Mercurial > libavcodec.hg
annotate i386/h264dsp_mmx.c @ 4693:86a2dfcc5eef libavcodec
Smacker Audio stereo fix
author | kostya |
---|---|
date | Wed, 21 Mar 2007 05:56:29 +0000 |
parents | 33c6fe489f62 |
children | 6c66ddbb054f |
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 | |
21 | |
22 /***********************************/ | |
23 /* IDCT */ | |
24 | |
25 /* in/out: mma=mma+mmb, mmb=mmb-mma */ | |
26 #define SUMSUB_BA( a, b ) \ | |
27 "paddw "#b", "#a" \n\t"\ | |
28 "paddw "#b", "#b" \n\t"\ | |
29 "psubw "#a", "#b" \n\t" | |
30 | |
31 #define SUMSUB_BADC( a, b, c, d ) \ | |
32 "paddw "#b", "#a" \n\t"\ | |
33 "paddw "#d", "#c" \n\t"\ | |
34 "paddw "#b", "#b" \n\t"\ | |
35 "paddw "#d", "#d" \n\t"\ | |
36 "psubw "#a", "#b" \n\t"\ | |
37 "psubw "#c", "#d" \n\t" | |
38 | |
39 #define SUMSUBD2_AB( a, b, t ) \ | |
40 "movq "#b", "#t" \n\t"\ | |
41 "psraw $1 , "#b" \n\t"\ | |
42 "paddw "#a", "#b" \n\t"\ | |
43 "psraw $1 , "#a" \n\t"\ | |
44 "psubw "#t", "#a" \n\t" | |
45 | |
46 #define IDCT4_1D( s02, s13, d02, d13, t ) \ | |
47 SUMSUB_BA ( s02, d02 )\ | |
48 SUMSUBD2_AB( s13, d13, t )\ | |
49 SUMSUB_BADC( d13, s02, s13, d02 ) | |
50 | |
51 #define TRANSPOSE4(a,b,c,d,t)\ | |
52 SBUTTERFLY(a,b,t,wd) /* a=aebf t=cgdh */\ | |
53 SBUTTERFLY(c,d,b,wd) /* c=imjn b=kolp */\ | |
54 SBUTTERFLY(a,c,d,dq) /* a=aeim d=bfjn */\ | |
55 SBUTTERFLY(t,b,c,dq) /* t=cgko c=dhlp */ | |
56 | |
57 #define STORE_DIFF_4P( p, t, z ) \ | |
58 "psraw $6, "#p" \n\t"\ | |
59 "movd (%0), "#t" \n\t"\ | |
60 "punpcklbw "#z", "#t" \n\t"\ | |
61 "paddsw "#t", "#p" \n\t"\ | |
62 "packuswb "#z", "#p" \n\t"\ | |
63 "movd "#p", (%0) \n\t" | |
64 | |
3173 | 65 static void ff_h264_idct_add_mmx(uint8_t *dst, int16_t *block, int stride) |
2754 | 66 { |
67 /* Load dct coeffs */ | |
68 asm volatile( | |
69 "movq (%0), %%mm0 \n\t" | |
70 "movq 8(%0), %%mm1 \n\t" | |
71 "movq 16(%0), %%mm2 \n\t" | |
72 "movq 24(%0), %%mm3 \n\t" | |
73 :: "r"(block) ); | |
74 | |
75 asm volatile( | |
76 /* mm1=s02+s13 mm2=s02-s13 mm4=d02+d13 mm0=d02-d13 */ | |
77 IDCT4_1D( %%mm2, %%mm1, %%mm0, %%mm3, %%mm4 ) | |
78 | |
79 "movq %0, %%mm6 \n\t" | |
80 /* in: 1,4,0,2 out: 1,2,3,0 */ | |
81 TRANSPOSE4( %%mm3, %%mm1, %%mm0, %%mm2, %%mm4 ) | |
82 | |
83 "paddw %%mm6, %%mm3 \n\t" | |
84 | |
85 /* mm2=s02+s13 mm3=s02-s13 mm4=d02+d13 mm1=d02-d13 */ | |
86 IDCT4_1D( %%mm4, %%mm2, %%mm3, %%mm0, %%mm1 ) | |
87 | |
88 "pxor %%mm7, %%mm7 \n\t" | |
89 :: "m"(ff_pw_32)); | |
90 | |
91 asm volatile( | |
92 STORE_DIFF_4P( %%mm0, %%mm1, %%mm7) | |
93 "add %1, %0 \n\t" | |
94 STORE_DIFF_4P( %%mm2, %%mm1, %%mm7) | |
95 "add %1, %0 \n\t" | |
96 STORE_DIFF_4P( %%mm3, %%mm1, %%mm7) | |
97 "add %1, %0 \n\t" | |
98 STORE_DIFF_4P( %%mm4, %%mm1, %%mm7) | |
99 : "+r"(dst) | |
100 : "r" ((long)stride) | |
101 ); | |
102 } | |
103 | |
3174 | 104 static inline void h264_idct8_1d(int16_t *block) |
105 { | |
106 asm volatile( | |
107 "movq 112(%0), %%mm7 \n\t" | |
108 "movq 80(%0), %%mm5 \n\t" | |
109 "movq 48(%0), %%mm3 \n\t" | |
110 "movq 16(%0), %%mm1 \n\t" | |
111 | |
112 "movq %%mm7, %%mm4 \n\t" | |
113 "movq %%mm3, %%mm6 \n\t" | |
114 "movq %%mm5, %%mm0 \n\t" | |
115 "movq %%mm7, %%mm2 \n\t" | |
116 "psraw $1, %%mm4 \n\t" | |
117 "psraw $1, %%mm6 \n\t" | |
118 "psubw %%mm7, %%mm0 \n\t" | |
119 "psubw %%mm6, %%mm2 \n\t" | |
120 "psubw %%mm4, %%mm0 \n\t" | |
121 "psubw %%mm3, %%mm2 \n\t" | |
122 "psubw %%mm3, %%mm0 \n\t" | |
123 "paddw %%mm1, %%mm2 \n\t" | |
124 | |
125 "movq %%mm5, %%mm4 \n\t" | |
126 "movq %%mm1, %%mm6 \n\t" | |
127 "psraw $1, %%mm4 \n\t" | |
128 "psraw $1, %%mm6 \n\t" | |
129 "paddw %%mm5, %%mm4 \n\t" | |
130 "paddw %%mm1, %%mm6 \n\t" | |
131 "paddw %%mm7, %%mm4 \n\t" | |
132 "paddw %%mm5, %%mm6 \n\t" | |
133 "psubw %%mm1, %%mm4 \n\t" | |
134 "paddw %%mm3, %%mm6 \n\t" | |
135 | |
136 "movq %%mm0, %%mm1 \n\t" | |
137 "movq %%mm4, %%mm3 \n\t" | |
138 "movq %%mm2, %%mm5 \n\t" | |
139 "movq %%mm6, %%mm7 \n\t" | |
140 "psraw $2, %%mm6 \n\t" | |
141 "psraw $2, %%mm3 \n\t" | |
142 "psraw $2, %%mm5 \n\t" | |
143 "psraw $2, %%mm0 \n\t" | |
144 "paddw %%mm6, %%mm1 \n\t" | |
145 "paddw %%mm2, %%mm3 \n\t" | |
146 "psubw %%mm4, %%mm5 \n\t" | |
147 "psubw %%mm0, %%mm7 \n\t" | |
148 | |
149 "movq 32(%0), %%mm2 \n\t" | |
150 "movq 96(%0), %%mm6 \n\t" | |
151 "movq %%mm2, %%mm4 \n\t" | |
152 "movq %%mm6, %%mm0 \n\t" | |
153 "psraw $1, %%mm4 \n\t" | |
154 "psraw $1, %%mm6 \n\t" | |
155 "psubw %%mm0, %%mm4 \n\t" | |
156 "paddw %%mm2, %%mm6 \n\t" | |
157 | |
158 "movq (%0), %%mm2 \n\t" | |
159 "movq 64(%0), %%mm0 \n\t" | |
160 SUMSUB_BA( %%mm0, %%mm2 ) | |
161 SUMSUB_BA( %%mm6, %%mm0 ) | |
162 SUMSUB_BA( %%mm4, %%mm2 ) | |
163 SUMSUB_BA( %%mm7, %%mm6 ) | |
164 SUMSUB_BA( %%mm5, %%mm4 ) | |
165 SUMSUB_BA( %%mm3, %%mm2 ) | |
166 SUMSUB_BA( %%mm1, %%mm0 ) | |
167 :: "r"(block) | |
168 ); | |
169 } | |
170 | |
171 static void ff_h264_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride) | |
172 { | |
173 int i; | |
174 int16_t __attribute__ ((aligned(8))) b2[64]; | |
175 | |
176 block[0] += 32; | |
177 | |
178 for(i=0; i<2; i++){ | |
4137 | 179 DECLARE_ALIGNED_8(uint64_t, tmp); |
3174 | 180 |
181 h264_idct8_1d(block+4*i); | |
182 | |
183 asm volatile( | |
184 "movq %%mm7, %0 \n\t" | |
185 TRANSPOSE4( %%mm0, %%mm2, %%mm4, %%mm6, %%mm7 ) | |
186 "movq %%mm0, 8(%1) \n\t" | |
187 "movq %%mm6, 24(%1) \n\t" | |
188 "movq %%mm7, 40(%1) \n\t" | |
189 "movq %%mm4, 56(%1) \n\t" | |
190 "movq %0, %%mm7 \n\t" | |
191 TRANSPOSE4( %%mm7, %%mm5, %%mm3, %%mm1, %%mm0 ) | |
192 "movq %%mm7, (%1) \n\t" | |
193 "movq %%mm1, 16(%1) \n\t" | |
194 "movq %%mm0, 32(%1) \n\t" | |
195 "movq %%mm3, 48(%1) \n\t" | |
196 : "=m"(tmp) | |
197 : "r"(b2+32*i) | |
198 : "memory" | |
199 ); | |
200 } | |
201 | |
202 for(i=0; i<2; i++){ | |
203 h264_idct8_1d(b2+4*i); | |
204 | |
205 asm volatile( | |
206 "psraw $6, %%mm7 \n\t" | |
207 "psraw $6, %%mm6 \n\t" | |
208 "psraw $6, %%mm5 \n\t" | |
209 "psraw $6, %%mm4 \n\t" | |
210 "psraw $6, %%mm3 \n\t" | |
211 "psraw $6, %%mm2 \n\t" | |
212 "psraw $6, %%mm1 \n\t" | |
213 "psraw $6, %%mm0 \n\t" | |
214 | |
215 "movq %%mm7, (%0) \n\t" | |
216 "movq %%mm5, 16(%0) \n\t" | |
217 "movq %%mm3, 32(%0) \n\t" | |
218 "movq %%mm1, 48(%0) \n\t" | |
219 "movq %%mm0, 64(%0) \n\t" | |
220 "movq %%mm2, 80(%0) \n\t" | |
221 "movq %%mm4, 96(%0) \n\t" | |
222 "movq %%mm6, 112(%0) \n\t" | |
223 :: "r"(b2+4*i) | |
224 : "memory" | |
225 ); | |
226 } | |
227 | |
228 add_pixels_clamped_mmx(b2, dst, stride); | |
229 } | |
230 | |
3173 | 231 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
|
232 { |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
233 int dc = (block[0] + 32) >> 6; |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
234 asm volatile( |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
235 "movd %0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
236 "pshufw $0, %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
237 "pxor %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
238 "psubw %%mm0, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
239 "packuswb %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
240 "packuswb %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
241 ::"r"(dc) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
242 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
243 asm volatile( |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
244 "movd %0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
245 "movd %1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
246 "movd %2, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
247 "movd %3, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
248 "paddusb %%mm0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
249 "paddusb %%mm0, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
250 "paddusb %%mm0, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
251 "paddusb %%mm0, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
252 "psubusb %%mm1, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
253 "psubusb %%mm1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
254 "psubusb %%mm1, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
255 "psubusb %%mm1, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
256 "movd %%mm2, %0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
257 "movd %%mm3, %1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
258 "movd %%mm4, %2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
259 "movd %%mm5, %3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
260 :"+m"(*(uint32_t*)(dst+0*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
261 "+m"(*(uint32_t*)(dst+1*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
262 "+m"(*(uint32_t*)(dst+2*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
263 "+m"(*(uint32_t*)(dst+3*stride)) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
264 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
265 } |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
266 |
3173 | 267 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
|
268 { |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
269 int dc = (block[0] + 32) >> 6; |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
270 int y; |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
271 asm volatile( |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
272 "movd %0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
273 "pshufw $0, %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
274 "pxor %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
275 "psubw %%mm0, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
276 "packuswb %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
277 "packuswb %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
278 ::"r"(dc) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
279 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
280 for(y=2; y--; dst += 4*stride){ |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
281 asm volatile( |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
282 "movq %0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
283 "movq %1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
284 "movq %2, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
285 "movq %3, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
286 "paddusb %%mm0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
287 "paddusb %%mm0, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
288 "paddusb %%mm0, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
289 "paddusb %%mm0, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
290 "psubusb %%mm1, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
291 "psubusb %%mm1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
292 "psubusb %%mm1, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
293 "psubusb %%mm1, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
294 "movq %%mm2, %0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
295 "movq %%mm3, %1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
296 "movq %%mm4, %2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
297 "movq %%mm5, %3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
298 :"+m"(*(uint64_t*)(dst+0*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
299 "+m"(*(uint64_t*)(dst+1*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
300 "+m"(*(uint64_t*)(dst+2*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
301 "+m"(*(uint64_t*)(dst+3*stride)) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
302 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
303 } |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
304 } |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
305 |
2754 | 306 |
307 /***********************************/ | |
308 /* deblocking */ | |
309 | |
310 // out: o = |x-y|>a | |
311 // clobbers: t | |
312 #define DIFF_GT_MMX(x,y,a,o,t)\ | |
313 "movq "#y", "#t" \n\t"\ | |
314 "movq "#x", "#o" \n\t"\ | |
315 "psubusb "#x", "#t" \n\t"\ | |
316 "psubusb "#y", "#o" \n\t"\ | |
317 "por "#t", "#o" \n\t"\ | |
318 "psubusb "#a", "#o" \n\t" | |
319 | |
4135 | 320 // out: o = |x-y|>a |
321 // clobbers: t | |
322 #define DIFF_GT2_MMX(x,y,a,o,t)\ | |
323 "movq "#y", "#t" \n\t"\ | |
324 "movq "#x", "#o" \n\t"\ | |
325 "psubusb "#x", "#t" \n\t"\ | |
326 "psubusb "#y", "#o" \n\t"\ | |
327 "psubusb "#a", "#t" \n\t"\ | |
328 "psubusb "#a", "#o" \n\t"\ | |
329 "pcmpeqb "#t", "#o" \n\t"\ | |
330 | |
2754 | 331 // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 |
332 // out: mm5=beta-1, mm7=mask | |
333 // clobbers: mm4,mm6 | |
334 #define H264_DEBLOCK_MASK(alpha1, beta1) \ | |
335 "pshufw $0, "#alpha1", %%mm4 \n\t"\ | |
336 "pshufw $0, "#beta1 ", %%mm5 \n\t"\ | |
337 "packuswb %%mm4, %%mm4 \n\t"\ | |
338 "packuswb %%mm5, %%mm5 \n\t"\ | |
339 DIFF_GT_MMX(%%mm1, %%mm2, %%mm4, %%mm7, %%mm6) /* |p0-q0| > alpha-1 */\ | |
340 DIFF_GT_MMX(%%mm0, %%mm1, %%mm5, %%mm4, %%mm6) /* |p1-p0| > beta-1 */\ | |
341 "por %%mm4, %%mm7 \n\t"\ | |
342 DIFF_GT_MMX(%%mm3, %%mm2, %%mm5, %%mm4, %%mm6) /* |q1-q0| > beta-1 */\ | |
343 "por %%mm4, %%mm7 \n\t"\ | |
344 "pxor %%mm6, %%mm6 \n\t"\ | |
345 "pcmpeqb %%mm6, %%mm7 \n\t" | |
346 | |
347 // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask) | |
348 // out: mm1=p0' mm2=q0' | |
349 // clobbers: mm0,3-6 | |
350 #define H264_DEBLOCK_P0_Q0(pb_01, pb_3f)\ | |
4127 | 351 "movq %%mm1 , %%mm5 \n\t"\ |
352 "pxor %%mm2 , %%mm5 \n\t" /* p0^q0*/\ | |
353 "pand "#pb_01" , %%mm5 \n\t" /* (p0^q0)&1*/\ | |
354 "pcmpeqb %%mm4 , %%mm4 \n\t"\ | |
355 "pxor %%mm4 , %%mm3 \n\t"\ | |
356 "pavgb %%mm0 , %%mm3 \n\t" /* (p1 - q1 + 256)>>1*/\ | |
357 "pavgb "MANGLE(ff_pb_3)" , %%mm3 \n\t" /*(((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2*/\ | |
358 "pxor %%mm1 , %%mm4 \n\t"\ | |
359 "pavgb %%mm2 , %%mm4 \n\t" /* (q0 - p0 + 256)>>1*/\ | |
360 "pavgb %%mm5 , %%mm3 \n\t"\ | |
4130 | 361 "paddusb %%mm4 , %%mm3 \n\t" /* d+128+33*/\ |
4129 | 362 "movq "MANGLE(ff_pb_A1)" , %%mm6 \n\t"\ |
363 "psubusb %%mm3 , %%mm6 \n\t"\ | |
4127 | 364 "psubusb "MANGLE(ff_pb_A1)" , %%mm3 \n\t"\ |
4129 | 365 "pminub %%mm7 , %%mm6 \n\t"\ |
4127 | 366 "pminub %%mm7 , %%mm3 \n\t"\ |
4129 | 367 "psubusb %%mm6 , %%mm1 \n\t"\ |
368 "psubusb %%mm3 , %%mm2 \n\t"\ | |
4127 | 369 "paddusb %%mm3 , %%mm1 \n\t"\ |
4129 | 370 "paddusb %%mm6 , %%mm2 \n\t" |
2754 | 371 |
372 // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask) %8=mm_bone | |
373 // out: (q1addr) = clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 ) | |
374 // clobbers: q2, tmp, tc0 | |
375 #define H264_DEBLOCK_Q1(p1, q2, q2addr, q1addr, tc0, tmp)\ | |
376 "movq %%mm1, "#tmp" \n\t"\ | |
377 "pavgb %%mm2, "#tmp" \n\t"\ | |
378 "pavgb "#tmp", "#q2" \n\t" /* avg(p2,avg(p0,q0)) */\ | |
379 "pxor "q2addr", "#tmp" \n\t"\ | |
380 "pand %8, "#tmp" \n\t" /* (p2^avg(p0,q0))&1 */\ | |
381 "psubusb "#tmp", "#q2" \n\t" /* (p2+((p0+q0+1)>>1))>>1 */\ | |
382 "movq "#p1", "#tmp" \n\t"\ | |
383 "psubusb "#tc0", "#tmp" \n\t"\ | |
384 "paddusb "#p1", "#tc0" \n\t"\ | |
385 "pmaxub "#tmp", "#q2" \n\t"\ | |
386 "pminub "#tc0", "#q2" \n\t"\ | |
387 "movq "#q2", "q1addr" \n\t" | |
388 | |
389 static inline void h264_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0) | |
390 { | |
4137 | 391 DECLARE_ALIGNED_8(uint64_t, tmp0[2]); |
2754 | 392 |
393 asm volatile( | |
394 "movq (%1,%3), %%mm0 \n\t" //p1 | |
395 "movq (%1,%3,2), %%mm1 \n\t" //p0 | |
396 "movq (%2), %%mm2 \n\t" //q0 | |
397 "movq (%2,%3), %%mm3 \n\t" //q1 | |
398 H264_DEBLOCK_MASK(%6, %7) | |
4133 | 399 |
400 "movd %5, %%mm4 \n\t" | |
401 "punpcklbw %%mm4, %%mm4 \n\t" | |
402 "punpcklwd %%mm4, %%mm4 \n\t" | |
403 "pcmpeqb %%mm3, %%mm3 \n\t" | |
404 "movq %%mm4, %%mm6 \n\t" | |
405 "pcmpgtb %%mm3, %%mm4 \n\t" | |
406 "movq %%mm6, 8+%0 \n\t" | |
407 "pand %%mm4, %%mm7 \n\t" | |
408 "movq %%mm7, %0 \n\t" | |
2754 | 409 |
410 /* filter p1 */ | |
411 "movq (%1), %%mm3 \n\t" //p2 | |
4135 | 412 DIFF_GT2_MMX(%%mm1, %%mm3, %%mm5, %%mm6, %%mm4) // |p2-p0|>beta-1 |
2754 | 413 "pand %%mm7, %%mm6 \n\t" // mask & |p2-p0|<beta |
4136 | 414 "pand 8+%0, %%mm7 \n\t" // mask & tc0 |
415 "movq %%mm7, %%mm4 \n\t" | |
4131
1a8e384d0463
2 instructions less in h264_loop_filter_luma_mmx2()
michael
parents:
4130
diff
changeset
|
416 "psubb %%mm6, %%mm7 \n\t" |
2754 | 417 "pand %%mm4, %%mm6 \n\t" // mask & |p2-p0|<beta & tc0 |
418 H264_DEBLOCK_Q1(%%mm0, %%mm3, "(%1)", "(%1,%3)", %%mm6, %%mm4) | |
419 | |
420 /* filter q1 */ | |
421 "movq (%2,%3,2), %%mm4 \n\t" //q2 | |
4135 | 422 DIFF_GT2_MMX(%%mm2, %%mm4, %%mm5, %%mm6, %%mm3) // |q2-q0|>beta-1 |
2754 | 423 "pand %0, %%mm6 \n\t" |
4134 | 424 "movq 8+%0, %%mm5 \n\t" // can be merged with the and below but is slower then |
2754 | 425 "pand %%mm6, %%mm5 \n\t" |
4131
1a8e384d0463
2 instructions less in h264_loop_filter_luma_mmx2()
michael
parents:
4130
diff
changeset
|
426 "psubb %%mm6, %%mm7 \n\t" |
2754 | 427 "movq (%2,%3), %%mm3 \n\t" |
428 H264_DEBLOCK_Q1(%%mm3, %%mm4, "(%2,%3,2)", "(%2,%3)", %%mm5, %%mm6) | |
429 | |
430 /* filter p0, q0 */ | |
4133 | 431 H264_DEBLOCK_P0_Q0(%8, unused) |
2754 | 432 "movq %%mm1, (%1,%3,2) \n\t" |
433 "movq %%mm2, (%2) \n\t" | |
434 | |
4133 | 435 : "=m"(*tmp0) |
2754 | 436 : "r"(pix-3*stride), "r"(pix), "r"((long)stride), |
4133 | 437 "m"(*tmp0/*unused*/), "m"(*(uint32_t*)tc0), "m"(alpha1), "m"(beta1), |
438 "m"(mm_bone) | |
2754 | 439 ); |
440 } | |
441 | |
442 static void h264_v_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
443 { | |
444 if((tc0[0] & tc0[1]) >= 0) | |
445 h264_loop_filter_luma_mmx2(pix, stride, alpha-1, beta-1, tc0); | |
446 if((tc0[2] & tc0[3]) >= 0) | |
447 h264_loop_filter_luma_mmx2(pix+8, stride, alpha-1, beta-1, tc0+2); | |
448 } | |
449 static void h264_h_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
450 { | |
451 //FIXME: could cut some load/stores by merging transpose with filter | |
452 // also, it only needs to transpose 6x8 | |
4137 | 453 DECLARE_ALIGNED_8(uint8_t, trans[8*8]); |
2754 | 454 int i; |
455 for(i=0; i<2; i++, pix+=8*stride, tc0+=2) { | |
456 if((tc0[0] & tc0[1]) < 0) | |
457 continue; | |
458 transpose4x4(trans, pix-4, 8, stride); | |
459 transpose4x4(trans +4*8, pix, 8, stride); | |
460 transpose4x4(trans+4, pix-4+4*stride, 8, stride); | |
461 transpose4x4(trans+4+4*8, pix +4*stride, 8, stride); | |
462 h264_loop_filter_luma_mmx2(trans+4*8, 8, alpha-1, beta-1, tc0); | |
463 transpose4x4(pix-2, trans +2*8, stride, 8); | |
464 transpose4x4(pix-2+4*stride, trans+4+2*8, stride, 8); | |
465 } | |
466 } | |
467 | |
468 static inline void h264_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0) | |
469 { | |
470 asm volatile( | |
471 "movq (%0), %%mm0 \n\t" //p1 | |
472 "movq (%0,%2), %%mm1 \n\t" //p0 | |
473 "movq (%1), %%mm2 \n\t" //q0 | |
474 "movq (%1,%2), %%mm3 \n\t" //q1 | |
475 H264_DEBLOCK_MASK(%4, %5) | |
476 "movd %3, %%mm6 \n\t" | |
477 "punpcklbw %%mm6, %%mm6 \n\t" | |
478 "pand %%mm6, %%mm7 \n\t" // mm7 = tc&mask | |
479 H264_DEBLOCK_P0_Q0(%6, %7) | |
480 "movq %%mm1, (%0,%2) \n\t" | |
481 "movq %%mm2, (%1) \n\t" | |
482 | |
483 :: "r"(pix-2*stride), "r"(pix), "r"((long)stride), | |
484 "r"(*(uint32_t*)tc0), | |
485 "m"(alpha1), "m"(beta1), "m"(mm_bone), "m"(ff_pb_3F) | |
486 ); | |
487 } | |
488 | |
489 static void h264_v_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
490 { | |
491 h264_loop_filter_chroma_mmx2(pix, stride, alpha-1, beta-1, tc0); | |
492 } | |
493 | |
494 static void h264_h_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
495 { | |
496 //FIXME: could cut some load/stores by merging transpose with filter | |
4137 | 497 DECLARE_ALIGNED_8(uint8_t, trans[8*4]); |
2754 | 498 transpose4x4(trans, pix-2, 8, stride); |
499 transpose4x4(trans+4, pix-2+4*stride, 8, stride); | |
500 h264_loop_filter_chroma_mmx2(trans+2*8, 8, alpha-1, beta-1, tc0); | |
501 transpose4x4(pix-2, trans, stride, 8); | |
502 transpose4x4(pix-2+4*stride, trans+4, stride, 8); | |
503 } | |
504 | |
505 // p0 = (p0 + q1 + 2*p1 + 2) >> 2 | |
506 #define H264_FILTER_CHROMA4(p0, p1, q1, one) \ | |
507 "movq "#p0", %%mm4 \n\t"\ | |
508 "pxor "#q1", %%mm4 \n\t"\ | |
509 "pand "#one", %%mm4 \n\t" /* mm4 = (p0^q1)&1 */\ | |
510 "pavgb "#q1", "#p0" \n\t"\ | |
511 "psubusb %%mm4, "#p0" \n\t"\ | |
512 "pavgb "#p1", "#p0" \n\t" /* dst = avg(p1, avg(p0,q1) - ((p0^q1)&1)) */\ | |
513 | |
514 static inline void h264_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha1, int beta1) | |
515 { | |
516 asm volatile( | |
517 "movq (%0), %%mm0 \n\t" | |
518 "movq (%0,%2), %%mm1 \n\t" | |
519 "movq (%1), %%mm2 \n\t" | |
520 "movq (%1,%2), %%mm3 \n\t" | |
521 H264_DEBLOCK_MASK(%3, %4) | |
522 "movq %%mm1, %%mm5 \n\t" | |
523 "movq %%mm2, %%mm6 \n\t" | |
524 H264_FILTER_CHROMA4(%%mm1, %%mm0, %%mm3, %5) //p0' | |
525 H264_FILTER_CHROMA4(%%mm2, %%mm3, %%mm0, %5) //q0' | |
526 "psubb %%mm5, %%mm1 \n\t" | |
527 "psubb %%mm6, %%mm2 \n\t" | |
528 "pand %%mm7, %%mm1 \n\t" | |
529 "pand %%mm7, %%mm2 \n\t" | |
530 "paddb %%mm5, %%mm1 \n\t" | |
531 "paddb %%mm6, %%mm2 \n\t" | |
532 "movq %%mm1, (%0,%2) \n\t" | |
533 "movq %%mm2, (%1) \n\t" | |
534 :: "r"(pix-2*stride), "r"(pix), "r"((long)stride), | |
535 "m"(alpha1), "m"(beta1), "m"(mm_bone) | |
536 ); | |
537 } | |
538 | |
539 static void h264_v_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta) | |
540 { | |
541 h264_loop_filter_chroma_intra_mmx2(pix, stride, alpha-1, beta-1); | |
542 } | |
543 | |
544 static void h264_h_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta) | |
545 { | |
546 //FIXME: could cut some load/stores by merging transpose with filter | |
4137 | 547 DECLARE_ALIGNED_8(uint8_t, trans[8*4]); |
2754 | 548 transpose4x4(trans, pix-2, 8, stride); |
549 transpose4x4(trans+4, pix-2+4*stride, 8, stride); | |
550 h264_loop_filter_chroma_intra_mmx2(trans+2*8, 8, alpha-1, beta-1); | |
551 transpose4x4(pix-2, trans, stride, 8); | |
552 transpose4x4(pix-2+4*stride, trans+4, stride, 8); | |
553 } | |
554 | |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
555 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
|
556 int bidir, int edges, int step, int mask_mv0, int mask_mv1 ) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
557 int dir; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
558 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
559 "pxor %%mm7, %%mm7 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
560 "movq %0, %%mm6 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
561 "movq %1, %%mm5 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
562 "movq %2, %%mm4 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
563 ::"m"(ff_pb_1), "m"(ff_pb_3), "m"(ff_pb_7) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
564 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
565 // 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
|
566 // average filter time by 1.2% |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
567 for( dir=1; dir>=0; dir-- ) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
568 const int d_idx = dir ? -8 : -1; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
569 const int mask_mv = dir ? mask_mv1 : mask_mv0; |
4137 | 570 DECLARE_ALIGNED_8(const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL; |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
571 int b_idx, edge, l; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
572 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
|
573 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
574 "pand %0, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
575 ::"m"(mask_dir) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
576 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
577 if(!(mask_mv & edge)) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
578 asm volatile("pxor %%mm0, %%mm0 \n\t":); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
579 for( l = bidir; l >= 0; l-- ) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
580 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
581 "movd %0, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
582 "punpckldq %1, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
583 "movq %%mm1, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
584 "psrlw $7, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
585 "pand %%mm6, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
586 "por %%mm2, %%mm1 \n\t" // ref_cache with -2 mapped to -1 |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
587 "punpckldq %%mm1, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
588 "pcmpeqb %%mm2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
589 "paddb %%mm6, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
590 "punpckhbw %%mm7, %%mm1 \n\t" // ref[b] != ref[bn] |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
591 "por %%mm1, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
592 |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
593 "movq %2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
594 "movq %3, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
595 "psubw %4, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
596 "psubw %5, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
597 "packsswb %%mm2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
598 "paddb %%mm5, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
599 "pminub %%mm4, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
600 "pcmpeqb %%mm4, %%mm1 \n\t" // abs(mv[b] - mv[bn]) >= limit |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
601 "por %%mm1, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
602 ::"m"(ref[l][b_idx]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
603 "m"(ref[l][b_idx+d_idx]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
604 "m"(mv[l][b_idx][0]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
605 "m"(mv[l][b_idx+2][0]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
606 "m"(mv[l][b_idx+d_idx][0]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
607 "m"(mv[l][b_idx+d_idx+2][0]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
608 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
609 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
610 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
611 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
612 "movd %0, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
613 "por %1, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
614 "punpcklbw %%mm7, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
615 "pcmpgtw %%mm7, %%mm1 \n\t" // nnz[b] || nnz[bn] |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
616 ::"m"(nnz[b_idx]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
617 "m"(nnz[b_idx+d_idx]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
618 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
619 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
620 "pcmpeqw %%mm7, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
621 "pcmpeqw %%mm7, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
622 "psrlw $15, %%mm0 \n\t" // nonzero -> 1 |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
623 "psrlw $14, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
624 "movq %%mm0, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
625 "por %%mm1, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
626 "psrlw $1, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
627 "pandn %%mm2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
628 "movq %%mm1, %0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
629 :"=m"(*bS[dir][edge]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
630 ::"memory" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
631 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
632 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
633 edges = 4; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
634 step = 1; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
635 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
636 asm volatile( |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
637 "movq (%0), %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
638 "movq 8(%0), %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
639 "movq 16(%0), %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
640 "movq 24(%0), %%mm3 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
641 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
642 "movq %%mm0, (%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
643 "movq %%mm3, 8(%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
644 "movq %%mm4, 16(%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
645 "movq %%mm2, 24(%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
646 ::"r"(bS[0]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
647 :"memory" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
648 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
649 } |
2754 | 650 |
651 /***********************************/ | |
652 /* motion compensation */ | |
653 | |
654 #define QPEL_H264V(A,B,C,D,E,F,OP)\ | |
2979 | 655 "movd (%0), "#F" \n\t"\ |
656 "movq "#C", %%mm6 \n\t"\ | |
657 "paddw "#D", %%mm6 \n\t"\ | |
658 "psllw $2, %%mm6 \n\t"\ | |
659 "psubw "#B", %%mm6 \n\t"\ | |
660 "psubw "#E", %%mm6 \n\t"\ | |
661 "pmullw %4, %%mm6 \n\t"\ | |
662 "add %2, %0 \n\t"\ | |
663 "punpcklbw %%mm7, "#F" \n\t"\ | |
664 "paddw %5, "#A" \n\t"\ | |
665 "paddw "#F", "#A" \n\t"\ | |
666 "paddw "#A", %%mm6 \n\t"\ | |
667 "psraw $5, %%mm6 \n\t"\ | |
668 "packuswb %%mm6, %%mm6 \n\t"\ | |
2754 | 669 OP(%%mm6, (%1), A, d)\ |
2979 | 670 "add %3, %1 \n\t" |
2754 | 671 |
672 #define QPEL_H264HV(A,B,C,D,E,F,OF)\ | |
2979 | 673 "movd (%0), "#F" \n\t"\ |
674 "movq "#C", %%mm6 \n\t"\ | |
675 "paddw "#D", %%mm6 \n\t"\ | |
676 "psllw $2, %%mm6 \n\t"\ | |
677 "psubw "#B", %%mm6 \n\t"\ | |
678 "psubw "#E", %%mm6 \n\t"\ | |
679 "pmullw %3, %%mm6 \n\t"\ | |
680 "add %2, %0 \n\t"\ | |
681 "punpcklbw %%mm7, "#F" \n\t"\ | |
682 "paddw "#F", "#A" \n\t"\ | |
683 "paddw "#A", %%mm6 \n\t"\ | |
684 "movq %%mm6, "#OF"(%1) \n\t" | |
2967 | 685 |
2754 | 686 #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
|
687 static av_noinline void OPNAME ## h264_qpel4_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 688 int h=4;\ |
689 \ | |
690 asm volatile(\ | |
2979 | 691 "pxor %%mm7, %%mm7 \n\t"\ |
692 "movq %5, %%mm4 \n\t"\ | |
693 "movq %6, %%mm5 \n\t"\ | |
694 "1: \n\t"\ | |
695 "movd -1(%0), %%mm1 \n\t"\ | |
696 "movd (%0), %%mm2 \n\t"\ | |
697 "movd 1(%0), %%mm3 \n\t"\ | |
698 "movd 2(%0), %%mm0 \n\t"\ | |
699 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
700 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
701 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
702 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
703 "paddw %%mm0, %%mm1 \n\t"\ | |
704 "paddw %%mm3, %%mm2 \n\t"\ | |
705 "movd -2(%0), %%mm0 \n\t"\ | |
706 "movd 3(%0), %%mm3 \n\t"\ | |
707 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
708 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
709 "paddw %%mm3, %%mm0 \n\t"\ | |
710 "psllw $2, %%mm2 \n\t"\ | |
711 "psubw %%mm1, %%mm2 \n\t"\ | |
712 "pmullw %%mm4, %%mm2 \n\t"\ | |
713 "paddw %%mm5, %%mm0 \n\t"\ | |
714 "paddw %%mm2, %%mm0 \n\t"\ | |
715 "psraw $5, %%mm0 \n\t"\ | |
716 "packuswb %%mm0, %%mm0 \n\t"\ | |
2754 | 717 OP(%%mm0, (%1),%%mm6, d)\ |
2979 | 718 "add %3, %0 \n\t"\ |
719 "add %4, %1 \n\t"\ | |
720 "decl %2 \n\t"\ | |
721 " jnz 1b \n\t"\ | |
2754 | 722 : "+a"(src), "+c"(dst), "+m"(h)\ |
723 : "d"((long)srcStride), "S"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
724 : "memory"\ | |
725 );\ | |
726 }\ | |
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
|
727 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 | 728 int h=4;\ |
729 asm volatile(\ | |
730 "pxor %%mm7, %%mm7 \n\t"\ | |
3165 | 731 "movq %0, %%mm4 \n\t"\ |
732 "movq %1, %%mm5 \n\t"\ | |
733 :: "m"(ff_pw_5), "m"(ff_pw_16)\ | |
734 );\ | |
735 do{\ | |
736 asm volatile(\ | |
3156 | 737 "movd -1(%0), %%mm1 \n\t"\ |
738 "movd (%0), %%mm2 \n\t"\ | |
739 "movd 1(%0), %%mm3 \n\t"\ | |
740 "movd 2(%0), %%mm0 \n\t"\ | |
741 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
742 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
743 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
744 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
745 "paddw %%mm0, %%mm1 \n\t"\ | |
746 "paddw %%mm3, %%mm2 \n\t"\ | |
747 "movd -2(%0), %%mm0 \n\t"\ | |
748 "movd 3(%0), %%mm3 \n\t"\ | |
749 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
750 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
751 "paddw %%mm3, %%mm0 \n\t"\ | |
752 "psllw $2, %%mm2 \n\t"\ | |
753 "psubw %%mm1, %%mm2 \n\t"\ | |
754 "pmullw %%mm4, %%mm2 \n\t"\ | |
755 "paddw %%mm5, %%mm0 \n\t"\ | |
756 "paddw %%mm2, %%mm0 \n\t"\ | |
757 "movd (%2), %%mm3 \n\t"\ | |
758 "psraw $5, %%mm0 \n\t"\ | |
759 "packuswb %%mm0, %%mm0 \n\t"\ | |
760 PAVGB" %%mm3, %%mm0 \n\t"\ | |
761 OP(%%mm0, (%1),%%mm6, d)\ | |
3165 | 762 "add %4, %0 \n\t"\ |
763 "add %4, %1 \n\t"\ | |
764 "add %3, %2 \n\t"\ | |
765 : "+a"(src), "+c"(dst), "+d"(src2)\ | |
766 : "D"((long)src2Stride), "S"((long)dstStride)\ | |
3156 | 767 : "memory"\ |
768 );\ | |
3165 | 769 }while(--h);\ |
3156 | 770 }\ |
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
|
771 static av_noinline void OPNAME ## h264_qpel4_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 772 src -= 2*srcStride;\ |
773 asm volatile(\ | |
2979 | 774 "pxor %%mm7, %%mm7 \n\t"\ |
775 "movd (%0), %%mm0 \n\t"\ | |
776 "add %2, %0 \n\t"\ | |
777 "movd (%0), %%mm1 \n\t"\ | |
778 "add %2, %0 \n\t"\ | |
779 "movd (%0), %%mm2 \n\t"\ | |
780 "add %2, %0 \n\t"\ | |
781 "movd (%0), %%mm3 \n\t"\ | |
782 "add %2, %0 \n\t"\ | |
783 "movd (%0), %%mm4 \n\t"\ | |
784 "add %2, %0 \n\t"\ | |
785 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
786 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
787 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
788 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
789 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
2754 | 790 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ |
791 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
792 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
793 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
794 \ | |
795 : "+a"(src), "+c"(dst)\ | |
796 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
797 : "memory"\ | |
798 );\ | |
799 }\ | |
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
|
800 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 | 801 int h=4;\ |
802 int w=3;\ | |
803 src -= 2*srcStride+2;\ | |
804 while(w--){\ | |
805 asm volatile(\ | |
2979 | 806 "pxor %%mm7, %%mm7 \n\t"\ |
807 "movd (%0), %%mm0 \n\t"\ | |
808 "add %2, %0 \n\t"\ | |
809 "movd (%0), %%mm1 \n\t"\ | |
810 "add %2, %0 \n\t"\ | |
811 "movd (%0), %%mm2 \n\t"\ | |
812 "add %2, %0 \n\t"\ | |
813 "movd (%0), %%mm3 \n\t"\ | |
814 "add %2, %0 \n\t"\ | |
815 "movd (%0), %%mm4 \n\t"\ | |
816 "add %2, %0 \n\t"\ | |
817 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
818 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
819 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
820 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
821 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
2754 | 822 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*8*3)\ |
823 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*8*3)\ | |
824 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*8*3)\ | |
825 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*8*3)\ | |
826 \ | |
827 : "+a"(src)\ | |
828 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5)\ | |
829 : "memory"\ | |
830 );\ | |
831 tmp += 4;\ | |
832 src += 4 - 9*srcStride;\ | |
833 }\ | |
834 tmp -= 3*4;\ | |
835 asm volatile(\ | |
2979 | 836 "movq %4, %%mm6 \n\t"\ |
837 "1: \n\t"\ | |
838 "movq (%0), %%mm0 \n\t"\ | |
839 "paddw 10(%0), %%mm0 \n\t"\ | |
840 "movq 2(%0), %%mm1 \n\t"\ | |
841 "paddw 8(%0), %%mm1 \n\t"\ | |
842 "movq 4(%0), %%mm2 \n\t"\ | |
843 "paddw 6(%0), %%mm2 \n\t"\ | |
844 "psubw %%mm1, %%mm0 \n\t"/*a-b (abccba)*/\ | |
845 "psraw $2, %%mm0 \n\t"/*(a-b)/4 */\ | |
846 "psubw %%mm1, %%mm0 \n\t"/*(a-b)/4-b */\ | |
847 "paddsw %%mm2, %%mm0 \n\t"\ | |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
848 "psraw $2, %%mm0 \n\t"/*((a-b)/4-b+c)/4 */\ |
2979 | 849 "paddw %%mm6, %%mm2 \n\t"\ |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
850 "paddw %%mm2, %%mm0 \n\t"/*(a-5*b+20*c)/16 +32 */\ |
2979 | 851 "psraw $6, %%mm0 \n\t"\ |
852 "packuswb %%mm0, %%mm0 \n\t"\ | |
2754 | 853 OP(%%mm0, (%1),%%mm7, d)\ |
2979 | 854 "add $24, %0 \n\t"\ |
855 "add %3, %1 \n\t"\ | |
856 "decl %2 \n\t"\ | |
857 " jnz 1b \n\t"\ | |
2754 | 858 : "+a"(tmp), "+c"(dst), "+m"(h)\ |
859 : "S"((long)dstStride), "m"(ff_pw_32)\ | |
860 : "memory"\ | |
861 );\ | |
862 }\ | |
863 \ | |
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
|
864 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 865 int h=8;\ |
866 asm volatile(\ | |
2979 | 867 "pxor %%mm7, %%mm7 \n\t"\ |
868 "movq %5, %%mm6 \n\t"\ | |
869 "1: \n\t"\ | |
870 "movq (%0), %%mm0 \n\t"\ | |
871 "movq 1(%0), %%mm2 \n\t"\ | |
872 "movq %%mm0, %%mm1 \n\t"\ | |
873 "movq %%mm2, %%mm3 \n\t"\ | |
874 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
875 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
876 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
877 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
878 "paddw %%mm2, %%mm0 \n\t"\ | |
879 "paddw %%mm3, %%mm1 \n\t"\ | |
880 "psllw $2, %%mm0 \n\t"\ | |
881 "psllw $2, %%mm1 \n\t"\ | |
882 "movq -1(%0), %%mm2 \n\t"\ | |
883 "movq 2(%0), %%mm4 \n\t"\ | |
884 "movq %%mm2, %%mm3 \n\t"\ | |
885 "movq %%mm4, %%mm5 \n\t"\ | |
886 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
887 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
888 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
889 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
890 "paddw %%mm4, %%mm2 \n\t"\ | |
891 "paddw %%mm3, %%mm5 \n\t"\ | |
892 "psubw %%mm2, %%mm0 \n\t"\ | |
893 "psubw %%mm5, %%mm1 \n\t"\ | |
894 "pmullw %%mm6, %%mm0 \n\t"\ | |
895 "pmullw %%mm6, %%mm1 \n\t"\ | |
896 "movd -2(%0), %%mm2 \n\t"\ | |
897 "movd 7(%0), %%mm5 \n\t"\ | |
898 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
899 "punpcklbw %%mm7, %%mm5 \n\t"\ | |
900 "paddw %%mm3, %%mm2 \n\t"\ | |
901 "paddw %%mm5, %%mm4 \n\t"\ | |
902 "movq %6, %%mm5 \n\t"\ | |
903 "paddw %%mm5, %%mm2 \n\t"\ | |
904 "paddw %%mm5, %%mm4 \n\t"\ | |
905 "paddw %%mm2, %%mm0 \n\t"\ | |
906 "paddw %%mm4, %%mm1 \n\t"\ | |
907 "psraw $5, %%mm0 \n\t"\ | |
908 "psraw $5, %%mm1 \n\t"\ | |
909 "packuswb %%mm1, %%mm0 \n\t"\ | |
2754 | 910 OP(%%mm0, (%1),%%mm5, q)\ |
2979 | 911 "add %3, %0 \n\t"\ |
912 "add %4, %1 \n\t"\ | |
913 "decl %2 \n\t"\ | |
914 " jnz 1b \n\t"\ | |
2754 | 915 : "+a"(src), "+c"(dst), "+m"(h)\ |
916 : "d"((long)srcStride), "S"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
917 : "memory"\ | |
918 );\ | |
919 }\ | |
920 \ | |
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
|
921 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 | 922 int h=8;\ |
923 asm volatile(\ | |
924 "pxor %%mm7, %%mm7 \n\t"\ | |
3165 | 925 "movq %0, %%mm6 \n\t"\ |
926 :: "m"(ff_pw_5)\ | |
927 );\ | |
928 do{\ | |
929 asm volatile(\ | |
3156 | 930 "movq (%0), %%mm0 \n\t"\ |
931 "movq 1(%0), %%mm2 \n\t"\ | |
932 "movq %%mm0, %%mm1 \n\t"\ | |
933 "movq %%mm2, %%mm3 \n\t"\ | |
934 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
935 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
936 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
937 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
938 "paddw %%mm2, %%mm0 \n\t"\ | |
939 "paddw %%mm3, %%mm1 \n\t"\ | |
940 "psllw $2, %%mm0 \n\t"\ | |
941 "psllw $2, %%mm1 \n\t"\ | |
942 "movq -1(%0), %%mm2 \n\t"\ | |
943 "movq 2(%0), %%mm4 \n\t"\ | |
944 "movq %%mm2, %%mm3 \n\t"\ | |
945 "movq %%mm4, %%mm5 \n\t"\ | |
946 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
947 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
948 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
949 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
950 "paddw %%mm4, %%mm2 \n\t"\ | |
951 "paddw %%mm3, %%mm5 \n\t"\ | |
952 "psubw %%mm2, %%mm0 \n\t"\ | |
953 "psubw %%mm5, %%mm1 \n\t"\ | |
954 "pmullw %%mm6, %%mm0 \n\t"\ | |
955 "pmullw %%mm6, %%mm1 \n\t"\ | |
956 "movd -2(%0), %%mm2 \n\t"\ | |
957 "movd 7(%0), %%mm5 \n\t"\ | |
958 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
959 "punpcklbw %%mm7, %%mm5 \n\t"\ | |
960 "paddw %%mm3, %%mm2 \n\t"\ | |
961 "paddw %%mm5, %%mm4 \n\t"\ | |
3165 | 962 "movq %5, %%mm5 \n\t"\ |
3156 | 963 "paddw %%mm5, %%mm2 \n\t"\ |
964 "paddw %%mm5, %%mm4 \n\t"\ | |
965 "paddw %%mm2, %%mm0 \n\t"\ | |
966 "paddw %%mm4, %%mm1 \n\t"\ | |
967 "psraw $5, %%mm0 \n\t"\ | |
968 "psraw $5, %%mm1 \n\t"\ | |
969 "movq (%2), %%mm4 \n\t"\ | |
970 "packuswb %%mm1, %%mm0 \n\t"\ | |
971 PAVGB" %%mm4, %%mm0 \n\t"\ | |
972 OP(%%mm0, (%1),%%mm5, q)\ | |
3165 | 973 "add %4, %0 \n\t"\ |
974 "add %4, %1 \n\t"\ | |
975 "add %3, %2 \n\t"\ | |
976 : "+a"(src), "+c"(dst), "+d"(src2)\ | |
3156 | 977 : "D"((long)src2Stride), "S"((long)dstStride),\ |
3165 | 978 "m"(ff_pw_16)\ |
3156 | 979 : "memory"\ |
980 );\ | |
3165 | 981 }while(--h);\ |
3156 | 982 }\ |
983 \ | |
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
|
984 static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
3094 | 985 int w= 2;\ |
2754 | 986 src -= 2*srcStride;\ |
987 \ | |
3094 | 988 while(w--){\ |
2754 | 989 asm volatile(\ |
2979 | 990 "pxor %%mm7, %%mm7 \n\t"\ |
991 "movd (%0), %%mm0 \n\t"\ | |
992 "add %2, %0 \n\t"\ | |
993 "movd (%0), %%mm1 \n\t"\ | |
994 "add %2, %0 \n\t"\ | |
995 "movd (%0), %%mm2 \n\t"\ | |
996 "add %2, %0 \n\t"\ | |
997 "movd (%0), %%mm3 \n\t"\ | |
998 "add %2, %0 \n\t"\ | |
999 "movd (%0), %%mm4 \n\t"\ | |
1000 "add %2, %0 \n\t"\ | |
1001 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1002 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
1003 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1004 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
1005 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
2754 | 1006 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ |
1007 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
1008 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
1009 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
1010 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\ | |
1011 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\ | |
1012 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ | |
1013 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
1014 \ | |
1015 : "+a"(src), "+c"(dst)\ | |
1016 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
1017 : "memory"\ | |
1018 );\ | |
3094 | 1019 if(h==16){\ |
1020 asm volatile(\ | |
1021 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
1022 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
1023 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\ | |
1024 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\ | |
1025 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ | |
1026 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
1027 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
1028 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
1029 \ | |
1030 : "+a"(src), "+c"(dst)\ | |
1031 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ | |
1032 : "memory"\ | |
1033 );\ | |
1034 }\ | |
1035 src += 4-(h+5)*srcStride;\ | |
1036 dst += 4-h*dstStride;\ | |
2754 | 1037 }\ |
1038 }\ | |
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
|
1039 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 | 1040 int h = size;\ |
1041 int w = (size+8)>>2;\ | |
2754 | 1042 src -= 2*srcStride+2;\ |
1043 while(w--){\ | |
1044 asm volatile(\ | |
2979 | 1045 "pxor %%mm7, %%mm7 \n\t"\ |
1046 "movd (%0), %%mm0 \n\t"\ | |
1047 "add %2, %0 \n\t"\ | |
1048 "movd (%0), %%mm1 \n\t"\ | |
1049 "add %2, %0 \n\t"\ | |
1050 "movd (%0), %%mm2 \n\t"\ | |
1051 "add %2, %0 \n\t"\ | |
1052 "movd (%0), %%mm3 \n\t"\ | |
1053 "add %2, %0 \n\t"\ | |
1054 "movd (%0), %%mm4 \n\t"\ | |
1055 "add %2, %0 \n\t"\ | |
1056 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1057 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
1058 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1059 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
1060 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
3093 | 1061 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*48)\ |
1062 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*48)\ | |
1063 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*48)\ | |
1064 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*48)\ | |
1065 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 4*48)\ | |
1066 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 5*48)\ | |
1067 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 6*48)\ | |
1068 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 7*48)\ | |
2754 | 1069 : "+a"(src)\ |
1070 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5)\ | |
1071 : "memory"\ | |
1072 );\ | |
3093 | 1073 if(size==16){\ |
1074 asm volatile(\ | |
1075 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 8*48)\ | |
1076 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 9*48)\ | |
1077 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 10*48)\ | |
1078 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 11*48)\ | |
1079 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 12*48)\ | |
1080 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 13*48)\ | |
1081 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 14*48)\ | |
1082 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 15*48)\ | |
1083 : "+a"(src)\ | |
1084 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5)\ | |
1085 : "memory"\ | |
1086 );\ | |
1087 }\ | |
2754 | 1088 tmp += 4;\ |
3093 | 1089 src += 4 - (size+5)*srcStride;\ |
2754 | 1090 }\ |
3093 | 1091 tmp -= size+8;\ |
1092 w = size>>4;\ | |
1093 do{\ | |
1094 h = size;\ | |
2754 | 1095 asm volatile(\ |
2979 | 1096 "movq %4, %%mm6 \n\t"\ |
1097 "1: \n\t"\ | |
1098 "movq (%0), %%mm0 \n\t"\ | |
1099 "movq 8(%0), %%mm3 \n\t"\ | |
1100 "movq 2(%0), %%mm1 \n\t"\ | |
1101 "movq 10(%0), %%mm4 \n\t"\ | |
1102 "paddw %%mm4, %%mm0 \n\t"\ | |
1103 "paddw %%mm3, %%mm1 \n\t"\ | |
1104 "paddw 18(%0), %%mm3 \n\t"\ | |
1105 "paddw 16(%0), %%mm4 \n\t"\ | |
1106 "movq 4(%0), %%mm2 \n\t"\ | |
1107 "movq 12(%0), %%mm5 \n\t"\ | |
1108 "paddw 6(%0), %%mm2 \n\t"\ | |
1109 "paddw 14(%0), %%mm5 \n\t"\ | |
1110 "psubw %%mm1, %%mm0 \n\t"\ | |
1111 "psubw %%mm4, %%mm3 \n\t"\ | |
1112 "psraw $2, %%mm0 \n\t"\ | |
1113 "psraw $2, %%mm3 \n\t"\ | |
1114 "psubw %%mm1, %%mm0 \n\t"\ | |
1115 "psubw %%mm4, %%mm3 \n\t"\ | |
1116 "paddsw %%mm2, %%mm0 \n\t"\ | |
1117 "paddsw %%mm5, %%mm3 \n\t"\ | |
1118 "psraw $2, %%mm0 \n\t"\ | |
1119 "psraw $2, %%mm3 \n\t"\ | |
1120 "paddw %%mm6, %%mm2 \n\t"\ | |
1121 "paddw %%mm6, %%mm5 \n\t"\ | |
1122 "paddw %%mm2, %%mm0 \n\t"\ | |
1123 "paddw %%mm5, %%mm3 \n\t"\ | |
1124 "psraw $6, %%mm0 \n\t"\ | |
1125 "psraw $6, %%mm3 \n\t"\ | |
1126 "packuswb %%mm3, %%mm0 \n\t"\ | |
2754 | 1127 OP(%%mm0, (%1),%%mm7, q)\ |
3093 | 1128 "add $48, %0 \n\t"\ |
2979 | 1129 "add %3, %1 \n\t"\ |
1130 "decl %2 \n\t"\ | |
1131 " jnz 1b \n\t"\ | |
2754 | 1132 : "+a"(tmp), "+c"(dst), "+m"(h)\ |
1133 : "S"((long)dstStride), "m"(ff_pw_32)\ | |
1134 : "memory"\ | |
1135 );\ | |
3093 | 1136 tmp += 8 - size*24;\ |
1137 dst += 8 - size*dstStride;\ | |
1138 }while(w--);\ | |
2754 | 1139 }\ |
3094 | 1140 \ |
1141 static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
1142 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 8);\ | |
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_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
3094 | 1145 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 16);\ |
1146 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\ | |
2754 | 1147 }\ |
1148 \ | |
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
|
1149 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 1150 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\ |
1151 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\ | |
1152 src += 8*srcStride;\ | |
1153 dst += 8*dstStride;\ | |
1154 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\ | |
1155 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\ | |
1156 }\ | |
1157 \ | |
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
|
1158 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 | 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 src += 8*dstStride;\ | |
1162 dst += 8*dstStride;\ | |
1163 src2 += 8*src2Stride;\ | |
1164 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\ | |
1165 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\ | |
1166 }\ | |
1167 \ | |
3093 | 1168 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ |
1169 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 8);\ | |
1170 }\ | |
1171 \ | |
2754 | 1172 static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ |
3093 | 1173 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 16);\ |
2754 | 1174 }\ |
3095 | 1175 \ |
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
|
1176 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 | 1177 {\ |
1178 asm volatile(\ | |
1179 "movq %5, %%mm6 \n\t"\ | |
1180 "movq (%1), %%mm0 \n\t"\ | |
1181 "movq 24(%1), %%mm1 \n\t"\ | |
1182 "paddw %%mm6, %%mm0 \n\t"\ | |
1183 "paddw %%mm6, %%mm1 \n\t"\ | |
3102 | 1184 "psraw $5, %%mm0 \n\t"\ |
1185 "psraw $5, %%mm1 \n\t"\ | |
3163 | 1186 "packuswb %%mm0, %%mm0 \n\t"\ |
1187 "packuswb %%mm1, %%mm1 \n\t"\ | |
3095 | 1188 PAVGB" (%0), %%mm0 \n\t"\ |
1189 PAVGB" (%0,%3), %%mm1 \n\t"\ | |
1190 OP(%%mm0, (%2), %%mm4, d)\ | |
1191 OP(%%mm1, (%2,%4), %%mm5, d)\ | |
1192 "lea (%0,%3,2), %0 \n\t"\ | |
1193 "lea (%2,%4,2), %2 \n\t"\ | |
1194 "movq 48(%1), %%mm0 \n\t"\ | |
1195 "movq 72(%1), %%mm1 \n\t"\ | |
1196 "paddw %%mm6, %%mm0 \n\t"\ | |
1197 "paddw %%mm6, %%mm1 \n\t"\ | |
3102 | 1198 "psraw $5, %%mm0 \n\t"\ |
1199 "psraw $5, %%mm1 \n\t"\ | |
3163 | 1200 "packuswb %%mm0, %%mm0 \n\t"\ |
1201 "packuswb %%mm1, %%mm1 \n\t"\ | |
3095 | 1202 PAVGB" (%0), %%mm0 \n\t"\ |
1203 PAVGB" (%0,%3), %%mm1 \n\t"\ | |
1204 OP(%%mm0, (%2), %%mm4, d)\ | |
1205 OP(%%mm1, (%2,%4), %%mm5, d)\ | |
1206 :"+a"(src8), "+c"(src16), "+d"(dst)\ | |
3096 | 1207 :"S"((long)src8Stride), "D"((long)dstStride), "m"(ff_pw_16)\ |
3095 | 1208 :"memory");\ |
1209 }\ | |
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
|
1210 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 | 1211 {\ |
1212 asm volatile(\ | |
1213 "movq %0, %%mm6 \n\t"\ | |
1214 ::"m"(ff_pw_16)\ | |
1215 );\ | |
1216 while(h--){\ | |
1217 asm volatile(\ | |
1218 "movq (%1), %%mm0 \n\t"\ | |
1219 "movq 8(%1), %%mm1 \n\t"\ | |
1220 "paddw %%mm6, %%mm0 \n\t"\ | |
1221 "paddw %%mm6, %%mm1 \n\t"\ | |
3102 | 1222 "psraw $5, %%mm0 \n\t"\ |
1223 "psraw $5, %%mm1 \n\t"\ | |
3095 | 1224 "packuswb %%mm1, %%mm0 \n\t"\ |
1225 PAVGB" (%0), %%mm0 \n\t"\ | |
1226 OP(%%mm0, (%2), %%mm5, q)\ | |
1227 ::"a"(src8), "c"(src16), "d"(dst)\ | |
1228 :"memory");\ | |
1229 src8 += src8Stride;\ | |
1230 src16 += 24;\ | |
1231 dst += dstStride;\ | |
1232 }\ | |
1233 }\ | |
1234 static void OPNAME ## pixels16_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\ | |
1235 {\ | |
1236 OPNAME ## pixels8_l2_shift5_ ## MMX(dst , src16 , src8 , dstStride, src8Stride, h);\ | |
1237 OPNAME ## pixels8_l2_shift5_ ## MMX(dst+8, src16+8, src8+8, dstStride, src8Stride, h);\ | |
1238 }\ | |
1239 | |
2754 | 1240 |
1241 #define H264_MC(OPNAME, SIZE, MMX) \ | |
1242 static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ | |
1243 OPNAME ## pixels ## SIZE ## _mmx(dst, src, stride, SIZE);\ | |
1244 }\ | |
1245 \ | |
1246 static void OPNAME ## h264_qpel ## SIZE ## _mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1247 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src, stride, stride);\ |
2754 | 1248 }\ |
1249 \ | |
1250 static void OPNAME ## h264_qpel ## SIZE ## _mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1251 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1252 }\ | |
1253 \ | |
1254 static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1255 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src+1, stride, stride);\ |
2754 | 1256 }\ |
1257 \ | |
1258 static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1259 uint64_t temp[SIZE*SIZE/8];\ | |
1260 uint8_t * const half= (uint8_t*)temp;\ | |
1261 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(half, src, SIZE, stride);\ | |
1262 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src, half, stride, stride, SIZE);\ | |
1263 }\ | |
1264 \ | |
1265 static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1266 OPNAME ## h264_qpel ## SIZE ## _v_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1267 }\ | |
1268 \ | |
1269 static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1270 uint64_t temp[SIZE*SIZE/8];\ | |
1271 uint8_t * const half= (uint8_t*)temp;\ | |
1272 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(half, src, SIZE, stride);\ | |
1273 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src+stride, half, stride, stride, SIZE);\ | |
1274 }\ | |
1275 \ | |
1276 static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1277 uint64_t temp[SIZE*SIZE/8];\ |
1278 uint8_t * const halfV= (uint8_t*)temp;\ | |
2754 | 1279 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src, SIZE, stride);\ |
3156 | 1280 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfV, stride, SIZE);\ |
2754 | 1281 }\ |
1282 \ | |
1283 static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1284 uint64_t temp[SIZE*SIZE/8];\ |
1285 uint8_t * const halfV= (uint8_t*)temp;\ | |
2754 | 1286 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src+1, SIZE, stride);\ |
3156 | 1287 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfV, stride, SIZE);\ |
2754 | 1288 }\ |
1289 \ | |
1290 static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1291 uint64_t temp[SIZE*SIZE/8];\ |
1292 uint8_t * const halfV= (uint8_t*)temp;\ | |
2754 | 1293 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src, SIZE, stride);\ |
3156 | 1294 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfV, stride, SIZE);\ |
2754 | 1295 }\ |
1296 \ | |
1297 static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1298 uint64_t temp[SIZE*SIZE/8];\ |
1299 uint8_t * const halfV= (uint8_t*)temp;\ | |
2754 | 1300 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src+1, SIZE, stride);\ |
3156 | 1301 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfV, stride, SIZE);\ |
2754 | 1302 }\ |
1303 \ | |
1304 static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3093 | 1305 uint64_t temp[SIZE*(SIZE<8?12:24)/4];\ |
2754 | 1306 int16_t * const tmp= (int16_t*)temp;\ |
1307 OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(dst, tmp, src, stride, SIZE, stride);\ | |
1308 }\ | |
1309 \ | |
1310 static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1311 uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\ |
1312 uint8_t * const halfHV= (uint8_t*)temp;\ | |
1313 int16_t * const tmp= ((int16_t*)temp) + SIZE*SIZE/2;\ | |
4528 | 1314 assert((int)temp & 7 == 0);\ |
2754 | 1315 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, tmp, src, SIZE, SIZE, stride);\ |
3156 | 1316 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfHV, stride, SIZE);\ |
2754 | 1317 }\ |
1318 \ | |
1319 static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1320 uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\ |
1321 uint8_t * const halfHV= (uint8_t*)temp;\ | |
1322 int16_t * const tmp= ((int16_t*)temp) + SIZE*SIZE/2;\ | |
4528 | 1323 assert((int)temp & 7 == 0);\ |
2754 | 1324 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, tmp, src, SIZE, SIZE, stride);\ |
3156 | 1325 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfHV, stride, SIZE);\ |
2754 | 1326 }\ |
1327 \ | |
1328 static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3095 | 1329 uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\ |
1330 int16_t * const halfV= ((int16_t*)temp) + SIZE*SIZE/2;\ | |
1331 uint8_t * const halfHV= ((uint8_t*)temp);\ | |
4528 | 1332 assert((int)temp & 7 == 0);\ |
3095 | 1333 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\ |
1334 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+2, halfHV, stride, SIZE, SIZE);\ | |
2754 | 1335 }\ |
1336 \ | |
1337 static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3095 | 1338 uint64_t temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8];\ |
1339 int16_t * const halfV= ((int16_t*)temp) + SIZE*SIZE/2;\ | |
1340 uint8_t * const halfHV= ((uint8_t*)temp);\ | |
4528 | 1341 assert((int)temp & 7 == 0);\ |
3095 | 1342 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\ |
1343 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+3, halfHV, stride, SIZE, SIZE);\ | |
2754 | 1344 }\ |
1345 | |
1346 | |
1347 #define AVG_3DNOW_OP(a,b,temp, size) \ | |
2979 | 1348 "mov" #size " " #b ", " #temp " \n\t"\ |
1349 "pavgusb " #temp ", " #a " \n\t"\ | |
1350 "mov" #size " " #a ", " #b " \n\t" | |
2754 | 1351 #define AVG_MMX2_OP(a,b,temp, size) \ |
2979 | 1352 "mov" #size " " #b ", " #temp " \n\t"\ |
1353 "pavgb " #temp ", " #a " \n\t"\ | |
1354 "mov" #size " " #a ", " #b " \n\t" | |
2754 | 1355 |
3095 | 1356 #define PAVGB "pavgusb" |
2754 | 1357 QPEL_H264(put_, PUT_OP, 3dnow) |
1358 QPEL_H264(avg_, AVG_3DNOW_OP, 3dnow) | |
3095 | 1359 #undef PAVGB |
1360 #define PAVGB "pavgb" | |
2754 | 1361 QPEL_H264(put_, PUT_OP, mmx2) |
1362 QPEL_H264(avg_, AVG_MMX2_OP, mmx2) | |
3095 | 1363 #undef PAVGB |
2754 | 1364 |
1365 H264_MC(put_, 4, 3dnow) | |
1366 H264_MC(put_, 8, 3dnow) | |
1367 H264_MC(put_, 16,3dnow) | |
1368 H264_MC(avg_, 4, 3dnow) | |
1369 H264_MC(avg_, 8, 3dnow) | |
1370 H264_MC(avg_, 16,3dnow) | |
1371 H264_MC(put_, 4, mmx2) | |
1372 H264_MC(put_, 8, mmx2) | |
1373 H264_MC(put_, 16,mmx2) | |
1374 H264_MC(avg_, 4, mmx2) | |
1375 H264_MC(avg_, 8, mmx2) | |
1376 H264_MC(avg_, 16,mmx2) | |
1377 | |
1378 | |
1379 #define H264_CHROMA_OP(S,D) | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1380 #define H264_CHROMA_OP4(S,D,T) |
2754 | 1381 #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
|
1382 #define H264_CHROMA_MC4_TMPL put_h264_chroma_mc4_mmx |
3213 | 1383 #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
|
1384 #define H264_CHROMA_MC8_MV0 put_pixels8_mmx |
2754 | 1385 #include "dsputil_h264_template_mmx.c" |
1386 #undef H264_CHROMA_OP | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1387 #undef H264_CHROMA_OP4 |
2754 | 1388 #undef H264_CHROMA_MC8_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1389 #undef H264_CHROMA_MC4_TMPL |
3213 | 1390 #undef H264_CHROMA_MC2_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1391 #undef H264_CHROMA_MC8_MV0 |
2754 | 1392 |
1393 #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
|
1394 #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
|
1395 "pavgb " #T ", " #D " \n\t" |
2754 | 1396 #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
|
1397 #define H264_CHROMA_MC4_TMPL avg_h264_chroma_mc4_mmx2 |
3213 | 1398 #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
|
1399 #define H264_CHROMA_MC8_MV0 avg_pixels8_mmx2 |
2754 | 1400 #include "dsputil_h264_template_mmx.c" |
1401 #undef H264_CHROMA_OP | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1402 #undef H264_CHROMA_OP4 |
2754 | 1403 #undef H264_CHROMA_MC8_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1404 #undef H264_CHROMA_MC4_TMPL |
3213 | 1405 #undef H264_CHROMA_MC2_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1406 #undef H264_CHROMA_MC8_MV0 |
2754 | 1407 |
1408 #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
|
1409 #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
|
1410 "pavgusb " #T ", " #D " \n\t" |
2754 | 1411 #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
|
1412 #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
|
1413 #define H264_CHROMA_MC8_MV0 avg_pixels8_3dnow |
2754 | 1414 #include "dsputil_h264_template_mmx.c" |
1415 #undef H264_CHROMA_OP | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1416 #undef H264_CHROMA_OP4 |
2754 | 1417 #undef H264_CHROMA_MC8_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1418 #undef H264_CHROMA_MC4_TMPL |
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
1419 #undef H264_CHROMA_MC8_MV0 |
2754 | 1420 |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1421 /***********************************/ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1422 /* weighted prediction */ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1423 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1424 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
|
1425 { |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1426 int x, y; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1427 offset <<= log2_denom; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1428 offset += (1 << log2_denom) >> 1; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1429 asm volatile( |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1430 "movd %0, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1431 "movd %1, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1432 "movd %2, %%mm6 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1433 "pshufw $0, %%mm4, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1434 "pshufw $0, %%mm5, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1435 "pxor %%mm7, %%mm7 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1436 :: "g"(weight), "g"(offset), "g"(log2_denom) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1437 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1438 for(y=0; y<h; y+=2){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1439 for(x=0; x<w; x+=4){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1440 asm volatile( |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1441 "movd %0, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1442 "movd %1, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1443 "punpcklbw %%mm7, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1444 "punpcklbw %%mm7, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1445 "pmullw %%mm4, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1446 "pmullw %%mm4, %%mm1 \n\t" |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
1447 "paddsw %%mm5, %%mm0 \n\t" |
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
1448 "paddsw %%mm5, %%mm1 \n\t" |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1449 "psraw %%mm6, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1450 "psraw %%mm6, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1451 "packuswb %%mm7, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1452 "packuswb %%mm7, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1453 "movd %%mm0, %0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1454 "movd %%mm1, %1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1455 : "+m"(*(uint32_t*)(dst+x)), |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1456 "+m"(*(uint32_t*)(dst+x+stride)) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1457 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1458 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1459 dst += 2*stride; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1460 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1461 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1462 |
3029 | 1463 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
|
1464 { |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1465 int x, y; |
3029 | 1466 offset = ((offset + 1) | 1) << log2_denom; |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1467 asm volatile( |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1468 "movd %0, %%mm3 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1469 "movd %1, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1470 "movd %2, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1471 "movd %3, %%mm6 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1472 "pshufw $0, %%mm3, %%mm3 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1473 "pshufw $0, %%mm4, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1474 "pshufw $0, %%mm5, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1475 "pxor %%mm7, %%mm7 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1476 :: "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
|
1477 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1478 for(y=0; y<h; y++){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1479 for(x=0; x<w; x+=4){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1480 asm volatile( |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1481 "movd %0, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1482 "movd %1, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1483 "punpcklbw %%mm7, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1484 "punpcklbw %%mm7, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1485 "pmullw %%mm3, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1486 "pmullw %%mm4, %%mm1 \n\t" |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
1487 "paddsw %%mm1, %%mm0 \n\t" |
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
1488 "paddsw %%mm5, %%mm0 \n\t" |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1489 "psraw %%mm6, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1490 "packuswb %%mm0, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1491 "movd %%mm0, %0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1492 : "+m"(*(uint32_t*)(dst+x)) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1493 : "m"(*(uint32_t*)(src+x)) |
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 src += stride; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1497 dst += stride; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1498 } |
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 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1501 #define H264_WEIGHT(W,H) \ |
3029 | 1502 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){ \ |
1503 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
|
1504 } \ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1505 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
|
1506 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
|
1507 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1508 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1509 H264_WEIGHT(16,16) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1510 H264_WEIGHT(16, 8) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1511 H264_WEIGHT( 8,16) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1512 H264_WEIGHT( 8, 8) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1513 H264_WEIGHT( 8, 4) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1514 H264_WEIGHT( 4, 8) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1515 H264_WEIGHT( 4, 4) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1516 H264_WEIGHT( 4, 2) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
1517 |