, line_size); | |
2707 } | |
2708 static void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block) | |
2709 { | |
2710 ff_idct_xvid_mmx2 (block); | |
2711 put_pixels_clamped_mmx(block, dest, line_size); | |
2712 } | |
2713 static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block) | |
2714 { | |
2715 ff_idct_xvid_mmx2 (block); | |
2716 add_pixels_clamped_mmx(block, dest, line_size); | |
2717 } | |
2718 #endif | |
2967 | 2719 |
3541
3fbddeb13686
10l, vorbis_inverse_coupling_sse() was really 3dnow
lorenm
parents:
3536
diff
changeset
|
2720 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
|
2721 { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2722 int i; |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2723 asm volatile("pxor %%mm7, %%mm7":); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2724 for(i=0; i<blocksize; i+=2) { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2725 asm volatile( |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2726 "movq %0, %%mm0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2727 "movq %1, %%mm1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2728 "movq %%mm0, %%mm2 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2729 "movq %%mm1, %%mm3 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2730 "pfcmpge %%mm7, %%mm2 \n\t" // m <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2731 "pfcmpge %%mm7, %%mm3 \n\t" // a <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2732 "pslld $31, %%mm2 \n\t" // keep only the sign bit |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2733 "pxor %%mm2, %%mm1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2734 "movq %%mm3, %%mm4 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2735 "pand %%mm1, %%mm3 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2736 "pandn %%mm1, %%mm4 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2737 "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
|
2738 "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
|
2739 "movq %%mm3, %1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2740 "movq %%mm0, %0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2741 :"+m"(mag[i]), "+m"(ang[i]) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2742 ::"memory" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2743 ); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2744 } |
3561 | 2745 asm volatile("femms"); |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2746 } |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2747 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
|
2748 { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2749 int i; |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2750 |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2751 asm volatile( |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2752 "movaps %0, %%xmm5 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2753 ::"m"(ff_pdw_80000000[0]) |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2754 ); |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2755 for(i=0; i<blocksize; i+=4) { |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2756 asm volatile( |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2757 "movaps %0, %%xmm0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2758 "movaps %1, %%xmm1 \n\t" |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2759 "xorps %%xmm2, %%xmm2 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2760 "xorps %%xmm3, %%xmm3 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2761 "cmpleps %%xmm0, %%xmm2 \n\t" // m <= 0.0 |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2762 "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
|
2763 "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
|
2764 "xorps %%xmm2, %%xmm1 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2765 "movaps %%xmm3, %%xmm4 \n\t" |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2766 "andps %%xmm1, %%xmm3 \n\t" |
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
2767 "andnps %%xmm1, %%xmm4 \n\t" |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2768 "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
|
2769 "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
|
2770 "movaps %%xmm3, %1 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2771 "movaps %%xmm0, %0 \n\t" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2772 :"+m"(mag[i]), "+m"(ang[i]) |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2773 ::"memory" |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2774 ); |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2775 } |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2776 } |
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
2777 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2778 static void vector_fmul_3dnow(float *dst, const float *src, int len){ |
3574 | 2779 long i = (len-4)*4; |
2780 asm volatile( | |
2781 "1: \n\t" | |
2782 "movq (%1,%0), %%mm0 \n\t" | |
2783 "movq 8(%1,%0), %%mm1 \n\t" | |
2784 "pfmul (%2,%0), %%mm0 \n\t" | |
2785 "pfmul 8(%2,%0), %%mm1 \n\t" | |
2786 "movq %%mm0, (%1,%0) \n\t" | |
2787 "movq %%mm1, 8(%1,%0) \n\t" | |
2788 "sub $16, %0 \n\t" | |
2789 "jge 1b \n\t" | |
2790 "femms \n\t" | |
2791 :"+r"(i) | |
2792 :"r"(dst), "r"(src) | |
2793 :"memory" | |
2794 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2795 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2796 static void vector_fmul_sse(float *dst, const float *src, int len){ |
3574 | 2797 long i = (len-8)*4; |
2798 asm volatile( | |
2799 "1: \n\t" | |
2800 "movaps (%1,%0), %%xmm0 \n\t" | |
2801 "movaps 16(%1,%0), %%xmm1 \n\t" | |
2802 "mulps (%2,%0), %%xmm0 \n\t" | |
2803 "mulps 16(%2,%0), %%xmm1 \n\t" | |
2804 "movaps %%xmm0, (%1,%0) \n\t" | |
2805 "movaps %%xmm1, 16(%1,%0) \n\t" | |
2806 "sub $32, %0 \n\t" | |
2807 "jge 1b \n\t" | |
2808 :"+r"(i) | |
2809 :"r"(dst), "r"(src) | |
2810 :"memory" | |
2811 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2812 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2813 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2814 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
|
2815 long i = len*4-16; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2816 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2817 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2818 "pswapd 8(%1), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2819 "pswapd (%1), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2820 "pfmul (%3,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2821 "pfmul 8(%3,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2822 "movq %%mm0, (%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2823 "movq %%mm1, 8(%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2824 "add $16, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2825 "sub $16, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2826 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2827 :"+r"(i), "+r"(src1) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2828 :"r"(dst), "r"(src0) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2829 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2830 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2831 } |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2832 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
|
2833 long i = len*4-32; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2834 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2835 "1: \n\t" |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2836 "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
|
2837 "movaps (%1), %%xmm1 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2838 "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
|
2839 "shufps $0x1b, %%xmm1, %%xmm1 \n\t" |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2840 "mulps (%3,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2841 "mulps 16(%3,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2842 "movaps %%xmm0, (%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2843 "movaps %%xmm1, 16(%2,%0) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2844 "add $32, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2845 "sub $32, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2846 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2847 :"+r"(i), "+r"(src1) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2848 :"r"(dst), "r"(src0) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2849 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2850 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2851 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2852 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
|
2853 const float *src2, int src3, int len, int step){ |
3574 | 2854 long i = (len-4)*4; |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2855 if(step == 2 && src3 == 0){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2856 dst += (len-4)*2; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2857 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2858 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2859 "movq (%2,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2860 "movq 8(%2,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2861 "pfmul (%3,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2862 "pfmul 8(%3,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2863 "pfadd (%4,%0), %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2864 "pfadd 8(%4,%0), %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2865 "movd %%mm0, (%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2866 "movd %%mm1, 16(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2867 "psrlq $32, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2868 "psrlq $32, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2869 "movd %%mm0, 8(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2870 "movd %%mm1, 24(%1) \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2871 "sub $32, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2872 "sub $16, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2873 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2874 :"+r"(i), "+r"(dst) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2875 :"r"(src0), "r"(src1), "r"(src2) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2876 :"memory" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2877 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2878 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2879 else if(step == 1 && src3 == 0){ |
3574 | 2880 asm volatile( |
2881 "1: \n\t" | |
2882 "movq (%2,%0), %%mm0 \n\t" | |
2883 "movq 8(%2,%0), %%mm1 \n\t" | |
2884 "pfmul (%3,%0), %%mm0 \n\t" | |
2885 "pfmul 8(%3,%0), %%mm1 \n\t" | |
2886 "pfadd (%4,%0), %%mm0 \n\t" | |
2887 "pfadd 8(%4,%0), %%mm1 \n\t" | |
2888 "movq %%mm0, (%1,%0) \n\t" | |
2889 "movq %%mm1, 8(%1,%0) \n\t" | |
2890 "sub $16, %0 \n\t" | |
2891 "jge 1b \n\t" | |
2892 :"+r"(i) | |
2893 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
2894 :"memory" | |
2895 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2896 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2897 else |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2898 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
|
2899 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2900 } |
3569
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2901 static void vector_fmul_add_add_sse(float *dst, const float *src0, const float *src1, |
3574 | 2902 const float *src2, int src3, int len, int step){ |
2903 long i = (len-8)*4; | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2904 if(step == 2 && src3 == 0){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2905 dst += (len-8)*2; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2906 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2907 "1: \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2908 "movaps (%2,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2909 "movaps 16(%2,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2910 "mulps (%3,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2911 "mulps 16(%3,%0), %%xmm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2912 "addps (%4,%0), %%xmm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2913 "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
|
2914 "movss %%xmm0, (%1) \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2915 "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
|
2916 "movhlps %%xmm0, %%xmm2 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2917 "movhlps %%xmm1, %%xmm3 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2918 "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
|
2919 "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
|
2920 "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
|
2921 "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
|
2922 "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
|
2923 "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
|
2924 "movhlps %%xmm0, %%xmm2 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2925 "movhlps %%xmm1, %%xmm3 \n\t" |
c42c03f3b402
convert vector_fmul_reverse_sse2 and vector_fmul_add_add_sse2 to sse
michael
parents:
3568
diff
changeset
|
2926 "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
|
2927 "movss %%xmm3, 56(%1) \n\t" |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2928 "sub $64, %1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2929 "sub $32, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2930 "jge 1b \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2931 :"+r"(i), "+r"(dst) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2932 :"r"(src0), "r"(src1), "r"(src2) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2933 :"memory" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2934 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2935 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2936 else if(step == 1 && src3 == 0){ |
3574 | 2937 asm volatile( |
2938 "1: \n\t" | |
2939 "movaps (%2,%0), %%xmm0 \n\t" | |
2940 "movaps 16(%2,%0), %%xmm1 \n\t" | |
2941 "mulps (%3,%0), %%xmm0 \n\t" | |
2942 "mulps 16(%3,%0), %%xmm1 \n\t" | |
2943 "addps (%4,%0), %%xmm0 \n\t" | |
2944 "addps 16(%4,%0), %%xmm1 \n\t" | |
2945 "movaps %%xmm0, (%1,%0) \n\t" | |
2946 "movaps %%xmm1, 16(%1,%0) \n\t" | |
2947 "sub $32, %0 \n\t" | |
2948 "jge 1b \n\t" | |
2949 :"+r"(i) | |
2950 :"r"(dst), "r"(src0), "r"(src1), "r"(src2) | |
2951 :"memory" | |
2952 ); | |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2953 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2954 else |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2955 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
|
2956 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2957 |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2958 void float_to_int16_3dnow(int16_t *dst, const float *src, int len){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2959 // 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
|
2960 int i; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2961 for(i=0; i<len; i+=4) { |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2962 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2963 "pf2id %1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2964 "pf2id %2, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2965 "packssdw %%mm1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2966 "movq %%mm0, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2967 :"=m"(dst[i]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2968 :"m"(src[i]), "m"(src[i+2]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2969 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2970 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2971 asm volatile("femms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2972 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2973 void float_to_int16_sse(int16_t *dst, const float *src, int len){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2974 int i; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2975 for(i=0; i<len; i+=4) { |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2976 asm volatile( |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2977 "cvtps2pi %1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2978 "cvtps2pi %2, %%mm1 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2979 "packssdw %%mm1, %%mm0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2980 "movq %%mm0, %0 \n\t" |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2981 :"=m"(dst[i]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2982 :"m"(src[i]), "m"(src[i+2]) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2983 ); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2984 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2985 asm volatile("emms"); |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2986 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
2987 |
3210 | 2988 #ifdef CONFIG_SNOW_ENCODER |
2989 extern void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width); | |
2990 extern void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width); | |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
2991 extern void ff_snow_vertical_compose97i_sse2(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width); |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
2992 extern void ff_snow_vertical_compose97i_mmx(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width); |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
2993 extern void ff_snow_inner_add_yblock_sse2(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
2994 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8); |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
2995 extern void ff_snow_inner_add_yblock_mmx(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
2996 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8); |
3210 | 2997 #endif |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
2998 |
1092 | 2999 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx) |
0 | 3000 { |
3001 mm_flags = mm_support(); | |
1115 | 3002 |
1122 | 3003 if (avctx->dsp_mask) { |
2979 | 3004 if (avctx->dsp_mask & FF_MM_FORCE) |
3005 mm_flags |= (avctx->dsp_mask & 0xffff); | |
3006 else | |
3007 mm_flags &= ~(avctx->dsp_mask & 0xffff); | |
1122 | 3008 } |
1115 | 3009 |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
3010 #if 0 |
1868 | 3011 av_log(avctx, AV_LOG_INFO, "libavcodec: CPU flags:"); |
0 | 3012 if (mm_flags & MM_MMX) |
1868 | 3013 av_log(avctx, AV_LOG_INFO, " mmx"); |
0 | 3014 if (mm_flags & MM_MMXEXT) |
1868 | 3015 av_log(avctx, AV_LOG_INFO, " mmxext"); |
0 | 3016 if (mm_flags & MM_3DNOW) |
1868 | 3017 av_log(avctx, AV_LOG_INFO, " 3dnow"); |
0 | 3018 if (mm_flags & MM_SSE) |
1868 | 3019 av_log(avctx, AV_LOG_INFO, " sse"); |
0 | 3020 if (mm_flags & MM_SSE2) |
1868 | 3021 av_log(avctx, AV_LOG_INFO, " sse2"); |
3022 av_log(avctx, AV_LOG_INFO, "\n"); | |
0 | 3023 #endif |
3024 | |
3025 if (mm_flags & MM_MMX) { | |
1092 | 3026 const int idct_algo= avctx->idct_algo; |
3027 | |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1186
diff
changeset
|
3028 #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
|
3029 const int dct_algo = avctx->dct_algo; |
1565 | 3030 if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){ |
1765
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
3031 if(mm_flags & MM_SSE2){ |
e31754bc5b65
SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents:
1739
diff
changeset
|
3032 c->fdct = ff_fdct_sse2; |
2979 | 3033 }else if(mm_flags & MM_MMXEXT){ |
1565 | 3034 c->fdct = ff_fdct_mmx2; |
3035 }else{ | |
3036 c->fdct = ff_fdct_mmx; | |
3037 } | |
3038 } | |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1186
diff
changeset
|
3039 #endif //CONFIG_ENCODERS |
2256 | 3040 if(avctx->lowres==0){ |
3041 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){ | |
3042 c->idct_put= ff_simple_idct_put_mmx; | |
3043 c->idct_add= ff_simple_idct_add_mmx; | |
3044 c->idct = ff_simple_idct_mmx; | |
3045 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM; | |
3046 }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){ | |
3047 if(mm_flags & MM_MMXEXT){ | |
3048 c->idct_put= ff_libmpeg2mmx2_idct_put; | |
3049 c->idct_add= ff_libmpeg2mmx2_idct_add; | |
3050 c->idct = ff_mmxext_idct; | |
3051 }else{ | |
3052 c->idct_put= ff_libmpeg2mmx_idct_put; | |
3053 c->idct_add= ff_libmpeg2mmx_idct_add; | |
3054 c->idct = ff_mmx_idct; | |
3055 } | |
3056 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM; | |
3496
341c83f0bebe
disable the vp3 mmx and sse2 idcts, their output doesnt match the c idct (tested with -f crc) and the theora spec does not allow different idcts not to mention the difference is quite vissible ...
michael
parents:
3416
diff
changeset
|
3057 #if 0 |
2696
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3058 }else if(idct_algo==FF_IDCT_VP3){ |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3059 if(mm_flags & MM_SSE2){ |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3060 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
|
3061 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
|
3062 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
|
3063 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
|
3064 }else{ |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3065 ff_vp3_dsp_init_mmx(); |
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
2691
diff
changeset
|
3066 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
|
3067 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
|
3068 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
|
3069 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
|
3070 } |
3496
341c83f0bebe
disable the vp3 mmx and sse2 idcts, their output doesnt match the c idct (tested with -f crc) and the theora spec does not allow different idcts not to mention the difference is quite vissible ...
michael
parents:
3416
diff
changeset
|
3071 #endif |
3524 | 3072 }else if(idct_algo==FF_IDCT_CAVS){ |
3073 c->idct_permutation_type= FF_TRANSPOSE_IDCT_PERM; | |
2868 | 3074 #ifdef CONFIG_GPL |
3075 }else if(idct_algo==FF_IDCT_XVIDMMX){ | |
3076 if(mm_flags & MM_MMXEXT){ | |
3077 c->idct_put= ff_idct_xvid_mmx2_put; | |
3078 c->idct_add= ff_idct_xvid_mmx2_add; | |
3079 c->idct = ff_idct_xvid_mmx2; | |
3080 }else{ | |
3081 c->idct_put= ff_idct_xvid_mmx_put; | |
3082 c->idct_add= ff_idct_xvid_mmx_add; | |
3083 c->idct = ff_idct_xvid_mmx; | |
3084 } | |
2871
f4afa3812818
Fix compile without CONFIG_GPL, misplaced #endif caused a missing }.
reimar
parents:
2868
diff
changeset
|
3085 #endif |
1092 | 3086 } |
3087 } | |
1868 | 3088 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3089 #ifdef CONFIG_ENCODERS |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3090 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
|
3091 c->diff_pixels = diff_pixels_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3092 #endif //CONFIG_ENCODERS |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3093 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
|
3094 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
|
3095 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
|
3096 c->clear_blocks = clear_blocks_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3097 #ifdef CONFIG_ENCODERS |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3098 c->pix_sum = pix_sum16_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3099 #endif //CONFIG_ENCODERS |
415 | 3100 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3101 c->put_pixels_tab[0][0] = put_pixels16_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3102 c->put_pixels_tab[0][1] = put_pixels16_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3103 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3104 c->put_pixels_tab[0][3] = put_pixels16_xy2_mmx; |
0 | 3105 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3106 c->put_no_rnd_pixels_tab[0][0] = put_pixels16_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3107 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3108 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3109 c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_mmx; |
651 | 3110 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3111 c->avg_pixels_tab[0][0] = avg_pixels16_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3112 c->avg_pixels_tab[0][1] = avg_pixels16_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3113 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3114 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx; |
415 | 3115 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3116 c->avg_no_rnd_pixels_tab[0][0] = avg_no_rnd_pixels16_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3117 c->avg_no_rnd_pixels_tab[0][1] = avg_no_rnd_pixels16_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3118 c->avg_no_rnd_pixels_tab[0][2] = avg_no_rnd_pixels16_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3119 c->avg_no_rnd_pixels_tab[0][3] = avg_no_rnd_pixels16_xy2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3120 |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3121 c->put_pixels_tab[1][0] = put_pixels8_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3122 c->put_pixels_tab[1][1] = put_pixels8_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3123 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3124 c->put_pixels_tab[1][3] = put_pixels8_xy2_mmx; |
0 | 3125 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3126 c->put_no_rnd_pixels_tab[1][0] = put_pixels8_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3127 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3128 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3129 c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_mmx; |
651 | 3130 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3131 c->avg_pixels_tab[1][0] = avg_pixels8_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3132 c->avg_pixels_tab[1][1] = avg_pixels8_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3133 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3134 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx; |
651 | 3135 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3136 c->avg_no_rnd_pixels_tab[1][0] = avg_no_rnd_pixels8_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3137 c->avg_no_rnd_pixels_tab[1][1] = avg_no_rnd_pixels8_x2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3138 c->avg_no_rnd_pixels_tab[1][2] = avg_no_rnd_pixels8_y2_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3139 c->avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels8_xy2_mmx; |
2967 | 3140 |
3248
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
3141 c->gmc= gmc_mmx; |
7aa9f80e7954
mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents:
3215
diff
changeset
|
3142 |
866 | 3143 c->add_bytes= add_bytes_mmx; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3144 #ifdef CONFIG_ENCODERS |
866 | 3145 c->diff_bytes= diff_bytes_mmx; |
2967 | 3146 |
936 | 3147 c->hadamard8_diff[0]= hadamard8_diff16_mmx; |
3148 c->hadamard8_diff[1]= hadamard8_diff_mmx; | |
2967 | 3149 |
2979 | 3150 c->pix_norm1 = pix_norm1_mmx; |
3151 c->sse[0] = (mm_flags & MM_SSE2) ? sse16_sse2 : sse16_mmx; | |
3152 c->sse[1] = sse8_mmx; | |
1729 | 3153 c->vsad[4]= vsad_intra16_mmx; |
3154 | |
2979 | 3155 c->nsse[0] = nsse16_mmx; |
3156 c->nsse[1] = nsse8_mmx; | |
1729 | 3157 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
3158 c->vsad[0] = vsad16_mmx; | |
3159 } | |
2967 | 3160 |
1784 | 3161 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
3162 c->try_8x8basis= try_8x8basis_mmx; | |
3163 } | |
3164 c->add_8x8basis= add_8x8basis_mmx; | |
2967 | 3165 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3166 #endif //CONFIG_ENCODERS |
1647 | 3167 |
3168 c->h263_v_loop_filter= h263_v_loop_filter_mmx; | |
2967 | 3169 c->h263_h_loop_filter= h263_h_loop_filter_mmx; |
2979 | 3170 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_mmx; |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
3171 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_mmx; |
2967 | 3172 |
3173 | 3173 c->h264_idct_dc_add= |
3174 c->h264_idct_add= ff_h264_idct_add_mmx; | |
3174 | 3175 c->h264_idct8_dc_add= |
3176 c->h264_idct8_add= ff_h264_idct8_add_mmx; | |
3173 | 3177 |
0 | 3178 if (mm_flags & MM_MMXEXT) { |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3179 c->prefetch = prefetch_mmx2; |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3180 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3181 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
|
3182 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2; |
651 | 3183 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3184 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
|
3185 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
|
3186 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2; |
415 | 3187 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3188 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
|
3189 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2; |
651 | 3190 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3191 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
|
3192 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
|
3193 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2; |
1092 | 3194 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3195 #ifdef CONFIG_ENCODERS |
1153 | 3196 c->hadamard8_diff[0]= hadamard8_diff16_mmx2; |
3197 c->hadamard8_diff[1]= hadamard8_diff_mmx2; | |
1729 | 3198 c->vsad[4]= vsad_intra16_mmx2; |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3199 #endif //CONFIG_ENCODERS |
1153 | 3200 |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3089
diff
changeset
|
3201 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
|
3202 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2; |
2745 | 3203 |
1092 | 3204 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ |
3205 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2; | |
3206 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2; | |
3207 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2; | |
3208 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2; | |
3209 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2; | |
3210 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
|
3211 #ifdef CONFIG_ENCODERS |
1729 | 3212 c->vsad[0] = vsad16_mmx2; |
1772
8cd5257195c9
vsad16_mmx2 only applies if encoders are turned on
melanson
parents:
1765
diff
changeset
|
3213 #endif //CONFIG_ENCODERS |
1092 | 3214 } |
959 | 3215 |
961 | 3216 #if 1 |
954 | 3217 SET_QPEL_FUNC(qpel_pixels_tab[0][ 0], qpel16_mc00_mmx2) |
3218 SET_QPEL_FUNC(qpel_pixels_tab[0][ 1], qpel16_mc10_mmx2) | |
3219 SET_QPEL_FUNC(qpel_pixels_tab[0][ 2], qpel16_mc20_mmx2) | |
3220 SET_QPEL_FUNC(qpel_pixels_tab[0][ 3], qpel16_mc30_mmx2) | |
3221 SET_QPEL_FUNC(qpel_pixels_tab[0][ 4], qpel16_mc01_mmx2) | |
3222 SET_QPEL_FUNC(qpel_pixels_tab[0][ 5], qpel16_mc11_mmx2) | |
3223 SET_QPEL_FUNC(qpel_pixels_tab[0][ 6], qpel16_mc21_mmx2) | |
3224 SET_QPEL_FUNC(qpel_pixels_tab[0][ 7], qpel16_mc31_mmx2) | |
3225 SET_QPEL_FUNC(qpel_pixels_tab[0][ 8], qpel16_mc02_mmx2) | |
3226 SET_QPEL_FUNC(qpel_pixels_tab[0][ 9], qpel16_mc12_mmx2) | |
3227 SET_QPEL_FUNC(qpel_pixels_tab[0][10], qpel16_mc22_mmx2) | |
3228 SET_QPEL_FUNC(qpel_pixels_tab[0][11], qpel16_mc32_mmx2) | |
3229 SET_QPEL_FUNC(qpel_pixels_tab[0][12], qpel16_mc03_mmx2) | |
3230 SET_QPEL_FUNC(qpel_pixels_tab[0][13], qpel16_mc13_mmx2) | |
3231 SET_QPEL_FUNC(qpel_pixels_tab[0][14], qpel16_mc23_mmx2) | |
3232 SET_QPEL_FUNC(qpel_pixels_tab[0][15], qpel16_mc33_mmx2) | |
3233 SET_QPEL_FUNC(qpel_pixels_tab[1][ 0], qpel8_mc00_mmx2) | |
3234 SET_QPEL_FUNC(qpel_pixels_tab[1][ 1], qpel8_mc10_mmx2) | |
3235 SET_QPEL_FUNC(qpel_pixels_tab[1][ 2], qpel8_mc20_mmx2) | |
3236 SET_QPEL_FUNC(qpel_pixels_tab[1][ 3], qpel8_mc30_mmx2) | |
3237 SET_QPEL_FUNC(qpel_pixels_tab[1][ 4], qpel8_mc01_mmx2) | |
3238 SET_QPEL_FUNC(qpel_pixels_tab[1][ 5], qpel8_mc11_mmx2) | |
3239 SET_QPEL_FUNC(qpel_pixels_tab[1][ 6], qpel8_mc21_mmx2) | |
3240 SET_QPEL_FUNC(qpel_pixels_tab[1][ 7], qpel8_mc31_mmx2) | |
3241 SET_QPEL_FUNC(qpel_pixels_tab[1][ 8], qpel8_mc02_mmx2) | |
3242 SET_QPEL_FUNC(qpel_pixels_tab[1][ 9], qpel8_mc12_mmx2) | |
3243 SET_QPEL_FUNC(qpel_pixels_tab[1][10], qpel8_mc22_mmx2) | |
3244 SET_QPEL_FUNC(qpel_pixels_tab[1][11], qpel8_mc32_mmx2) | |
3245 SET_QPEL_FUNC(qpel_pixels_tab[1][12], qpel8_mc03_mmx2) | |
3246 SET_QPEL_FUNC(qpel_pixels_tab[1][13], qpel8_mc13_mmx2) | |
3247 SET_QPEL_FUNC(qpel_pixels_tab[1][14], qpel8_mc23_mmx2) | |
3248 SET_QPEL_FUNC(qpel_pixels_tab[1][15], qpel8_mc33_mmx2) | |
961 | 3249 #endif |
1527 | 3250 |
2209 | 3251 //FIXME 3dnow too |
3252 #define dspfunc(PFX, IDX, NUM) \ | |
3253 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_mmx2; \ | |
3254 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_mmx2; \ | |
3255 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_mmx2; \ | |
3256 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_mmx2; \ | |
3257 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_mmx2; \ | |
3258 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_mmx2; \ | |
3259 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_mmx2; \ | |
3260 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_mmx2; \ | |
3261 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_mmx2; \ | |
3262 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_mmx2; \ | |
3263 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_mmx2; \ | |
3264 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_mmx2; \ | |
3265 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_mmx2; \ | |
3266 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_mmx2; \ | |
3267 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_mmx2; \ | |
3268 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_mmx2 | |
3269 | |
3270 dspfunc(put_h264_qpel, 0, 16); | |
3271 dspfunc(put_h264_qpel, 1, 8); | |
3272 dspfunc(put_h264_qpel, 2, 4); | |
3273 dspfunc(avg_h264_qpel, 0, 16); | |
3274 dspfunc(avg_h264_qpel, 1, 8); | |
3275 dspfunc(avg_h264_qpel, 2, 4); | |
3276 #undef dspfunc | |
3277 | |
2979 | 3278 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_mmx2; |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
3279 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_mmx2; |
3213 | 3280 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_mmx2; |
3281 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_mmx2; | |
2633 | 3282 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_mmx2; |
3283 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_mmx2; | |
3284 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_mmx2; | |
3285 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
|
3286 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
|
3287 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
|
3288 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2; |
2633 | 3289 |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3290 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
|
3291 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
|
3292 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
|
3293 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
|
3294 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
|
3295 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
|
3296 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
|
3297 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
|
3298 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2899
diff
changeset
|
3299 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
|
3300 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
|
3301 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
|
3302 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
|
3303 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
|
3304 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
|
3305 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
|
3306 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
|
3307 |
3524 | 3308 #ifdef CONFIG_CAVS_DECODER |
3309 ff_cavsdsp_init_mmx2(c, avctx); | |
3310 #endif | |
3311 | |
1686
68abbec33289
Here are juste two added #ifdef CONFIG_ENCODERS to allow
michael
parents:
1648
diff
changeset
|
3312 #ifdef CONFIG_ENCODERS |
1527 | 3313 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
|
3314 #endif //CONFIG_ENCODERS |
0 | 3315 } else if (mm_flags & MM_3DNOW) { |
3215
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3316 c->prefetch = prefetch_3dnow; |
06f98047ff26
prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents:
3213
diff
changeset
|
3317 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3318 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
|
3319 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow; |
393 | 3320 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3321 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
|
3322 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
|
3323 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow; |
651 | 3324 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3325 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
|
3326 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
|
3327 |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
3328 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
|
3329 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
|
3330 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow; |
1092 | 3331 |
3332 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
3333 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow; | |
3334 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow; | |
3335 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow; | |
3336 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow; | |
3337 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow; | |
3338 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow; | |
3339 } | |
984 | 3340 |
954 | 3341 SET_QPEL_FUNC(qpel_pixels_tab[0][ 0], qpel16_mc00_3dnow) |
3342 SET_QPEL_FUNC(qpel_pixels_tab[0][ 1], qpel16_mc10_3dnow) | |
3343 SET_QPEL_FUNC(qpel_pixels_tab[0][ 2], qpel16_mc20_3dnow) | |
3344 SET_QPEL_FUNC(qpel_pixels_tab[0][ 3], qpel16_mc30_3dnow) | |
3345 SET_QPEL_FUNC(qpel_pixels_tab[0][ 4], qpel16_mc01_3dnow) | |
3346 SET_QPEL_FUNC(qpel_pixels_tab[0][ 5], qpel16_mc11_3dnow) | |
3347 SET_QPEL_FUNC(qpel_pixels_tab[0][ 6], qpel16_mc21_3dnow) | |
3348 SET_QPEL_FUNC(qpel_pixels_tab[0][ 7], qpel16_mc31_3dnow) | |
3349 SET_QPEL_FUNC(qpel_pixels_tab[0][ 8], qpel16_mc02_3dnow) | |
3350 SET_QPEL_FUNC(qpel_pixels_tab[0][ 9], qpel16_mc12_3dnow) | |
3351 SET_QPEL_FUNC(qpel_pixels_tab[0][10], qpel16_mc22_3dnow) | |
3352 SET_QPEL_FUNC(qpel_pixels_tab[0][11], qpel16_mc32_3dnow) | |
3353 SET_QPEL_FUNC(qpel_pixels_tab[0][12], qpel16_mc03_3dnow) | |
3354 SET_QPEL_FUNC(qpel_pixels_tab[0][13], qpel16_mc13_3dnow) | |
3355 SET_QPEL_FUNC(qpel_pixels_tab[0][14], qpel16_mc23_3dnow) | |
3356 SET_QPEL_FUNC(qpel_pixels_tab[0][15], qpel16_mc33_3dnow) | |
3357 SET_QPEL_FUNC(qpel_pixels_tab[1][ 0], qpel8_mc00_3dnow) | |
3358 SET_QPEL_FUNC(qpel_pixels_tab[1][ 1], qpel8_mc10_3dnow) | |
3359 SET_QPEL_FUNC(qpel_pixels_tab[1][ 2], qpel8_mc20_3dnow) | |
3360 SET_QPEL_FUNC(qpel_pixels_tab[1][ 3], qpel8_mc30_3dnow) | |
3361 SET_QPEL_FUNC(qpel_pixels_tab[1][ 4], qpel8_mc01_3dnow) | |
3362 SET_QPEL_FUNC(qpel_pixels_tab[1][ 5], qpel8_mc11_3dnow) | |
3363 SET_QPEL_FUNC(qpel_pixels_tab[1][ 6], qpel8_mc21_3dnow) | |
3364 SET_QPEL_FUNC(qpel_pixels_tab[1][ 7], qpel8_mc31_3dnow) | |
3365 SET_QPEL_FUNC(qpel_pixels_tab[1][ 8], qpel8_mc02_3dnow) | |
3366 SET_QPEL_FUNC(qpel_pixels_tab[1][ 9], qpel8_mc12_3dnow) | |
3367 SET_QPEL_FUNC(qpel_pixels_tab[1][10], qpel8_mc22_3dnow) | |
3368 SET_QPEL_FUNC(qpel_pixels_tab[1][11], qpel8_mc32_3dnow) | |
3369 SET_QPEL_FUNC(qpel_pixels_tab[1][12], qpel8_mc03_3dnow) | |
3370 SET_QPEL_FUNC(qpel_pixels_tab[1][13], qpel8_mc13_3dnow) | |
3371 SET_QPEL_FUNC(qpel_pixels_tab[1][14], qpel8_mc23_3dnow) | |
3372 SET_QPEL_FUNC(qpel_pixels_tab[1][15], qpel8_mc33_3dnow) | |
2209 | 3373 |
3374 #define dspfunc(PFX, IDX, NUM) \ | |
3375 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_3dnow; \ | |
3376 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_3dnow; \ | |
3377 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_3dnow; \ | |
3378 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_3dnow; \ | |
3379 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_3dnow; \ | |
3380 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_3dnow; \ | |
3381 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_3dnow; \ | |
3382 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_3dnow; \ | |
3383 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_3dnow; \ | |
3384 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_3dnow; \ | |
3385 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_3dnow; \ | |
3386 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_3dnow; \ | |
3387 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_3dnow; \ | |
3388 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_3dnow; \ | |
3389 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_3dnow; \ | |
3390 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_3dnow | |
3391 | |
3392 dspfunc(put_h264_qpel, 0, 16); | |
3393 dspfunc(put_h264_qpel, 1, 8); | |
3394 dspfunc(put_h264_qpel, 2, 4); | |
3395 dspfunc(avg_h264_qpel, 0, 16); | |
3396 dspfunc(avg_h264_qpel, 1, 8); | |
3397 dspfunc(avg_h264_qpel, 2, 4); | |
2732 | 3398 |
2979 | 3399 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_3dnow; |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
3400 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_3dnow; |
0 | 3401 } |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3402 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3403 #ifdef CONFIG_SNOW_ENCODER |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3404 if(mm_flags & MM_SSE2){ |
3210 | 3405 c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2; |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3406 c->vertical_compose97i = ff_snow_vertical_compose97i_sse2; |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
3407 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
|
3408 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3409 else{ |
3210 | 3410 c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx; |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3411 c->vertical_compose97i = ff_snow_vertical_compose97i_mmx; |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
3412 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
|
3413 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
3174
diff
changeset
|
3414 #endif |
3536
545a15c19c91
sse & sse2 implementations of vorbis channel coupling.
lorenm
parents:
3524
diff
changeset
|
3415 |
3568
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3416 if(mm_flags & MM_3DNOW){ |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3417 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
|
3418 c->vector_fmul = vector_fmul_3dnow; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3419 if(!(avctx->flags & CODEC_FLAG_BITEXACT)) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3420 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
|
3421 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3422 if(mm_flags & MM_3DNOWEXT) |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3423 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3424 if(mm_flags & MM_SSE){ |
3557
8e13ec0f8aa3
change vorbis_inverse_coupling_sse2() so it works on sse1 cpus
michael
parents:
3541
diff
changeset
|
3425 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
|
3426 c->vector_fmul = vector_fmul_sse; |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3427 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
|
3428 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
|
3429 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
|
3430 } |
945caa35ee9a
sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents:
3561
diff
changeset
|
3431 if(mm_flags & MM_3DNOW) |
3574 | 3432 c->vector_fmul_add_add = vector_fmul_add_add_3dnow; // faster than sse |
0 | 3433 } |
2967 | 3434 |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3435 #ifdef CONFIG_ENCODERS |
1092 | 3436 dsputil_init_pix_mmx(c, avctx); |
1530
3b31998fe22f
disable encoders where appropriate (patch courtesy of BERO
melanson
parents:
1527
diff
changeset
|
3437 #endif //CONFIG_ENCODERS |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3438 #if 0 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3439 // for speed testing |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3440 get_pixels = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3441 put_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3442 add_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3443 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3444 pix_abs16x16 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3445 pix_abs16x16_x2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3446 pix_abs16x16_y2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3447 pix_abs16x16_xy2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3448 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3449 put_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3450 put_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3451 put_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3452 put_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3453 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3454 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
|
3455 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
|
3456 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
|
3457 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
|
3458 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3459 avg_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3460 avg_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3461 avg_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3462 avg_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3463 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3464 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
|
3465 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
|
3466 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
|
3467 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
|
3468 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3469 //av_fdct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3470 //ff_idct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
3471 #endif |
0 | 3472 } |