Mercurial > libavcodec.hg
annotate i386/h264dsp_mmx.c @ 6014:33b0514769b2 libavcodec
Fix non_zero_count_cache for deblocking in field pictures.
Based on original patch by Martin Zlomek martin.zlomek a email D cz
ffmpeg-devel thread: H264: Fix non_zero_count_cache for deblocking in fields
Fri, 30 Nov 2007 9:58:23
author | heydowns |
---|---|
date | Thu, 13 Dec 2007 00:50:47 +0000 |
parents | 37a03989871b |
children | 03febc8f506f |
rev | line source |
---|---|
2754 | 1 /* |
2 * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt | |
3 * | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3645
diff
changeset
|
4 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3645
diff
changeset
|
5 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3645
diff
changeset
|
6 * FFmpeg is free software; you can redistribute it and/or |
2754 | 7 * modify it under the terms of the GNU Lesser General Public |
8 * License as published by the Free Software Foundation; either | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3645
diff
changeset
|
9 * version 2.1 of the License, or (at your option) any later version. |
2754 | 10 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3645
diff
changeset
|
11 * FFmpeg is distributed in the hope that it will be useful, |
2754 | 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 * Lesser General Public License for more details. | |
15 * | |
16 * You should have received a copy of the GNU Lesser General Public | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3645
diff
changeset
|
17 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
3029
diff
changeset
|
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
2754 | 19 */ |
20 | |
5946
55251379b5b1
make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents:
5809
diff
changeset
|
21 #include "dsputil_mmx.h" |
2754 | 22 |
23 /***********************************/ | |
24 /* IDCT */ | |
25 | |
26 /* in/out: mma=mma+mmb, mmb=mmb-mma */ | |
27 #define SUMSUB_BA( a, b ) \ | |
28 "paddw "#b", "#a" \n\t"\ | |
29 "paddw "#b", "#b" \n\t"\ | |
30 "psubw "#a", "#b" \n\t" | |
31 | |
32 #define SUMSUB_BADC( a, b, c, d ) \ | |
33 "paddw "#b", "#a" \n\t"\ | |
34 "paddw "#d", "#c" \n\t"\ | |
35 "paddw "#b", "#b" \n\t"\ | |
36 "paddw "#d", "#d" \n\t"\ | |
37 "psubw "#a", "#b" \n\t"\ | |
38 "psubw "#c", "#d" \n\t" | |
39 | |
40 #define SUMSUBD2_AB( a, b, t ) \ | |
41 "movq "#b", "#t" \n\t"\ | |
42 "psraw $1 , "#b" \n\t"\ | |
43 "paddw "#a", "#b" \n\t"\ | |
44 "psraw $1 , "#a" \n\t"\ | |
45 "psubw "#t", "#a" \n\t" | |
46 | |
47 #define IDCT4_1D( s02, s13, d02, d13, t ) \ | |
48 SUMSUB_BA ( s02, d02 )\ | |
49 SUMSUBD2_AB( s13, d13, t )\ | |
50 SUMSUB_BADC( d13, s02, s13, d02 ) | |
51 | |
52 #define STORE_DIFF_4P( p, t, z ) \ | |
53 "psraw $6, "#p" \n\t"\ | |
54 "movd (%0), "#t" \n\t"\ | |
55 "punpcklbw "#z", "#t" \n\t"\ | |
56 "paddsw "#t", "#p" \n\t"\ | |
57 "packuswb "#z", "#p" \n\t"\ | |
58 "movd "#p", (%0) \n\t" | |
59 | |
3173 | 60 static void ff_h264_idct_add_mmx(uint8_t *dst, int16_t *block, int stride) |
2754 | 61 { |
62 /* Load dct coeffs */ | |
63 asm volatile( | |
64 "movq (%0), %%mm0 \n\t" | |
65 "movq 8(%0), %%mm1 \n\t" | |
66 "movq 16(%0), %%mm2 \n\t" | |
67 "movq 24(%0), %%mm3 \n\t" | |
68 :: "r"(block) ); | |
69 | |
70 asm volatile( | |
71 /* mm1=s02+s13 mm2=s02-s13 mm4=d02+d13 mm0=d02-d13 */ | |
72 IDCT4_1D( %%mm2, %%mm1, %%mm0, %%mm3, %%mm4 ) | |
73 | |
74 "movq %0, %%mm6 \n\t" | |
75 /* in: 1,4,0,2 out: 1,2,3,0 */ | |
76 TRANSPOSE4( %%mm3, %%mm1, %%mm0, %%mm2, %%mm4 ) | |
77 | |
78 "paddw %%mm6, %%mm3 \n\t" | |
79 | |
80 /* mm2=s02+s13 mm3=s02-s13 mm4=d02+d13 mm1=d02-d13 */ | |
81 IDCT4_1D( %%mm4, %%mm2, %%mm3, %%mm0, %%mm1 ) | |
82 | |
83 "pxor %%mm7, %%mm7 \n\t" | |
84 :: "m"(ff_pw_32)); | |
85 | |
86 asm volatile( | |
87 STORE_DIFF_4P( %%mm0, %%mm1, %%mm7) | |
88 "add %1, %0 \n\t" | |
89 STORE_DIFF_4P( %%mm2, %%mm1, %%mm7) | |
90 "add %1, %0 \n\t" | |
91 STORE_DIFF_4P( %%mm3, %%mm1, %%mm7) | |
92 "add %1, %0 \n\t" | |
93 STORE_DIFF_4P( %%mm4, %%mm1, %%mm7) | |
94 : "+r"(dst) | |
95 : "r" ((long)stride) | |
96 ); | |
97 } | |
98 | |
3174 | 99 static inline void h264_idct8_1d(int16_t *block) |
100 { | |
101 asm volatile( | |
102 "movq 112(%0), %%mm7 \n\t" | |
103 "movq 80(%0), %%mm5 \n\t" | |
104 "movq 48(%0), %%mm3 \n\t" | |
105 "movq 16(%0), %%mm1 \n\t" | |
106 | |
107 "movq %%mm7, %%mm4 \n\t" | |
108 "movq %%mm3, %%mm6 \n\t" | |
109 "movq %%mm5, %%mm0 \n\t" | |
110 "movq %%mm7, %%mm2 \n\t" | |
111 "psraw $1, %%mm4 \n\t" | |
112 "psraw $1, %%mm6 \n\t" | |
113 "psubw %%mm7, %%mm0 \n\t" | |
114 "psubw %%mm6, %%mm2 \n\t" | |
115 "psubw %%mm4, %%mm0 \n\t" | |
116 "psubw %%mm3, %%mm2 \n\t" | |
117 "psubw %%mm3, %%mm0 \n\t" | |
118 "paddw %%mm1, %%mm2 \n\t" | |
119 | |
120 "movq %%mm5, %%mm4 \n\t" | |
121 "movq %%mm1, %%mm6 \n\t" | |
122 "psraw $1, %%mm4 \n\t" | |
123 "psraw $1, %%mm6 \n\t" | |
124 "paddw %%mm5, %%mm4 \n\t" | |
125 "paddw %%mm1, %%mm6 \n\t" | |
126 "paddw %%mm7, %%mm4 \n\t" | |
127 "paddw %%mm5, %%mm6 \n\t" | |
128 "psubw %%mm1, %%mm4 \n\t" | |
129 "paddw %%mm3, %%mm6 \n\t" | |
130 | |
131 "movq %%mm0, %%mm1 \n\t" | |
132 "movq %%mm4, %%mm3 \n\t" | |
133 "movq %%mm2, %%mm5 \n\t" | |
134 "movq %%mm6, %%mm7 \n\t" | |
135 "psraw $2, %%mm6 \n\t" | |
136 "psraw $2, %%mm3 \n\t" | |
137 "psraw $2, %%mm5 \n\t" | |
138 "psraw $2, %%mm0 \n\t" | |
139 "paddw %%mm6, %%mm1 \n\t" | |
140 "paddw %%mm2, %%mm3 \n\t" | |
141 "psubw %%mm4, %%mm5 \n\t" | |
142 "psubw %%mm0, %%mm7 \n\t" | |
143 | |
144 "movq 32(%0), %%mm2 \n\t" | |
145 "movq 96(%0), %%mm6 \n\t" | |
146 "movq %%mm2, %%mm4 \n\t" | |
147 "movq %%mm6, %%mm0 \n\t" | |
148 "psraw $1, %%mm4 \n\t" | |
149 "psraw $1, %%mm6 \n\t" | |
150 "psubw %%mm0, %%mm4 \n\t" | |
151 "paddw %%mm2, %%mm6 \n\t" | |
152 | |
153 "movq (%0), %%mm2 \n\t" | |
154 "movq 64(%0), %%mm0 \n\t" | |
155 SUMSUB_BA( %%mm0, %%mm2 ) | |
156 SUMSUB_BA( %%mm6, %%mm0 ) | |
157 SUMSUB_BA( %%mm4, %%mm2 ) | |
158 SUMSUB_BA( %%mm7, %%mm6 ) | |
159 SUMSUB_BA( %%mm5, %%mm4 ) | |
160 SUMSUB_BA( %%mm3, %%mm2 ) | |
161 SUMSUB_BA( %%mm1, %%mm0 ) | |
162 :: "r"(block) | |
163 ); | |
164 } | |
165 | |
166 static void ff_h264_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride) | |
167 { | |
168 int i; | |
169 int16_t __attribute__ ((aligned(8))) b2[64]; | |
170 | |
171 block[0] += 32; | |
172 | |
173 for(i=0; i<2; i++){ | |
4137 | 174 DECLARE_ALIGNED_8(uint64_t, tmp); |
3174 | 175 |
176 h264_idct8_1d(block+4*i); | |
177 | |
178 asm volatile( | |
179 "movq %%mm7, %0 \n\t" | |
180 TRANSPOSE4( %%mm0, %%mm2, %%mm4, %%mm6, %%mm7 ) | |
181 "movq %%mm0, 8(%1) \n\t" | |
182 "movq %%mm6, 24(%1) \n\t" | |
183 "movq %%mm7, 40(%1) \n\t" | |
184 "movq %%mm4, 56(%1) \n\t" | |
185 "movq %0, %%mm7 \n\t" | |
186 TRANSPOSE4( %%mm7, %%mm5, %%mm3, %%mm1, %%mm0 ) | |
187 "movq %%mm7, (%1) \n\t" | |
188 "movq %%mm1, 16(%1) \n\t" | |
189 "movq %%mm0, 32(%1) \n\t" | |
190 "movq %%mm3, 48(%1) \n\t" | |
191 : "=m"(tmp) | |
192 : "r"(b2+32*i) | |
193 : "memory" | |
194 ); | |
195 } | |
196 | |
197 for(i=0; i<2; i++){ | |
198 h264_idct8_1d(b2+4*i); | |
199 | |
200 asm volatile( | |
201 "psraw $6, %%mm7 \n\t" | |
202 "psraw $6, %%mm6 \n\t" | |
203 "psraw $6, %%mm5 \n\t" | |
204 "psraw $6, %%mm4 \n\t" | |
205 "psraw $6, %%mm3 \n\t" | |
206 "psraw $6, %%mm2 \n\t" | |
207 "psraw $6, %%mm1 \n\t" | |
208 "psraw $6, %%mm0 \n\t" | |
209 | |
210 "movq %%mm7, (%0) \n\t" | |
211 "movq %%mm5, 16(%0) \n\t" | |
212 "movq %%mm3, 32(%0) \n\t" | |
213 "movq %%mm1, 48(%0) \n\t" | |
214 "movq %%mm0, 64(%0) \n\t" | |
215 "movq %%mm2, 80(%0) \n\t" | |
216 "movq %%mm4, 96(%0) \n\t" | |
217 "movq %%mm6, 112(%0) \n\t" | |
218 :: "r"(b2+4*i) | |
219 : "memory" | |
220 ); | |
221 } | |
222 | |
223 add_pixels_clamped_mmx(b2, dst, stride); | |
224 } | |
225 | |
3173 | 226 static void ff_h264_idct_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride) |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
227 { |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
228 int dc = (block[0] + 32) >> 6; |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
229 asm volatile( |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
230 "movd %0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
231 "pshufw $0, %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
232 "pxor %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
233 "psubw %%mm0, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
234 "packuswb %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
235 "packuswb %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
236 ::"r"(dc) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
237 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
238 asm volatile( |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
239 "movd %0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
240 "movd %1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
241 "movd %2, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
242 "movd %3, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
243 "paddusb %%mm0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
244 "paddusb %%mm0, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
245 "paddusb %%mm0, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
246 "paddusb %%mm0, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
247 "psubusb %%mm1, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
248 "psubusb %%mm1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
249 "psubusb %%mm1, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
250 "psubusb %%mm1, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
251 "movd %%mm2, %0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
252 "movd %%mm3, %1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
253 "movd %%mm4, %2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
254 "movd %%mm5, %3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
255 :"+m"(*(uint32_t*)(dst+0*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
256 "+m"(*(uint32_t*)(dst+1*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
257 "+m"(*(uint32_t*)(dst+2*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
258 "+m"(*(uint32_t*)(dst+3*stride)) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
259 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
260 } |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
261 |
3173 | 262 static void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride) |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
263 { |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
264 int dc = (block[0] + 32) >> 6; |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
265 int y; |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
266 asm volatile( |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
267 "movd %0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
268 "pshufw $0, %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
269 "pxor %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
270 "psubw %%mm0, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
271 "packuswb %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
272 "packuswb %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
273 ::"r"(dc) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
274 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
275 for(y=2; y--; dst += 4*stride){ |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
276 asm volatile( |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
277 "movq %0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
278 "movq %1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
279 "movq %2, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
280 "movq %3, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
281 "paddusb %%mm0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
282 "paddusb %%mm0, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
283 "paddusb %%mm0, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
284 "paddusb %%mm0, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
285 "psubusb %%mm1, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
286 "psubusb %%mm1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
287 "psubusb %%mm1, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
288 "psubusb %%mm1, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
289 "movq %%mm2, %0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
290 "movq %%mm3, %1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
291 "movq %%mm4, %2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
292 "movq %%mm5, %3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
293 :"+m"(*(uint64_t*)(dst+0*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
294 "+m"(*(uint64_t*)(dst+1*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
295 "+m"(*(uint64_t*)(dst+2*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
296 "+m"(*(uint64_t*)(dst+3*stride)) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
297 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
298 } |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
299 } |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
300 |
2754 | 301 |
302 /***********************************/ | |
303 /* deblocking */ | |
304 | |
305 // out: o = |x-y|>a | |
306 // clobbers: t | |
307 #define DIFF_GT_MMX(x,y,a,o,t)\ | |
308 "movq "#y", "#t" \n\t"\ | |
309 "movq "#x", "#o" \n\t"\ | |
310 "psubusb "#x", "#t" \n\t"\ | |
311 "psubusb "#y", "#o" \n\t"\ | |
312 "por "#t", "#o" \n\t"\ | |
313 "psubusb "#a", "#o" \n\t" | |
314 | |
4135 | 315 // out: o = |x-y|>a |
316 // clobbers: t | |
317 #define DIFF_GT2_MMX(x,y,a,o,t)\ | |
318 "movq "#y", "#t" \n\t"\ | |
319 "movq "#x", "#o" \n\t"\ | |
320 "psubusb "#x", "#t" \n\t"\ | |
321 "psubusb "#y", "#o" \n\t"\ | |
322 "psubusb "#a", "#t" \n\t"\ | |
323 "psubusb "#a", "#o" \n\t"\ | |
324 "pcmpeqb "#t", "#o" \n\t"\ | |
325 | |
2754 | 326 // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 |
327 // out: mm5=beta-1, mm7=mask | |
328 // clobbers: mm4,mm6 | |
329 #define H264_DEBLOCK_MASK(alpha1, beta1) \ | |
330 "pshufw $0, "#alpha1", %%mm4 \n\t"\ | |
331 "pshufw $0, "#beta1 ", %%mm5 \n\t"\ | |
332 "packuswb %%mm4, %%mm4 \n\t"\ | |
333 "packuswb %%mm5, %%mm5 \n\t"\ | |
334 DIFF_GT_MMX(%%mm1, %%mm2, %%mm4, %%mm7, %%mm6) /* |p0-q0| > alpha-1 */\ | |
335 DIFF_GT_MMX(%%mm0, %%mm1, %%mm5, %%mm4, %%mm6) /* |p1-p0| > beta-1 */\ | |
336 "por %%mm4, %%mm7 \n\t"\ | |
337 DIFF_GT_MMX(%%mm3, %%mm2, %%mm5, %%mm4, %%mm6) /* |q1-q0| > beta-1 */\ | |
338 "por %%mm4, %%mm7 \n\t"\ | |
339 "pxor %%mm6, %%mm6 \n\t"\ | |
340 "pcmpeqb %%mm6, %%mm7 \n\t" | |
341 | |
342 // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask) | |
343 // out: mm1=p0' mm2=q0' | |
344 // clobbers: mm0,3-6 | |
345 #define H264_DEBLOCK_P0_Q0(pb_01, pb_3f)\ | |
4127 | 346 "movq %%mm1 , %%mm5 \n\t"\ |
347 "pxor %%mm2 , %%mm5 \n\t" /* p0^q0*/\ | |
348 "pand "#pb_01" , %%mm5 \n\t" /* (p0^q0)&1*/\ | |
349 "pcmpeqb %%mm4 , %%mm4 \n\t"\ | |
350 "pxor %%mm4 , %%mm3 \n\t"\ | |
351 "pavgb %%mm0 , %%mm3 \n\t" /* (p1 - q1 + 256)>>1*/\ | |
352 "pavgb "MANGLE(ff_pb_3)" , %%mm3 \n\t" /*(((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2*/\ | |
353 "pxor %%mm1 , %%mm4 \n\t"\ | |
354 "pavgb %%mm2 , %%mm4 \n\t" /* (q0 - p0 + 256)>>1*/\ | |
355 "pavgb %%mm5 , %%mm3 \n\t"\ | |
4130 | 356 "paddusb %%mm4 , %%mm3 \n\t" /* d+128+33*/\ |
4129 | 357 "movq "MANGLE(ff_pb_A1)" , %%mm6 \n\t"\ |
358 "psubusb %%mm3 , %%mm6 \n\t"\ | |
4127 | 359 "psubusb "MANGLE(ff_pb_A1)" , %%mm3 \n\t"\ |
4129 | 360 "pminub %%mm7 , %%mm6 \n\t"\ |
4127 | 361 "pminub %%mm7 , %%mm3 \n\t"\ |
4129 | 362 "psubusb %%mm6 , %%mm1 \n\t"\ |
363 "psubusb %%mm3 , %%mm2 \n\t"\ | |
4127 | 364 "paddusb %%mm3 , %%mm1 \n\t"\ |
4129 | 365 "paddusb %%mm6 , %%mm2 \n\t" |
2754 | 366 |
5947 | 367 // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask) %8=ff_bone |
4866 | 368 // out: (q1addr) = av_clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 ) |
2754 | 369 // clobbers: q2, tmp, tc0 |
370 #define H264_DEBLOCK_Q1(p1, q2, q2addr, q1addr, tc0, tmp)\ | |
371 "movq %%mm1, "#tmp" \n\t"\ | |
372 "pavgb %%mm2, "#tmp" \n\t"\ | |
373 "pavgb "#tmp", "#q2" \n\t" /* avg(p2,avg(p0,q0)) */\ | |
374 "pxor "q2addr", "#tmp" \n\t"\ | |
375 "pand %8, "#tmp" \n\t" /* (p2^avg(p0,q0))&1 */\ | |
376 "psubusb "#tmp", "#q2" \n\t" /* (p2+((p0+q0+1)>>1))>>1 */\ | |
377 "movq "#p1", "#tmp" \n\t"\ | |
378 "psubusb "#tc0", "#tmp" \n\t"\ | |
379 "paddusb "#p1", "#tc0" \n\t"\ | |
380 "pmaxub "#tmp", "#q2" \n\t"\ | |
381 "pminub "#tc0", "#q2" \n\t"\ | |
382 "movq "#q2", "q1addr" \n\t" | |
383 | |
384 static inline void h264_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0) | |
385 { | |
4137 | 386 DECLARE_ALIGNED_8(uint64_t, tmp0[2]); |
2754 | 387 |
388 asm volatile( | |
389 "movq (%1,%3), %%mm0 \n\t" //p1 | |
390 "movq (%1,%3,2), %%mm1 \n\t" //p0 | |
391 "movq (%2), %%mm2 \n\t" //q0 | |
392 "movq (%2,%3), %%mm3 \n\t" //q1 | |
393 H264_DEBLOCK_MASK(%6, %7) | |
4133 | 394 |
395 "movd %5, %%mm4 \n\t" | |
396 "punpcklbw %%mm4, %%mm4 \n\t" | |
397 "punpcklwd %%mm4, %%mm4 \n\t" | |
398 "pcmpeqb %%mm3, %%mm3 \n\t" | |
399 "movq %%mm4, %%mm6 \n\t" | |
400 "pcmpgtb %%mm3, %%mm4 \n\t" | |
401 "movq %%mm6, 8+%0 \n\t" | |
402 "pand %%mm4, %%mm7 \n\t" | |
403 "movq %%mm7, %0 \n\t" | |
2754 | 404 |
405 /* filter p1 */ | |
406 "movq (%1), %%mm3 \n\t" //p2 | |
4135 | 407 DIFF_GT2_MMX(%%mm1, %%mm3, %%mm5, %%mm6, %%mm4) // |p2-p0|>beta-1 |
2754 | 408 "pand %%mm7, %%mm6 \n\t" // mask & |p2-p0|<beta |
4136 | 409 "pand 8+%0, %%mm7 \n\t" // mask & tc0 |
410 "movq %%mm7, %%mm4 \n\t" | |
4131
1a8e384d0463
2 instructions less in h264_loop_filter_luma_mmx2()
michael
parents:
4130
diff
changeset
|
411 "psubb %%mm6, %%mm7 \n\t" |
2754 | 412 "pand %%mm4, %%mm6 \n\t" // mask & |p2-p0|<beta & tc0 |
413 H264_DEBLOCK_Q1(%%mm0, %%mm3, "(%1)", "(%1,%3)", %%mm6, %%mm4) | |
414 | |
415 /* filter q1 */ | |
416 "movq (%2,%3,2), %%mm4 \n\t" //q2 | |
4135 | 417 DIFF_GT2_MMX(%%mm2, %%mm4, %%mm5, %%mm6, %%mm3) // |q2-q0|>beta-1 |
2754 | 418 "pand %0, %%mm6 \n\t" |
4134 | 419 "movq 8+%0, %%mm5 \n\t" // can be merged with the and below but is slower then |
2754 | 420 "pand %%mm6, %%mm5 \n\t" |
4131
1a8e384d0463
2 instructions less in h264_loop_filter_luma_mmx2()
michael
parents:
4130
diff
changeset
|
421 "psubb %%mm6, %%mm7 \n\t" |
2754 | 422 "movq (%2,%3), %%mm3 \n\t" |
423 H264_DEBLOCK_Q1(%%mm3, %%mm4, "(%2,%3,2)", "(%2,%3)", %%mm5, %%mm6) | |
424 | |
425 /* filter p0, q0 */ | |
4133 | 426 H264_DEBLOCK_P0_Q0(%8, unused) |
2754 | 427 "movq %%mm1, (%1,%3,2) \n\t" |
428 "movq %%mm2, (%2) \n\t" | |
429 | |
4133 | 430 : "=m"(*tmp0) |
2754 | 431 : "r"(pix-3*stride), "r"(pix), "r"((long)stride), |
4133 | 432 "m"(*tmp0/*unused*/), "m"(*(uint32_t*)tc0), "m"(alpha1), "m"(beta1), |
5947 | 433 "m"(ff_bone) |
2754 | 434 ); |
435 } | |
436 | |
437 static void h264_v_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
438 { | |
439 if((tc0[0] & tc0[1]) >= 0) | |
440 h264_loop_filter_luma_mmx2(pix, stride, alpha-1, beta-1, tc0); | |
441 if((tc0[2] & tc0[3]) >= 0) | |
442 h264_loop_filter_luma_mmx2(pix+8, stride, alpha-1, beta-1, tc0+2); | |
443 } | |
444 static void h264_h_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
445 { | |
446 //FIXME: could cut some load/stores by merging transpose with filter | |
447 // also, it only needs to transpose 6x8 | |
4137 | 448 DECLARE_ALIGNED_8(uint8_t, trans[8*8]); |
2754 | 449 int i; |
450 for(i=0; i<2; i++, pix+=8*stride, tc0+=2) { | |
451 if((tc0[0] & tc0[1]) < 0) | |
452 continue; | |
453 transpose4x4(trans, pix-4, 8, stride); | |
454 transpose4x4(trans +4*8, pix, 8, stride); | |
455 transpose4x4(trans+4, pix-4+4*stride, 8, stride); | |
456 transpose4x4(trans+4+4*8, pix +4*stride, 8, stride); | |
457 h264_loop_filter_luma_mmx2(trans+4*8, 8, alpha-1, beta-1, tc0); | |
458 transpose4x4(pix-2, trans +2*8, stride, 8); | |
459 transpose4x4(pix-2+4*stride, trans+4+2*8, stride, 8); | |
460 } | |
461 } | |
462 | |
463 static inline void h264_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0) | |
464 { | |
465 asm volatile( | |
466 "movq (%0), %%mm0 \n\t" //p1 | |
467 "movq (%0,%2), %%mm1 \n\t" //p0 | |
468 "movq (%1), %%mm2 \n\t" //q0 | |
469 "movq (%1,%2), %%mm3 \n\t" //q1 | |
470 H264_DEBLOCK_MASK(%4, %5) | |
471 "movd %3, %%mm6 \n\t" | |
472 "punpcklbw %%mm6, %%mm6 \n\t" | |
473 "pand %%mm6, %%mm7 \n\t" // mm7 = tc&mask | |
474 H264_DEBLOCK_P0_Q0(%6, %7) | |
475 "movq %%mm1, (%0,%2) \n\t" | |
476 "movq %%mm2, (%1) \n\t" | |
477 | |
478 :: "r"(pix-2*stride), "r"(pix), "r"((long)stride), | |
479 "r"(*(uint32_t*)tc0), | |
5947 | 480 "m"(alpha1), "m"(beta1), "m"(ff_bone), "m"(ff_pb_3F) |
2754 | 481 ); |
482 } | |
483 | |
484 static void h264_v_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
485 { | |
486 h264_loop_filter_chroma_mmx2(pix, stride, alpha-1, beta-1, tc0); | |
487 } | |
488 | |
489 static void h264_h_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
490 { | |
491 //FIXME: could cut some load/stores by merging transpose with filter | |
4137 | 492 DECLARE_ALIGNED_8(uint8_t, trans[8*4]); |
2754 | 493 transpose4x4(trans, pix-2, 8, stride); |
494 transpose4x4(trans+4, pix-2+4*stride, 8, stride); | |
495 h264_loop_filter_chroma_mmx2(trans+2*8, 8, alpha-1, beta-1, tc0); | |
496 transpose4x4(pix-2, trans, stride, 8); | |
497 transpose4x4(pix-2+4*stride, trans+4, stride, 8); | |
498 } | |
499 | |
500 // p0 = (p0 + q1 + 2*p1 + 2) >> 2 | |
501 #define H264_FILTER_CHROMA4(p0, p1, q1, one) \ | |
502 "movq "#p0", %%mm4 \n\t"\ | |
503 "pxor "#q1", %%mm4 \n\t"\ | |
504 "pand "#one", %%mm4 \n\t" /* mm4 = (p0^q1)&1 */\ | |
505 "pavgb "#q1", "#p0" \n\t"\ | |
506 "psubusb %%mm4, "#p0" \n\t"\ | |
507 "pavgb "#p1", "#p0" \n\t" /* dst = avg(p1, avg(p0,q1) - ((p0^q1)&1)) */\ | |
508 | |
509 static inline void h264_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha1, int beta1) | |
510 { | |
511 asm volatile( | |
512 "movq (%0), %%mm0 \n\t" | |
513 "movq (%0,%2), %%mm1 \n\t" | |
514 "movq (%1), %%mm2 \n\t" | |
515 "movq (%1,%2), %%mm3 \n\t" | |
516 H264_DEBLOCK_MASK(%3, %4) | |
517 "movq %%mm1, %%mm5 \n\t" | |
518 "movq %%mm2, %%mm6 \n\t" | |
519 H264_FILTER_CHROMA4(%%mm1, %%mm0, %%mm3, %5) //p0' | |
520 H264_FILTER_CHROMA4(%%mm2, %%mm3, %%mm0, %5) //q0' | |
521 "psubb %%mm5, %%mm1 \n\t" | |
522 "psubb %%mm6, %%mm2 \n\t" | |
523 "pand %%mm7, %%mm1 \n\t" | |
524 "pand %%mm7, %%mm2 \n\t" | |
525 "paddb %%mm5, %%mm1 \n\t" | |
526 "paddb %%mm6, %%mm2 \n\t" | |
527 "movq %%mm1, (%0,%2) \n\t" | |
528 "movq %%mm2, (%1) \n\t" | |
529 :: "r"(pix-2*stride), "r"(pix), "r"((long)stride), | |
5947 | 530 "m"(alpha1), "m"(beta1), "m"(ff_bone) |
2754 | 531 ); |
532 } | |
533 | |
534 static void h264_v_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta) | |
535 { | |
536 h264_loop_filter_chroma_intra_mmx2(pix, stride, alpha-1, beta-1); | |
537 } | |
538 | |
539 static void h264_h_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta) | |
540 { | |
541 //FIXME: could cut some load/stores by merging transpose with filter | |
4137 | 542 DECLARE_ALIGNED_8(uint8_t, trans[8*4]); |
2754 | 543 transpose4x4(trans, pix-2, 8, stride); |
544 transpose4x4(trans+4, pix-2+4*stride, 8, stride); | |
545 h264_loop_filter_chroma_intra_mmx2(trans+2*8, 8, alpha-1, beta-1); | |
546 transpose4x4(pix-2, trans, stride, 8); | |
547 transpose4x4(pix-2+4*stride, trans+4, stride, 8); | |
548 } | |
549 | |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
550 static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2], |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
551 int bidir, int edges, int step, int mask_mv0, int mask_mv1 ) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
552 int dir; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
553 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
554 "pxor %%mm7, %%mm7 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
555 "movq %0, %%mm6 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
556 "movq %1, %%mm5 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
557 "movq %2, %%mm4 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
558 ::"m"(ff_pb_1), "m"(ff_pb_3), "m"(ff_pb_7) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
559 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
560 // could do a special case for dir==0 && edges==1, but it only reduces the |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
561 // average filter time by 1.2% |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
562 for( dir=1; dir>=0; dir-- ) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
563 const int d_idx = dir ? -8 : -1; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
564 const int mask_mv = dir ? mask_mv1 : mask_mv0; |
4137 | 565 DECLARE_ALIGNED_8(const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL; |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
566 int b_idx, edge, l; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
567 for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
568 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
569 "pand %0, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
570 ::"m"(mask_dir) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
571 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
572 if(!(mask_mv & edge)) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
573 asm volatile("pxor %%mm0, %%mm0 \n\t":); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
574 for( l = bidir; l >= 0; l-- ) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
575 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
576 "movd %0, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
577 "punpckldq %1, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
578 "movq %%mm1, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
579 "psrlw $7, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
580 "pand %%mm6, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
581 "por %%mm2, %%mm1 \n\t" // ref_cache with -2 mapped to -1 |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
582 "punpckldq %%mm1, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
583 "pcmpeqb %%mm2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
584 "paddb %%mm6, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
585 "punpckhbw %%mm7, %%mm1 \n\t" // ref[b] != ref[bn] |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
586 "por %%mm1, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
587 |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
588 "movq %2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
589 "movq %3, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
590 "psubw %4, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
591 "psubw %5, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
592 "packsswb %%mm2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
593 "paddb %%mm5, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
594 "pminub %%mm4, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
595 "pcmpeqb %%mm4, %%mm1 \n\t" // abs(mv[b] - mv[bn]) >= limit |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
596 "por %%mm1, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
597 ::"m"(ref[l][b_idx]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
598 "m"(ref[l][b_idx+d_idx]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
599 "m"(mv[l][b_idx][0]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
600 "m"(mv[l][b_idx+2][0]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
601 "m"(mv[l][b_idx+d_idx][0]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
602 "m"(mv[l][b_idx+d_idx+2][0]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
603 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
604 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
605 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
606 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
607 "movd %0, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
608 "por %1, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
609 "punpcklbw %%mm7, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
610 "pcmpgtw %%mm7, %%mm1 \n\t" // nnz[b] || nnz[bn] |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
611 ::"m"(nnz[b_idx]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
612 "m"(nnz[b_idx+d_idx]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
613 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
614 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
615 "pcmpeqw %%mm7, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
616 "pcmpeqw %%mm7, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
617 "psrlw $15, %%mm0 \n\t" // nonzero -> 1 |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
618 "psrlw $14, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
619 "movq %%mm0, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
620 "por %%mm1, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
621 "psrlw $1, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
622 "pandn %%mm2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
623 "movq %%mm1, %0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
624 :"=m"(*bS[dir][edge]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
625 ::"memory" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
626 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
627 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
628 edges = 4; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
629 step = 1; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
630 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
631 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
632 "movq (%0), %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
633 "movq 8(%0), %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
634 "movq 16(%0), %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
635 "movq 24(%0), %%mm3 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
636 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
637 "movq %%mm0, (%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
638 "movq %%mm3, 8(%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
639 "movq %%mm4, 16(%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
640 "movq %%mm2, 24(%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
641 ::"r"(bS[0]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
642 :"memory" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
643 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
644 } |
2754 | 645 |
646 /***********************************/ | |
647 /* motion compensation */ | |
648 | |
649 #define QPEL_H264V(A,B,C,D,E,F,OP)\ | |
2979 | 650 "movd (%0), "#F" \n\t"\ |
651 "movq "#C", %%mm6 \n\t"\ | |
652 "paddw "#D", %%mm6 \n\t"\ | |
653 "psllw $2, %%mm6 \n\t"\ | |
654 "psubw "#B", %%mm6 \n\t"\ | |
655 "psubw "#E", %%mm6 \n\t"\ | |
656 "pmullw %4, %%mm6 \n\t"\ | |
657 "add %2, %0 \n\t"\ | |
658 "punpcklbw %%mm7, "#F" \n\t"\ | |
659 "paddw %5, "#A" \n\t"\ | |
660 "paddw "#F", "#A" \n\t"\ | |
661 "paddw "#A", %%mm6 \n\t"\ | |
662 "psraw $5, %%mm6 \n\t"\ | |
663 "packuswb %%mm6, %%mm6 \n\t"\ | |
2754 | 664 OP(%%mm6, (%1), A, d)\ |
2979 | 665 "add %3, %1 \n\t" |
2754 | 666 |
667 #define QPEL_H264HV(A,B,C,D,E,F,OF)\ | |
2979 | 668 "movd (%0), "#F" \n\t"\ |
669 "movq "#C", %%mm6 \n\t"\ | |
670 "paddw "#D", %%mm6 \n\t"\ | |
671 "psllw $2, %%mm6 \n\t"\ | |
672 "psubw "#B", %%mm6 \n\t"\ | |
673 "psubw "#E", %%mm6 \n\t"\ | |
674 "pmullw %3, %%mm6 \n\t"\ | |
675 "add %2, %0 \n\t"\ | |
676 "punpcklbw %%mm7, "#F" \n\t"\ | |
677 "paddw "#F", "#A" \n\t"\ | |
678 "paddw "#A", %%mm6 \n\t"\ | |
679 "movq %%mm6, "#OF"(%1) \n\t" | |
2967 | 680 |
2754 | 681 #define QPEL_H264(OPNAME, OP, MMX)\ |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
682 static av_noinline void OPNAME ## h264_qpel4_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 683 int h=4;\ |
684 \ | |
685 asm volatile(\ | |
2979 | 686 "pxor %%mm7, %%mm7 \n\t"\ |
687 "movq %5, %%mm4 \n\t"\ | |
688 "movq %6, %%mm5 \n\t"\ | |
689 "1: \n\t"\ | |
690 "movd -1(%0), %%mm1 \n\t"\ | |
691 "movd (%0), %%mm2 \n\t"\ | |
692 "movd 1(%0), %%mm3 \n\t"\ | |
693 "movd 2(%0), %%mm0 \n\t"\ | |
694 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
695 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
696 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
697 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
698 "paddw %%mm0, %%mm1 \n\t"\ | |
699 "paddw %%mm3, %%mm2 \n\t"\ | |
700 "movd -2(%0), %%mm0 \n\t"\ | |
701 "movd 3(%0), %%mm3 \n\t"\ | |
702 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
703 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
704 "paddw %%mm3, %%mm0 \n\t"\ | |
705 "psllw $2, %%mm2 \n\t"\ | |
706 "psubw %%mm1, %%mm2 \n\t"\ | |
707 "pmullw %%mm4, %%mm2 \n\t"\ | |
708 "paddw %%mm5, %%mm0 \n\t"\ | |
709 "paddw %%mm2, %%mm0 \n\t"\ | |
710 "psraw $5, %%mm0 \n\t"\ | |
711 "packuswb %%mm0, %%mm0 \n\t"\ | |
2754 | 712 OP(%%mm0, (%1),%%mm6, d)\ |
2979 | 713 "add %3, %0 \n\t"\ |
714 "add %4, %1 \n\t"\ | |
715 "decl %2 \n\t"\ | |
716 " jnz 1b \n\t"\ | |
2754 | 717 : "+a"(src), "+c"(dst), "+m"(h)\ |
718 : "d"((long)srcStride), "S"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
719 : "memory"\ | |
720 );\ | |
721 }\ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
722 static av_noinline void OPNAME ## h264_qpel4_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\ |
3156 | 723 int h=4;\ |
724 asm volatile(\ | |
725 "pxor %%mm7, %%mm7 \n\t"\ | |
3165 | 726 "movq %0, %%mm4 \n\t"\ |
727 "movq %1, %%mm5 \n\t"\ | |
728 :: "m"(ff_pw_5), "m"(ff_pw_16)\ | |
729 );\ | |
730 do{\ | |
731 asm volatile(\ | |
3156 | 732 "movd -1(%0), %%mm1 \n\t"\ |
733 "movd (%0), %%mm2 \n\t"\ | |
734 "movd 1(%0), %%mm3 \n\t"\ | |
735 "movd 2(%0), %%mm0 \n\t"\ | |
736 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
737 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
738 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
739 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
740 "paddw %%mm0, %%mm1 \n\t"\ | |
741 "paddw %%mm3, %%mm2 \n\t"\ | |
742 "movd -2(%0), %%mm0 \n\t"\ | |
743 "movd 3(%0), %%mm3 \n\t"\ | |
744 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
745 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
746 "paddw %%mm3, %%mm0 \n\t"\ | |
747 "psllw $2, %%mm2 \n\t"\ | |
748 "psubw %%mm1, %%mm2 \n\t"\ | |
749 "pmullw %%mm4, %%mm2 \n\t"\ | |
750 "paddw %%mm5, %%mm0 \n\t"\ | |
751 "paddw %%mm2, %%mm0 \n\t"\ | |
752 "movd (%2), %%mm3 \n\t"\ | |
753 "psraw $5, %%mm0 \n\t"\ | |
754 "packuswb %%mm0, %%mm0 \n\t"\ | |
755 PAVGB" %%mm3, %%mm0 \n\t"\ | |
756 OP(%%mm0, (%1),%%mm6, d)\ | |
3165 | 757 "add %4, %0 \n\t"\ |
758 "add %4, %1 \n\t"\ | |
759 "add %3, %2 \n\t"\ | |
760 : "+a"(src), "+c"(dst), "+d"(src2)\ | |
761 : "D"((long)src2Stride), "S"((long)dstStride)\ | |
3156 | 762 : "memory"\ |
763 );\ | |
3165 | 764 }while(--h);\ |
3156 | 765 }\ |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
766 static av_noinline void OPNAME ## h264_qpel4_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 767 src -= 2*srcStride;\ |
768 asm volatile(\ | |
2979 | 769 "pxor %%mm7, %%mm7 \n\t"\ |
770 "movd (%0), %%mm0 \n\t"\ | |
771 "add %2, %0 \n\t"\ | |
772 "movd (%0), %%mm1 \n\t"\ | |
773 "add %2, %0 \n\t"\ | |
774 "movd (%0), %%mm2 \n\t"\ | |
775 "add %2, %0 \n\t"\ | |
776 "movd (%0), %%mm3 \n\t"\ | |
777 "add %2, %0 \n\t"\ | |
778 "movd (%0), %%mm4 \n\t"\ | |
779 "add %2, %0 \n\t"\ | |
780 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
781 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
782 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
783 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
784 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
2754 | 785 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ |
786 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
787 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
788 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
789 \ | |
790 : "+a"(src), "+c"(dst)\ | |
791 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
792 : "memory"\ | |
793 );\ | |
794 }\ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
795 static av_noinline void OPNAME ## h264_qpel4_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ |
2754 | 796 int h=4;\ |
797 int w=3;\ | |
798 src -= 2*srcStride+2;\ | |
799 while(w--){\ | |
800 asm volatile(\ | |
2979 | 801 "pxor %%mm7, %%mm7 \n\t"\ |
802 "movd (%0), %%mm0 \n\t"\ | |
803 "add %2, %0 \n\t"\ | |
804 "movd (%0), %%mm1 \n\t"\ | |
805 "add %2, %0 \n\t"\ | |
806 "movd (%0), %%mm2 \n\t"\ | |
807 "add %2, %0 \n\t"\ | |
808 "movd (%0), %%mm3 \n\t"\ | |
809 "add %2, %0 \n\t"\ | |
810 "movd (%0), %%mm4 \n\t"\ | |
811 "add %2, %0 \n\t"\ | |
812 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
813 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
814 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
815 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
816 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
2754 | 817 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*8*3)\ |
818 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*8*3)\ | |
819 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*8*3)\ | |
820 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*8*3)\ | |
821 \ | |
822 : "+a"(src)\ | |
823 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5)\ | |
824 : "memory"\ | |
825 );\ | |
826 tmp += 4;\ | |
827 src += 4 - 9*srcStride;\ | |
828 }\ | |
829 tmp -= 3*4;\ | |
830 asm volatile(\ | |
2979 | 831 "movq %4, %%mm6 \n\t"\ |
832 "1: \n\t"\ | |
833 "movq (%0), %%mm0 \n\t"\ | |
834 "paddw 10(%0), %%mm0 \n\t"\ | |
835 "movq 2(%0), %%mm1 \n\t"\ | |
836 "paddw 8(%0), %%mm1 \n\t"\ | |
837 "movq 4(%0), %%mm2 \n\t"\ | |
838 "paddw 6(%0), %%mm2 \n\t"\ | |
839 "psubw %%mm1, %%mm0 \n\t"/*a-b (abccba)*/\ | |
840 "psraw $2, %%mm0 \n\t"/*(a-b)/4 */\ | |
841 "psubw %%mm1, %%mm0 \n\t"/*(a-b)/4-b */\ | |
842 "paddsw %%mm2, %%mm0 \n\t"\ | |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
843 "psraw $2, %%mm0 \n\t"/*((a-b)/4-b+c)/4 */\ |
2979 | 844 "paddw %%mm6, %%mm2 \n\t"\ |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
845 "paddw %%mm2, %%mm0 \n\t"/*(a-5*b+20*c)/16 +32 */\ |
2979 | 846 "psraw $6, %%mm0 \n\t"\ |
847 "packuswb %%mm0, %%mm0 \n\t"\ | |
2754 | 848 OP(%%mm0, (%1),%%mm7, d)\ |
2979 | 849 "add $24, %0 \n\t"\ |
850 "add %3, %1 \n\t"\ | |
851 "decl %2 \n\t"\ | |
852 " jnz 1b \n\t"\ | |
2754 | 853 : "+a"(tmp), "+c"(dst), "+m"(h)\ |
854 : "S"((long)dstStride), "m"(ff_pw_32)\ | |
855 : "memory"\ | |
856 );\ | |
857 }\ | |
858 \ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
859 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 860 int h=8;\ |
861 asm volatile(\ | |
2979 | 862 "pxor %%mm7, %%mm7 \n\t"\ |
863 "movq %5, %%mm6 \n\t"\ | |
864 "1: \n\t"\ | |
865 "movq (%0), %%mm0 \n\t"\ | |
866 "movq 1(%0), %%mm2 \n\t"\ | |
867 "movq %%mm0, %%mm1 \n\t"\ | |
868 "movq %%mm2, %%mm3 \n\t"\ | |
869 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
870 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
871 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
872 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
873 "paddw %%mm2, %%mm0 \n\t"\ | |
874 "paddw %%mm3, %%mm1 \n\t"\ | |
875 "psllw $2, %%mm0 \n\t"\ | |
876 "psllw $2, %%mm1 \n\t"\ | |
877 "movq -1(%0), %%mm2 \n\t"\ | |
878 "movq 2(%0), %%mm4 \n\t"\ | |
879 "movq %%mm2, %%mm3 \n\t"\ | |
880 "movq %%mm4, %%mm5 \n\t"\ | |
881 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
882 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
883 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
884 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
885 "paddw %%mm4, %%mm2 \n\t"\ | |
886 "paddw %%mm3, %%mm5 \n\t"\ | |
887 "psubw %%mm2, %%mm0 \n\t"\ | |
888 "psubw %%mm5, %%mm1 \n\t"\ | |
889 "pmullw %%mm6, %%mm0 \n\t"\ | |
890 "pmullw %%mm6, %%mm1 \n\t"\ | |
891 "movd -2(%0), %%mm2 \n\t"\ | |
892 "movd 7(%0), %%mm5 \n\t"\ | |
893 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
894 "punpcklbw %%mm7, %%mm5 \n\t"\ | |
895 "paddw %%mm3, %%mm2 \n\t"\ | |
896 "paddw %%mm5, %%mm4 \n\t"\ | |
897 "movq %6, %%mm5 \n\t"\ | |
898 "paddw %%mm5, %%mm2 \n\t"\ | |
899 "paddw %%mm5, %%mm4 \n\t"\ | |
900 "paddw %%mm2, %%mm0 \n\t"\ | |
901 "paddw %%mm4, %%mm1 \n\t"\ | |
902 "psraw $5, %%mm0 \n\t"\ | |
903 "psraw $5, %%mm1 \n\t"\ | |
904 "packuswb %%mm1, %%mm0 \n\t"\ | |
2754 | 905 OP(%%mm0, (%1),%%mm5, q)\ |
2979 | 906 "add %3, %0 \n\t"\ |
907 "add %4, %1 \n\t"\ | |
908 "decl %2 \n\t"\ | |
909 " jnz 1b \n\t"\ | |
2754 | 910 : "+a"(src), "+c"(dst), "+m"(h)\ |
911 : "d"((long)srcStride), "S"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
912 : "memory"\ | |
913 );\ | |
914 }\ | |
915 \ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
916 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\ |
3156 | 917 int h=8;\ |
918 asm volatile(\ | |
919 "pxor %%mm7, %%mm7 \n\t"\ | |
3165 | 920 "movq %0, %%mm6 \n\t"\ |
921 :: "m"(ff_pw_5)\ | |
922 );\ | |
923 do{\ | |
924 asm volatile(\ | |
3156 | 925 "movq (%0), %%mm0 \n\t"\ |
926 "movq 1(%0), %%mm2 \n\t"\ | |
927 "movq %%mm0, %%mm1 \n\t"\ | |
928 "movq %%mm2, %%mm3 \n\t"\ | |
929 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
930 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
931 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
932 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
933 "paddw %%mm2, %%mm0 \n\t"\ | |
934 "paddw %%mm3, %%mm1 \n\t"\ | |
935 "psllw $2, %%mm0 \n\t"\ | |
936 "psllw $2, %%mm1 \n\t"\ | |
937 "movq -1(%0), %%mm2 \n\t"\ | |
938 "movq 2(%0), %%mm4 \n\t"\ | |
939 "movq %%mm2, %%mm3 \n\t"\ | |
940 "movq %%mm4, %%mm5 \n\t"\ | |
941 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
942 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
943 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
944 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
945 "paddw %%mm4, %%mm2 \n\t"\ | |
946 "paddw %%mm3, %%mm5 \n\t"\ | |
947 "psubw %%mm2, %%mm0 \n\t"\ | |
948 "psubw %%mm5, %%mm1 \n\t"\ | |
949 "pmullw %%mm6, %%mm0 \n\t"\ | |
950 "pmullw %%mm6, %%mm1 \n\t"\ | |
951 "movd -2(%0), %%mm2 \n\t"\ | |
952 "movd 7(%0), %%mm5 \n\t"\ | |
953 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
954 "punpcklbw %%mm7, %%mm5 \n\t"\ | |
955 "paddw %%mm3, %%mm2 \n\t"\ | |
956 "paddw %%mm5, %%mm4 \n\t"\ | |
3165 | 957 "movq %5, %%mm5 \n\t"\ |
3156 | 958 "paddw %%mm5, %%mm2 \n\t"\ |
959 "paddw %%mm5, %%mm4 \n\t"\ | |
960 "paddw %%mm2, %%mm0 \n\t"\ | |
961 "paddw %%mm4, %%mm1 \n\t"\ | |
962 "psraw $5, %%mm0 \n\t"\ | |
963 "psraw $5, %%mm1 \n\t"\ | |
964 "movq (%2), %%mm4 \n\t"\ | |
965 "packuswb %%mm1, %%mm0 \n\t"\ | |
966 PAVGB" %%mm4, %%mm0 \n\t"\ | |
967 OP(%%mm0, (%1),%%mm5, q)\ | |
3165 | 968 "add %4, %0 \n\t"\ |
969 "add %4, %1 \n\t"\ | |
970 "add %3, %2 \n\t"\ | |
971 : "+a"(src), "+c"(dst), "+d"(src2)\ | |
3156 | 972 : "D"((long)src2Stride), "S"((long)dstStride),\ |
3165 | 973 "m"(ff_pw_16)\ |
3156 | 974 : "memory"\ |
975 );\ | |
3165 | 976 }while(--h);\ |
3156 | 977 }\ |
978 \ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
979 static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
3094 | 980 int w= 2;\ |
2754 | 981 src -= 2*srcStride;\ |
982 \ | |
3094 | 983 while(w--){\ |
2754 | 984 asm volatile(\ |
2979 | 985 "pxor %%mm7, %%mm7 \n\t"\ |
986 "movd (%0), %%mm0 \n\t"\ | |
987 "add %2, %0 \n\t"\ | |
988 "movd (%0), %%mm1 \n\t"\ | |
989 "add %2, %0 \n\t"\ | |
990 "movd (%0), %%mm2 \n\t"\ | |
991 "add %2, %0 \n\t"\ | |
992 "movd (%0), %%mm3 \n\t"\ | |
993 "add %2, %0 \n\t"\ | |
994 "movd (%0), %%mm4 \n\t"\ | |
995 "add %2, %0 \n\t"\ | |
996 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
997 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
998 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
999 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
1000 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
2754 | 1001 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ |
1002 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
1003 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
1004 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
1005 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\ | |
1006 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\ | |
1007 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ | |
1008 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
1009 \ | |
1010 : "+a"(src), "+c"(dst)\ | |
1011 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
1012 : "memory"\ | |
1013 );\ | |
3094 | 1014 if(h==16){\ |
1015 asm volatile(\ | |
1016 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
1017 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
1018 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\ | |
1019 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\ | |
1020 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ | |
1021 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
1022 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
1023 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
1024 \ | |
1025 : "+a"(src), "+c"(dst)\ | |
1026 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
1027 : "memory"\ | |
1028 );\ | |
1029 }\ | |
1030 src += 4-(h+5)*srcStride;\ | |
1031 dst += 4-h*dstStride;\ | |
2754 | 1032 }\ |
1033 }\ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
1034 static av_noinline void OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride, int size){\ |
3093 | 1035 int h = size;\ |
1036 int w = (size+8)>>2;\ | |
2754 | 1037 src -= 2*srcStride+2;\ |
1038 while(w--){\ | |
1039 asm volatile(\ | |
2979 | 1040 "pxor %%mm7, %%mm7 \n\t"\ |
1041 "movd (%0), %%mm0 \n\t"\ | |
1042 "add %2, %0 \n\t"\ | |
1043 "movd (%0), %%mm1 \n\t"\ | |
1044 "add %2, %0 \n\t"\ | |
1045 "movd (%0), %%mm2 \n\t"\ | |
1046 "add %2, %0 \n\t"\ | |
1047 "movd (%0), %%mm3 \n\t"\ | |
1048 "add %2, %0 \n\t"\ | |
1049 "movd (%0), %%mm4 \n\t"\ | |
1050 "add %2, %0 \n\t"\ | |
1051 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1052 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
1053 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1054 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
1055 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
3093 | 1056 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*48)\ |
1057 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*48)\ | |
1058 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*48)\ | |
1059 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*48)\ | |
1060 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 4*48)\ | |
1061 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 5*48)\ | |
1062 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 6*48)\ | |
1063 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 7*48)\ | |
2754 | 1064 : "+a"(src)\ |
1065 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5)\ | |
1066 : "memory"\ | |
1067 );\ | |
3093 | 1068 if(size==16){\ |
1069 asm volatile(\ | |
1070 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 8*48)\ | |
1071 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 9*48)\ | |
1072 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 10*48)\ | |
1073 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 11*48)\ | |
1074 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 12*48)\ | |
1075 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 13*48)\ | |
1076 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 14*48)\ | |
1077 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 15*48)\ | |
1078 : "+a"(src)\ | |
1079 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5)\ | |
1080 : "memory"\ | |
1081 );\ | |
1082 }\ | |
2754 | 1083 tmp += 4;\ |
3093 | 1084 src += 4 - (size+5)*srcStride;\ |
2754 | 1085 }\ |
3093 | 1086 tmp -= size+8;\ |
1087 w = size>>4;\ | |
1088 do{\ | |
1089 h = size;\ | |
2754 | 1090 asm volatile(\ |
2979 | 1091 "movq %4, %%mm6 \n\t"\ |
1092 "1: \n\t"\ | |
1093 "movq (%0), %%mm0 \n\t"\ | |
1094 "movq 8(%0), %%mm3 \n\t"\ | |
1095 "movq 2(%0), %%mm1 \n\t"\ | |
1096 "movq 10(%0), %%mm4 \n\t"\ | |
1097 "paddw %%mm4, %%mm0 \n\t"\ | |
1098 "paddw %%mm3, %%mm1 \n\t"\ | |
1099 "paddw 18(%0), %%mm3 \n\t"\ | |
1100 "paddw 16(%0), %%mm4 \n\t"\ | |
1101 "movq 4(%0), %%mm2 \n\t"\ | |
1102 "movq 12(%0), %%mm5 \n\t"\ | |
1103 "paddw 6(%0), %%mm2 \n\t"\ | |
1104 "paddw 14(%0), %%mm5 \n\t"\ | |
1105 "psubw %%mm1, %%mm0 \n\t"\ | |
1106 "psubw %%mm4, %%mm3 \n\t"\ | |
1107 "psraw $2, %%mm0 \n\t"\ | |
1108 "psraw $2, %%mm3 \n\t"\ | |
1109 "psubw %%mm1, %%mm0 \n\t"\ | |
1110 "psubw %%mm4, %%mm3 \n\t"\ | |
1111 "paddsw %%mm2, %%mm0 \n\t"\ | |
1112 "paddsw %%mm5, %%mm3 \n\t"\ | |
1113 "psraw $2, %%mm0 \n\t"\ | |
1114 "psraw $2, %%mm3 \n\t"\ | |
1115 "paddw %%mm6, %%mm2 \n\t"\ | |
1116 "paddw %%mm6, %%mm5 \n\t"\ | |
1117 "paddw %%mm2, %%mm0 \n\t"\ | |
1118 "paddw %%mm5, %%mm3 \n\t"\ | |
1119 "psraw $6, %%mm0 \n\t"\ | |
1120 "psraw $6, %%mm3 \n\t"\ | |
1121 "packuswb %%mm3, %%mm0 \n\t"\ | |
2754 | 1122 OP(%%mm0, (%1),%%mm7, q)\ |
3093 | 1123 "add $48, %0 \n\t"\ |
2979 | 1124 "add %3, %1 \n\t"\ |
1125 "decl %2 \n\t"\ | |
1126 " jnz 1b \n\t"\ | |
2754 | 1127 : "+a"(tmp), "+c"(dst), "+m"(h)\ |
1128 : "S"((long)dstStride), "m"(ff_pw_32)\ | |
1129 : "memory"\ | |
1130 );\ | |
3093 | 1131 tmp += 8 - size*24;\ |
1132 dst += 8 - size*dstStride;\ | |
1133 }while(w--);\ | |
2754 | 1134 }\ |
3094 | 1135 \ |
1136 static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
1137 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 8);\ | |
1138 }\ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
1139 static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
3094 | 1140 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 16);\ |
1141 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\ | |
2754 | 1142 }\ |
1143 \ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
1144 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 1145 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\ |
1146 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\ | |
1147 src += 8*srcStride;\ | |
1148 dst += 8*dstStride;\ | |
1149 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\ | |
1150 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\ | |
1151 }\ | |
1152 \ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
1153 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\ |
3156 | 1154 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\ |
1155 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\ | |
1156 src += 8*dstStride;\ | |
1157 dst += 8*dstStride;\ | |
1158 src2 += 8*src2Stride;\ | |
1159 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\ | |
1160 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\ | |
1161 }\ | |
1162 \ | |
3093 | 1163 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ |
1164 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 8);\ | |
1165 }\ | |
1166 \ | |
2754 | 1167 static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ |
3093 | 1168 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 16);\ |
2754 | 1169 }\ |
3095 | 1170 \ |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
1171 static av_noinline void OPNAME ## pixels4_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\ |
3095 | 1172 {\ |
1173 asm volatile(\ | |
1174 "movq %5, %%mm6 \n\t"\ | |
1175 "movq (%1), %%mm0 \n\t"\ | |
1176 "movq 24(%1), %%mm1 \n\t"\ | |
1177 "paddw %%mm6, %%mm0 \n\t"\ | |
1178 "paddw %%mm6, %%mm1 \n\t"\ | |
3102 | 1179 "psraw $5, %%mm0 \n\t"\ |
1180 "psraw $5, %%mm1 \n\t"\ | |
3163 | 1181 "packuswb %%mm0, %%mm0 \n\t"\ |
1182 "packuswb %%mm1, %%mm1 \n\t"\ | |
3095 | 1183 PAVGB" (%0), %%mm0 \n\t"\ |
1184 PAVGB" (%0,%3), %%mm1 \n\t"\ | |
1185 OP(%%mm0, (%2), %%mm4, d)\ | |
1186 OP(%%mm1, (%2,%4), %%mm5, d)\ | |
1187 "lea (%0,%3,2), %0 \n\t"\ | |
1188 "lea (%2,%4,2), %2 \n\t"\ | |
1189 "movq 48(%1), %%mm0 \n\t"\ | |
1190 "movq 72(%1), %%mm1 \n\t"\ | |
1191 "paddw %%mm6, %%mm0 \n\t"\ | |
1192 "paddw %%mm6, %%mm1 \n\t"\ | |
3102 | 1193 "psraw $5, %%mm0 \n\t"\ |
1194 "psraw $5, %%mm1 \n\t"\ | |
3163 | 1195 "packuswb %%mm0, %%mm0 \n\t"\ |
1196 "packuswb %%mm1, %%mm1 \n\t"\ | |
3095 | 1197 PAVGB" (%0), %%mm0 \n\t"\ |
1198 PAVGB" (%0,%3), %%mm1 \n\t"\ | |
1199 OP(%%mm0, (%2), %%mm4, d)\ | |
1200 OP(%%mm1, (%2,%4), %%mm5, d)\ | |
1201 :"+a"(src8), "+c"(src16), "+d"(dst)\ | |
3096 | 1202 :"S"((long)src8Stride), "D"((long)dstStride), "m"(ff_pw_16)\ |
3095 | 1203 :"memory");\ |
1204 }\ | |
4527
481763d70193
prevent h.264 MC related functions from being inlined (yes this is much faster the code just doesnt fit in the code cache otherwise)
michael
parents:
4137
diff
changeset
|
1205 static av_noinline void OPNAME ## pixels8_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\ |
3095 | 1206 {\ |
1207 asm volatile(\ | |
1208 "movq %0, %%mm6 \n\t"\ | |
1209 ::"m"(ff_pw_16)\ | |
1210 );\ | |
1211 while(h--){\ | |
1212 asm volatile(\ | |
1213 "movq (%1), %%mm0 \n\t"\ | |
1214 "movq 8(%1), %%mm1 \n\t"\ | |
1215 "paddw %%mm6, %%mm0 \n\t"\ | |
1216 "paddw %%mm6, %%mm1 \n\t"\ | |
3102 | 1217 "psraw $5, %%mm0 \n\t"\ |
1218 "psraw $5, %%mm1 \n\t"\ | |
3095 | 1219 "packuswb %%mm1, %%mm0 \n\t"\ |
1220 PAVGB" (%0), %%mm0 \n\t"\ | |
1221 OP(%%mm0, (%2), %%mm5, q)\ | |
1222 ::"a"(src8), "c"(src16), "d"(dst)\ | |
1223 :"memory");\ | |
1224 src8 += src8Stride;\ | |
1225 src16 += 24;\ | |
1226 dst += dstStride;\ | |
1227 }\ | |
1228 }\ | |
1229 static void OPNAME ## pixels16_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\ | |
1230 {\ | |
1231 OPNAME ## pixels8_l2_shift5_ ## MMX(dst , src16 , src8 , dstStride, src8Stride, h);\ | |
1232 OPNAME ## pixels8_l2_shift5_ ## MMX(dst+8, src16+8, src8+8, dstStride, src8Stride, h);\ | |
1233 }\ | |
1234 | |
2754 | 1235 |
1236 #define H264_MC(OPNAME, SIZE, MMX) \ | |
1237 static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ | |
1238 OPNAME ## pixels ## SIZE ## _mmx(dst, src, stride, SIZE);\ | |
1239 }\ | |
1240 \ | |
1241 static void OPNAME ## h264_qpel ## SIZE ## _mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1242 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src, stride, stride);\ |
2754 | 1243 }\ |
1244 \ | |
1245 static void OPNAME ## h264_qpel ## SIZE ## _mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1246 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1247 }\ | |
1248 \ | |
1249 static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1250 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src+1, stride, stride);\ |
2754 | 1251 }\ |
1252 \ | |
1253 static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1254 uint64_t temp[SIZE*SIZE/8];\ | |
1255 uint8_t * const half= (uint8_t*)temp;\ | |
1256 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(half, src, SIZE, stride);\ | |
1257 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src, half, stride, stride, SIZE);\ | |
1258 }\ | |
1259 \ | |
1260 static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1261 OPNAME ## h264_qpel ## SIZE ## _v_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1262 }\ | |
1263 \ | |
1264 static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1265 uint64_t temp[SIZE*SIZE/8];\ | |
1266 uint8_t * const half= (uint8_t*)temp;\ | |
1267 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(half, src, SIZE, stride);\ | |
1268 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src+stride, half, stride, stride, SIZE);\ | |
1269 }\ | |
1270 \ | |
1271 static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1272 uint64_t temp[SIZE*SIZE/8];\ |
1273 uint8_t * const halfV= (uint8_t*)temp;\ | |
2754 | 1274 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src, SIZE, stride);\ |
3156 | 1275 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfV, stride, SIZE);\ |
2754 | 1276 }\ |
1277 \ | |
1278 static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1279 uint64_t temp[SIZE*SIZE/8];\ |
1280 uint8_t * const halfV= (uint8_t*)temp;\ | |
2754 | 1281 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src+1, SIZE, stride);\ |
3156 | 1282 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfV, stride, SIZE);\ |
2754 | 1283 }\ |
1284 \ | |
1285 static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1286 uint64_t temp[SIZE*SIZE/8];\ |
1287 uint8_t * const halfV= (uint8_t*)temp;\ | |
2754 | 1288 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src, SIZE, stride);\ |
3156 | 1289 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfV, stride, SIZE);\ |
2754 | 1290 }\ |
1291 \ | |
1292 static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1293 uint64_t temp[SIZE*SIZE/8];\ |
1294 uint8_t * const halfV= (uint8_t*)temp;\ | |
2754 | 1295 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src+1, SIZE, stride);\ |
3156 | 1296 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfV, stride, SIZE);\ |
2754 | 1297 }\ |
1298 \ | |
1299 static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3093 | 1300 uint64_t temp[SIZE*(SIZE<8?12:24)/4];\ |
2754 | 1301 int16_t * const tmp= (int16_t*)temp;\ |
1302 OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(dst, tmp, src, stride, SIZE, stride);\ | |
1303 }\ | |
1304 \ | |
1305 static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1306 uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\ |
1307 uint8_t * const halfHV= (uint8_t*)temp;\ | |
1308 int16_t * const tmp= ((int16_t*)temp) + SIZE*SIZE/2;\ | |
5809
7301ea0ae221
Fix intended order of operations for 4 assert() checks.
cehoyos
parents:
4939
diff
changeset
|
1309 assert(((int)temp & 7) == 0);\ |
2754 | 1310 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, tmp, src, SIZE, SIZE, stride);\ |
3156 | 1311 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfHV, stride, SIZE);\ |
2754 | 1312 }\ |
1313 \ | |
1314 static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1315 uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\ |
1316 uint8_t * const halfHV= (uint8_t*)temp;\ | |
1317 int16_t * const tmp= ((int16_t*)temp) + SIZE*SIZE/2;\ | |
5809
7301ea0ae221
Fix intended order of operations for 4 assert() checks.
cehoyos
parents:
4939
diff
changeset
|
1318 assert(((int)temp & 7) == 0);\ |
2754 | 1319 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, tmp, src, SIZE, SIZE, stride);\ |
3156 | 1320 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfHV, stride, SIZE);\ |
2754 | 1321 }\ |
1322 \ | |
1323 static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3095 | 1324 uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\ |
1325 int16_t * const halfV= ((int16_t*)temp) + SIZE*SIZE/2;\ | |
1326 uint8_t * const halfHV= ((uint8_t*)temp);\ | |
5809
7301ea0ae221
Fix intended order of operations for 4 assert() checks.
cehoyos
parents:
4939
diff
changeset
|
1327 assert(((int)temp & 7) == 0);\ |
3095 | 1328 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\ |
1329 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+2, halfHV, stride, SIZE, SIZE);\ | |
2754 | 1330 }\ |
1331 \ | |
1332 static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3095 | 1333 uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\ |
1334 int16_t * const halfV= ((int16_t*)temp) + SIZE*SIZE/2;\ | |
1335 uint8_t * const halfHV= ((uint8_t*)temp);\ | |
5809
7301ea0ae221
Fix intended order of operations for 4 assert() checks.
cehoyos
parents:
4939
diff
changeset
|
1336 assert(((int)temp & 7) == 0);\ |
3095 | 1337 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\ |
1338 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+3, halfHV, stride, SIZE, SIZE);\ | |
2754 | 1339 }\ |
1340 | |
1341 | |
1342 #define AVG_3DNOW_OP(a,b,temp, size) \ | |
2979 | 1343 "mov" #size " " #b ", " #temp " \n\t"\ |
1344 "pavgusb " #temp ", " #a " \n\t"\ | |
1345 "mov" #size " " #a ", " #b " \n\t" | |
2754 | 1346 #define AVG_MMX2_OP(a,b,temp, size) \ |
2979 | 1347 "mov" #size " " #b ", " #temp " \n\t"\ |
1348 "pavgb " #temp ", " #a " \n\t"\ | |
1349 "mov" #size " " #a ", " #b " \n\t" | |
2754 | 1350 |
3095 | 1351 #define PAVGB "pavgusb" |
2754 | 1352 QPEL_H264(put_, PUT_OP, 3dnow) |
1353 QPEL_H264(avg_, AVG_3DNOW_OP, 3dnow) | |
3095 | 1354 #undef PAVGB |
1355 #define PAVGB "pavgb" | |
2754 | 1356 QPEL_H264(put_, PUT_OP, mmx2) |
1357 QPEL_H264(avg_, AVG_MMX2_OP, mmx2) | |
3095 | 1358 #undef PAVGB |
2754 | 1359 |
1360 H264_MC(put_, 4, 3dnow) | |
1361 H264_MC(put_, 8, 3dnow) | |
1362 H264_MC(put_, 16,3dnow) | |
1363 H264_MC(avg_, 4, 3dnow) | |
1364 H264_MC(avg_, 8, 3dnow) | |
1365 H264_MC(avg_, 16,3dnow) | |
1366 H264_MC(put_, 4, mmx2) | |
1367 H264_MC(put_, 8, mmx2) | |
1368 H264_MC(put_, 16,mmx2) | |
1369 H264_MC(avg_, 4, mmx2) | |
1370 H264_MC(avg_, 8, mmx2) | |
1371 H264_MC(avg_, 16,mmx2) | |
1372 | |
1373 | |
1374 #define H264_CHROMA_OP(S,D) | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1375 #define H264_CHROMA_OP4(S,D,T) |
2754 | 1376 #define H264_CHROMA_MC8_TMPL put_h264_chroma_mc8_mmx |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1377 #define H264_CHROMA_MC4_TMPL put_h264_chroma_mc4_mmx |
3213 | 1378 #define H264_CHROMA_MC2_TMPL put_h264_chroma_mc2_mmx2 |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1379 #define H264_CHROMA_MC8_MV0 put_pixels8_mmx |
2754 | 1380 #include "dsputil_h264_template_mmx.c" |
1381 #undef H264_CHROMA_OP | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1382 #undef H264_CHROMA_OP4 |
2754 | 1383 #undef H264_CHROMA_MC8_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1384 #undef H264_CHROMA_MC4_TMPL |
3213 | 1385 #undef H264_CHROMA_MC2_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1386 #undef H264_CHROMA_MC8_MV0 |
2754 | 1387 |
1388 #define H264_CHROMA_OP(S,D) "pavgb " #S ", " #D " \n\t" | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1389 #define H264_CHROMA_OP4(S,D,T) "movd " #S ", " #T " \n\t"\ |
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1390 "pavgb " #T ", " #D " \n\t" |
2754 | 1391 #define H264_CHROMA_MC8_TMPL avg_h264_chroma_mc8_mmx2 |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1392 #define H264_CHROMA_MC4_TMPL avg_h264_chroma_mc4_mmx2 |
3213 | 1393 #define H264_CHROMA_MC2_TMPL avg_h264_chroma_mc2_mmx2 |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1394 #define H264_CHROMA_MC8_MV0 avg_pixels8_mmx2 |
2754 | 1395 #include "dsputil_h264_template_mmx.c" |
1396 #undef H264_CHROMA_OP | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1397 #undef H264_CHROMA_OP4 |
2754 | 1398 #undef H264_CHROMA_MC8_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1399 #undef H264_CHROMA_MC4_TMPL |
3213 | 1400 #undef H264_CHROMA_MC2_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1401 #undef H264_CHROMA_MC8_MV0 |
2754 | 1402 |
1403 #define H264_CHROMA_OP(S,D) "pavgusb " #S ", " #D " \n\t" | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1404 #define H264_CHROMA_OP4(S,D,T) "movd " #S ", " #T " \n\t"\ |
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1405 "pavgusb " #T ", " #D " \n\t" |
2754 | 1406 #define H264_CHROMA_MC8_TMPL avg_h264_chroma_mc8_3dnow |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1407 #define H264_CHROMA_MC4_TMPL avg_h264_chroma_mc4_3dnow |
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1408 #define H264_CHROMA_MC8_MV0 avg_pixels8_3dnow |
2754 | 1409 #include "dsputil_h264_template_mmx.c" |
1410 #undef H264_CHROMA_OP | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1411 #undef H264_CHROMA_OP4 |
2754 | 1412 #undef H264_CHROMA_MC8_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1413 #undef H264_CHROMA_MC4_TMPL |
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1414 #undef H264_CHROMA_MC8_MV0 |
2754 | 1415 |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1416 /***********************************/ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1417 /* weighted prediction */ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1418 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1419 static inline void ff_h264_weight_WxH_mmx2(uint8_t *dst, int stride, int log2_denom, int weight, int offset, int w, int h) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1420 { |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1421 int x, y; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1422 offset <<= log2_denom; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1423 offset += (1 << log2_denom) >> 1; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1424 asm volatile( |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1425 "movd %0, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1426 "movd %1, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1427 "movd %2, %%mm6 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1428 "pshufw $0, %%mm4, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1429 "pshufw $0, %%mm5, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1430 "pxor %%mm7, %%mm7 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1431 :: "g"(weight), "g"(offset), "g"(log2_denom) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1432 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1433 for(y=0; y<h; y+=2){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1434 for(x=0; x<w; x+=4){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1435 asm volatile( |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1436 "movd %0, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1437 "movd %1, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1438 "punpcklbw %%mm7, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1439 "punpcklbw %%mm7, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1440 "pmullw %%mm4, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1441 "pmullw %%mm4, %%mm1 \n\t" |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
1442 "paddsw %%mm5, %%mm0 \n\t" |
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
1443 "paddsw %%mm5, %%mm1 \n\t" |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1444 "psraw %%mm6, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1445 "psraw %%mm6, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1446 "packuswb %%mm7, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1447 "packuswb %%mm7, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1448 "movd %%mm0, %0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1449 "movd %%mm1, %1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1450 : "+m"(*(uint32_t*)(dst+x)), |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1451 "+m"(*(uint32_t*)(dst+x+stride)) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1452 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1453 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1454 dst += 2*stride; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1455 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1456 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1457 |
3029 | 1458 static inline void ff_h264_biweight_WxH_mmx2(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset, int w, int h) |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1459 { |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1460 int x, y; |
3029 | 1461 offset = ((offset + 1) | 1) << log2_denom; |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1462 asm volatile( |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1463 "movd %0, %%mm3 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1464 "movd %1, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1465 "movd %2, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1466 "movd %3, %%mm6 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1467 "pshufw $0, %%mm3, %%mm3 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1468 "pshufw $0, %%mm4, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1469 "pshufw $0, %%mm5, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1470 "pxor %%mm7, %%mm7 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1471 :: "g"(weightd), "g"(weights), "g"(offset), "g"(log2_denom+1) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1472 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1473 for(y=0; y<h; y++){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1474 for(x=0; x<w; x+=4){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1475 asm volatile( |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1476 "movd %0, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1477 "movd %1, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1478 "punpcklbw %%mm7, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1479 "punpcklbw %%mm7, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1480 "pmullw %%mm3, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1481 "pmullw %%mm4, %%mm1 \n\t" |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
1482 "paddsw %%mm1, %%mm0 \n\t" |
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
1483 "paddsw %%mm5, %%mm0 \n\t" |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1484 "psraw %%mm6, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1485 "packuswb %%mm0, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1486 "movd %%mm0, %0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1487 : "+m"(*(uint32_t*)(dst+x)) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1488 : "m"(*(uint32_t*)(src+x)) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1489 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1490 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1491 src += stride; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1492 dst += stride; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1493 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1494 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1495 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1496 #define H264_WEIGHT(W,H) \ |
3029 | 1497 static void ff_h264_biweight_ ## W ## x ## H ## _mmx2(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset){ \ |
1498 ff_h264_biweight_WxH_mmx2(dst, src, stride, log2_denom, weightd, weights, offset, W, H); \ | |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1499 } \ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1500 static void ff_h264_weight_ ## W ## x ## H ## _mmx2(uint8_t *dst, int stride, int log2_denom, int weight, int offset){ \ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1501 ff_h264_weight_WxH_mmx2(dst, stride, log2_denom, weight, offset, W, H); \ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1502 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1503 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1504 H264_WEIGHT(16,16) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1505 H264_WEIGHT(16, 8) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1506 H264_WEIGHT( 8,16) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1507 H264_WEIGHT( 8, 8) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1508 H264_WEIGHT( 8, 4) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1509 H264_WEIGHT( 4, 8) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1510 H264_WEIGHT( 4, 4) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1511 H264_WEIGHT( 4, 2) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1512 |