Mercurial > libavcodec.hg
annotate i386/dsputil_mmx.c @ 1092:f59c3f66363b libavcodec
MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
bitexact cleanup
author | michaelni |
---|---|
date | Mon, 03 Mar 2003 14:54:00 +0000 |
parents | 92fb44eae6b6 |
children | 74a46d77e061 |
rev | line source |
---|---|
0 | 1 /* |
2 * MMX optimized DSP utils | |
429 | 3 * Copyright (c) 2000, 2001 Fabrice Bellard. |
0 | 4 * |
429 | 5 * This library is free software; you can redistribute it and/or |
6 * modify it under the terms of the GNU Lesser General Public | |
7 * License as published by the Free Software Foundation; either | |
8 * version 2 of the License, or (at your option) any later version. | |
0 | 9 * |
429 | 10 * This library is distributed in the hope that it will be useful, |
0 | 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
429 | 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
13 * Lesser General Public License for more details. | |
0 | 14 * |
429 | 15 * You should have received a copy of the GNU Lesser General Public |
16 * License along with this library; if not, write to the Free Software | |
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
0 | 18 * |
19 * MMX optimization by Nick Kurshev <nickols_k@mail.ru> | |
20 */ | |
21 | |
22 #include "../dsputil.h" | |
1092 | 23 #include "../simple_idct.h" |
0 | 24 |
5 | 25 int mm_flags; /* multimedia extension flags */ |
936 | 26 |
0 | 27 /* pixel operations */ |
387 | 28 static const uint64_t mm_bone __attribute__ ((aligned(8))) = 0x0101010101010101ULL; |
29 static const uint64_t mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL; | |
30 static const uint64_t mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002ULL; | |
0 | 31 |
954 | 32 static const uint64_t ff_pw_20 __attribute__ ((aligned(8))) = 0x0014001400140014ULL; |
33 static const uint64_t ff_pw_3 __attribute__ ((aligned(8))) = 0x0003000300030003ULL; | |
34 static const uint64_t ff_pw_16 __attribute__ ((aligned(8))) = 0x0010001000100010ULL; | |
35 static const uint64_t ff_pw_15 __attribute__ ((aligned(8))) = 0x000F000F000F000FULL; | |
36 | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
37 #define JUMPALIGN() __asm __volatile (".balign 8"::) |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
38 #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
|
39 |
448 | 40 #define MOVQ_WONE(regd) \ |
41 __asm __volatile ( \ | |
42 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ | |
43 "psrlw $15, %%" #regd ::) | |
44 | |
45 #define MOVQ_BFE(regd) \ | |
46 __asm __volatile ( \ | |
47 "pcmpeqd %%" #regd ", %%" #regd " \n\t"\ | |
48 "paddb %%" #regd ", %%" #regd " \n\t" ::) | |
49 | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
50 #ifndef PIC |
448 | 51 #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
|
52 #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
|
53 #else |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
54 // 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
|
55 // pcmpeqd -> -1 |
448 | 56 #define MOVQ_BONE(regd) \ |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
57 __asm __volatile ( \ |
448 | 58 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
59 "psrlw $15, %%" #regd " \n\t" \ | |
60 "packuswb %%" #regd ", %%" #regd " \n\t" ::) | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
61 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
62 #define MOVQ_WTWO(regd) \ |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
63 __asm __volatile ( \ |
448 | 64 "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ |
65 "psrlw $15, %%" #regd " \n\t" \ | |
66 "psllw $1, %%" #regd " \n\t"::) | |
387 | 67 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
68 #endif |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
69 |
448 | 70 // 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
|
71 // first argument is unmodifed and second is trashed |
471 | 72 // regfe is supposed to contain 0xfefefefefefefefe |
73 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe) \ | |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
74 "movq " #rega ", " #regr " \n\t"\ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
75 "pand " #regb ", " #regr " \n\t"\ |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
76 "pxor " #rega ", " #regb " \n\t"\ |
471 | 77 "pand " #regfe "," #regb " \n\t"\ |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
78 "psrlq $1, " #regb " \n\t"\ |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
79 "paddb " #regb ", " #regr " \n\t" |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
80 |
471 | 81 #define PAVGB_MMX(rega, regb, regr, regfe) \ |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
82 "movq " #rega ", " #regr " \n\t"\ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
83 "por " #regb ", " #regr " \n\t"\ |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
84 "pxor " #rega ", " #regb " \n\t"\ |
471 | 85 "pand " #regfe "," #regb " \n\t"\ |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
86 "psrlq $1, " #regb " \n\t"\ |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
87 "psubb " #regb ", " #regr " \n\t" |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
88 |
471 | 89 // mm6 is supposed to contain 0xfefefefefefefefe |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
90 #define PAVGBP_MMX_NO_RND(rega, regb, regr, regc, regd, regp) \ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
91 "movq " #rega ", " #regr " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
92 "movq " #regc ", " #regp " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
93 "pand " #regb ", " #regr " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
94 "pand " #regd ", " #regp " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
95 "pxor " #rega ", " #regb " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
96 "pxor " #regc ", " #regd " \n\t"\ |
448 | 97 "pand %%mm6, " #regb " \n\t"\ |
98 "pand %%mm6, " #regd " \n\t"\ | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
99 "psrlq $1, " #regb " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
100 "psrlq $1, " #regd " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
101 "paddb " #regb ", " #regr " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
102 "paddb " #regd ", " #regp " \n\t" |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
103 |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
104 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
105 "movq " #rega ", " #regr " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
106 "movq " #regc ", " #regp " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
107 "por " #regb ", " #regr " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
108 "por " #regd ", " #regp " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
109 "pxor " #rega ", " #regb " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
110 "pxor " #regc ", " #regd " \n\t"\ |
448 | 111 "pand %%mm6, " #regb " \n\t"\ |
112 "pand %%mm6, " #regd " \n\t"\ | |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
113 "psrlq $1, " #regd " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
114 "psrlq $1, " #regb " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
115 "psubb " #regb ", " #regr " \n\t"\ |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
116 "psubb " #regd ", " #regp " \n\t" |
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
117 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
118 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
119 /* MMX no rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
120 #define DEF(x, y) x ## _no_rnd_ ## y ##_mmx |
448 | 121 #define SET_RND MOVQ_WONE |
122 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX_NO_RND(a, b, c, d, e, f) | |
471 | 123 #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
|
124 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
125 #include "dsputil_mmx_rnd.h" |
444
a5edef76dac6
* new mmx code - based upon http://aggregate.org/MAGIC
kabi
parents:
438
diff
changeset
|
126 |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
127 #undef DEF |
448 | 128 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
129 #undef PAVGBP |
471 | 130 #undef PAVGB |
445
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
131 /***********************************/ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
132 /* MMX rounding */ |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
133 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
134 #define DEF(x, y) x ## _ ## y ##_mmx |
448 | 135 #define SET_RND MOVQ_WTWO |
136 #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f) | |
471 | 137 #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
|
138 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
139 #include "dsputil_mmx_rnd.h" |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
140 |
62c01dbdc1e0
* code with new PAVGB for MMX only CPU splited into separate file
kabi
parents:
444
diff
changeset
|
141 #undef DEF |
448 | 142 #undef SET_RND |
446
efe0c0d40577
* reenabled original xy2 put routine - rounding error is really bad with
kabi
parents:
445
diff
changeset
|
143 #undef PAVGBP |
471 | 144 #undef PAVGB |
387 | 145 |
0 | 146 /***********************************/ |
147 /* 3Dnow specific */ | |
148 | |
149 #define DEF(x) x ## _3dnow | |
150 /* for Athlons PAVGUSB is prefered */ | |
151 #define PAVGB "pavgusb" | |
152 | |
153 #include "dsputil_mmx_avg.h" | |
154 | |
155 #undef DEF | |
156 #undef PAVGB | |
157 | |
158 /***********************************/ | |
159 /* MMX2 specific */ | |
160 | |
386 | 161 #define DEF(x) x ## _mmx2 |
0 | 162 |
163 /* Introduced only in MMX2 set */ | |
164 #define PAVGB "pavgb" | |
165 | |
166 #include "dsputil_mmx_avg.h" | |
167 | |
168 #undef DEF | |
169 #undef PAVGB | |
170 | |
171 /***********************************/ | |
172 /* standard MMX */ | |
173 | |
1064 | 174 static void get_pixels_mmx(DCTELEM *block, const uint8_t *pixels, int line_size) |
0 | 175 { |
386 | 176 asm volatile( |
177 "movl $-128, %%eax \n\t" | |
178 "pxor %%mm7, %%mm7 \n\t" | |
179 ".balign 16 \n\t" | |
180 "1: \n\t" | |
181 "movq (%0), %%mm0 \n\t" | |
182 "movq (%0, %2), %%mm2 \n\t" | |
183 "movq %%mm0, %%mm1 \n\t" | |
184 "movq %%mm2, %%mm3 \n\t" | |
185 "punpcklbw %%mm7, %%mm0 \n\t" | |
186 "punpckhbw %%mm7, %%mm1 \n\t" | |
187 "punpcklbw %%mm7, %%mm2 \n\t" | |
188 "punpckhbw %%mm7, %%mm3 \n\t" | |
189 "movq %%mm0, (%1, %%eax)\n\t" | |
190 "movq %%mm1, 8(%1, %%eax)\n\t" | |
191 "movq %%mm2, 16(%1, %%eax)\n\t" | |
192 "movq %%mm3, 24(%1, %%eax)\n\t" | |
193 "addl %3, %0 \n\t" | |
194 "addl $32, %%eax \n\t" | |
195 "js 1b \n\t" | |
196 : "+r" (pixels) | |
197 : "r" (block+64), "r" (line_size), "r" (line_size*2) | |
198 : "%eax" | |
199 ); | |
0 | 200 } |
201 | |
1064 | 202 static inline void diff_pixels_mmx(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride) |
324 | 203 { |
204 asm volatile( | |
386 | 205 "pxor %%mm7, %%mm7 \n\t" |
206 "movl $-128, %%eax \n\t" | |
324 | 207 ".balign 16 \n\t" |
208 "1: \n\t" | |
209 "movq (%0), %%mm0 \n\t" | |
210 "movq (%1), %%mm2 \n\t" | |
211 "movq %%mm0, %%mm1 \n\t" | |
212 "movq %%mm2, %%mm3 \n\t" | |
213 "punpcklbw %%mm7, %%mm0 \n\t" | |
214 "punpckhbw %%mm7, %%mm1 \n\t" | |
215 "punpcklbw %%mm7, %%mm2 \n\t" | |
216 "punpckhbw %%mm7, %%mm3 \n\t" | |
217 "psubw %%mm2, %%mm0 \n\t" | |
218 "psubw %%mm3, %%mm1 \n\t" | |
219 "movq %%mm0, (%2, %%eax)\n\t" | |
220 "movq %%mm1, 8(%2, %%eax)\n\t" | |
221 "addl %3, %0 \n\t" | |
222 "addl %3, %1 \n\t" | |
223 "addl $16, %%eax \n\t" | |
224 "jnz 1b \n\t" | |
225 : "+r" (s1), "+r" (s2) | |
226 : "r" (block+64), "r" (stride) | |
227 : "%eax" | |
228 ); | |
229 } | |
230 | |
1064 | 231 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 232 { |
233 const DCTELEM *p; | |
1064 | 234 uint8_t *pix; |
0 | 235 |
236 /* read the pixels */ | |
237 p = block; | |
238 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
239 /* unrolled loop */ |
0 | 240 __asm __volatile( |
151
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
241 "movq %3, %%mm0\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
242 "movq 8%3, %%mm1\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
243 "movq 16%3, %%mm2\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
244 "movq 24%3, %%mm3\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
245 "movq 32%3, %%mm4\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
246 "movq 40%3, %%mm5\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
247 "movq 48%3, %%mm6\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
248 "movq 56%3, %%mm7\n\t" |
0 | 249 "packuswb %%mm1, %%mm0\n\t" |
250 "packuswb %%mm3, %%mm2\n\t" | |
251 "packuswb %%mm5, %%mm4\n\t" | |
252 "packuswb %%mm7, %%mm6\n\t" | |
151
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
253 "movq %%mm0, (%0)\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
254 "movq %%mm2, (%0, %1)\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
255 "movq %%mm4, (%0, %1, 2)\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
256 "movq %%mm6, (%0, %2)\n\t" |
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
257 ::"r" (pix), "r" (line_size), "r" (line_size*3), "m"(*p) |
0 | 258 :"memory"); |
259 pix += line_size*4; | |
260 p += 32; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
261 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
262 // 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
|
263 // compiler would generate some very strange code |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
264 // thus using "r" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
265 __asm __volatile( |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
266 "movq (%3), %%mm0\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
267 "movq 8(%3), %%mm1\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
268 "movq 16(%3), %%mm2\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
269 "movq 24(%3), %%mm3\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
270 "movq 32(%3), %%mm4\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
271 "movq 40(%3), %%mm5\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
272 "movq 48(%3), %%mm6\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
273 "movq 56(%3), %%mm7\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
274 "packuswb %%mm1, %%mm0\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
275 "packuswb %%mm3, %%mm2\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
276 "packuswb %%mm5, %%mm4\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
277 "packuswb %%mm7, %%mm6\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
278 "movq %%mm0, (%0)\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
279 "movq %%mm2, (%0, %1)\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
280 "movq %%mm4, (%0, %1, 2)\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
281 "movq %%mm6, (%0, %2)\n\t" |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
282 ::"r" (pix), "r" (line_size), "r" (line_size*3), "r"(p) |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
283 :"memory"); |
0 | 284 } |
285 | |
1064 | 286 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size) |
0 | 287 { |
288 const DCTELEM *p; | |
1064 | 289 uint8_t *pix; |
0 | 290 int i; |
291 | |
292 /* read the pixels */ | |
293 p = block; | |
294 pix = pixels; | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
295 MOVQ_ZERO(mm7); |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
296 i = 4; |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
297 do { |
0 | 298 __asm __volatile( |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
299 "movq (%2), %%mm0\n\t" |
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
300 "movq 8(%2), %%mm1\n\t" |
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
301 "movq 16(%2), %%mm2\n\t" |
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
302 "movq 24(%2), %%mm3\n\t" |
0 | 303 "movq %0, %%mm4\n\t" |
304 "movq %1, %%mm6\n\t" | |
305 "movq %%mm4, %%mm5\n\t" | |
306 "punpcklbw %%mm7, %%mm4\n\t" | |
307 "punpckhbw %%mm7, %%mm5\n\t" | |
308 "paddsw %%mm4, %%mm0\n\t" | |
309 "paddsw %%mm5, %%mm1\n\t" | |
310 "movq %%mm6, %%mm5\n\t" | |
311 "punpcklbw %%mm7, %%mm6\n\t" | |
312 "punpckhbw %%mm7, %%mm5\n\t" | |
313 "paddsw %%mm6, %%mm2\n\t" | |
314 "paddsw %%mm5, %%mm3\n\t" | |
315 "packuswb %%mm1, %%mm0\n\t" | |
316 "packuswb %%mm3, %%mm2\n\t" | |
317 "movq %%mm0, %0\n\t" | |
318 "movq %%mm2, %1\n\t" | |
151
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
42
diff
changeset
|
319 :"+m"(*pix), "+m"(*(pix+line_size)) |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
320 :"r"(p) |
0 | 321 :"memory"); |
322 pix += line_size*2; | |
323 p += 16; | |
342
8635a7036395
* fixes problem with -funroll-loops and buggy gcc compiler
kabi
parents:
324
diff
changeset
|
324 } while (--i); |
0 | 325 } |
326 | |
1064 | 327 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 328 { |
471 | 329 __asm __volatile( |
420 | 330 "lea (%3, %3), %%eax \n\t" |
422 | 331 ".balign 8 \n\t" |
420 | 332 "1: \n\t" |
333 "movq (%1), %%mm0 \n\t" | |
334 "movq (%1, %3), %%mm1 \n\t" | |
335 "movq %%mm0, (%2) \n\t" | |
336 "movq %%mm1, (%2, %3) \n\t" | |
337 "addl %%eax, %1 \n\t" | |
338 "addl %%eax, %2 \n\t" | |
339 "movq (%1), %%mm0 \n\t" | |
340 "movq (%1, %3), %%mm1 \n\t" | |
341 "movq %%mm0, (%2) \n\t" | |
342 "movq %%mm1, (%2, %3) \n\t" | |
343 "addl %%eax, %1 \n\t" | |
344 "addl %%eax, %2 \n\t" | |
345 "subl $4, %0 \n\t" | |
346 "jnz 1b \n\t" | |
347 : "+g"(h), "+r" (pixels), "+r" (block) | |
348 : "r"(line_size) | |
349 : "%eax", "memory" | |
350 ); | |
0 | 351 } |
352 | |
1064 | 353 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
651 | 354 { |
355 __asm __volatile( | |
356 "lea (%3, %3), %%eax \n\t" | |
357 ".balign 8 \n\t" | |
358 "1: \n\t" | |
359 "movq (%1), %%mm0 \n\t" | |
360 "movq 8(%1), %%mm4 \n\t" | |
361 "movq (%1, %3), %%mm1 \n\t" | |
362 "movq 8(%1, %3), %%mm5 \n\t" | |
363 "movq %%mm0, (%2) \n\t" | |
364 "movq %%mm4, 8(%2) \n\t" | |
365 "movq %%mm1, (%2, %3) \n\t" | |
366 "movq %%mm5, 8(%2, %3) \n\t" | |
367 "addl %%eax, %1 \n\t" | |
368 "addl %%eax, %2 \n\t" | |
369 "movq (%1), %%mm0 \n\t" | |
370 "movq 8(%1), %%mm4 \n\t" | |
371 "movq (%1, %3), %%mm1 \n\t" | |
372 "movq 8(%1, %3), %%mm5 \n\t" | |
373 "movq %%mm0, (%2) \n\t" | |
374 "movq %%mm4, 8(%2) \n\t" | |
375 "movq %%mm1, (%2, %3) \n\t" | |
376 "movq %%mm5, 8(%2, %3) \n\t" | |
377 "addl %%eax, %1 \n\t" | |
378 "addl %%eax, %2 \n\t" | |
379 "subl $4, %0 \n\t" | |
380 "jnz 1b \n\t" | |
381 : "+g"(h), "+r" (pixels), "+r" (block) | |
382 : "r"(line_size) | |
383 : "%eax", "memory" | |
384 ); | |
385 } | |
386 | |
296 | 387 static void clear_blocks_mmx(DCTELEM *blocks) |
388 { | |
471 | 389 __asm __volatile( |
296 | 390 "pxor %%mm7, %%mm7 \n\t" |
391 "movl $-128*6, %%eax \n\t" | |
392 "1: \n\t" | |
393 "movq %%mm7, (%0, %%eax) \n\t" | |
394 "movq %%mm7, 8(%0, %%eax) \n\t" | |
395 "movq %%mm7, 16(%0, %%eax) \n\t" | |
396 "movq %%mm7, 24(%0, %%eax) \n\t" | |
397 "addl $32, %%eax \n\t" | |
398 " js 1b \n\t" | |
399 : : "r" (((int)blocks)+128*6) | |
400 : "%eax" | |
401 ); | |
402 } | |
403 | |
1064 | 404 static int pix_sum16_mmx(uint8_t * pix, int line_size){ |
688 | 405 const int h=16; |
406 int sum; | |
407 int index= -line_size*h; | |
408 | |
409 __asm __volatile( | |
410 "pxor %%mm7, %%mm7 \n\t" | |
411 "pxor %%mm6, %%mm6 \n\t" | |
412 "1: \n\t" | |
413 "movq (%2, %1), %%mm0 \n\t" | |
414 "movq (%2, %1), %%mm1 \n\t" | |
415 "movq 8(%2, %1), %%mm2 \n\t" | |
416 "movq 8(%2, %1), %%mm3 \n\t" | |
417 "punpcklbw %%mm7, %%mm0 \n\t" | |
418 "punpckhbw %%mm7, %%mm1 \n\t" | |
419 "punpcklbw %%mm7, %%mm2 \n\t" | |
420 "punpckhbw %%mm7, %%mm3 \n\t" | |
421 "paddw %%mm0, %%mm1 \n\t" | |
422 "paddw %%mm2, %%mm3 \n\t" | |
423 "paddw %%mm1, %%mm3 \n\t" | |
424 "paddw %%mm3, %%mm6 \n\t" | |
425 "addl %3, %1 \n\t" | |
426 " js 1b \n\t" | |
427 "movq %%mm6, %%mm5 \n\t" | |
428 "psrlq $32, %%mm6 \n\t" | |
429 "paddw %%mm5, %%mm6 \n\t" | |
430 "movq %%mm6, %%mm5 \n\t" | |
431 "psrlq $16, %%mm6 \n\t" | |
432 "paddw %%mm5, %%mm6 \n\t" | |
433 "movd %%mm6, %0 \n\t" | |
434 "andl $0xFFFF, %0 \n\t" | |
435 : "=&r" (sum), "+r" (index) | |
436 : "r" (pix - index), "r" (line_size) | |
437 ); | |
438 | |
439 return sum; | |
440 } | |
441 | |
866 | 442 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w){ |
443 int i=0; | |
444 asm volatile( | |
445 "1: \n\t" | |
446 "movq (%1, %0), %%mm0 \n\t" | |
447 "movq (%2, %0), %%mm1 \n\t" | |
448 "paddb %%mm0, %%mm1 \n\t" | |
449 "movq %%mm1, (%2, %0) \n\t" | |
450 "movq 8(%1, %0), %%mm0 \n\t" | |
451 "movq 8(%2, %0), %%mm1 \n\t" | |
452 "paddb %%mm0, %%mm1 \n\t" | |
453 "movq %%mm1, 8(%2, %0) \n\t" | |
454 "addl $16, %0 \n\t" | |
455 "cmpl %3, %0 \n\t" | |
456 " jb 1b \n\t" | |
457 : "+r" (i) | |
458 : "r"(src), "r"(dst), "r"(w-15) | |
459 ); | |
460 for(; i<w; i++) | |
461 dst[i+0] += src[i+0]; | |
462 } | |
463 | |
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
|
464 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
|
465 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
|
466 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
|
467 "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
|
468 "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
|
469 "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
|
470 "1:\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
|
471 "movq (%0),%%mm2\n" /* mm2 = pix[0-7] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
472 "movq 8(%0),%%mm3\n" /* mm3 = pix[8-15] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
473 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
474 "movq %%mm2,%%mm1\n" /* mm1 = mm2 = pix[0-7] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
475 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
476 "punpckhbw %%mm0,%%mm1\n" /* mm1 = [pix4-7] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
477 "punpcklbw %%mm0,%%mm2\n" /* mm2 = [pix0-3] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
478 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
479 "movq %%mm3,%%mm4\n" /* mm4 = mm3 = pix[8-15] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
480 "punpckhbw %%mm0,%%mm3\n" /* mm3 = [pix12-15] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
481 "punpcklbw %%mm0,%%mm4\n" /* mm4 = [pix8-11] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
482 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
483 "pmaddwd %%mm1,%%mm1\n" /* mm1 = (pix0^2+pix1^2,pix2^2+pix3^2) */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
484 "pmaddwd %%mm2,%%mm2\n" /* mm2 = (pix4^2+pix5^2,pix6^2+pix7^2) */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
485 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
486 "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
|
487 "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
|
488 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
489 "paddd %%mm1,%%mm2\n" /* mm2 = (pix0^2+pix1^2+pix4^2+pix5^2, |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
490 pix2^2+pix3^2+pix6^2+pix7^2) */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
491 "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
|
492 "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
|
493 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
494 "addl %2, %0\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
|
495 "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
|
496 "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
|
497 "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
|
498 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
499 "movq %%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
|
500 "psrlq $32, %%mm7\n" /* shift hi dword to lo */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
501 "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
|
502 "movd %%mm1,%1\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
|
503 : "+r" (pix), "=r"(tmp) : "r" (line_size) : "%ecx" ); |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
504 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
|
505 } |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
506 |
1064 | 507 static int sse16_mmx(void *v, uint8_t * pix1, uint8_t * pix2, int line_size) { |
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
|
508 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
|
509 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
|
510 "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
|
511 "pxor %%mm0,%%mm0\n" /* mm0 = 0 */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
512 "pxor %%mm7,%%mm7\n" /* mm7 holds the sum */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
513 "1:\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
|
514 "movq (%0),%%mm1\n" /* mm1 = pix1[0-7] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
515 "movq (%1),%%mm2\n" /* mm2 = pix2[0-7] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
516 "movq 8(%0),%%mm3\n" /* mm3 = pix1[8-15] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
517 "movq 8(%1),%%mm4\n" /* mm4 = pix2[8-15] */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
518 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
519 /* 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
|
520 /* 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
|
521 /* 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
|
522 "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
|
523 "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
|
524 "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
|
525 "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
|
526 "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
|
527 "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
|
528 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
529 "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
|
530 "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
|
531 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
532 /* 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
|
533 "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
|
534 "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
|
535 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
536 "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
|
537 "punpckhbw %%mm0,%%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
|
538 "punpcklbw %%mm0,%%mm1\n" /* mm1 now spread over (mm1,mm2) */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
539 "punpcklbw %%mm0,%%mm3\n" /* mm4 now spread over (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
|
540 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
541 "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
|
542 "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
|
543 "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
|
544 "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
|
545 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
546 "addl %3,%0\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
|
547 "addl %3,%1\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
|
548 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
549 "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
|
550 "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
|
551 "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
|
552 "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
|
553 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
554 "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
|
555 "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
|
556 |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
557 "movq %%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
|
558 "psrlq $32, %%mm7\n" /* shift hi dword to lo */ |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
559 "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
|
560 "movd %%mm1,%2\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
|
561 : "+r" (pix1), "+r" (pix2), "=r"(tmp) : "r" (line_size) : "ecx"); |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
562 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
|
563 } |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
564 |
866 | 565 static void diff_bytes_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){ |
566 int i=0; | |
567 asm volatile( | |
568 "1: \n\t" | |
569 "movq (%2, %0), %%mm0 \n\t" | |
570 "movq (%1, %0), %%mm1 \n\t" | |
571 "psubb %%mm0, %%mm1 \n\t" | |
572 "movq %%mm1, (%3, %0) \n\t" | |
573 "movq 8(%2, %0), %%mm0 \n\t" | |
574 "movq 8(%1, %0), %%mm1 \n\t" | |
575 "psubb %%mm0, %%mm1 \n\t" | |
576 "movq %%mm1, 8(%3, %0) \n\t" | |
577 "addl $16, %0 \n\t" | |
578 "cmpl %4, %0 \n\t" | |
579 " jb 1b \n\t" | |
580 : "+r" (i) | |
581 : "r"(src1), "r"(src2), "r"(dst), "r"(w-15) | |
582 ); | |
583 for(; i<w; i++) | |
584 dst[i+0] = src1[i+0]-src2[i+0]; | |
585 } | |
936 | 586 #define LBUTTERFLY(a,b)\ |
587 "paddw " #b ", " #a " \n\t"\ | |
588 "paddw " #b ", " #b " \n\t"\ | |
589 "psubw " #a ", " #b " \n\t" | |
866 | 590 |
936 | 591 #define HADAMARD48\ |
592 LBUTTERFLY(%%mm0, %%mm1)\ | |
593 LBUTTERFLY(%%mm2, %%mm3)\ | |
594 LBUTTERFLY(%%mm4, %%mm5)\ | |
595 LBUTTERFLY(%%mm6, %%mm7)\ | |
596 \ | |
597 LBUTTERFLY(%%mm0, %%mm2)\ | |
598 LBUTTERFLY(%%mm1, %%mm3)\ | |
599 LBUTTERFLY(%%mm4, %%mm6)\ | |
600 LBUTTERFLY(%%mm5, %%mm7)\ | |
601 \ | |
602 LBUTTERFLY(%%mm0, %%mm4)\ | |
603 LBUTTERFLY(%%mm1, %%mm5)\ | |
604 LBUTTERFLY(%%mm2, %%mm6)\ | |
605 LBUTTERFLY(%%mm3, %%mm7) | |
606 | |
607 #define MMABS(a,z)\ | |
608 "pxor " #z ", " #z " \n\t"\ | |
609 "pcmpgtw " #a ", " #z " \n\t"\ | |
610 "pxor " #z ", " #a " \n\t"\ | |
611 "psubw " #z ", " #a " \n\t" | |
612 | |
613 #define MMABS_SUM(a,z, sum)\ | |
614 "pxor " #z ", " #z " \n\t"\ | |
615 "pcmpgtw " #a ", " #z " \n\t"\ | |
616 "pxor " #z ", " #a " \n\t"\ | |
617 "psubw " #z ", " #a " \n\t"\ | |
618 "paddusw " #a ", " #sum " \n\t" | |
619 | |
620 | |
621 #define SBUTTERFLY(a,b,t,n)\ | |
622 "movq " #a ", " #t " \n\t" /* abcd */\ | |
623 "punpckl" #n " " #b ", " #a " \n\t" /* aebf */\ | |
624 "punpckh" #n " " #b ", " #t " \n\t" /* cgdh */\ | |
625 | |
626 #define TRANSPOSE4(a,b,c,d,t)\ | |
627 SBUTTERFLY(a,b,t,wd) /* a=aebf t=cgdh */\ | |
628 SBUTTERFLY(c,d,b,wd) /* c=imjn b=kolp */\ | |
629 SBUTTERFLY(a,c,d,dq) /* a=aeim d=bfjn */\ | |
630 SBUTTERFLY(t,b,c,dq) /* t=cgko c=dhlp */ | |
631 | |
632 #define LOAD4(o, a, b, c, d)\ | |
633 "movq "#o"(%1), " #a " \n\t"\ | |
634 "movq "#o"+16(%1), " #b " \n\t"\ | |
635 "movq "#o"+32(%1), " #c " \n\t"\ | |
636 "movq "#o"+48(%1), " #d " \n\t" | |
637 | |
638 #define STORE4(o, a, b, c, d)\ | |
639 "movq "#a", "#o"(%1) \n\t"\ | |
640 "movq "#b", "#o"+16(%1) \n\t"\ | |
641 "movq "#c", "#o"+32(%1) \n\t"\ | |
642 "movq "#d", "#o"+48(%1) \n\t"\ | |
643 | |
644 static int hadamard8_diff_mmx(void *s, uint8_t *src1, uint8_t *src2, int stride){ | |
645 uint64_t temp[16] __align8; | |
646 int sum=0; | |
647 | |
648 diff_pixels_mmx((DCTELEM*)temp, src1, src2, stride); | |
649 | |
650 asm volatile( | |
651 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3) | |
652 LOAD4(64, %%mm4, %%mm5, %%mm6, %%mm7) | |
653 | |
654 HADAMARD48 | |
655 | |
656 "movq %%mm7, 112(%1) \n\t" | |
657 | |
658 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7) | |
659 STORE4(0 , %%mm0, %%mm3, %%mm7, %%mm2) | |
660 | |
661 "movq 112(%1), %%mm7 \n\t" | |
662 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0) | |
663 STORE4(64, %%mm4, %%mm7, %%mm0, %%mm6) | |
664 | |
665 LOAD4(8 , %%mm0, %%mm1, %%mm2, %%mm3) | |
666 LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7) | |
667 | |
668 HADAMARD48 | |
669 | |
670 "movq %%mm7, 120(%1) \n\t" | |
671 | |
672 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm7) | |
673 STORE4(8 , %%mm0, %%mm3, %%mm7, %%mm2) | |
674 | |
675 "movq 120(%1), %%mm7 \n\t" | |
676 TRANSPOSE4(%%mm4, %%mm5, %%mm6, %%mm7, %%mm0) | |
677 "movq %%mm7, %%mm5 \n\t"//FIXME remove | |
678 "movq %%mm6, %%mm7 \n\t" | |
679 "movq %%mm0, %%mm6 \n\t" | |
680 // STORE4(72, %%mm4, %%mm7, %%mm0, %%mm6) //FIXME remove | |
681 | |
682 LOAD4(64, %%mm0, %%mm1, %%mm2, %%mm3) | |
683 // LOAD4(72, %%mm4, %%mm5, %%mm6, %%mm7) | |
684 | |
685 HADAMARD48 | |
686 "movq %%mm7, 64(%1) \n\t" | |
687 MMABS(%%mm0, %%mm7) | |
688 MMABS_SUM(%%mm1, %%mm7, %%mm0) | |
689 MMABS_SUM(%%mm2, %%mm7, %%mm0) | |
690 MMABS_SUM(%%mm3, %%mm7, %%mm0) | |
691 MMABS_SUM(%%mm4, %%mm7, %%mm0) | |
692 MMABS_SUM(%%mm5, %%mm7, %%mm0) | |
693 MMABS_SUM(%%mm6, %%mm7, %%mm0) | |
694 "movq 64(%1), %%mm1 \n\t" | |
695 MMABS_SUM(%%mm1, %%mm7, %%mm0) | |
696 "movq %%mm0, 64(%1) \n\t" | |
697 | |
698 LOAD4(0 , %%mm0, %%mm1, %%mm2, %%mm3) | |
699 LOAD4(8 , %%mm4, %%mm5, %%mm6, %%mm7) | |
700 | |
701 HADAMARD48 | |
702 "movq %%mm7, (%1) \n\t" | |
703 MMABS(%%mm0, %%mm7) | |
704 MMABS_SUM(%%mm1, %%mm7, %%mm0) | |
705 MMABS_SUM(%%mm2, %%mm7, %%mm0) | |
706 MMABS_SUM(%%mm3, %%mm7, %%mm0) | |
707 MMABS_SUM(%%mm4, %%mm7, %%mm0) | |
708 MMABS_SUM(%%mm5, %%mm7, %%mm0) | |
709 MMABS_SUM(%%mm6, %%mm7, %%mm0) | |
710 "movq (%1), %%mm1 \n\t" | |
711 MMABS_SUM(%%mm1, %%mm7, %%mm0) | |
712 "movq 64(%1), %%mm1 \n\t" | |
713 MMABS_SUM(%%mm1, %%mm7, %%mm0) | |
714 | |
715 "movq %%mm0, %%mm1 \n\t" | |
716 "psrlq $32, %%mm0 \n\t" | |
717 "paddusw %%mm1, %%mm0 \n\t" | |
718 "movq %%mm0, %%mm1 \n\t" | |
719 "psrlq $16, %%mm0 \n\t" | |
720 "paddusw %%mm1, %%mm0 \n\t" | |
721 "movd %%mm0, %0 \n\t" | |
722 | |
723 : "=r" (sum) | |
724 : "r"(temp) | |
725 ); | |
726 return sum&0xFFFF; | |
727 } | |
728 | |
729 WARPER88_1616(hadamard8_diff_mmx, hadamard8_diff16_mmx) | |
866 | 730 |
959 | 731 #define put_no_rnd_pixels8_mmx(a,b,c,d) put_pixels8_mmx(a,b,c,d) |
732 #define put_no_rnd_pixels16_mmx(a,b,c,d) put_pixels16_mmx(a,b,c,d) | |
733 | |
954 | 734 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\ |
735 "paddw " #m4 ", " #m3 " \n\t" /* x1 */\ | |
961 | 736 "movq "MANGLE(ff_pw_20)", %%mm4 \n\t" /* 20 */\ |
954 | 737 "pmullw " #m3 ", %%mm4 \n\t" /* 20x1 */\ |
738 "movq "#in7", " #m3 " \n\t" /* d */\ | |
739 "movq "#in0", %%mm5 \n\t" /* D */\ | |
740 "paddw " #m3 ", %%mm5 \n\t" /* x4 */\ | |
741 "psubw %%mm5, %%mm4 \n\t" /* 20x1 - x4 */\ | |
742 "movq "#in1", %%mm5 \n\t" /* C */\ | |
743 "movq "#in2", %%mm6 \n\t" /* B */\ | |
744 "paddw " #m6 ", %%mm5 \n\t" /* x3 */\ | |
745 "paddw " #m5 ", %%mm6 \n\t" /* x2 */\ | |
746 "paddw %%mm6, %%mm6 \n\t" /* 2x2 */\ | |
747 "psubw %%mm6, %%mm5 \n\t" /* -2x2 + x3 */\ | |
961 | 748 "pmullw "MANGLE(ff_pw_3)", %%mm5 \n\t" /* -6x2 + 3x3 */\ |
954 | 749 "paddw " #rnd ", %%mm4 \n\t" /* x2 */\ |
750 "paddw %%mm4, %%mm5 \n\t" /* 20x1 - 6x2 + 3x3 - x4 */\ | |
751 "psraw $5, %%mm5 \n\t"\ | |
752 "packuswb %%mm5, %%mm5 \n\t"\ | |
753 OP(%%mm5, out, %%mm7, d) | |
754 | |
959 | 755 #define QPEL_BASE(OPNAME, ROUNDER, RND, OP_MMX2, OP_3DNOW)\ |
1057 | 756 static void OPNAME ## mpeg4_qpel16_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
954 | 757 uint64_t temp;\ |
758 \ | |
759 asm volatile(\ | |
760 "pxor %%mm7, %%mm7 \n\t"\ | |
761 "1: \n\t"\ | |
762 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
763 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
764 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
765 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
766 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
767 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
768 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
769 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
770 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
771 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
772 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
773 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
774 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
775 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
776 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
777 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
778 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
779 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
780 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
781 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
961 | 782 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ |
954 | 783 "paddw %%mm4, %%mm0 \n\t" /* a */\ |
784 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
961 | 785 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ |
954 | 786 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ |
961 | 787 "paddw %6, %%mm6 \n\t"\ |
954 | 788 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ |
789 "psraw $5, %%mm0 \n\t"\ | |
961 | 790 "movq %%mm0, %5 \n\t"\ |
954 | 791 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ |
792 \ | |
793 "movq 5(%0), %%mm0 \n\t" /* FGHIJKLM */\ | |
794 "movq %%mm0, %%mm5 \n\t" /* FGHIJKLM */\ | |
795 "movq %%mm0, %%mm6 \n\t" /* FGHIJKLM */\ | |
796 "psrlq $8, %%mm0 \n\t" /* GHIJKLM0 */\ | |
797 "psrlq $16, %%mm5 \n\t" /* HIJKLM00 */\ | |
798 "punpcklbw %%mm7, %%mm0 \n\t" /* 0G0H0I0J */\ | |
799 "punpcklbw %%mm7, %%mm5 \n\t" /* 0H0I0J0K */\ | |
800 "paddw %%mm0, %%mm2 \n\t" /* b */\ | |
801 "paddw %%mm5, %%mm3 \n\t" /* c */\ | |
802 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
803 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
804 "movq %%mm6, %%mm2 \n\t" /* FGHIJKLM */\ | |
805 "psrlq $24, %%mm6 \n\t" /* IJKLM000 */\ | |
806 "punpcklbw %%mm7, %%mm2 \n\t" /* 0F0G0H0I */\ | |
807 "punpcklbw %%mm7, %%mm6 \n\t" /* 0I0J0K0L */\ | |
961 | 808 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ |
954 | 809 "paddw %%mm2, %%mm1 \n\t" /* a */\ |
810 "paddw %%mm6, %%mm4 \n\t" /* d */\ | |
961 | 811 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ |
954 | 812 "psubw %%mm4, %%mm3 \n\t" /* - 6b +3c - d */\ |
961 | 813 "paddw %6, %%mm1 \n\t"\ |
954 | 814 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b +3c - d */\ |
815 "psraw $5, %%mm3 \n\t"\ | |
961 | 816 "movq %5, %%mm1 \n\t"\ |
954 | 817 "packuswb %%mm3, %%mm1 \n\t"\ |
959 | 818 OP_MMX2(%%mm1, (%1),%%mm4, q)\ |
954 | 819 /* mm0= GHIJ, mm2=FGHI, mm5=HIJK, mm6=IJKL, mm7=0 */\ |
820 \ | |
821 "movq 9(%0), %%mm1 \n\t" /* JKLMNOPQ */\ | |
822 "movq %%mm1, %%mm4 \n\t" /* JKLMNOPQ */\ | |
823 "movq %%mm1, %%mm3 \n\t" /* JKLMNOPQ */\ | |
824 "psrlq $8, %%mm1 \n\t" /* KLMNOPQ0 */\ | |
825 "psrlq $16, %%mm4 \n\t" /* LMNOPQ00 */\ | |
826 "punpcklbw %%mm7, %%mm1 \n\t" /* 0K0L0M0N */\ | |
827 "punpcklbw %%mm7, %%mm4 \n\t" /* 0L0M0N0O */\ | |
828 "paddw %%mm1, %%mm5 \n\t" /* b */\ | |
829 "paddw %%mm4, %%mm0 \n\t" /* c */\ | |
830 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
831 "psubw %%mm5, %%mm0 \n\t" /* c - 2b */\ | |
832 "movq %%mm3, %%mm5 \n\t" /* JKLMNOPQ */\ | |
833 "psrlq $24, %%mm3 \n\t" /* MNOPQ000 */\ | |
961 | 834 "pmullw "MANGLE(ff_pw_3)", %%mm0 \n\t" /* 3c - 6b */\ |
954 | 835 "punpcklbw %%mm7, %%mm3 \n\t" /* 0M0N0O0P */\ |
836 "paddw %%mm3, %%mm2 \n\t" /* d */\ | |
837 "psubw %%mm2, %%mm0 \n\t" /* -6b + 3c - d */\ | |
838 "movq %%mm5, %%mm2 \n\t" /* JKLMNOPQ */\ | |
839 "punpcklbw %%mm7, %%mm2 \n\t" /* 0J0K0L0M */\ | |
840 "punpckhbw %%mm7, %%mm5 \n\t" /* 0N0O0P0Q */\ | |
841 "paddw %%mm2, %%mm6 \n\t" /* a */\ | |
961 | 842 "pmullw "MANGLE(ff_pw_20)", %%mm6 \n\t" /* 20a */\ |
843 "paddw %6, %%mm0 \n\t"\ | |
954 | 844 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ |
845 "psraw $5, %%mm0 \n\t"\ | |
846 /* mm1=KLMN, mm2=JKLM, mm3=MNOP, mm4=LMNO, mm5=NOPQ mm7=0 */\ | |
847 \ | |
848 "paddw %%mm5, %%mm3 \n\t" /* a */\ | |
849 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0O0P0Q0Q */\ | |
850 "paddw %%mm4, %%mm6 \n\t" /* b */\ | |
851 "pshufw $0xBE, %%mm5, %%mm4 \n\t" /* 0P0Q0Q0P */\ | |
852 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0Q0Q0P0O */\ | |
853 "paddw %%mm1, %%mm4 \n\t" /* c */\ | |
854 "paddw %%mm2, %%mm5 \n\t" /* d */\ | |
855 "paddw %%mm6, %%mm6 \n\t" /* 2b */\ | |
856 "psubw %%mm6, %%mm4 \n\t" /* c - 2b */\ | |
961 | 857 "pmullw "MANGLE(ff_pw_20)", %%mm3 \n\t" /* 20a */\ |
858 "pmullw "MANGLE(ff_pw_3)", %%mm4 \n\t" /* 3c - 6b */\ | |
954 | 859 "psubw %%mm5, %%mm3 \n\t" /* -6b + 3c - d */\ |
961 | 860 "paddw %6, %%mm4 \n\t"\ |
954 | 861 "paddw %%mm3, %%mm4 \n\t" /* 20a - 6b + 3c - d */\ |
862 "psraw $5, %%mm4 \n\t"\ | |
863 "packuswb %%mm4, %%mm0 \n\t"\ | |
959 | 864 OP_MMX2(%%mm0, 8(%1), %%mm4, q)\ |
954 | 865 \ |
866 "addl %3, %0 \n\t"\ | |
867 "addl %4, %1 \n\t"\ | |
868 "decl %2 \n\t"\ | |
869 " jnz 1b \n\t"\ | |
967 | 870 : "+a"(src), "+c"(dst), "+m"(h)\ |
966 | 871 : "d"(srcStride), "S"(dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
872 : "memory"\ | |
954 | 873 );\ |
874 }\ | |
875 \ | |
876 static void OPNAME ## mpeg4_qpel16_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
877 int i;\ | |
878 int16_t temp[16];\ | |
879 /* quick HACK, XXX FIXME MUST be optimized */\ | |
880 for(i=0; i<h; i++)\ | |
881 {\ | |
882 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
883 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
884 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
885 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
886 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
887 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]);\ | |
888 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]);\ | |
889 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]);\ | |
890 temp[ 8]= (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]);\ | |
891 temp[ 9]= (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]);\ | |
892 temp[10]= (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]);\ | |
893 temp[11]= (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]);\ | |
894 temp[12]= (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]);\ | |
895 temp[13]= (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]);\ | |
896 temp[14]= (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]);\ | |
897 temp[15]= (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]);\ | |
898 asm volatile(\ | |
899 "movq (%0), %%mm0 \n\t"\ | |
900 "movq 8(%0), %%mm1 \n\t"\ | |
901 "paddw %2, %%mm0 \n\t"\ | |
902 "paddw %2, %%mm1 \n\t"\ | |
903 "psraw $5, %%mm0 \n\t"\ | |
904 "psraw $5, %%mm1 \n\t"\ | |
905 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 906 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ |
954 | 907 "movq 16(%0), %%mm0 \n\t"\ |
908 "movq 24(%0), %%mm1 \n\t"\ | |
909 "paddw %2, %%mm0 \n\t"\ | |
910 "paddw %2, %%mm1 \n\t"\ | |
911 "psraw $5, %%mm0 \n\t"\ | |
912 "psraw $5, %%mm1 \n\t"\ | |
913 "packuswb %%mm1, %%mm0 \n\t"\ | |
959 | 914 OP_3DNOW(%%mm0, 8(%1), %%mm1, q)\ |
954 | 915 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ |
966 | 916 : "memory"\ |
954 | 917 );\ |
918 dst+=dstStride;\ | |
919 src+=srcStride;\ | |
920 }\ | |
921 }\ | |
922 \ | |
1057 | 923 static void OPNAME ## mpeg4_qpel8_h_lowpass_mmx2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
959 | 924 uint64_t temp;\ |
925 \ | |
926 asm volatile(\ | |
927 "pxor %%mm7, %%mm7 \n\t"\ | |
928 "1: \n\t"\ | |
929 "movq (%0), %%mm0 \n\t" /* ABCDEFGH */\ | |
930 "movq %%mm0, %%mm1 \n\t" /* ABCDEFGH */\ | |
931 "movq %%mm0, %%mm2 \n\t" /* ABCDEFGH */\ | |
932 "punpcklbw %%mm7, %%mm0 \n\t" /* 0A0B0C0D */\ | |
933 "punpckhbw %%mm7, %%mm1 \n\t" /* 0E0F0G0H */\ | |
934 "pshufw $0x90, %%mm0, %%mm5 \n\t" /* 0A0A0B0C */\ | |
935 "pshufw $0x41, %%mm0, %%mm6 \n\t" /* 0B0A0A0B */\ | |
936 "movq %%mm2, %%mm3 \n\t" /* ABCDEFGH */\ | |
937 "movq %%mm2, %%mm4 \n\t" /* ABCDEFGH */\ | |
938 "psllq $8, %%mm2 \n\t" /* 0ABCDEFG */\ | |
939 "psllq $16, %%mm3 \n\t" /* 00ABCDEF */\ | |
940 "psllq $24, %%mm4 \n\t" /* 000ABCDE */\ | |
941 "punpckhbw %%mm7, %%mm2 \n\t" /* 0D0E0F0G */\ | |
942 "punpckhbw %%mm7, %%mm3 \n\t" /* 0C0D0E0F */\ | |
943 "punpckhbw %%mm7, %%mm4 \n\t" /* 0B0C0D0E */\ | |
944 "paddw %%mm3, %%mm5 \n\t" /* b */\ | |
945 "paddw %%mm2, %%mm6 \n\t" /* c */\ | |
946 "paddw %%mm5, %%mm5 \n\t" /* 2b */\ | |
947 "psubw %%mm5, %%mm6 \n\t" /* c - 2b */\ | |
948 "pshufw $0x06, %%mm0, %%mm5 \n\t" /* 0C0B0A0A */\ | |
961 | 949 "pmullw "MANGLE(ff_pw_3)", %%mm6 \n\t" /* 3c - 6b */\ |
959 | 950 "paddw %%mm4, %%mm0 \n\t" /* a */\ |
951 "paddw %%mm1, %%mm5 \n\t" /* d */\ | |
961 | 952 "pmullw "MANGLE(ff_pw_20)", %%mm0 \n\t" /* 20a */\ |
959 | 953 "psubw %%mm5, %%mm0 \n\t" /* 20a - d */\ |
961 | 954 "paddw %6, %%mm6 \n\t"\ |
959 | 955 "paddw %%mm6, %%mm0 \n\t" /* 20a - 6b + 3c - d */\ |
956 "psraw $5, %%mm0 \n\t"\ | |
957 /* mm1=EFGH, mm2=DEFG, mm3=CDEF, mm4=BCDE, mm7=0 */\ | |
958 \ | |
959 "movd 5(%0), %%mm5 \n\t" /* FGHI */\ | |
960 "punpcklbw %%mm7, %%mm5 \n\t" /* 0F0G0H0I */\ | |
961 "pshufw $0xF9, %%mm5, %%mm6 \n\t" /* 0G0H0I0I */\ | |
962 "paddw %%mm5, %%mm1 \n\t" /* a */\ | |
963 "paddw %%mm6, %%mm2 \n\t" /* b */\ | |
964 "pshufw $0xBE, %%mm5, %%mm6 \n\t" /* 0H0I0I0H */\ | |
965 "pshufw $0x6F, %%mm5, %%mm5 \n\t" /* 0I0I0H0G */\ | |
966 "paddw %%mm6, %%mm3 \n\t" /* c */\ | |
967 "paddw %%mm5, %%mm4 \n\t" /* d */\ | |
968 "paddw %%mm2, %%mm2 \n\t" /* 2b */\ | |
969 "psubw %%mm2, %%mm3 \n\t" /* c - 2b */\ | |
961 | 970 "pmullw "MANGLE(ff_pw_20)", %%mm1 \n\t" /* 20a */\ |
971 "pmullw "MANGLE(ff_pw_3)", %%mm3 \n\t" /* 3c - 6b */\ | |
959 | 972 "psubw %%mm4, %%mm3 \n\t" /* -6b + 3c - d */\ |
961 | 973 "paddw %6, %%mm1 \n\t"\ |
959 | 974 "paddw %%mm1, %%mm3 \n\t" /* 20a - 6b + 3c - d */\ |
975 "psraw $5, %%mm3 \n\t"\ | |
976 "packuswb %%mm3, %%mm0 \n\t"\ | |
977 OP_MMX2(%%mm0, (%1), %%mm4, q)\ | |
978 \ | |
979 "addl %3, %0 \n\t"\ | |
980 "addl %4, %1 \n\t"\ | |
981 "decl %2 \n\t"\ | |
961 | 982 " jnz 1b \n\t"\ |
967 | 983 : "+a"(src), "+c"(dst), "+m"(h)\ |
966 | 984 : "S"(srcStride), "D"(dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(temp), "m"(ROUNDER)\ |
985 : "memory"\ | |
959 | 986 );\ |
987 }\ | |
988 \ | |
989 static void OPNAME ## mpeg4_qpel8_h_lowpass_3dnow(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ | |
990 int i;\ | |
991 int16_t temp[8];\ | |
992 /* quick HACK, XXX FIXME MUST be optimized */\ | |
993 for(i=0; i<h; i++)\ | |
994 {\ | |
995 temp[ 0]= (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]);\ | |
996 temp[ 1]= (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]);\ | |
997 temp[ 2]= (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]);\ | |
998 temp[ 3]= (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]);\ | |
999 temp[ 4]= (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]);\ | |
1000 temp[ 5]= (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 8]);\ | |
1001 temp[ 6]= (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 8])*3 - (src[ 3]+src[ 7]);\ | |
1002 temp[ 7]= (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 8])*6 + (src[ 5]+src[ 7])*3 - (src[ 4]+src[ 6]);\ | |
1003 asm volatile(\ | |
1004 "movq (%0), %%mm0 \n\t"\ | |
1005 "movq 8(%0), %%mm1 \n\t"\ | |
1006 "paddw %2, %%mm0 \n\t"\ | |
1007 "paddw %2, %%mm1 \n\t"\ | |
1008 "psraw $5, %%mm0 \n\t"\ | |
1009 "psraw $5, %%mm1 \n\t"\ | |
1010 "packuswb %%mm1, %%mm0 \n\t"\ | |
1011 OP_3DNOW(%%mm0, (%1), %%mm1, q)\ | |
1012 :: "r"(temp), "r"(dst), "m"(ROUNDER)\ | |
966 | 1013 :"memory"\ |
959 | 1014 );\ |
1015 dst+=dstStride;\ | |
1016 src+=srcStride;\ | |
1017 }\ | |
1018 } | |
1019 | |
1020 #define QPEL_OP(OPNAME, ROUNDER, RND, OP, MMX)\ | |
1021 \ | |
1022 static void OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
954 | 1023 uint64_t temp[17*4];\ |
1024 uint64_t *temp_ptr= temp;\ | |
1025 int count= 17;\ | |
1026 \ | |
1027 /*FIXME unroll */\ | |
1028 asm volatile(\ | |
1029 "pxor %%mm7, %%mm7 \n\t"\ | |
1030 "1: \n\t"\ | |
1031 "movq (%0), %%mm0 \n\t"\ | |
1032 "movq (%0), %%mm1 \n\t"\ | |
1033 "movq 8(%0), %%mm2 \n\t"\ | |
1034 "movq 8(%0), %%mm3 \n\t"\ | |
1035 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1036 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1037 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1038 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1039 "movq %%mm0, (%1) \n\t"\ | |
1040 "movq %%mm1, 17*8(%1) \n\t"\ | |
967 | 1041 "movq %%mm2, 2*17*8(%1) \n\t"\ |
1042 "movq %%mm3, 3*17*8(%1) \n\t"\ | |
954 | 1043 "addl $8, %1 \n\t"\ |
1044 "addl %3, %0 \n\t"\ | |
1045 "decl %2 \n\t"\ | |
1046 " jnz 1b \n\t"\ | |
1047 : "+r" (src), "+r" (temp_ptr), "+r"(count)\ | |
967 | 1048 : "r" (srcStride)\ |
966 | 1049 : "memory"\ |
954 | 1050 );\ |
1051 \ | |
1052 temp_ptr= temp;\ | |
1053 count=4;\ | |
1054 \ | |
1055 /*FIXME reorder for speed */\ | |
1056 asm volatile(\ | |
1057 /*"pxor %%mm7, %%mm7 \n\t"*/\ | |
1058 "1: \n\t"\ | |
1059 "movq (%0), %%mm0 \n\t"\ | |
1060 "movq 8(%0), %%mm1 \n\t"\ | |
1061 "movq 16(%0), %%mm2 \n\t"\ | |
1062 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 1063 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
1064 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
954 | 1065 "addl %4, %1 \n\t"\ |
961 | 1066 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 1067 \ |
961 | 1068 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
954 | 1069 "addl %4, %1 \n\t"\ |
961 | 1070 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
1071 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 72(%0), (%1, %3), OP)\ | |
954 | 1072 "addl %4, %1 \n\t"\ |
961 | 1073 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 80(%0), (%1), OP)\ |
1074 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 88(%0), (%1, %3), OP)\ | |
954 | 1075 "addl %4, %1 \n\t"\ |
961 | 1076 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 40(%0), 48(%0), 56(%0), 96(%0), (%1), OP)\ |
1077 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 48(%0), 56(%0), 64(%0),104(%0), (%1, %3), OP)\ | |
954 | 1078 "addl %4, %1 \n\t"\ |
961 | 1079 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 56(%0), 64(%0), 72(%0),112(%0), (%1), OP)\ |
1080 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 64(%0), 72(%0), 80(%0),120(%0), (%1, %3), OP)\ | |
954 | 1081 "addl %4, %1 \n\t"\ |
961 | 1082 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 72(%0), 80(%0), 88(%0),128(%0), (%1), OP)\ |
954 | 1083 \ |
961 | 1084 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 80(%0), 88(%0), 96(%0),128(%0), (%1, %3), OP)\ |
954 | 1085 "addl %4, %1 \n\t" \ |
961 | 1086 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 88(%0), 96(%0),104(%0),120(%0), (%1), OP)\ |
1087 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 96(%0),104(%0),112(%0),112(%0), (%1, %3), OP)\ | |
954 | 1088 \ |
1089 "addl $136, %0 \n\t"\ | |
961 | 1090 "addl %6, %1 \n\t"\ |
954 | 1091 "decl %2 \n\t"\ |
1092 " 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
|
1093 \ |
967 | 1094 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\ |
961 | 1095 : "r"(dstStride), "r"(2*dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-14*dstStride)\ |
966 | 1096 :"memory"\ |
954 | 1097 );\ |
1098 }\ | |
1099 \ | |
1057 | 1100 static void OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
954 | 1101 uint64_t temp[9*4];\ |
1102 uint64_t *temp_ptr= temp;\ | |
1103 int count= 9;\ | |
1104 \ | |
1105 /*FIXME unroll */\ | |
1106 asm volatile(\ | |
1107 "pxor %%mm7, %%mm7 \n\t"\ | |
1108 "1: \n\t"\ | |
1109 "movq (%0), %%mm0 \n\t"\ | |
1110 "movq (%0), %%mm1 \n\t"\ | |
1111 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1112 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1113 "movq %%mm0, (%1) \n\t"\ | |
1114 "movq %%mm1, 9*8(%1) \n\t"\ | |
1115 "addl $8, %1 \n\t"\ | |
1116 "addl %3, %0 \n\t"\ | |
1117 "decl %2 \n\t"\ | |
1118 " jnz 1b \n\t"\ | |
1119 : "+r" (src), "+r" (temp_ptr), "+r"(count)\ | |
1120 : "r" (srcStride)\ | |
966 | 1121 : "memory"\ |
954 | 1122 );\ |
1123 \ | |
1124 temp_ptr= temp;\ | |
1125 count=2;\ | |
1126 \ | |
1127 /*FIXME reorder for speed */\ | |
1128 asm volatile(\ | |
1129 /*"pxor %%mm7, %%mm7 \n\t"*/\ | |
1130 "1: \n\t"\ | |
1131 "movq (%0), %%mm0 \n\t"\ | |
1132 "movq 8(%0), %%mm1 \n\t"\ | |
1133 "movq 16(%0), %%mm2 \n\t"\ | |
1134 "movq 24(%0), %%mm3 \n\t"\ | |
961 | 1135 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 16(%0), 8(%0), (%0), 32(%0), (%1), OP)\ |
1136 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 8(%0), (%0), (%0), 40(%0), (%1, %3), OP)\ | |
954 | 1137 "addl %4, %1 \n\t"\ |
961 | 1138 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, (%0), (%0), 8(%0), 48(%0), (%1), OP)\ |
954 | 1139 \ |
961 | 1140 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, (%0), 8(%0), 16(%0), 56(%0), (%1, %3), OP)\ |
954 | 1141 "addl %4, %1 \n\t"\ |
961 | 1142 QPEL_V_LOW(%%mm0, %%mm1, %%mm2, %%mm3, %5, %6, %5, 8(%0), 16(%0), 24(%0), 64(%0), (%1), OP)\ |
954 | 1143 \ |
961 | 1144 QPEL_V_LOW(%%mm1, %%mm2, %%mm3, %%mm0, %5, %6, %5, 16(%0), 24(%0), 32(%0), 64(%0), (%1, %3), OP)\ |
954 | 1145 "addl %4, %1 \n\t"\ |
961 | 1146 QPEL_V_LOW(%%mm2, %%mm3, %%mm0, %%mm1, %5, %6, %5, 24(%0), 32(%0), 40(%0), 56(%0), (%1), OP)\ |
1147 QPEL_V_LOW(%%mm3, %%mm0, %%mm1, %%mm2, %5, %6, %5, 32(%0), 40(%0), 48(%0), 48(%0), (%1, %3), OP)\ | |
954 | 1148 \ |
1149 "addl $72, %0 \n\t"\ | |
961 | 1150 "addl %6, %1 \n\t"\ |
954 | 1151 "decl %2 \n\t"\ |
1152 " jnz 1b \n\t"\ | |
1153 \ | |
961 | 1154 : "+r"(temp_ptr), "+r"(dst), "+g"(count)\ |
1155 : "r"(dstStride), "r"(2*dstStride), /*"m"(ff_pw_20), "m"(ff_pw_3),*/ "m"(ROUNDER), "g"(4-6*dstStride)\ | |
966 | 1156 : "memory"\ |
1157 );\ | |
959 | 1158 }\ |
954 | 1159 \ |
1064 | 1160 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
959 | 1161 OPNAME ## pixels8_mmx(dst, src, stride, 8);\ |
954 | 1162 }\ |
1163 \ | |
1064 | 1164 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1165 uint64_t temp[8];\ |
954 | 1166 uint8_t * const half= (uint8_t*)temp;\ |
1167 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\ | |
1168 OPNAME ## pixels8_l2_mmx(dst, src, half, stride, stride, 8);\ | |
1169 }\ | |
1170 \ | |
1064 | 1171 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1172 OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, stride, 8);\ |
1173 }\ | |
1174 \ | |
1064 | 1175 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1176 uint64_t temp[8];\ |
954 | 1177 uint8_t * const half= (uint8_t*)temp;\ |
1178 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\ | |
1179 OPNAME ## pixels8_l2_mmx(dst, src+1, half, stride, stride, 8);\ | |
1180 }\ | |
1181 \ | |
1064 | 1182 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1183 uint64_t temp[8];\ |
954 | 1184 uint8_t * const half= (uint8_t*)temp;\ |
959 | 1185 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\ |
954 | 1186 OPNAME ## pixels8_l2_mmx(dst, src, half, stride, stride, 8);\ |
1187 }\ | |
1188 \ | |
1064 | 1189 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 1190 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 1191 }\ |
1192 \ | |
1064 | 1193 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1194 uint64_t temp[8];\ |
954 | 1195 uint8_t * const half= (uint8_t*)temp;\ |
959 | 1196 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\ |
954 | 1197 OPNAME ## pixels8_l2_mmx(dst, src+stride, half, stride, stride, 8);\ |
1198 }\ | |
1064 | 1199 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1200 uint64_t half[8 + 9];\ |
1201 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1202 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1203 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
984 | 1204 put ## RND ## pixels8_l2_mmx(halfH, src, halfH, 8, stride, 9);\ |
959 | 1205 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
984 | 1206 OPNAME ## pixels8_l2_mmx(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 1207 }\ |
1064 | 1208 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1209 uint64_t half[8 + 9];\ |
1210 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1211 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1212 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
984 | 1213 put ## RND ## pixels8_l2_mmx(halfH, src+1, halfH, 8, stride, 9);\ |
959 | 1214 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
984 | 1215 OPNAME ## pixels8_l2_mmx(dst, halfH, halfHV, stride, 8, 8);\ |
954 | 1216 }\ |
1064 | 1217 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1218 uint64_t half[8 + 9];\ |
1219 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1220 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1221 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
984 | 1222 put ## RND ## pixels8_l2_mmx(halfH, src, halfH, 8, stride, 9);\ |
959 | 1223 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
984 | 1224 OPNAME ## pixels8_l2_mmx(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 1225 }\ |
1064 | 1226 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1227 uint64_t half[8 + 9];\ |
1228 uint8_t * const halfH= ((uint8_t*)half) + 64;\ | |
1229 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1230 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
1231 put ## RND ## pixels8_l2_mmx(halfH, src+1, halfH, 8, stride, 9);\ | |
959 | 1232 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
984 | 1233 OPNAME ## pixels8_l2_mmx(dst, halfH+8, halfHV, stride, 8, 8);\ |
954 | 1234 }\ |
1064 | 1235 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1236 uint64_t half[8 + 9];\ |
954 | 1237 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
1238 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1239 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 1240 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
954 | 1241 OPNAME ## pixels8_l2_mmx(dst, halfH, halfHV, stride, 8, 8);\ |
1242 }\ | |
1064 | 1243 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1244 uint64_t half[8 + 9];\ |
954 | 1245 uint8_t * const halfH= ((uint8_t*)half) + 64;\ |
1246 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1247 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 1248 put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ |
954 | 1249 OPNAME ## pixels8_l2_mmx(dst, halfH+8, halfHV, stride, 8, 8);\ |
1250 }\ | |
1064 | 1251 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1252 uint64_t half[8 + 9];\ |
1253 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 1254 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
984 | 1255 put ## RND ## pixels8_l2_mmx(halfH, src, halfH, 8, stride, 9);\ |
1256 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ | |
954 | 1257 }\ |
1064 | 1258 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1259 uint64_t half[8 + 9];\ |
1260 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 1261 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ |
984 | 1262 put ## RND ## pixels8_l2_mmx(halfH, src+1, halfH, 8, stride, 9);\ |
1263 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ | |
954 | 1264 }\ |
1064 | 1265 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1266 uint64_t half[9];\ |
954 | 1267 uint8_t * const halfH= ((uint8_t*)half);\ |
1268 put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\ | |
959 | 1269 OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\ |
954 | 1270 }\ |
1064 | 1271 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
959 | 1272 OPNAME ## pixels16_mmx(dst, src, stride, 16);\ |
954 | 1273 }\ |
1274 \ | |
1064 | 1275 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1276 uint64_t temp[32];\ |
1277 uint8_t * const half= (uint8_t*)temp;\ | |
1278 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\ | |
1279 OPNAME ## pixels16_l2_mmx(dst, src, half, stride, stride, 16);\ | |
1280 }\ | |
1281 \ | |
1064 | 1282 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1283 OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, stride, stride, 16);\ |
1284 }\ | |
1285 \ | |
1064 | 1286 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1287 uint64_t temp[32];\ |
1288 uint8_t * const half= (uint8_t*)temp;\ | |
1289 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\ | |
1290 OPNAME ## pixels16_l2_mmx(dst, src+1, half, stride, stride, 16);\ | |
1291 }\ | |
1292 \ | |
1064 | 1293 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1294 uint64_t temp[32];\ |
1295 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 1296 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\ |
954 | 1297 OPNAME ## pixels16_l2_mmx(dst, src, half, stride, stride, 16);\ |
1298 }\ | |
1299 \ | |
1064 | 1300 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
959 | 1301 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, stride, stride);\ |
954 | 1302 }\ |
1303 \ | |
1064 | 1304 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1305 uint64_t temp[32];\ |
1306 uint8_t * const half= (uint8_t*)temp;\ | |
959 | 1307 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\ |
954 | 1308 OPNAME ## pixels16_l2_mmx(dst, src+stride, half, stride, stride, 16);\ |
1309 }\ | |
1064 | 1310 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1311 uint64_t half[16*2 + 17*2];\ |
1312 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1313 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1314 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
984 | 1315 put ## RND ## pixels16_l2_mmx(halfH, src, halfH, 16, stride, 17);\ |
959 | 1316 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
984 | 1317 OPNAME ## pixels16_l2_mmx(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 1318 }\ |
1064 | 1319 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1320 uint64_t half[16*2 + 17*2];\ |
1321 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1322 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1323 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
984 | 1324 put ## RND ## pixels16_l2_mmx(halfH, src+1, halfH, 16, stride, 17);\ |
959 | 1325 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
984 | 1326 OPNAME ## pixels16_l2_mmx(dst, halfH, halfHV, stride, 16, 16);\ |
954 | 1327 }\ |
1064 | 1328 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1329 uint64_t half[16*2 + 17*2];\ |
1330 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1331 uint8_t * const halfHV= ((uint8_t*)half);\ | |
954 | 1332 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
984 | 1333 put ## RND ## pixels16_l2_mmx(halfH, src, halfH, 16, stride, 17);\ |
959 | 1334 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
984 | 1335 OPNAME ## pixels16_l2_mmx(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 1336 }\ |
1064 | 1337 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1338 uint64_t half[16*2 + 17*2];\ |
1339 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1340 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1341 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
1342 put ## RND ## pixels16_l2_mmx(halfH, src+1, halfH, 16, stride, 17);\ | |
959 | 1343 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
984 | 1344 OPNAME ## pixels16_l2_mmx(dst, halfH+16, halfHV, stride, 16, 16);\ |
954 | 1345 }\ |
1064 | 1346 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1347 uint64_t half[16*2 + 17*2];\ |
1348 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1349 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1350 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 1351 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
954 | 1352 OPNAME ## pixels16_l2_mmx(dst, halfH, halfHV, stride, 16, 16);\ |
1353 }\ | |
1064 | 1354 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1355 uint64_t half[16*2 + 17*2];\ |
1356 uint8_t * const halfH= ((uint8_t*)half) + 256;\ | |
1357 uint8_t * const halfHV= ((uint8_t*)half);\ | |
1358 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 1359 put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\ |
954 | 1360 OPNAME ## pixels16_l2_mmx(dst, halfH+16, halfHV, stride, 16, 16);\ |
1361 }\ | |
1064 | 1362 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1363 uint64_t half[17*2];\ |
1364 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 1365 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
984 | 1366 put ## RND ## pixels16_l2_mmx(halfH, src, halfH, 16, stride, 17);\ |
1367 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ | |
954 | 1368 }\ |
1064 | 1369 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
984 | 1370 uint64_t half[17*2];\ |
1371 uint8_t * const halfH= ((uint8_t*)half);\ | |
954 | 1372 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ |
984 | 1373 put ## RND ## pixels16_l2_mmx(halfH, src+1, halfH, 16, stride, 17);\ |
1374 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ | |
954 | 1375 }\ |
1064 | 1376 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
954 | 1377 uint64_t half[17*2];\ |
1378 uint8_t * const halfH= ((uint8_t*)half);\ | |
1379 put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\ | |
959 | 1380 OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\ |
954 | 1381 } |
1382 | |
1383 | |
1384 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t" | |
959 | 1385 #define AVG_3DNOW_OP(a,b,temp, size) \ |
954 | 1386 "mov" #size " " #b ", " #temp " \n\t"\ |
1387 "pavgusb " #temp ", " #a " \n\t"\ | |
1388 "mov" #size " " #a ", " #b " \n\t" | |
959 | 1389 #define AVG_MMX2_OP(a,b,temp, size) \ |
954 | 1390 "mov" #size " " #b ", " #temp " \n\t"\ |
1391 "pavgb " #temp ", " #a " \n\t"\ | |
1392 "mov" #size " " #a ", " #b " \n\t" | |
959 | 1393 |
1394 QPEL_BASE(put_ , ff_pw_16, _ , PUT_OP, PUT_OP) | |
1395 QPEL_BASE(avg_ , ff_pw_16, _ , AVG_MMX2_OP, AVG_3DNOW_OP) | |
1396 QPEL_BASE(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, PUT_OP) | |
1397 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, 3dnow) | |
1398 QPEL_OP(avg_ , ff_pw_16, _ , AVG_3DNOW_OP, 3dnow) | |
1399 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, 3dnow) | |
954 | 1400 QPEL_OP(put_ , ff_pw_16, _ , PUT_OP, mmx2) |
959 | 1401 QPEL_OP(avg_ , ff_pw_16, _ , AVG_MMX2_OP, mmx2) |
954 | 1402 QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, PUT_OP, mmx2) |
1403 | |
393 | 1404 #if 0 |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1405 static void just_return() { return; } |
393 | 1406 #endif |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1407 |
954 | 1408 #define SET_QPEL_FUNC(postfix1, postfix2) \ |
1409 c->put_ ## postfix1 = put_ ## postfix2;\ | |
1410 c->put_no_rnd_ ## postfix1 = put_no_rnd_ ## postfix2;\ | |
1411 c->avg_ ## postfix1 = avg_ ## postfix2; | |
1092 | 1412 |
1413 /* external functions, from idct_mmx.c */ | |
1414 void ff_mmx_idct(DCTELEM *block); | |
1415 void ff_mmxext_idct(DCTELEM *block); | |
1416 | |
1417 /* XXX: those functions should be suppressed ASAP when all IDCTs are | |
1418 converted */ | |
1419 static void ff_libmpeg2mmx_idct_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1420 { | |
1421 ff_mmx_idct (block); | |
1422 put_pixels_clamped_mmx(block, dest, line_size); | |
1423 } | |
1424 static void ff_libmpeg2mmx_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1425 { | |
1426 ff_mmx_idct (block); | |
1427 add_pixels_clamped_mmx(block, dest, line_size); | |
1428 } | |
1429 static void ff_libmpeg2mmx2_idct_put(uint8_t *dest, int line_size, DCTELEM *block) | |
1430 { | |
1431 ff_mmxext_idct (block); | |
1432 put_pixels_clamped_mmx(block, dest, line_size); | |
1433 } | |
1434 static void ff_libmpeg2mmx2_idct_add(uint8_t *dest, int line_size, DCTELEM *block) | |
1435 { | |
1436 ff_mmxext_idct (block); | |
1437 add_pixels_clamped_mmx(block, dest, line_size); | |
1438 } | |
954 | 1439 |
1092 | 1440 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx) |
0 | 1441 { |
1442 mm_flags = mm_support(); | |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
1443 #if 0 |
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
1444 fprintf(stderr, "libavcodec: CPU flags:"); |
0 | 1445 if (mm_flags & MM_MMX) |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
1446 fprintf(stderr, " mmx"); |
0 | 1447 if (mm_flags & MM_MMXEXT) |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
1448 fprintf(stderr, " mmxext"); |
0 | 1449 if (mm_flags & MM_3DNOW) |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
1450 fprintf(stderr, " 3dnow"); |
0 | 1451 if (mm_flags & MM_SSE) |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
1452 fprintf(stderr, " sse"); |
0 | 1453 if (mm_flags & MM_SSE2) |
631
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
1454 fprintf(stderr, " sse2"); |
47a8964ba5cd
be less verbose patch by (Lennert Buytenhek <buytenh at math dot leidenuniv dot nl>)
michaelni
parents:
629
diff
changeset
|
1455 fprintf(stderr, "\n"); |
0 | 1456 #endif |
1457 | |
1458 if (mm_flags & MM_MMX) { | |
1092 | 1459 const int dct_algo = avctx->dct_algo; |
1460 const int idct_algo= avctx->idct_algo; | |
1461 | |
1462 if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX) | |
1463 c->fdct = ff_fdct_mmx; | |
1464 | |
1465 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){ | |
1466 c->idct_put= ff_simple_idct_put_mmx; | |
1467 c->idct_add= ff_simple_idct_add_mmx; | |
1468 c->idct_permutation_type= FF_SIMPLE_IDCT_PERM; | |
1469 }else if(idct_algo==FF_IDCT_LIBMPEG2MMX){ | |
1470 if(mm_flags & MM_MMXEXT){ | |
1471 c->idct_put= ff_libmpeg2mmx2_idct_put; | |
1472 c->idct_add= ff_libmpeg2mmx2_idct_add; | |
1473 }else{ | |
1474 c->idct_put= ff_libmpeg2mmx_idct_put; | |
1475 c->idct_add= ff_libmpeg2mmx_idct_add; | |
1476 } | |
1477 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM; | |
1478 } | |
1479 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1480 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
|
1481 c->diff_pixels = diff_pixels_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1482 c->put_pixels_clamped = put_pixels_clamped_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1483 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
|
1484 c->clear_blocks = clear_blocks_mmx; |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1485 c->pix_sum = pix_sum16_mmx; |
415 | 1486 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1487 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
|
1488 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
|
1489 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
|
1490 c->put_pixels_tab[0][3] = put_pixels16_xy2_mmx; |
0 | 1491 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1492 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
|
1493 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
|
1494 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
|
1495 c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_mmx; |
651 | 1496 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1497 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
|
1498 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
|
1499 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
|
1500 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx; |
415 | 1501 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1502 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
|
1503 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
|
1504 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
|
1505 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
|
1506 |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1507 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
|
1508 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
|
1509 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
|
1510 c->put_pixels_tab[1][3] = put_pixels8_xy2_mmx; |
0 | 1511 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1512 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
|
1513 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
|
1514 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
|
1515 c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_mmx; |
651 | 1516 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1517 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
|
1518 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
|
1519 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
|
1520 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx; |
651 | 1521 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1522 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
|
1523 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
|
1524 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
|
1525 c->avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels8_xy2_mmx; |
954 | 1526 |
866 | 1527 c->add_bytes= add_bytes_mmx; |
1528 c->diff_bytes= diff_bytes_mmx; | |
936 | 1529 |
1530 c->hadamard8_diff[0]= hadamard8_diff16_mmx; | |
1531 c->hadamard8_diff[1]= hadamard8_diff_mmx; | |
1532 | |
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
|
1533 c->pix_norm1 = pix_norm1_mmx; |
4dfe15ae0078
sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents:
984
diff
changeset
|
1534 c->sse[0] = sse16_mmx; |
936 | 1535 |
0 | 1536 if (mm_flags & MM_MMXEXT) { |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1537 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
|
1538 c->put_pixels_tab[0][2] = put_pixels16_y2_mmx2; |
651 | 1539 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1540 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
|
1541 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
|
1542 c->avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2; |
415 | 1543 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1544 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
|
1545 c->put_pixels_tab[1][2] = put_pixels8_y2_mmx2; |
651 | 1546 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1547 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
|
1548 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
|
1549 c->avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2; |
1092 | 1550 |
1551 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
1552 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2; | |
1553 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2; | |
1554 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2; | |
1555 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2; | |
1556 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2; | |
1557 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2; | |
1558 } | |
959 | 1559 |
961 | 1560 #if 1 |
954 | 1561 SET_QPEL_FUNC(qpel_pixels_tab[0][ 0], qpel16_mc00_mmx2) |
1562 SET_QPEL_FUNC(qpel_pixels_tab[0][ 1], qpel16_mc10_mmx2) | |
1563 SET_QPEL_FUNC(qpel_pixels_tab[0][ 2], qpel16_mc20_mmx2) | |
1564 SET_QPEL_FUNC(qpel_pixels_tab[0][ 3], qpel16_mc30_mmx2) | |
1565 SET_QPEL_FUNC(qpel_pixels_tab[0][ 4], qpel16_mc01_mmx2) | |
1566 SET_QPEL_FUNC(qpel_pixels_tab[0][ 5], qpel16_mc11_mmx2) | |
1567 SET_QPEL_FUNC(qpel_pixels_tab[0][ 6], qpel16_mc21_mmx2) | |
1568 SET_QPEL_FUNC(qpel_pixels_tab[0][ 7], qpel16_mc31_mmx2) | |
1569 SET_QPEL_FUNC(qpel_pixels_tab[0][ 8], qpel16_mc02_mmx2) | |
1570 SET_QPEL_FUNC(qpel_pixels_tab[0][ 9], qpel16_mc12_mmx2) | |
1571 SET_QPEL_FUNC(qpel_pixels_tab[0][10], qpel16_mc22_mmx2) | |
1572 SET_QPEL_FUNC(qpel_pixels_tab[0][11], qpel16_mc32_mmx2) | |
1573 SET_QPEL_FUNC(qpel_pixels_tab[0][12], qpel16_mc03_mmx2) | |
1574 SET_QPEL_FUNC(qpel_pixels_tab[0][13], qpel16_mc13_mmx2) | |
1575 SET_QPEL_FUNC(qpel_pixels_tab[0][14], qpel16_mc23_mmx2) | |
1576 SET_QPEL_FUNC(qpel_pixels_tab[0][15], qpel16_mc33_mmx2) | |
1577 SET_QPEL_FUNC(qpel_pixels_tab[1][ 0], qpel8_mc00_mmx2) | |
1578 SET_QPEL_FUNC(qpel_pixels_tab[1][ 1], qpel8_mc10_mmx2) | |
1579 SET_QPEL_FUNC(qpel_pixels_tab[1][ 2], qpel8_mc20_mmx2) | |
1580 SET_QPEL_FUNC(qpel_pixels_tab[1][ 3], qpel8_mc30_mmx2) | |
1581 SET_QPEL_FUNC(qpel_pixels_tab[1][ 4], qpel8_mc01_mmx2) | |
1582 SET_QPEL_FUNC(qpel_pixels_tab[1][ 5], qpel8_mc11_mmx2) | |
1583 SET_QPEL_FUNC(qpel_pixels_tab[1][ 6], qpel8_mc21_mmx2) | |
1584 SET_QPEL_FUNC(qpel_pixels_tab[1][ 7], qpel8_mc31_mmx2) | |
1585 SET_QPEL_FUNC(qpel_pixels_tab[1][ 8], qpel8_mc02_mmx2) | |
1586 SET_QPEL_FUNC(qpel_pixels_tab[1][ 9], qpel8_mc12_mmx2) | |
1587 SET_QPEL_FUNC(qpel_pixels_tab[1][10], qpel8_mc22_mmx2) | |
1588 SET_QPEL_FUNC(qpel_pixels_tab[1][11], qpel8_mc32_mmx2) | |
1589 SET_QPEL_FUNC(qpel_pixels_tab[1][12], qpel8_mc03_mmx2) | |
1590 SET_QPEL_FUNC(qpel_pixels_tab[1][13], qpel8_mc13_mmx2) | |
1591 SET_QPEL_FUNC(qpel_pixels_tab[1][14], qpel8_mc23_mmx2) | |
1592 SET_QPEL_FUNC(qpel_pixels_tab[1][15], qpel8_mc33_mmx2) | |
961 | 1593 #endif |
0 | 1594 } else if (mm_flags & MM_3DNOW) { |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1595 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
|
1596 c->put_pixels_tab[0][2] = put_pixels16_y2_3dnow; |
393 | 1597 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1598 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
|
1599 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
|
1600 c->avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow; |
651 | 1601 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1602 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
|
1603 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
|
1604 |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
706
diff
changeset
|
1605 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
|
1606 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
|
1607 c->avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow; |
1092 | 1608 |
1609 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){ | |
1610 c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow; | |
1611 c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow; | |
1612 c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow; | |
1613 c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow; | |
1614 c->avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow; | |
1615 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow; | |
1616 } | |
984 | 1617 |
954 | 1618 SET_QPEL_FUNC(qpel_pixels_tab[0][ 0], qpel16_mc00_3dnow) |
1619 SET_QPEL_FUNC(qpel_pixels_tab[0][ 1], qpel16_mc10_3dnow) | |
1620 SET_QPEL_FUNC(qpel_pixels_tab[0][ 2], qpel16_mc20_3dnow) | |
1621 SET_QPEL_FUNC(qpel_pixels_tab[0][ 3], qpel16_mc30_3dnow) | |
1622 SET_QPEL_FUNC(qpel_pixels_tab[0][ 4], qpel16_mc01_3dnow) | |
1623 SET_QPEL_FUNC(qpel_pixels_tab[0][ 5], qpel16_mc11_3dnow) | |
1624 SET_QPEL_FUNC(qpel_pixels_tab[0][ 6], qpel16_mc21_3dnow) | |
1625 SET_QPEL_FUNC(qpel_pixels_tab[0][ 7], qpel16_mc31_3dnow) | |
1626 SET_QPEL_FUNC(qpel_pixels_tab[0][ 8], qpel16_mc02_3dnow) | |
1627 SET_QPEL_FUNC(qpel_pixels_tab[0][ 9], qpel16_mc12_3dnow) | |
1628 SET_QPEL_FUNC(qpel_pixels_tab[0][10], qpel16_mc22_3dnow) | |
1629 SET_QPEL_FUNC(qpel_pixels_tab[0][11], qpel16_mc32_3dnow) | |
1630 SET_QPEL_FUNC(qpel_pixels_tab[0][12], qpel16_mc03_3dnow) | |
1631 SET_QPEL_FUNC(qpel_pixels_tab[0][13], qpel16_mc13_3dnow) | |
1632 SET_QPEL_FUNC(qpel_pixels_tab[0][14], qpel16_mc23_3dnow) | |
1633 SET_QPEL_FUNC(qpel_pixels_tab[0][15], qpel16_mc33_3dnow) | |
1634 SET_QPEL_FUNC(qpel_pixels_tab[1][ 0], qpel8_mc00_3dnow) | |
1635 SET_QPEL_FUNC(qpel_pixels_tab[1][ 1], qpel8_mc10_3dnow) | |
1636 SET_QPEL_FUNC(qpel_pixels_tab[1][ 2], qpel8_mc20_3dnow) | |
1637 SET_QPEL_FUNC(qpel_pixels_tab[1][ 3], qpel8_mc30_3dnow) | |
1638 SET_QPEL_FUNC(qpel_pixels_tab[1][ 4], qpel8_mc01_3dnow) | |
1639 SET_QPEL_FUNC(qpel_pixels_tab[1][ 5], qpel8_mc11_3dnow) | |
1640 SET_QPEL_FUNC(qpel_pixels_tab[1][ 6], qpel8_mc21_3dnow) | |
1641 SET_QPEL_FUNC(qpel_pixels_tab[1][ 7], qpel8_mc31_3dnow) | |
1642 SET_QPEL_FUNC(qpel_pixels_tab[1][ 8], qpel8_mc02_3dnow) | |
1643 SET_QPEL_FUNC(qpel_pixels_tab[1][ 9], qpel8_mc12_3dnow) | |
1644 SET_QPEL_FUNC(qpel_pixels_tab[1][10], qpel8_mc22_3dnow) | |
1645 SET_QPEL_FUNC(qpel_pixels_tab[1][11], qpel8_mc32_3dnow) | |
1646 SET_QPEL_FUNC(qpel_pixels_tab[1][12], qpel8_mc03_3dnow) | |
1647 SET_QPEL_FUNC(qpel_pixels_tab[1][13], qpel8_mc13_3dnow) | |
1648 SET_QPEL_FUNC(qpel_pixels_tab[1][14], qpel8_mc23_3dnow) | |
1649 SET_QPEL_FUNC(qpel_pixels_tab[1][15], qpel8_mc33_3dnow) | |
0 | 1650 } |
1651 } | |
1092 | 1652 |
1653 dsputil_init_pix_mmx(c, avctx); | |
247
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1654 #if 0 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1655 // for speed testing |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1656 get_pixels = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1657 put_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1658 add_pixels_clamped = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1659 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1660 pix_abs16x16 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1661 pix_abs16x16_x2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1662 pix_abs16x16_y2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1663 pix_abs16x16_xy2 = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1664 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1665 put_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1666 put_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1667 put_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1668 put_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1669 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1670 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
|
1671 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
|
1672 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
|
1673 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
|
1674 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1675 avg_pixels_tab[0] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1676 avg_pixels_tab[1] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1677 avg_pixels_tab[2] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1678 avg_pixels_tab[3] = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1679 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1680 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
|
1681 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
|
1682 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
|
1683 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
|
1684 |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1685 //av_fdct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1686 //ff_idct = just_return; |
6f48cacd9ed9
* some modifications to allow gcc to compile same code for -fPIC
kabi
parents:
188
diff
changeset
|
1687 #endif |
0 | 1688 } |