Mercurial > libavcodec.hg
annotate i386/h264dsp_mmx.c @ 8375:de2509cf3c44 libavcodec
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
thus avoiding the calling overhead.
New functions are not yet used.
author | michael |
---|---|
date | Thu, 18 Dec 2008 02:36:48 +0000 |
parents | eebc7209c47f |
children |
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 |
7327
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
23 DECLARE_ALIGNED_8 (static const uint64_t, ff_pb_3_1 ) = 0x0103010301030103ULL; |
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
24 DECLARE_ALIGNED_8 (static const uint64_t, ff_pb_7_3 ) = 0x0307030703070307ULL; |
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
25 |
2754 | 26 /***********************************/ |
27 /* IDCT */ | |
28 | |
29 #define SUMSUB_BADC( a, b, c, d ) \ | |
30 "paddw "#b", "#a" \n\t"\ | |
31 "paddw "#d", "#c" \n\t"\ | |
32 "paddw "#b", "#b" \n\t"\ | |
33 "paddw "#d", "#d" \n\t"\ | |
34 "psubw "#a", "#b" \n\t"\ | |
35 "psubw "#c", "#d" \n\t" | |
36 | |
37 #define SUMSUBD2_AB( a, b, t ) \ | |
38 "movq "#b", "#t" \n\t"\ | |
39 "psraw $1 , "#b" \n\t"\ | |
40 "paddw "#a", "#b" \n\t"\ | |
41 "psraw $1 , "#a" \n\t"\ | |
42 "psubw "#t", "#a" \n\t" | |
43 | |
44 #define IDCT4_1D( s02, s13, d02, d13, t ) \ | |
45 SUMSUB_BA ( s02, d02 )\ | |
46 SUMSUBD2_AB( s13, d13, t )\ | |
47 SUMSUB_BADC( d13, s02, s13, d02 ) | |
48 | |
49 #define STORE_DIFF_4P( p, t, z ) \ | |
50 "psraw $6, "#p" \n\t"\ | |
51 "movd (%0), "#t" \n\t"\ | |
52 "punpcklbw "#z", "#t" \n\t"\ | |
53 "paddsw "#t", "#p" \n\t"\ | |
54 "packuswb "#z", "#p" \n\t"\ | |
55 "movd "#p", (%0) \n\t" | |
56 | |
3173 | 57 static void ff_h264_idct_add_mmx(uint8_t *dst, int16_t *block, int stride) |
2754 | 58 { |
59 /* Load dct coeffs */ | |
8031 | 60 __asm__ volatile( |
2754 | 61 "movq (%0), %%mm0 \n\t" |
62 "movq 8(%0), %%mm1 \n\t" | |
63 "movq 16(%0), %%mm2 \n\t" | |
64 "movq 24(%0), %%mm3 \n\t" | |
65 :: "r"(block) ); | |
66 | |
8031 | 67 __asm__ volatile( |
2754 | 68 /* mm1=s02+s13 mm2=s02-s13 mm4=d02+d13 mm0=d02-d13 */ |
69 IDCT4_1D( %%mm2, %%mm1, %%mm0, %%mm3, %%mm4 ) | |
70 | |
71 "movq %0, %%mm6 \n\t" | |
72 /* in: 1,4,0,2 out: 1,2,3,0 */ | |
73 TRANSPOSE4( %%mm3, %%mm1, %%mm0, %%mm2, %%mm4 ) | |
74 | |
75 "paddw %%mm6, %%mm3 \n\t" | |
76 | |
77 /* mm2=s02+s13 mm3=s02-s13 mm4=d02+d13 mm1=d02-d13 */ | |
78 IDCT4_1D( %%mm4, %%mm2, %%mm3, %%mm0, %%mm1 ) | |
79 | |
80 "pxor %%mm7, %%mm7 \n\t" | |
6329
5969caa9190d
clean up an ugliness introduced in r11826. this syntax will require fewer changes when adding future sse2 code.
lorenm
parents:
6321
diff
changeset
|
81 :: "m"(ff_pw_32)); |
2754 | 82 |
8031 | 83 __asm__ volatile( |
2754 | 84 STORE_DIFF_4P( %%mm0, %%mm1, %%mm7) |
85 "add %1, %0 \n\t" | |
86 STORE_DIFF_4P( %%mm2, %%mm1, %%mm7) | |
87 "add %1, %0 \n\t" | |
88 STORE_DIFF_4P( %%mm3, %%mm1, %%mm7) | |
89 "add %1, %0 \n\t" | |
90 STORE_DIFF_4P( %%mm4, %%mm1, %%mm7) | |
91 : "+r"(dst) | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
92 : "r" ((x86_reg)stride) |
2754 | 93 ); |
94 } | |
95 | |
3174 | 96 static inline void h264_idct8_1d(int16_t *block) |
97 { | |
8031 | 98 __asm__ volatile( |
3174 | 99 "movq 112(%0), %%mm7 \n\t" |
6319
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
100 "movq 80(%0), %%mm0 \n\t" |
3174 | 101 "movq 48(%0), %%mm3 \n\t" |
6319
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
102 "movq 16(%0), %%mm5 \n\t" |
3174 | 103 |
6319
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
104 "movq %%mm0, %%mm4 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
105 "movq %%mm5, %%mm1 \n\t" |
3174 | 106 "psraw $1, %%mm4 \n\t" |
6319
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
107 "psraw $1, %%mm1 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
108 "paddw %%mm0, %%mm4 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
109 "paddw %%mm5, %%mm1 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
110 "paddw %%mm7, %%mm4 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
111 "paddw %%mm0, %%mm1 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
112 "psubw %%mm5, %%mm4 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
113 "paddw %%mm3, %%mm1 \n\t" |
3174 | 114 |
6319
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
115 "psubw %%mm3, %%mm5 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
116 "psubw %%mm3, %%mm0 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
117 "paddw %%mm7, %%mm5 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
118 "psubw %%mm7, %%mm0 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
119 "psraw $1, %%mm3 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
120 "psraw $1, %%mm7 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
121 "psubw %%mm3, %%mm5 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
122 "psubw %%mm7, %%mm0 \n\t" |
3174 | 123 |
124 "movq %%mm4, %%mm3 \n\t" | |
6319
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
125 "movq %%mm1, %%mm7 \n\t" |
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
126 "psraw $2, %%mm1 \n\t" |
3174 | 127 "psraw $2, %%mm3 \n\t" |
6319
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
128 "paddw %%mm5, %%mm3 \n\t" |
3174 | 129 "psraw $2, %%mm5 \n\t" |
6319
4089a1ae6558
remove some movq in ff_h264_idct8_add_mmx. 225->217 cycles on core2.
lorenm
parents:
6135
diff
changeset
|
130 "paddw %%mm0, %%mm1 \n\t" |
3174 | 131 "psraw $2, %%mm0 \n\t" |
132 "psubw %%mm4, %%mm5 \n\t" | |
133 "psubw %%mm0, %%mm7 \n\t" | |
134 | |
135 "movq 32(%0), %%mm2 \n\t" | |
136 "movq 96(%0), %%mm6 \n\t" | |
137 "movq %%mm2, %%mm4 \n\t" | |
138 "movq %%mm6, %%mm0 \n\t" | |
139 "psraw $1, %%mm4 \n\t" | |
140 "psraw $1, %%mm6 \n\t" | |
141 "psubw %%mm0, %%mm4 \n\t" | |
142 "paddw %%mm2, %%mm6 \n\t" | |
143 | |
144 "movq (%0), %%mm2 \n\t" | |
145 "movq 64(%0), %%mm0 \n\t" | |
146 SUMSUB_BA( %%mm0, %%mm2 ) | |
147 SUMSUB_BA( %%mm6, %%mm0 ) | |
148 SUMSUB_BA( %%mm4, %%mm2 ) | |
149 SUMSUB_BA( %%mm7, %%mm6 ) | |
150 SUMSUB_BA( %%mm5, %%mm4 ) | |
151 SUMSUB_BA( %%mm3, %%mm2 ) | |
152 SUMSUB_BA( %%mm1, %%mm0 ) | |
153 :: "r"(block) | |
154 ); | |
155 } | |
156 | |
157 static void ff_h264_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride) | |
158 { | |
159 int i; | |
160 int16_t __attribute__ ((aligned(8))) b2[64]; | |
161 | |
162 block[0] += 32; | |
163 | |
164 for(i=0; i<2; i++){ | |
4137 | 165 DECLARE_ALIGNED_8(uint64_t, tmp); |
3174 | 166 |
167 h264_idct8_1d(block+4*i); | |
168 | |
8031 | 169 __asm__ volatile( |
3174 | 170 "movq %%mm7, %0 \n\t" |
171 TRANSPOSE4( %%mm0, %%mm2, %%mm4, %%mm6, %%mm7 ) | |
172 "movq %%mm0, 8(%1) \n\t" | |
173 "movq %%mm6, 24(%1) \n\t" | |
174 "movq %%mm7, 40(%1) \n\t" | |
175 "movq %%mm4, 56(%1) \n\t" | |
176 "movq %0, %%mm7 \n\t" | |
177 TRANSPOSE4( %%mm7, %%mm5, %%mm3, %%mm1, %%mm0 ) | |
178 "movq %%mm7, (%1) \n\t" | |
179 "movq %%mm1, 16(%1) \n\t" | |
180 "movq %%mm0, 32(%1) \n\t" | |
181 "movq %%mm3, 48(%1) \n\t" | |
182 : "=m"(tmp) | |
183 : "r"(b2+32*i) | |
184 : "memory" | |
185 ); | |
186 } | |
187 | |
188 for(i=0; i<2; i++){ | |
189 h264_idct8_1d(b2+4*i); | |
190 | |
8031 | 191 __asm__ volatile( |
3174 | 192 "psraw $6, %%mm7 \n\t" |
193 "psraw $6, %%mm6 \n\t" | |
194 "psraw $6, %%mm5 \n\t" | |
195 "psraw $6, %%mm4 \n\t" | |
196 "psraw $6, %%mm3 \n\t" | |
197 "psraw $6, %%mm2 \n\t" | |
198 "psraw $6, %%mm1 \n\t" | |
199 "psraw $6, %%mm0 \n\t" | |
200 | |
201 "movq %%mm7, (%0) \n\t" | |
202 "movq %%mm5, 16(%0) \n\t" | |
203 "movq %%mm3, 32(%0) \n\t" | |
204 "movq %%mm1, 48(%0) \n\t" | |
205 "movq %%mm0, 64(%0) \n\t" | |
206 "movq %%mm2, 80(%0) \n\t" | |
207 "movq %%mm4, 96(%0) \n\t" | |
208 "movq %%mm6, 112(%0) \n\t" | |
209 :: "r"(b2+4*i) | |
210 : "memory" | |
211 ); | |
212 } | |
213 | |
214 add_pixels_clamped_mmx(b2, dst, stride); | |
215 } | |
216 | |
6320 | 217 #define STORE_DIFF_8P( p, d, t, z )\ |
218 "movq "#d", "#t" \n"\ | |
219 "psraw $6, "#p" \n"\ | |
220 "punpcklbw "#z", "#t" \n"\ | |
221 "paddsw "#t", "#p" \n"\ | |
222 "packuswb "#p", "#p" \n"\ | |
223 "movq "#p", "#d" \n" | |
224 | |
225 #define H264_IDCT8_1D_SSE2(a,b,c,d,e,f,g,h)\ | |
226 "movdqa "#c", "#a" \n"\ | |
227 "movdqa "#g", "#e" \n"\ | |
228 "psraw $1, "#c" \n"\ | |
229 "psraw $1, "#g" \n"\ | |
230 "psubw "#e", "#c" \n"\ | |
231 "paddw "#a", "#g" \n"\ | |
232 "movdqa "#b", "#e" \n"\ | |
233 "psraw $1, "#e" \n"\ | |
234 "paddw "#b", "#e" \n"\ | |
235 "paddw "#d", "#e" \n"\ | |
236 "paddw "#f", "#e" \n"\ | |
237 "movdqa "#f", "#a" \n"\ | |
238 "psraw $1, "#a" \n"\ | |
239 "paddw "#f", "#a" \n"\ | |
240 "paddw "#h", "#a" \n"\ | |
241 "psubw "#b", "#a" \n"\ | |
242 "psubw "#d", "#b" \n"\ | |
243 "psubw "#d", "#f" \n"\ | |
244 "paddw "#h", "#b" \n"\ | |
245 "psubw "#h", "#f" \n"\ | |
246 "psraw $1, "#d" \n"\ | |
247 "psraw $1, "#h" \n"\ | |
248 "psubw "#d", "#b" \n"\ | |
249 "psubw "#h", "#f" \n"\ | |
250 "movdqa "#e", "#d" \n"\ | |
251 "movdqa "#a", "#h" \n"\ | |
252 "psraw $2, "#d" \n"\ | |
253 "psraw $2, "#h" \n"\ | |
254 "paddw "#f", "#d" \n"\ | |
255 "paddw "#b", "#h" \n"\ | |
256 "psraw $2, "#f" \n"\ | |
257 "psraw $2, "#b" \n"\ | |
258 "psubw "#f", "#e" \n"\ | |
259 "psubw "#a", "#b" \n"\ | |
260 "movdqa 0x00(%1), "#a" \n"\ | |
261 "movdqa 0x40(%1), "#f" \n"\ | |
262 SUMSUB_BA(f, a)\ | |
263 SUMSUB_BA(g, f)\ | |
264 SUMSUB_BA(c, a)\ | |
265 SUMSUB_BA(e, g)\ | |
266 SUMSUB_BA(b, c)\ | |
267 SUMSUB_BA(h, a)\ | |
268 SUMSUB_BA(d, f) | |
269 | |
270 static void ff_h264_idct8_add_sse2(uint8_t *dst, int16_t *block, int stride) | |
271 { | |
8031 | 272 __asm__ volatile( |
6320 | 273 "movdqa 0x10(%1), %%xmm1 \n" |
274 "movdqa 0x20(%1), %%xmm2 \n" | |
275 "movdqa 0x30(%1), %%xmm3 \n" | |
276 "movdqa 0x50(%1), %%xmm5 \n" | |
277 "movdqa 0x60(%1), %%xmm6 \n" | |
278 "movdqa 0x70(%1), %%xmm7 \n" | |
279 H264_IDCT8_1D_SSE2(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm6, %%xmm7) | |
280 TRANSPOSE8(%%xmm4, %%xmm1, %%xmm7, %%xmm3, %%xmm5, %%xmm0, %%xmm2, %%xmm6, (%1)) | |
281 "paddw %4, %%xmm4 \n" | |
282 "movdqa %%xmm4, 0x00(%1) \n" | |
283 "movdqa %%xmm2, 0x40(%1) \n" | |
284 H264_IDCT8_1D_SSE2(%%xmm4, %%xmm0, %%xmm6, %%xmm3, %%xmm2, %%xmm5, %%xmm7, %%xmm1) | |
285 "movdqa %%xmm6, 0x60(%1) \n" | |
286 "movdqa %%xmm7, 0x70(%1) \n" | |
287 "pxor %%xmm7, %%xmm7 \n" | |
288 STORE_DIFF_8P(%%xmm2, (%0), %%xmm6, %%xmm7) | |
289 STORE_DIFF_8P(%%xmm0, (%0,%2), %%xmm6, %%xmm7) | |
290 STORE_DIFF_8P(%%xmm1, (%0,%2,2), %%xmm6, %%xmm7) | |
291 STORE_DIFF_8P(%%xmm3, (%0,%3), %%xmm6, %%xmm7) | |
292 "lea (%0,%2,4), %0 \n" | |
293 STORE_DIFF_8P(%%xmm5, (%0), %%xmm6, %%xmm7) | |
294 STORE_DIFF_8P(%%xmm4, (%0,%2), %%xmm6, %%xmm7) | |
295 "movdqa 0x60(%1), %%xmm0 \n" | |
296 "movdqa 0x70(%1), %%xmm1 \n" | |
297 STORE_DIFF_8P(%%xmm0, (%0,%2,2), %%xmm6, %%xmm7) | |
298 STORE_DIFF_8P(%%xmm1, (%0,%3), %%xmm6, %%xmm7) | |
299 :"+r"(dst) | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
300 :"r"(block), "r"((x86_reg)stride), "r"((x86_reg)3L*stride), "m"(ff_pw_32) |
6320 | 301 ); |
302 } | |
303 | |
3173 | 304 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
|
305 { |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
306 int dc = (block[0] + 32) >> 6; |
8031 | 307 __asm__ volatile( |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
308 "movd %0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
309 "pshufw $0, %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
310 "pxor %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
311 "psubw %%mm0, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
312 "packuswb %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
313 "packuswb %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
314 ::"r"(dc) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
315 ); |
8031 | 316 __asm__ volatile( |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
317 "movd %0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
318 "movd %1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
319 "movd %2, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
320 "movd %3, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
321 "paddusb %%mm0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
322 "paddusb %%mm0, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
323 "paddusb %%mm0, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
324 "paddusb %%mm0, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
325 "psubusb %%mm1, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
326 "psubusb %%mm1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
327 "psubusb %%mm1, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
328 "psubusb %%mm1, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
329 "movd %%mm2, %0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
330 "movd %%mm3, %1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
331 "movd %%mm4, %2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
332 "movd %%mm5, %3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
333 :"+m"(*(uint32_t*)(dst+0*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
334 "+m"(*(uint32_t*)(dst+1*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
335 "+m"(*(uint32_t*)(dst+2*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
336 "+m"(*(uint32_t*)(dst+3*stride)) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
337 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
338 } |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
339 |
3173 | 340 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
|
341 { |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
342 int dc = (block[0] + 32) >> 6; |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
343 int y; |
8031 | 344 __asm__ volatile( |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
345 "movd %0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
346 "pshufw $0, %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
347 "pxor %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
348 "psubw %%mm0, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
349 "packuswb %%mm0, %%mm0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
350 "packuswb %%mm1, %%mm1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
351 ::"r"(dc) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
352 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
353 for(y=2; y--; dst += 4*stride){ |
8031 | 354 __asm__ volatile( |
3105
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
355 "movq %0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
356 "movq %1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
357 "movq %2, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
358 "movq %3, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
359 "paddusb %%mm0, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
360 "paddusb %%mm0, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
361 "paddusb %%mm0, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
362 "paddusb %%mm0, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
363 "psubusb %%mm1, %%mm2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
364 "psubusb %%mm1, %%mm3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
365 "psubusb %%mm1, %%mm4 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
366 "psubusb %%mm1, %%mm5 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
367 "movq %%mm2, %0 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
368 "movq %%mm3, %1 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
369 "movq %%mm4, %2 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
370 "movq %%mm5, %3 \n\t" |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
371 :"+m"(*(uint64_t*)(dst+0*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
372 "+m"(*(uint64_t*)(dst+1*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
373 "+m"(*(uint64_t*)(dst+2*stride)), |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
374 "+m"(*(uint64_t*)(dst+3*stride)) |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
375 ); |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
376 } |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
377 } |
2d35fb3cb940
h264: special case dc-only idct. ~1% faster overall
lorenm
parents:
3102
diff
changeset
|
378 |
8375
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
379 //FIXME this table is a duplicate from h264data.h, and will be removed once the tables from, h264 have been split |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
380 static const uint8_t scan8[16 + 2*4]={ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
381 4+1*8, 5+1*8, 4+2*8, 5+2*8, |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
382 6+1*8, 7+1*8, 6+2*8, 7+2*8, |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
383 4+3*8, 5+3*8, 4+4*8, 5+4*8, |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
384 6+3*8, 7+3*8, 6+4*8, 7+4*8, |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
385 1+1*8, 2+1*8, |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
386 1+2*8, 2+2*8, |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
387 1+4*8, 2+4*8, |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
388 1+5*8, 2+5*8, |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
389 }; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
390 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
391 static void ff_h264_idct_add16_mmx(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
392 int i; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
393 for(i=0; i<16; i++){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
394 if(nnzc[ scan8[i] ]) |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
395 ff_h264_idct_add_mmx(dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
396 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
397 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
398 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
399 static void ff_h264_idct8_add4_mmx(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
400 int i; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
401 for(i=0; i<16; i+=4){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
402 if(nnzc[ scan8[i] ]) |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
403 ff_h264_idct8_add_mmx(dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
404 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
405 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
406 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
407 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
408 static void ff_h264_idct_add16_mmx2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
409 int i; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
410 for(i=0; i<16; i++){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
411 int nnz = nnzc[ scan8[i] ]; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
412 if(nnz){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
413 if(nnz==1 && block[i*16]) ff_h264_idct_dc_add_mmx2(dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
414 else ff_h264_idct_add_mmx (dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
415 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
416 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
417 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
418 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
419 static void ff_h264_idct_add16intra_mmx(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
420 int i; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
421 for(i=0; i<16; i++){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
422 if(nnzc[ scan8[i] ] || block[i*16]) |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
423 ff_h264_idct_add_mmx(dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
424 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
425 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
426 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
427 static void ff_h264_idct_add16intra_mmx2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
428 int i; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
429 for(i=0; i<16; i++){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
430 if(nnzc[ scan8[i] ]) ff_h264_idct_add_mmx (dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
431 else if(block[i*16]) ff_h264_idct_dc_add_mmx2(dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
432 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
433 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
434 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
435 static void ff_h264_idct8_add4_mmx2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
436 int i; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
437 for(i=0; i<16; i+=4){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
438 int nnz = nnzc[ scan8[i] ]; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
439 if(nnz){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
440 if(nnz==1 && block[i*16]) ff_h264_idct8_dc_add_mmx2(dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
441 else ff_h264_idct8_add_mmx (dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
442 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
443 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
444 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
445 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
446 static void ff_h264_idct8_add4_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
447 int i; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
448 for(i=0; i<16; i+=4){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
449 int nnz = nnzc[ scan8[i] ]; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
450 if(nnz){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
451 if(nnz==1 && block[i*16]) ff_h264_idct8_dc_add_mmx2(dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
452 else ff_h264_idct8_add_sse2 (dst + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
453 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
454 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
455 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
456 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
457 static void ff_h264_idct_add8_mmx(uint8_t **dest, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
458 int i; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
459 for(i=16; i<16+8; i++){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
460 if(nnzc[ scan8[i] ] || block[i*16]) |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
461 ff_h264_idct_add_mmx (dest[(i&4)>>2] + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
462 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
463 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
464 |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
465 static void ff_h264_idct_add8_mmx2(uint8_t **dest, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
466 int i; |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
467 for(i=16; i<16+8; i++){ |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
468 if(nnzc[ scan8[i] ]) |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
469 ff_h264_idct_add_mmx (dest[(i&4)>>2] + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
470 else if(block[i*16]) |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
471 ff_h264_idct_dc_add_mmx2(dest[(i&4)>>2] + block_offset[i], block + i*16, stride); |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
472 } |
de2509cf3c44
H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents:
8031
diff
changeset
|
473 } |
2754 | 474 |
475 /***********************************/ | |
476 /* deblocking */ | |
477 | |
478 // out: o = |x-y|>a | |
479 // clobbers: t | |
480 #define DIFF_GT_MMX(x,y,a,o,t)\ | |
481 "movq "#y", "#t" \n\t"\ | |
482 "movq "#x", "#o" \n\t"\ | |
483 "psubusb "#x", "#t" \n\t"\ | |
484 "psubusb "#y", "#o" \n\t"\ | |
485 "por "#t", "#o" \n\t"\ | |
486 "psubusb "#a", "#o" \n\t" | |
487 | |
4135 | 488 // out: o = |x-y|>a |
489 // clobbers: t | |
490 #define DIFF_GT2_MMX(x,y,a,o,t)\ | |
491 "movq "#y", "#t" \n\t"\ | |
492 "movq "#x", "#o" \n\t"\ | |
493 "psubusb "#x", "#t" \n\t"\ | |
494 "psubusb "#y", "#o" \n\t"\ | |
495 "psubusb "#a", "#t" \n\t"\ | |
496 "psubusb "#a", "#o" \n\t"\ | |
497 "pcmpeqb "#t", "#o" \n\t"\ | |
498 | |
2754 | 499 // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 |
500 // out: mm5=beta-1, mm7=mask | |
501 // clobbers: mm4,mm6 | |
502 #define H264_DEBLOCK_MASK(alpha1, beta1) \ | |
503 "pshufw $0, "#alpha1", %%mm4 \n\t"\ | |
504 "pshufw $0, "#beta1 ", %%mm5 \n\t"\ | |
505 "packuswb %%mm4, %%mm4 \n\t"\ | |
506 "packuswb %%mm5, %%mm5 \n\t"\ | |
507 DIFF_GT_MMX(%%mm1, %%mm2, %%mm4, %%mm7, %%mm6) /* |p0-q0| > alpha-1 */\ | |
508 DIFF_GT_MMX(%%mm0, %%mm1, %%mm5, %%mm4, %%mm6) /* |p1-p0| > beta-1 */\ | |
509 "por %%mm4, %%mm7 \n\t"\ | |
510 DIFF_GT_MMX(%%mm3, %%mm2, %%mm5, %%mm4, %%mm6) /* |q1-q0| > beta-1 */\ | |
511 "por %%mm4, %%mm7 \n\t"\ | |
512 "pxor %%mm6, %%mm6 \n\t"\ | |
513 "pcmpeqb %%mm6, %%mm7 \n\t" | |
514 | |
515 // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask) | |
516 // out: mm1=p0' mm2=q0' | |
517 // clobbers: mm0,3-6 | |
518 #define H264_DEBLOCK_P0_Q0(pb_01, pb_3f)\ | |
4127 | 519 "movq %%mm1 , %%mm5 \n\t"\ |
520 "pxor %%mm2 , %%mm5 \n\t" /* p0^q0*/\ | |
521 "pand "#pb_01" , %%mm5 \n\t" /* (p0^q0)&1*/\ | |
522 "pcmpeqb %%mm4 , %%mm4 \n\t"\ | |
523 "pxor %%mm4 , %%mm3 \n\t"\ | |
524 "pavgb %%mm0 , %%mm3 \n\t" /* (p1 - q1 + 256)>>1*/\ | |
525 "pavgb "MANGLE(ff_pb_3)" , %%mm3 \n\t" /*(((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2*/\ | |
526 "pxor %%mm1 , %%mm4 \n\t"\ | |
527 "pavgb %%mm2 , %%mm4 \n\t" /* (q0 - p0 + 256)>>1*/\ | |
528 "pavgb %%mm5 , %%mm3 \n\t"\ | |
4130 | 529 "paddusb %%mm4 , %%mm3 \n\t" /* d+128+33*/\ |
4129 | 530 "movq "MANGLE(ff_pb_A1)" , %%mm6 \n\t"\ |
531 "psubusb %%mm3 , %%mm6 \n\t"\ | |
4127 | 532 "psubusb "MANGLE(ff_pb_A1)" , %%mm3 \n\t"\ |
4129 | 533 "pminub %%mm7 , %%mm6 \n\t"\ |
4127 | 534 "pminub %%mm7 , %%mm3 \n\t"\ |
4129 | 535 "psubusb %%mm6 , %%mm1 \n\t"\ |
536 "psubusb %%mm3 , %%mm2 \n\t"\ | |
4127 | 537 "paddusb %%mm3 , %%mm1 \n\t"\ |
4129 | 538 "paddusb %%mm6 , %%mm2 \n\t" |
2754 | 539 |
5947 | 540 // in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask) %8=ff_bone |
4866 | 541 // out: (q1addr) = av_clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 ) |
2754 | 542 // clobbers: q2, tmp, tc0 |
543 #define H264_DEBLOCK_Q1(p1, q2, q2addr, q1addr, tc0, tmp)\ | |
544 "movq %%mm1, "#tmp" \n\t"\ | |
545 "pavgb %%mm2, "#tmp" \n\t"\ | |
546 "pavgb "#tmp", "#q2" \n\t" /* avg(p2,avg(p0,q0)) */\ | |
547 "pxor "q2addr", "#tmp" \n\t"\ | |
548 "pand %8, "#tmp" \n\t" /* (p2^avg(p0,q0))&1 */\ | |
549 "psubusb "#tmp", "#q2" \n\t" /* (p2+((p0+q0+1)>>1))>>1 */\ | |
550 "movq "#p1", "#tmp" \n\t"\ | |
551 "psubusb "#tc0", "#tmp" \n\t"\ | |
552 "paddusb "#p1", "#tc0" \n\t"\ | |
553 "pmaxub "#tmp", "#q2" \n\t"\ | |
554 "pminub "#tc0", "#q2" \n\t"\ | |
555 "movq "#q2", "q1addr" \n\t" | |
556 | |
557 static inline void h264_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0) | |
558 { | |
4137 | 559 DECLARE_ALIGNED_8(uint64_t, tmp0[2]); |
2754 | 560 |
8031 | 561 __asm__ volatile( |
2754 | 562 "movq (%1,%3), %%mm0 \n\t" //p1 |
563 "movq (%1,%3,2), %%mm1 \n\t" //p0 | |
564 "movq (%2), %%mm2 \n\t" //q0 | |
565 "movq (%2,%3), %%mm3 \n\t" //q1 | |
566 H264_DEBLOCK_MASK(%6, %7) | |
4133 | 567 |
568 "movd %5, %%mm4 \n\t" | |
569 "punpcklbw %%mm4, %%mm4 \n\t" | |
570 "punpcklwd %%mm4, %%mm4 \n\t" | |
571 "pcmpeqb %%mm3, %%mm3 \n\t" | |
572 "movq %%mm4, %%mm6 \n\t" | |
573 "pcmpgtb %%mm3, %%mm4 \n\t" | |
574 "movq %%mm6, 8+%0 \n\t" | |
575 "pand %%mm4, %%mm7 \n\t" | |
576 "movq %%mm7, %0 \n\t" | |
2754 | 577 |
578 /* filter p1 */ | |
579 "movq (%1), %%mm3 \n\t" //p2 | |
4135 | 580 DIFF_GT2_MMX(%%mm1, %%mm3, %%mm5, %%mm6, %%mm4) // |p2-p0|>beta-1 |
2754 | 581 "pand %%mm7, %%mm6 \n\t" // mask & |p2-p0|<beta |
4136 | 582 "pand 8+%0, %%mm7 \n\t" // mask & tc0 |
583 "movq %%mm7, %%mm4 \n\t" | |
4131
1a8e384d0463
2 instructions less in h264_loop_filter_luma_mmx2()
michael
parents:
4130
diff
changeset
|
584 "psubb %%mm6, %%mm7 \n\t" |
2754 | 585 "pand %%mm4, %%mm6 \n\t" // mask & |p2-p0|<beta & tc0 |
586 H264_DEBLOCK_Q1(%%mm0, %%mm3, "(%1)", "(%1,%3)", %%mm6, %%mm4) | |
587 | |
588 /* filter q1 */ | |
589 "movq (%2,%3,2), %%mm4 \n\t" //q2 | |
4135 | 590 DIFF_GT2_MMX(%%mm2, %%mm4, %%mm5, %%mm6, %%mm3) // |q2-q0|>beta-1 |
2754 | 591 "pand %0, %%mm6 \n\t" |
4134 | 592 "movq 8+%0, %%mm5 \n\t" // can be merged with the and below but is slower then |
2754 | 593 "pand %%mm6, %%mm5 \n\t" |
4131
1a8e384d0463
2 instructions less in h264_loop_filter_luma_mmx2()
michael
parents:
4130
diff
changeset
|
594 "psubb %%mm6, %%mm7 \n\t" |
2754 | 595 "movq (%2,%3), %%mm3 \n\t" |
596 H264_DEBLOCK_Q1(%%mm3, %%mm4, "(%2,%3,2)", "(%2,%3)", %%mm5, %%mm6) | |
597 | |
598 /* filter p0, q0 */ | |
4133 | 599 H264_DEBLOCK_P0_Q0(%8, unused) |
2754 | 600 "movq %%mm1, (%1,%3,2) \n\t" |
601 "movq %%mm2, (%2) \n\t" | |
602 | |
4133 | 603 : "=m"(*tmp0) |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
604 : "r"(pix-3*stride), "r"(pix), "r"((x86_reg)stride), |
4133 | 605 "m"(*tmp0/*unused*/), "m"(*(uint32_t*)tc0), "m"(alpha1), "m"(beta1), |
5947 | 606 "m"(ff_bone) |
2754 | 607 ); |
608 } | |
609 | |
610 static void h264_v_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
611 { | |
612 if((tc0[0] & tc0[1]) >= 0) | |
613 h264_loop_filter_luma_mmx2(pix, stride, alpha-1, beta-1, tc0); | |
614 if((tc0[2] & tc0[3]) >= 0) | |
615 h264_loop_filter_luma_mmx2(pix+8, stride, alpha-1, beta-1, tc0+2); | |
616 } | |
617 static void h264_h_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
618 { | |
619 //FIXME: could cut some load/stores by merging transpose with filter | |
620 // also, it only needs to transpose 6x8 | |
4137 | 621 DECLARE_ALIGNED_8(uint8_t, trans[8*8]); |
2754 | 622 int i; |
623 for(i=0; i<2; i++, pix+=8*stride, tc0+=2) { | |
624 if((tc0[0] & tc0[1]) < 0) | |
625 continue; | |
626 transpose4x4(trans, pix-4, 8, stride); | |
627 transpose4x4(trans +4*8, pix, 8, stride); | |
628 transpose4x4(trans+4, pix-4+4*stride, 8, stride); | |
629 transpose4x4(trans+4+4*8, pix +4*stride, 8, stride); | |
630 h264_loop_filter_luma_mmx2(trans+4*8, 8, alpha-1, beta-1, tc0); | |
631 transpose4x4(pix-2, trans +2*8, stride, 8); | |
632 transpose4x4(pix-2+4*stride, trans+4+2*8, stride, 8); | |
633 } | |
634 } | |
635 | |
636 static inline void h264_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0) | |
637 { | |
8031 | 638 __asm__ volatile( |
2754 | 639 "movq (%0), %%mm0 \n\t" //p1 |
640 "movq (%0,%2), %%mm1 \n\t" //p0 | |
641 "movq (%1), %%mm2 \n\t" //q0 | |
642 "movq (%1,%2), %%mm3 \n\t" //q1 | |
643 H264_DEBLOCK_MASK(%4, %5) | |
644 "movd %3, %%mm6 \n\t" | |
645 "punpcklbw %%mm6, %%mm6 \n\t" | |
646 "pand %%mm6, %%mm7 \n\t" // mm7 = tc&mask | |
647 H264_DEBLOCK_P0_Q0(%6, %7) | |
648 "movq %%mm1, (%0,%2) \n\t" | |
649 "movq %%mm2, (%1) \n\t" | |
650 | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
651 :: "r"(pix-2*stride), "r"(pix), "r"((x86_reg)stride), |
2754 | 652 "r"(*(uint32_t*)tc0), |
5947 | 653 "m"(alpha1), "m"(beta1), "m"(ff_bone), "m"(ff_pb_3F) |
2754 | 654 ); |
655 } | |
656 | |
657 static void h264_v_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
658 { | |
659 h264_loop_filter_chroma_mmx2(pix, stride, alpha-1, beta-1, tc0); | |
660 } | |
661 | |
662 static void h264_h_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) | |
663 { | |
664 //FIXME: could cut some load/stores by merging transpose with filter | |
4137 | 665 DECLARE_ALIGNED_8(uint8_t, trans[8*4]); |
2754 | 666 transpose4x4(trans, pix-2, 8, stride); |
667 transpose4x4(trans+4, pix-2+4*stride, 8, stride); | |
668 h264_loop_filter_chroma_mmx2(trans+2*8, 8, alpha-1, beta-1, tc0); | |
669 transpose4x4(pix-2, trans, stride, 8); | |
670 transpose4x4(pix-2+4*stride, trans+4, stride, 8); | |
671 } | |
672 | |
673 // p0 = (p0 + q1 + 2*p1 + 2) >> 2 | |
674 #define H264_FILTER_CHROMA4(p0, p1, q1, one) \ | |
675 "movq "#p0", %%mm4 \n\t"\ | |
676 "pxor "#q1", %%mm4 \n\t"\ | |
677 "pand "#one", %%mm4 \n\t" /* mm4 = (p0^q1)&1 */\ | |
678 "pavgb "#q1", "#p0" \n\t"\ | |
679 "psubusb %%mm4, "#p0" \n\t"\ | |
680 "pavgb "#p1", "#p0" \n\t" /* dst = avg(p1, avg(p0,q1) - ((p0^q1)&1)) */\ | |
681 | |
682 static inline void h264_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha1, int beta1) | |
683 { | |
8031 | 684 __asm__ volatile( |
2754 | 685 "movq (%0), %%mm0 \n\t" |
686 "movq (%0,%2), %%mm1 \n\t" | |
687 "movq (%1), %%mm2 \n\t" | |
688 "movq (%1,%2), %%mm3 \n\t" | |
689 H264_DEBLOCK_MASK(%3, %4) | |
690 "movq %%mm1, %%mm5 \n\t" | |
691 "movq %%mm2, %%mm6 \n\t" | |
692 H264_FILTER_CHROMA4(%%mm1, %%mm0, %%mm3, %5) //p0' | |
693 H264_FILTER_CHROMA4(%%mm2, %%mm3, %%mm0, %5) //q0' | |
694 "psubb %%mm5, %%mm1 \n\t" | |
695 "psubb %%mm6, %%mm2 \n\t" | |
696 "pand %%mm7, %%mm1 \n\t" | |
697 "pand %%mm7, %%mm2 \n\t" | |
698 "paddb %%mm5, %%mm1 \n\t" | |
699 "paddb %%mm6, %%mm2 \n\t" | |
700 "movq %%mm1, (%0,%2) \n\t" | |
701 "movq %%mm2, (%1) \n\t" | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
702 :: "r"(pix-2*stride), "r"(pix), "r"((x86_reg)stride), |
5947 | 703 "m"(alpha1), "m"(beta1), "m"(ff_bone) |
2754 | 704 ); |
705 } | |
706 | |
707 static void h264_v_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta) | |
708 { | |
709 h264_loop_filter_chroma_intra_mmx2(pix, stride, alpha-1, beta-1); | |
710 } | |
711 | |
712 static void h264_h_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta) | |
713 { | |
714 //FIXME: could cut some load/stores by merging transpose with filter | |
4137 | 715 DECLARE_ALIGNED_8(uint8_t, trans[8*4]); |
2754 | 716 transpose4x4(trans, pix-2, 8, stride); |
717 transpose4x4(trans+4, pix-2+4*stride, 8, stride); | |
718 h264_loop_filter_chroma_intra_mmx2(trans+2*8, 8, alpha-1, beta-1); | |
719 transpose4x4(pix-2, trans, stride, 8); | |
720 transpose4x4(pix-2+4*stride, trans+4, stride, 8); | |
721 } | |
722 | |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
723 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], |
7327
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
724 int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) { |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
725 int dir; |
8031 | 726 __asm__ volatile( |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
727 "pxor %%mm7, %%mm7 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
728 "movq %0, %%mm6 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
729 "movq %1, %%mm5 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
730 "movq %2, %%mm4 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
731 ::"m"(ff_pb_1), "m"(ff_pb_3), "m"(ff_pb_7) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
732 ); |
7327
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
733 if(field) |
8031 | 734 __asm__ volatile( |
7327
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
735 "movq %0, %%mm5 \n\t" |
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
736 "movq %1, %%mm4 \n\t" |
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
737 ::"m"(ff_pb_3_1), "m"(ff_pb_7_3) |
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
738 ); |
483421b11d98
Fix h264_loop_filter_strength_mmx2() so it works with PAFF.
michael
parents:
6755
diff
changeset
|
739 |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
740 // 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
|
741 // average filter time by 1.2% |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
742 for( dir=1; dir>=0; dir-- ) { |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
743 const int d_idx = dir ? -8 : -1; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
744 const int mask_mv = dir ? mask_mv1 : mask_mv0; |
4137 | 745 DECLARE_ALIGNED_8(const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL; |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
746 int b_idx, edge, l; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
747 for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) { |
8031 | 748 __asm__ volatile( |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
749 "pand %0, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
750 ::"m"(mask_dir) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
751 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
752 if(!(mask_mv & edge)) { |
8031 | 753 __asm__ volatile("pxor %%mm0, %%mm0 \n\t":); |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
754 for( l = bidir; l >= 0; l-- ) { |
8031 | 755 __asm__ volatile( |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
756 "movd %0, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
757 "punpckldq %1, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
758 "movq %%mm1, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
759 "psrlw $7, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
760 "pand %%mm6, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
761 "por %%mm2, %%mm1 \n\t" // ref_cache with -2 mapped to -1 |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
762 "punpckldq %%mm1, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
763 "pcmpeqb %%mm2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
764 "paddb %%mm6, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
765 "punpckhbw %%mm7, %%mm1 \n\t" // ref[b] != ref[bn] |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
766 "por %%mm1, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
767 |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
768 "movq %2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
769 "movq %3, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
770 "psubw %4, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
771 "psubw %5, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
772 "packsswb %%mm2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
773 "paddb %%mm5, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
774 "pminub %%mm4, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
775 "pcmpeqb %%mm4, %%mm1 \n\t" // abs(mv[b] - mv[bn]) >= limit |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
776 "por %%mm1, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
777 ::"m"(ref[l][b_idx]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
778 "m"(ref[l][b_idx+d_idx]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
779 "m"(mv[l][b_idx][0]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
780 "m"(mv[l][b_idx+2][0]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
781 "m"(mv[l][b_idx+d_idx][0]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
782 "m"(mv[l][b_idx+d_idx+2][0]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
783 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
784 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
785 } |
8031 | 786 __asm__ volatile( |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
787 "movd %0, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
788 "por %1, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
789 "punpcklbw %%mm7, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
790 "pcmpgtw %%mm7, %%mm1 \n\t" // nnz[b] || nnz[bn] |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
791 ::"m"(nnz[b_idx]), |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
792 "m"(nnz[b_idx+d_idx]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
793 ); |
8031 | 794 __asm__ volatile( |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
795 "pcmpeqw %%mm7, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
796 "pcmpeqw %%mm7, %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
797 "psrlw $15, %%mm0 \n\t" // nonzero -> 1 |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
798 "psrlw $14, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
799 "movq %%mm0, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
800 "por %%mm1, %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
801 "psrlw $1, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
802 "pandn %%mm2, %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
803 "movq %%mm1, %0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
804 :"=m"(*bS[dir][edge]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
805 ::"memory" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
806 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
807 } |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
808 edges = 4; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
809 step = 1; |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
810 } |
8031 | 811 __asm__ volatile( |
3645
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
812 "movq (%0), %%mm0 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
813 "movq 8(%0), %%mm1 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
814 "movq 16(%0), %%mm2 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
815 "movq 24(%0), %%mm3 \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
816 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
817 "movq %%mm0, (%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
818 "movq %%mm3, 8(%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
819 "movq %%mm4, 16(%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
820 "movq %%mm2, 24(%0) \n\t" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
821 ::"r"(bS[0]) |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
822 :"memory" |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
823 ); |
47821be55b6c
mmx implementation of deblocking strength decision.
lorenm
parents:
3394
diff
changeset
|
824 } |
2754 | 825 |
826 /***********************************/ | |
827 /* motion compensation */ | |
828 | |
6331 | 829 #define QPEL_H264V_MM(A,B,C,D,E,F,OP,T,Z,d,q)\ |
830 "mov"#q" "#C", "#T" \n\t"\ | |
831 "mov"#d" (%0), "#F" \n\t"\ | |
832 "paddw "#D", "#T" \n\t"\ | |
833 "psllw $2, "#T" \n\t"\ | |
834 "psubw "#B", "#T" \n\t"\ | |
835 "psubw "#E", "#T" \n\t"\ | |
836 "punpcklbw "#Z", "#F" \n\t"\ | |
837 "pmullw %4, "#T" \n\t"\ | |
838 "paddw %5, "#A" \n\t"\ | |
2979 | 839 "add %2, %0 \n\t"\ |
840 "paddw "#F", "#A" \n\t"\ | |
6331 | 841 "paddw "#A", "#T" \n\t"\ |
842 "psraw $5, "#T" \n\t"\ | |
843 "packuswb "#T", "#T" \n\t"\ | |
844 OP(T, (%1), A, d)\ | |
2979 | 845 "add %3, %1 \n\t" |
2754 | 846 |
6331 | 847 #define QPEL_H264HV_MM(A,B,C,D,E,F,OF,T,Z,d,q)\ |
848 "mov"#q" "#C", "#T" \n\t"\ | |
849 "mov"#d" (%0), "#F" \n\t"\ | |
850 "paddw "#D", "#T" \n\t"\ | |
851 "psllw $2, "#T" \n\t"\ | |
852 "paddw %4, "#A" \n\t"\ | |
853 "psubw "#B", "#T" \n\t"\ | |
854 "psubw "#E", "#T" \n\t"\ | |
855 "punpcklbw "#Z", "#F" \n\t"\ | |
856 "pmullw %3, "#T" \n\t"\ | |
857 "paddw "#F", "#A" \n\t"\ | |
2979 | 858 "add %2, %0 \n\t"\ |
6331 | 859 "paddw "#A", "#T" \n\t"\ |
860 "mov"#q" "#T", "#OF"(%1) \n\t" | |
861 | |
862 #define QPEL_H264V(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%mm6,%%mm7,d,q) | |
863 #define QPEL_H264HV(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%mm6,%%mm7,d,q) | |
864 #define QPEL_H264V_XMM(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%xmm6,%%xmm7,q,dqa) | |
865 #define QPEL_H264HV_XMM(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%xmm6,%%xmm7,q,dqa) | |
866 | |
2967 | 867 |
2754 | 868 #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
|
869 static av_noinline void OPNAME ## h264_qpel4_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 870 int h=4;\ |
871 \ | |
8031 | 872 __asm__ volatile(\ |
2979 | 873 "pxor %%mm7, %%mm7 \n\t"\ |
874 "movq %5, %%mm4 \n\t"\ | |
875 "movq %6, %%mm5 \n\t"\ | |
876 "1: \n\t"\ | |
877 "movd -1(%0), %%mm1 \n\t"\ | |
878 "movd (%0), %%mm2 \n\t"\ | |
879 "movd 1(%0), %%mm3 \n\t"\ | |
880 "movd 2(%0), %%mm0 \n\t"\ | |
881 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
882 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
883 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
884 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
885 "paddw %%mm0, %%mm1 \n\t"\ | |
886 "paddw %%mm3, %%mm2 \n\t"\ | |
887 "movd -2(%0), %%mm0 \n\t"\ | |
888 "movd 3(%0), %%mm3 \n\t"\ | |
889 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
890 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
891 "paddw %%mm3, %%mm0 \n\t"\ | |
892 "psllw $2, %%mm2 \n\t"\ | |
893 "psubw %%mm1, %%mm2 \n\t"\ | |
894 "pmullw %%mm4, %%mm2 \n\t"\ | |
895 "paddw %%mm5, %%mm0 \n\t"\ | |
896 "paddw %%mm2, %%mm0 \n\t"\ | |
897 "psraw $5, %%mm0 \n\t"\ | |
898 "packuswb %%mm0, %%mm0 \n\t"\ | |
2754 | 899 OP(%%mm0, (%1),%%mm6, d)\ |
2979 | 900 "add %3, %0 \n\t"\ |
901 "add %4, %1 \n\t"\ | |
902 "decl %2 \n\t"\ | |
903 " jnz 1b \n\t"\ | |
6335
950811a14eb3
put loop counter in a register if possible. makes some of the qpel functions 3% faster.
lorenm
parents:
6334
diff
changeset
|
904 : "+a"(src), "+c"(dst), "+g"(h)\ |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
905 : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
2754 | 906 : "memory"\ |
907 );\ | |
908 }\ | |
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
|
909 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 | 910 int h=4;\ |
8031 | 911 __asm__ volatile(\ |
3156 | 912 "pxor %%mm7, %%mm7 \n\t"\ |
3165 | 913 "movq %0, %%mm4 \n\t"\ |
914 "movq %1, %%mm5 \n\t"\ | |
915 :: "m"(ff_pw_5), "m"(ff_pw_16)\ | |
916 );\ | |
917 do{\ | |
8031 | 918 __asm__ volatile(\ |
3156 | 919 "movd -1(%0), %%mm1 \n\t"\ |
920 "movd (%0), %%mm2 \n\t"\ | |
921 "movd 1(%0), %%mm3 \n\t"\ | |
922 "movd 2(%0), %%mm0 \n\t"\ | |
923 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
924 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
925 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
926 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
927 "paddw %%mm0, %%mm1 \n\t"\ | |
928 "paddw %%mm3, %%mm2 \n\t"\ | |
929 "movd -2(%0), %%mm0 \n\t"\ | |
930 "movd 3(%0), %%mm3 \n\t"\ | |
931 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
932 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
933 "paddw %%mm3, %%mm0 \n\t"\ | |
934 "psllw $2, %%mm2 \n\t"\ | |
935 "psubw %%mm1, %%mm2 \n\t"\ | |
936 "pmullw %%mm4, %%mm2 \n\t"\ | |
937 "paddw %%mm5, %%mm0 \n\t"\ | |
938 "paddw %%mm2, %%mm0 \n\t"\ | |
939 "movd (%2), %%mm3 \n\t"\ | |
940 "psraw $5, %%mm0 \n\t"\ | |
941 "packuswb %%mm0, %%mm0 \n\t"\ | |
942 PAVGB" %%mm3, %%mm0 \n\t"\ | |
943 OP(%%mm0, (%1),%%mm6, d)\ | |
3165 | 944 "add %4, %0 \n\t"\ |
945 "add %4, %1 \n\t"\ | |
946 "add %3, %2 \n\t"\ | |
947 : "+a"(src), "+c"(dst), "+d"(src2)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
948 : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride)\ |
3156 | 949 : "memory"\ |
950 );\ | |
3165 | 951 }while(--h);\ |
3156 | 952 }\ |
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
|
953 static av_noinline void OPNAME ## h264_qpel4_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 954 src -= 2*srcStride;\ |
8031 | 955 __asm__ volatile(\ |
2979 | 956 "pxor %%mm7, %%mm7 \n\t"\ |
957 "movd (%0), %%mm0 \n\t"\ | |
958 "add %2, %0 \n\t"\ | |
959 "movd (%0), %%mm1 \n\t"\ | |
960 "add %2, %0 \n\t"\ | |
961 "movd (%0), %%mm2 \n\t"\ | |
962 "add %2, %0 \n\t"\ | |
963 "movd (%0), %%mm3 \n\t"\ | |
964 "add %2, %0 \n\t"\ | |
965 "movd (%0), %%mm4 \n\t"\ | |
966 "add %2, %0 \n\t"\ | |
967 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
968 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
969 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
970 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
971 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
2754 | 972 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ |
973 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
974 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
975 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
976 \ | |
977 : "+a"(src), "+c"(dst)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
978 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
2754 | 979 : "memory"\ |
980 );\ | |
981 }\ | |
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
|
982 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 | 983 int h=4;\ |
984 int w=3;\ | |
985 src -= 2*srcStride+2;\ | |
986 while(w--){\ | |
8031 | 987 __asm__ volatile(\ |
2979 | 988 "pxor %%mm7, %%mm7 \n\t"\ |
989 "movd (%0), %%mm0 \n\t"\ | |
990 "add %2, %0 \n\t"\ | |
991 "movd (%0), %%mm1 \n\t"\ | |
992 "add %2, %0 \n\t"\ | |
993 "movd (%0), %%mm2 \n\t"\ | |
994 "add %2, %0 \n\t"\ | |
995 "movd (%0), %%mm3 \n\t"\ | |
996 "add %2, %0 \n\t"\ | |
997 "movd (%0), %%mm4 \n\t"\ | |
998 "add %2, %0 \n\t"\ | |
999 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1000 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
1001 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1002 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
1003 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
2754 | 1004 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*8*3)\ |
1005 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*8*3)\ | |
1006 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*8*3)\ | |
1007 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*8*3)\ | |
1008 \ | |
1009 : "+a"(src)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1010 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
2754 | 1011 : "memory"\ |
1012 );\ | |
1013 tmp += 4;\ | |
1014 src += 4 - 9*srcStride;\ | |
1015 }\ | |
1016 tmp -= 3*4;\ | |
8031 | 1017 __asm__ volatile(\ |
2979 | 1018 "1: \n\t"\ |
1019 "movq (%0), %%mm0 \n\t"\ | |
1020 "paddw 10(%0), %%mm0 \n\t"\ | |
1021 "movq 2(%0), %%mm1 \n\t"\ | |
1022 "paddw 8(%0), %%mm1 \n\t"\ | |
1023 "movq 4(%0), %%mm2 \n\t"\ | |
1024 "paddw 6(%0), %%mm2 \n\t"\ | |
1025 "psubw %%mm1, %%mm0 \n\t"/*a-b (abccba)*/\ | |
1026 "psraw $2, %%mm0 \n\t"/*(a-b)/4 */\ | |
1027 "psubw %%mm1, %%mm0 \n\t"/*(a-b)/4-b */\ | |
1028 "paddsw %%mm2, %%mm0 \n\t"\ | |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
1029 "psraw $2, %%mm0 \n\t"/*((a-b)/4-b+c)/4 */\ |
6330
5070b3157fcf
add qpel rounder once during hv rather than twice during hv and whatever it's averaged with
lorenm
parents:
6329
diff
changeset
|
1030 "paddw %%mm2, %%mm0 \n\t"/*(a-5*b+20*c)/16 */\ |
2979 | 1031 "psraw $6, %%mm0 \n\t"\ |
1032 "packuswb %%mm0, %%mm0 \n\t"\ | |
2754 | 1033 OP(%%mm0, (%1),%%mm7, d)\ |
2979 | 1034 "add $24, %0 \n\t"\ |
1035 "add %3, %1 \n\t"\ | |
1036 "decl %2 \n\t"\ | |
1037 " jnz 1b \n\t"\ | |
6335
950811a14eb3
put loop counter in a register if possible. makes some of the qpel functions 3% faster.
lorenm
parents:
6334
diff
changeset
|
1038 : "+a"(tmp), "+c"(dst), "+g"(h)\ |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1039 : "S"((x86_reg)dstStride)\ |
2754 | 1040 : "memory"\ |
1041 );\ | |
1042 }\ | |
1043 \ | |
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
|
1044 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 1045 int h=8;\ |
8031 | 1046 __asm__ volatile(\ |
2979 | 1047 "pxor %%mm7, %%mm7 \n\t"\ |
1048 "movq %5, %%mm6 \n\t"\ | |
1049 "1: \n\t"\ | |
1050 "movq (%0), %%mm0 \n\t"\ | |
1051 "movq 1(%0), %%mm2 \n\t"\ | |
1052 "movq %%mm0, %%mm1 \n\t"\ | |
1053 "movq %%mm2, %%mm3 \n\t"\ | |
1054 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1055 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1056 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1057 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1058 "paddw %%mm2, %%mm0 \n\t"\ | |
1059 "paddw %%mm3, %%mm1 \n\t"\ | |
1060 "psllw $2, %%mm0 \n\t"\ | |
1061 "psllw $2, %%mm1 \n\t"\ | |
1062 "movq -1(%0), %%mm2 \n\t"\ | |
1063 "movq 2(%0), %%mm4 \n\t"\ | |
1064 "movq %%mm2, %%mm3 \n\t"\ | |
1065 "movq %%mm4, %%mm5 \n\t"\ | |
1066 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1067 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1068 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
1069 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
1070 "paddw %%mm4, %%mm2 \n\t"\ | |
1071 "paddw %%mm3, %%mm5 \n\t"\ | |
1072 "psubw %%mm2, %%mm0 \n\t"\ | |
1073 "psubw %%mm5, %%mm1 \n\t"\ | |
1074 "pmullw %%mm6, %%mm0 \n\t"\ | |
1075 "pmullw %%mm6, %%mm1 \n\t"\ | |
1076 "movd -2(%0), %%mm2 \n\t"\ | |
1077 "movd 7(%0), %%mm5 \n\t"\ | |
1078 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1079 "punpcklbw %%mm7, %%mm5 \n\t"\ | |
1080 "paddw %%mm3, %%mm2 \n\t"\ | |
1081 "paddw %%mm5, %%mm4 \n\t"\ | |
1082 "movq %6, %%mm5 \n\t"\ | |
1083 "paddw %%mm5, %%mm2 \n\t"\ | |
1084 "paddw %%mm5, %%mm4 \n\t"\ | |
1085 "paddw %%mm2, %%mm0 \n\t"\ | |
1086 "paddw %%mm4, %%mm1 \n\t"\ | |
1087 "psraw $5, %%mm0 \n\t"\ | |
1088 "psraw $5, %%mm1 \n\t"\ | |
1089 "packuswb %%mm1, %%mm0 \n\t"\ | |
2754 | 1090 OP(%%mm0, (%1),%%mm5, q)\ |
2979 | 1091 "add %3, %0 \n\t"\ |
1092 "add %4, %1 \n\t"\ | |
1093 "decl %2 \n\t"\ | |
1094 " jnz 1b \n\t"\ | |
6335
950811a14eb3
put loop counter in a register if possible. makes some of the qpel functions 3% faster.
lorenm
parents:
6334
diff
changeset
|
1095 : "+a"(src), "+c"(dst), "+g"(h)\ |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1096 : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
2754 | 1097 : "memory"\ |
1098 );\ | |
1099 }\ | |
1100 \ | |
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
|
1101 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 | 1102 int h=8;\ |
8031 | 1103 __asm__ volatile(\ |
3156 | 1104 "pxor %%mm7, %%mm7 \n\t"\ |
3165 | 1105 "movq %0, %%mm6 \n\t"\ |
1106 :: "m"(ff_pw_5)\ | |
1107 );\ | |
1108 do{\ | |
8031 | 1109 __asm__ volatile(\ |
3156 | 1110 "movq (%0), %%mm0 \n\t"\ |
1111 "movq 1(%0), %%mm2 \n\t"\ | |
1112 "movq %%mm0, %%mm1 \n\t"\ | |
1113 "movq %%mm2, %%mm3 \n\t"\ | |
1114 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1115 "punpckhbw %%mm7, %%mm1 \n\t"\ | |
1116 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1117 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1118 "paddw %%mm2, %%mm0 \n\t"\ | |
1119 "paddw %%mm3, %%mm1 \n\t"\ | |
1120 "psllw $2, %%mm0 \n\t"\ | |
1121 "psllw $2, %%mm1 \n\t"\ | |
1122 "movq -1(%0), %%mm2 \n\t"\ | |
1123 "movq 2(%0), %%mm4 \n\t"\ | |
1124 "movq %%mm2, %%mm3 \n\t"\ | |
1125 "movq %%mm4, %%mm5 \n\t"\ | |
1126 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1127 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1128 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
1129 "punpckhbw %%mm7, %%mm5 \n\t"\ | |
1130 "paddw %%mm4, %%mm2 \n\t"\ | |
1131 "paddw %%mm3, %%mm5 \n\t"\ | |
1132 "psubw %%mm2, %%mm0 \n\t"\ | |
1133 "psubw %%mm5, %%mm1 \n\t"\ | |
1134 "pmullw %%mm6, %%mm0 \n\t"\ | |
1135 "pmullw %%mm6, %%mm1 \n\t"\ | |
1136 "movd -2(%0), %%mm2 \n\t"\ | |
1137 "movd 7(%0), %%mm5 \n\t"\ | |
1138 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1139 "punpcklbw %%mm7, %%mm5 \n\t"\ | |
1140 "paddw %%mm3, %%mm2 \n\t"\ | |
1141 "paddw %%mm5, %%mm4 \n\t"\ | |
3165 | 1142 "movq %5, %%mm5 \n\t"\ |
3156 | 1143 "paddw %%mm5, %%mm2 \n\t"\ |
1144 "paddw %%mm5, %%mm4 \n\t"\ | |
1145 "paddw %%mm2, %%mm0 \n\t"\ | |
1146 "paddw %%mm4, %%mm1 \n\t"\ | |
1147 "psraw $5, %%mm0 \n\t"\ | |
1148 "psraw $5, %%mm1 \n\t"\ | |
1149 "movq (%2), %%mm4 \n\t"\ | |
1150 "packuswb %%mm1, %%mm0 \n\t"\ | |
1151 PAVGB" %%mm4, %%mm0 \n\t"\ | |
1152 OP(%%mm0, (%1),%%mm5, q)\ | |
3165 | 1153 "add %4, %0 \n\t"\ |
1154 "add %4, %1 \n\t"\ | |
1155 "add %3, %2 \n\t"\ | |
1156 : "+a"(src), "+c"(dst), "+d"(src2)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1157 : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride),\ |
3165 | 1158 "m"(ff_pw_16)\ |
3156 | 1159 : "memory"\ |
1160 );\ | |
3165 | 1161 }while(--h);\ |
3156 | 1162 }\ |
1163 \ | |
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
|
1164 static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
3094 | 1165 int w= 2;\ |
2754 | 1166 src -= 2*srcStride;\ |
1167 \ | |
3094 | 1168 while(w--){\ |
8031 | 1169 __asm__ volatile(\ |
2979 | 1170 "pxor %%mm7, %%mm7 \n\t"\ |
1171 "movd (%0), %%mm0 \n\t"\ | |
1172 "add %2, %0 \n\t"\ | |
1173 "movd (%0), %%mm1 \n\t"\ | |
1174 "add %2, %0 \n\t"\ | |
1175 "movd (%0), %%mm2 \n\t"\ | |
1176 "add %2, %0 \n\t"\ | |
1177 "movd (%0), %%mm3 \n\t"\ | |
1178 "add %2, %0 \n\t"\ | |
1179 "movd (%0), %%mm4 \n\t"\ | |
1180 "add %2, %0 \n\t"\ | |
1181 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1182 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
1183 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1184 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
1185 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
2754 | 1186 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ |
1187 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
1188 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
1189 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
1190 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\ | |
1191 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\ | |
1192 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ | |
1193 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
1194 \ | |
1195 : "+a"(src), "+c"(dst)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1196 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
2754 | 1197 : "memory"\ |
1198 );\ | |
3094 | 1199 if(h==16){\ |
8031 | 1200 __asm__ volatile(\ |
3094 | 1201 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ |
1202 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
1203 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\ | |
1204 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\ | |
1205 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\ | |
1206 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\ | |
1207 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\ | |
1208 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\ | |
1209 \ | |
1210 : "+a"(src), "+c"(dst)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1211 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
3094 | 1212 : "memory"\ |
1213 );\ | |
1214 }\ | |
1215 src += 4-(h+5)*srcStride;\ | |
1216 dst += 4-h*dstStride;\ | |
2754 | 1217 }\ |
1218 }\ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1219 static av_always_inline void OPNAME ## h264_qpel8or16_hv1_lowpass_ ## MMX(int16_t *tmp, uint8_t *src, int tmpStride, int srcStride, int size){\ |
3093 | 1220 int w = (size+8)>>2;\ |
2754 | 1221 src -= 2*srcStride+2;\ |
1222 while(w--){\ | |
8031 | 1223 __asm__ volatile(\ |
2979 | 1224 "pxor %%mm7, %%mm7 \n\t"\ |
1225 "movd (%0), %%mm0 \n\t"\ | |
1226 "add %2, %0 \n\t"\ | |
1227 "movd (%0), %%mm1 \n\t"\ | |
1228 "add %2, %0 \n\t"\ | |
1229 "movd (%0), %%mm2 \n\t"\ | |
1230 "add %2, %0 \n\t"\ | |
1231 "movd (%0), %%mm3 \n\t"\ | |
1232 "add %2, %0 \n\t"\ | |
1233 "movd (%0), %%mm4 \n\t"\ | |
1234 "add %2, %0 \n\t"\ | |
1235 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1236 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
1237 "punpcklbw %%mm7, %%mm2 \n\t"\ | |
1238 "punpcklbw %%mm7, %%mm3 \n\t"\ | |
1239 "punpcklbw %%mm7, %%mm4 \n\t"\ | |
3093 | 1240 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*48)\ |
1241 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*48)\ | |
1242 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*48)\ | |
1243 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*48)\ | |
1244 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 4*48)\ | |
1245 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 5*48)\ | |
1246 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 6*48)\ | |
1247 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 7*48)\ | |
2754 | 1248 : "+a"(src)\ |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1249 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
2754 | 1250 : "memory"\ |
1251 );\ | |
3093 | 1252 if(size==16){\ |
8031 | 1253 __asm__ volatile(\ |
3093 | 1254 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 8*48)\ |
1255 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 9*48)\ | |
1256 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 10*48)\ | |
1257 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 11*48)\ | |
1258 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 12*48)\ | |
1259 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 13*48)\ | |
1260 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 14*48)\ | |
1261 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 15*48)\ | |
1262 : "+a"(src)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1263 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
3093 | 1264 : "memory"\ |
1265 );\ | |
1266 }\ | |
2754 | 1267 tmp += 4;\ |
3093 | 1268 src += 4 - (size+5)*srcStride;\ |
2754 | 1269 }\ |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1270 }\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1271 static av_always_inline void OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, int dstStride, int tmpStride, int size){\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1272 int w = size>>4;\ |
3093 | 1273 do{\ |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1274 int h = size;\ |
8031 | 1275 __asm__ volatile(\ |
2979 | 1276 "1: \n\t"\ |
1277 "movq (%0), %%mm0 \n\t"\ | |
1278 "movq 8(%0), %%mm3 \n\t"\ | |
1279 "movq 2(%0), %%mm1 \n\t"\ | |
1280 "movq 10(%0), %%mm4 \n\t"\ | |
1281 "paddw %%mm4, %%mm0 \n\t"\ | |
1282 "paddw %%mm3, %%mm1 \n\t"\ | |
1283 "paddw 18(%0), %%mm3 \n\t"\ | |
1284 "paddw 16(%0), %%mm4 \n\t"\ | |
1285 "movq 4(%0), %%mm2 \n\t"\ | |
1286 "movq 12(%0), %%mm5 \n\t"\ | |
1287 "paddw 6(%0), %%mm2 \n\t"\ | |
1288 "paddw 14(%0), %%mm5 \n\t"\ | |
1289 "psubw %%mm1, %%mm0 \n\t"\ | |
1290 "psubw %%mm4, %%mm3 \n\t"\ | |
1291 "psraw $2, %%mm0 \n\t"\ | |
1292 "psraw $2, %%mm3 \n\t"\ | |
1293 "psubw %%mm1, %%mm0 \n\t"\ | |
1294 "psubw %%mm4, %%mm3 \n\t"\ | |
1295 "paddsw %%mm2, %%mm0 \n\t"\ | |
1296 "paddsw %%mm5, %%mm3 \n\t"\ | |
1297 "psraw $2, %%mm0 \n\t"\ | |
1298 "psraw $2, %%mm3 \n\t"\ | |
1299 "paddw %%mm2, %%mm0 \n\t"\ | |
1300 "paddw %%mm5, %%mm3 \n\t"\ | |
1301 "psraw $6, %%mm0 \n\t"\ | |
1302 "psraw $6, %%mm3 \n\t"\ | |
1303 "packuswb %%mm3, %%mm0 \n\t"\ | |
2754 | 1304 OP(%%mm0, (%1),%%mm7, q)\ |
3093 | 1305 "add $48, %0 \n\t"\ |
2979 | 1306 "add %3, %1 \n\t"\ |
1307 "decl %2 \n\t"\ | |
1308 " jnz 1b \n\t"\ | |
6335
950811a14eb3
put loop counter in a register if possible. makes some of the qpel functions 3% faster.
lorenm
parents:
6334
diff
changeset
|
1309 : "+a"(tmp), "+c"(dst), "+g"(h)\ |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1310 : "S"((x86_reg)dstStride)\ |
2754 | 1311 : "memory"\ |
1312 );\ | |
3093 | 1313 tmp += 8 - size*24;\ |
1314 dst += 8 - size*dstStride;\ | |
1315 }while(w--);\ | |
2754 | 1316 }\ |
3094 | 1317 \ |
1318 static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ | |
1319 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 8);\ | |
1320 }\ | |
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
|
1321 static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
3094 | 1322 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 16);\ |
1323 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\ | |
2754 | 1324 }\ |
1325 \ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1326 static void OPNAME ## h264_qpel16_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
2754 | 1327 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\ |
1328 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\ | |
1329 src += 8*srcStride;\ | |
1330 dst += 8*dstStride;\ | |
1331 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\ | |
1332 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\ | |
1333 }\ | |
1334 \ | |
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
|
1335 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 | 1336 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\ |
1337 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\ | |
1338 src += 8*dstStride;\ | |
1339 dst += 8*dstStride;\ | |
1340 src2 += 8*src2Stride;\ | |
1341 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\ | |
1342 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\ | |
1343 }\ | |
1344 \ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1345 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){\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1346 put_h264_qpel8or16_hv1_lowpass_ ## MMX(tmp, src, tmpStride, srcStride, size);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1347 OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(dst, tmp, dstStride, tmpStride, size);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1348 }\ |
3093 | 1349 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ |
1350 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 8);\ | |
1351 }\ | |
1352 \ | |
2754 | 1353 static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ |
3093 | 1354 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 16);\ |
2754 | 1355 }\ |
3095 | 1356 \ |
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
|
1357 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 | 1358 {\ |
8031 | 1359 __asm__ volatile(\ |
3095 | 1360 "movq (%1), %%mm0 \n\t"\ |
1361 "movq 24(%1), %%mm1 \n\t"\ | |
3102 | 1362 "psraw $5, %%mm0 \n\t"\ |
1363 "psraw $5, %%mm1 \n\t"\ | |
3163 | 1364 "packuswb %%mm0, %%mm0 \n\t"\ |
1365 "packuswb %%mm1, %%mm1 \n\t"\ | |
3095 | 1366 PAVGB" (%0), %%mm0 \n\t"\ |
1367 PAVGB" (%0,%3), %%mm1 \n\t"\ | |
1368 OP(%%mm0, (%2), %%mm4, d)\ | |
1369 OP(%%mm1, (%2,%4), %%mm5, d)\ | |
1370 "lea (%0,%3,2), %0 \n\t"\ | |
1371 "lea (%2,%4,2), %2 \n\t"\ | |
1372 "movq 48(%1), %%mm0 \n\t"\ | |
1373 "movq 72(%1), %%mm1 \n\t"\ | |
3102 | 1374 "psraw $5, %%mm0 \n\t"\ |
1375 "psraw $5, %%mm1 \n\t"\ | |
3163 | 1376 "packuswb %%mm0, %%mm0 \n\t"\ |
1377 "packuswb %%mm1, %%mm1 \n\t"\ | |
3095 | 1378 PAVGB" (%0), %%mm0 \n\t"\ |
1379 PAVGB" (%0,%3), %%mm1 \n\t"\ | |
1380 OP(%%mm0, (%2), %%mm4, d)\ | |
1381 OP(%%mm1, (%2,%4), %%mm5, d)\ | |
1382 :"+a"(src8), "+c"(src16), "+d"(dst)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1383 :"S"((x86_reg)src8Stride), "D"((x86_reg)dstStride)\ |
3095 | 1384 :"memory");\ |
1385 }\ | |
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
|
1386 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 | 1387 {\ |
6331 | 1388 do{\ |
8031 | 1389 __asm__ volatile(\ |
3095 | 1390 "movq (%1), %%mm0 \n\t"\ |
1391 "movq 8(%1), %%mm1 \n\t"\ | |
6331 | 1392 "movq 48(%1), %%mm2 \n\t"\ |
1393 "movq 8+48(%1), %%mm3 \n\t"\ | |
3102 | 1394 "psraw $5, %%mm0 \n\t"\ |
1395 "psraw $5, %%mm1 \n\t"\ | |
6331 | 1396 "psraw $5, %%mm2 \n\t"\ |
1397 "psraw $5, %%mm3 \n\t"\ | |
3095 | 1398 "packuswb %%mm1, %%mm0 \n\t"\ |
6331 | 1399 "packuswb %%mm3, %%mm2 \n\t"\ |
3095 | 1400 PAVGB" (%0), %%mm0 \n\t"\ |
6331 | 1401 PAVGB" (%0,%3), %%mm2 \n\t"\ |
3095 | 1402 OP(%%mm0, (%2), %%mm5, q)\ |
6331 | 1403 OP(%%mm2, (%2,%4), %%mm5, q)\ |
1404 ::"a"(src8), "c"(src16), "d"(dst),\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1405 "r"((x86_reg)src8Stride), "r"((x86_reg)dstStride)\ |
3095 | 1406 :"memory");\ |
6331 | 1407 src8 += 2L*src8Stride;\ |
1408 src16 += 48;\ | |
1409 dst += 2L*dstStride;\ | |
1410 }while(h-=2);\ | |
3095 | 1411 }\ |
1412 static void OPNAME ## pixels16_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\ | |
1413 {\ | |
1414 OPNAME ## pixels8_l2_shift5_ ## MMX(dst , src16 , src8 , dstStride, src8Stride, h);\ | |
1415 OPNAME ## pixels8_l2_shift5_ ## MMX(dst+8, src16+8, src8+8, dstStride, src8Stride, h);\ | |
1416 }\ | |
1417 | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1418 |
6331 | 1419 #ifdef ARCH_X86_64 |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1420 #define QPEL_H264_H16_XMM(OPNAME, OP, MMX)\ |
6331 | 1421 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\ |
1422 int h=16;\ | |
8031 | 1423 __asm__ volatile(\ |
6331 | 1424 "pxor %%xmm15, %%xmm15 \n\t"\ |
1425 "movdqa %6, %%xmm14 \n\t"\ | |
1426 "movdqa %7, %%xmm13 \n\t"\ | |
1427 "1: \n\t"\ | |
1428 "lddqu 3(%0), %%xmm1 \n\t"\ | |
1429 "lddqu -5(%0), %%xmm7 \n\t"\ | |
1430 "movdqa %%xmm1, %%xmm0 \n\t"\ | |
1431 "punpckhbw %%xmm15, %%xmm1 \n\t"\ | |
1432 "punpcklbw %%xmm15, %%xmm0 \n\t"\ | |
1433 "punpcklbw %%xmm15, %%xmm7 \n\t"\ | |
1434 "movdqa %%xmm1, %%xmm2 \n\t"\ | |
1435 "movdqa %%xmm0, %%xmm6 \n\t"\ | |
1436 "movdqa %%xmm1, %%xmm3 \n\t"\ | |
1437 "movdqa %%xmm0, %%xmm8 \n\t"\ | |
1438 "movdqa %%xmm1, %%xmm4 \n\t"\ | |
1439 "movdqa %%xmm0, %%xmm9 \n\t"\ | |
1440 "movdqa %%xmm1, %%xmm5 \n\t"\ | |
1441 "movdqa %%xmm0, %%xmm10 \n\t"\ | |
1442 "palignr $6, %%xmm0, %%xmm5 \n\t"\ | |
1443 "palignr $6, %%xmm7, %%xmm10\n\t"\ | |
1444 "palignr $8, %%xmm0, %%xmm4 \n\t"\ | |
1445 "palignr $8, %%xmm7, %%xmm9 \n\t"\ | |
1446 "palignr $10,%%xmm0, %%xmm3 \n\t"\ | |
1447 "palignr $10,%%xmm7, %%xmm8 \n\t"\ | |
1448 "paddw %%xmm1, %%xmm5 \n\t"\ | |
1449 "paddw %%xmm0, %%xmm10 \n\t"\ | |
1450 "palignr $12,%%xmm0, %%xmm2 \n\t"\ | |
1451 "palignr $12,%%xmm7, %%xmm6 \n\t"\ | |
1452 "palignr $14,%%xmm0, %%xmm1 \n\t"\ | |
1453 "palignr $14,%%xmm7, %%xmm0 \n\t"\ | |
1454 "paddw %%xmm3, %%xmm2 \n\t"\ | |
1455 "paddw %%xmm8, %%xmm6 \n\t"\ | |
1456 "paddw %%xmm4, %%xmm1 \n\t"\ | |
1457 "paddw %%xmm9, %%xmm0 \n\t"\ | |
1458 "psllw $2, %%xmm2 \n\t"\ | |
1459 "psllw $2, %%xmm6 \n\t"\ | |
1460 "psubw %%xmm1, %%xmm2 \n\t"\ | |
1461 "psubw %%xmm0, %%xmm6 \n\t"\ | |
1462 "paddw %%xmm13,%%xmm5 \n\t"\ | |
1463 "paddw %%xmm13,%%xmm10 \n\t"\ | |
1464 "pmullw %%xmm14,%%xmm2 \n\t"\ | |
1465 "pmullw %%xmm14,%%xmm6 \n\t"\ | |
1466 "lddqu (%2), %%xmm3 \n\t"\ | |
1467 "paddw %%xmm5, %%xmm2 \n\t"\ | |
1468 "paddw %%xmm10,%%xmm6 \n\t"\ | |
1469 "psraw $5, %%xmm2 \n\t"\ | |
1470 "psraw $5, %%xmm6 \n\t"\ | |
1471 "packuswb %%xmm2,%%xmm6 \n\t"\ | |
1472 "pavgb %%xmm3, %%xmm6 \n\t"\ | |
1473 OP(%%xmm6, (%1), %%xmm4, dqa)\ | |
1474 "add %5, %0 \n\t"\ | |
1475 "add %5, %1 \n\t"\ | |
1476 "add %4, %2 \n\t"\ | |
6345 | 1477 "decl %3 \n\t"\ |
6331 | 1478 "jg 1b \n\t"\ |
1479 : "+a"(src), "+c"(dst), "+d"(src2), "+g"(h)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1480 : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride),\ |
6331 | 1481 "m"(ff_pw_5), "m"(ff_pw_16)\ |
1482 : "memory"\ | |
1483 );\ | |
1484 } | |
1485 #else // ARCH_X86_64 | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1486 #define QPEL_H264_H16_XMM(OPNAME, OP, MMX)\ |
6331 | 1487 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\ |
1488 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\ | |
1489 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\ | |
1490 src += 8*dstStride;\ | |
1491 dst += 8*dstStride;\ | |
1492 src2 += 8*src2Stride;\ | |
1493 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\ | |
1494 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\ | |
1495 } | |
1496 #endif // ARCH_X86_64 | |
2754 | 1497 |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1498 #define QPEL_H264_H_XMM(OPNAME, OP, MMX)\ |
6331 | 1499 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\ |
1500 int h=8;\ | |
8031 | 1501 __asm__ volatile(\ |
6331 | 1502 "pxor %%xmm7, %%xmm7 \n\t"\ |
1503 "movdqa %0, %%xmm6 \n\t"\ | |
1504 :: "m"(ff_pw_5)\ | |
1505 );\ | |
1506 do{\ | |
8031 | 1507 __asm__ volatile(\ |
6331 | 1508 "lddqu -5(%0), %%xmm1 \n\t"\ |
1509 "movdqa %%xmm1, %%xmm0 \n\t"\ | |
1510 "punpckhbw %%xmm7, %%xmm1 \n\t"\ | |
1511 "punpcklbw %%xmm7, %%xmm0 \n\t"\ | |
1512 "movdqa %%xmm1, %%xmm2 \n\t"\ | |
1513 "movdqa %%xmm1, %%xmm3 \n\t"\ | |
1514 "movdqa %%xmm1, %%xmm4 \n\t"\ | |
1515 "movdqa %%xmm1, %%xmm5 \n\t"\ | |
1516 "palignr $6, %%xmm0, %%xmm5 \n\t"\ | |
1517 "palignr $8, %%xmm0, %%xmm4 \n\t"\ | |
1518 "palignr $10,%%xmm0, %%xmm3 \n\t"\ | |
1519 "paddw %%xmm1, %%xmm5 \n\t"\ | |
1520 "palignr $12,%%xmm0, %%xmm2 \n\t"\ | |
1521 "palignr $14,%%xmm0, %%xmm1 \n\t"\ | |
1522 "paddw %%xmm3, %%xmm2 \n\t"\ | |
1523 "paddw %%xmm4, %%xmm1 \n\t"\ | |
1524 "psllw $2, %%xmm2 \n\t"\ | |
1525 "movq (%2), %%xmm3 \n\t"\ | |
1526 "psubw %%xmm1, %%xmm2 \n\t"\ | |
1527 "paddw %5, %%xmm5 \n\t"\ | |
1528 "pmullw %%xmm6, %%xmm2 \n\t"\ | |
1529 "paddw %%xmm5, %%xmm2 \n\t"\ | |
1530 "psraw $5, %%xmm2 \n\t"\ | |
1531 "packuswb %%xmm2, %%xmm2 \n\t"\ | |
1532 "pavgb %%xmm3, %%xmm2 \n\t"\ | |
1533 OP(%%xmm2, (%1), %%xmm4, q)\ | |
1534 "add %4, %0 \n\t"\ | |
1535 "add %4, %1 \n\t"\ | |
1536 "add %3, %2 \n\t"\ | |
1537 : "+a"(src), "+c"(dst), "+d"(src2)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1538 : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride),\ |
6331 | 1539 "m"(ff_pw_16)\ |
1540 : "memory"\ | |
1541 );\ | |
1542 }while(--h);\ | |
1543 }\ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1544 QPEL_H264_H16_XMM(OPNAME, OP, MMX)\ |
6331 | 1545 \ |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1546 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1547 int h=8;\ |
8031 | 1548 __asm__ volatile(\ |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1549 "pxor %%xmm7, %%xmm7 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1550 "movdqa %5, %%xmm6 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1551 "1: \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1552 "lddqu -5(%0), %%xmm1 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1553 "movdqa %%xmm1, %%xmm0 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1554 "punpckhbw %%xmm7, %%xmm1 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1555 "punpcklbw %%xmm7, %%xmm0 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1556 "movdqa %%xmm1, %%xmm2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1557 "movdqa %%xmm1, %%xmm3 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1558 "movdqa %%xmm1, %%xmm4 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1559 "movdqa %%xmm1, %%xmm5 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1560 "palignr $6, %%xmm0, %%xmm5 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1561 "palignr $8, %%xmm0, %%xmm4 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1562 "palignr $10,%%xmm0, %%xmm3 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1563 "paddw %%xmm1, %%xmm5 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1564 "palignr $12,%%xmm0, %%xmm2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1565 "palignr $14,%%xmm0, %%xmm1 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1566 "paddw %%xmm3, %%xmm2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1567 "paddw %%xmm4, %%xmm1 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1568 "psllw $2, %%xmm2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1569 "psubw %%xmm1, %%xmm2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1570 "paddw %6, %%xmm5 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1571 "pmullw %%xmm6, %%xmm2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1572 "paddw %%xmm5, %%xmm2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1573 "psraw $5, %%xmm2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1574 "packuswb %%xmm2, %%xmm2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1575 OP(%%xmm2, (%1), %%xmm4, q)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1576 "add %3, %0 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1577 "add %4, %1 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1578 "decl %2 \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1579 " jnz 1b \n\t"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1580 : "+a"(src), "+c"(dst), "+g"(h)\ |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1581 : "D"((x86_reg)srcStride), "S"((x86_reg)dstStride),\ |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1582 "m"(ff_pw_5), "m"(ff_pw_16)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1583 : "memory"\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1584 );\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1585 }\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1586 static void OPNAME ## h264_qpel16_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1587 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1588 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1589 src += 8*srcStride;\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1590 dst += 8*dstStride;\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1591 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1592 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1593 }\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1594 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1595 #define QPEL_H264_V_XMM(OPNAME, OP, MMX)\ |
6331 | 1596 static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\ |
1597 src -= 2*srcStride;\ | |
1598 \ | |
8031 | 1599 __asm__ volatile(\ |
6331 | 1600 "pxor %%xmm7, %%xmm7 \n\t"\ |
1601 "movq (%0), %%xmm0 \n\t"\ | |
1602 "add %2, %0 \n\t"\ | |
1603 "movq (%0), %%xmm1 \n\t"\ | |
1604 "add %2, %0 \n\t"\ | |
1605 "movq (%0), %%xmm2 \n\t"\ | |
1606 "add %2, %0 \n\t"\ | |
1607 "movq (%0), %%xmm3 \n\t"\ | |
1608 "add %2, %0 \n\t"\ | |
1609 "movq (%0), %%xmm4 \n\t"\ | |
1610 "add %2, %0 \n\t"\ | |
1611 "punpcklbw %%xmm7, %%xmm0 \n\t"\ | |
1612 "punpcklbw %%xmm7, %%xmm1 \n\t"\ | |
1613 "punpcklbw %%xmm7, %%xmm2 \n\t"\ | |
1614 "punpcklbw %%xmm7, %%xmm3 \n\t"\ | |
1615 "punpcklbw %%xmm7, %%xmm4 \n\t"\ | |
1616 QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\ | |
1617 QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\ | |
1618 QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\ | |
1619 QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\ | |
1620 QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\ | |
1621 QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\ | |
1622 QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\ | |
1623 QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\ | |
1624 \ | |
1625 : "+a"(src), "+c"(dst)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1626 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
6331 | 1627 : "memory"\ |
1628 );\ | |
1629 if(h==16){\ | |
8031 | 1630 __asm__ volatile(\ |
6331 | 1631 QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\ |
1632 QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\ | |
1633 QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\ | |
1634 QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\ | |
1635 QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\ | |
1636 QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\ | |
1637 QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\ | |
1638 QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\ | |
1639 \ | |
1640 : "+a"(src), "+c"(dst)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1641 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\ |
6331 | 1642 : "memory"\ |
1643 );\ | |
1644 }\ | |
1645 }\ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1646 static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1647 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 8);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1648 }\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1649 static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1650 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 16);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1651 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1652 } |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1653 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1654 static av_always_inline void put_h264_qpel8or16_hv1_lowpass_sse2(int16_t *tmp, uint8_t *src, int tmpStride, int srcStride, int size){ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1655 int w = (size+8)>>3; |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1656 src -= 2*srcStride+2; |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1657 while(w--){ |
8031 | 1658 __asm__ volatile( |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1659 "pxor %%xmm7, %%xmm7 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1660 "movq (%0), %%xmm0 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1661 "add %2, %0 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1662 "movq (%0), %%xmm1 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1663 "add %2, %0 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1664 "movq (%0), %%xmm2 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1665 "add %2, %0 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1666 "movq (%0), %%xmm3 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1667 "add %2, %0 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1668 "movq (%0), %%xmm4 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1669 "add %2, %0 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1670 "punpcklbw %%xmm7, %%xmm0 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1671 "punpcklbw %%xmm7, %%xmm1 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1672 "punpcklbw %%xmm7, %%xmm2 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1673 "punpcklbw %%xmm7, %%xmm3 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1674 "punpcklbw %%xmm7, %%xmm4 \n\t" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1675 QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 0*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1676 QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 1*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1677 QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 2*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1678 QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 3*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1679 QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 4*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1680 QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 5*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1681 QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 6*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1682 QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 7*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1683 : "+a"(src) |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1684 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16) |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1685 : "memory" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1686 ); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1687 if(size==16){ |
8031 | 1688 __asm__ volatile( |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1689 QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 8*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1690 QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 9*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1691 QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 10*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1692 QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 11*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1693 QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 12*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1694 QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 13*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1695 QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 14*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1696 QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 15*48) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1697 : "+a"(src) |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1698 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16) |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1699 : "memory" |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1700 ); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1701 } |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1702 tmp += 8; |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1703 src += 8 - (size+5)*srcStride; |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1704 } |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1705 } |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1706 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1707 #define QPEL_H264_HV2_XMM(OPNAME, OP, MMX)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1708 static av_always_inline void OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, int dstStride, int tmpStride, int size){\ |
6331 | 1709 int h = size;\ |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1710 if(size == 16){\ |
8031 | 1711 __asm__ volatile(\ |
6331 | 1712 "1: \n\t"\ |
1713 "movdqa 32(%0), %%xmm4 \n\t"\ | |
1714 "movdqa 16(%0), %%xmm5 \n\t"\ | |
1715 "movdqa (%0), %%xmm7 \n\t"\ | |
1716 "movdqa %%xmm4, %%xmm3 \n\t"\ | |
1717 "movdqa %%xmm4, %%xmm2 \n\t"\ | |
1718 "movdqa %%xmm4, %%xmm1 \n\t"\ | |
1719 "movdqa %%xmm4, %%xmm0 \n\t"\ | |
1720 "palignr $10, %%xmm5, %%xmm0 \n\t"\ | |
1721 "palignr $8, %%xmm5, %%xmm1 \n\t"\ | |
1722 "palignr $6, %%xmm5, %%xmm2 \n\t"\ | |
1723 "palignr $4, %%xmm5, %%xmm3 \n\t"\ | |
1724 "palignr $2, %%xmm5, %%xmm4 \n\t"\ | |
1725 "paddw %%xmm5, %%xmm0 \n\t"\ | |
1726 "paddw %%xmm4, %%xmm1 \n\t"\ | |
1727 "paddw %%xmm3, %%xmm2 \n\t"\ | |
1728 "movdqa %%xmm5, %%xmm6 \n\t"\ | |
1729 "movdqa %%xmm5, %%xmm4 \n\t"\ | |
1730 "movdqa %%xmm5, %%xmm3 \n\t"\ | |
1731 "palignr $8, %%xmm7, %%xmm4 \n\t"\ | |
1732 "palignr $2, %%xmm7, %%xmm6 \n\t"\ | |
1733 "palignr $10, %%xmm7, %%xmm3 \n\t"\ | |
1734 "paddw %%xmm6, %%xmm4 \n\t"\ | |
1735 "movdqa %%xmm5, %%xmm6 \n\t"\ | |
1736 "palignr $6, %%xmm7, %%xmm5 \n\t"\ | |
1737 "palignr $4, %%xmm7, %%xmm6 \n\t"\ | |
1738 "paddw %%xmm7, %%xmm3 \n\t"\ | |
1739 "paddw %%xmm6, %%xmm5 \n\t"\ | |
1740 \ | |
1741 "psubw %%xmm1, %%xmm0 \n\t"\ | |
1742 "psubw %%xmm4, %%xmm3 \n\t"\ | |
1743 "psraw $2, %%xmm0 \n\t"\ | |
1744 "psraw $2, %%xmm3 \n\t"\ | |
1745 "psubw %%xmm1, %%xmm0 \n\t"\ | |
1746 "psubw %%xmm4, %%xmm3 \n\t"\ | |
1747 "paddw %%xmm2, %%xmm0 \n\t"\ | |
1748 "paddw %%xmm5, %%xmm3 \n\t"\ | |
1749 "psraw $2, %%xmm0 \n\t"\ | |
1750 "psraw $2, %%xmm3 \n\t"\ | |
1751 "paddw %%xmm2, %%xmm0 \n\t"\ | |
1752 "paddw %%xmm5, %%xmm3 \n\t"\ | |
1753 "psraw $6, %%xmm0 \n\t"\ | |
1754 "psraw $6, %%xmm3 \n\t"\ | |
1755 "packuswb %%xmm0, %%xmm3 \n\t"\ | |
1756 OP(%%xmm3, (%1), %%xmm7, dqa)\ | |
1757 "add $48, %0 \n\t"\ | |
1758 "add %3, %1 \n\t"\ | |
1759 "decl %2 \n\t"\ | |
1760 " jnz 1b \n\t"\ | |
1761 : "+a"(tmp), "+c"(dst), "+g"(h)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1762 : "S"((x86_reg)dstStride)\ |
6331 | 1763 : "memory"\ |
1764 );\ | |
1765 }else{\ | |
8031 | 1766 __asm__ volatile(\ |
6331 | 1767 "1: \n\t"\ |
1768 "movdqa 16(%0), %%xmm1 \n\t"\ | |
1769 "movdqa (%0), %%xmm0 \n\t"\ | |
1770 "movdqa %%xmm1, %%xmm2 \n\t"\ | |
1771 "movdqa %%xmm1, %%xmm3 \n\t"\ | |
1772 "movdqa %%xmm1, %%xmm4 \n\t"\ | |
1773 "movdqa %%xmm1, %%xmm5 \n\t"\ | |
1774 "palignr $10, %%xmm0, %%xmm5 \n\t"\ | |
1775 "palignr $8, %%xmm0, %%xmm4 \n\t"\ | |
1776 "palignr $6, %%xmm0, %%xmm3 \n\t"\ | |
1777 "palignr $4, %%xmm0, %%xmm2 \n\t"\ | |
1778 "palignr $2, %%xmm0, %%xmm1 \n\t"\ | |
1779 "paddw %%xmm5, %%xmm0 \n\t"\ | |
1780 "paddw %%xmm4, %%xmm1 \n\t"\ | |
1781 "paddw %%xmm3, %%xmm2 \n\t"\ | |
1782 "psubw %%xmm1, %%xmm0 \n\t"\ | |
1783 "psraw $2, %%xmm0 \n\t"\ | |
1784 "psubw %%xmm1, %%xmm0 \n\t"\ | |
1785 "paddw %%xmm2, %%xmm0 \n\t"\ | |
1786 "psraw $2, %%xmm0 \n\t"\ | |
1787 "paddw %%xmm2, %%xmm0 \n\t"\ | |
1788 "psraw $6, %%xmm0 \n\t"\ | |
1789 "packuswb %%xmm0, %%xmm0 \n\t"\ | |
1790 OP(%%xmm0, (%1), %%xmm7, q)\ | |
1791 "add $48, %0 \n\t"\ | |
1792 "add %3, %1 \n\t"\ | |
1793 "decl %2 \n\t"\ | |
1794 " jnz 1b \n\t"\ | |
1795 : "+a"(tmp), "+c"(dst), "+g"(h)\ | |
6755
33896780c612
Do not misuse long as the size of a register in x86.
ramiro
parents:
6557
diff
changeset
|
1796 : "S"((x86_reg)dstStride)\ |
6331 | 1797 : "memory"\ |
1798 );\ | |
1799 }\ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1800 } |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1801 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1802 #define QPEL_H264_HV_XMM(OPNAME, OP, MMX)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1803 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){\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1804 put_h264_qpel8or16_hv1_lowpass_sse2(tmp, src, tmpStride, srcStride, size);\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1805 OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(dst, tmp, dstStride, tmpStride, size);\ |
6331 | 1806 }\ |
1807 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1808 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst, tmp, src, dstStride, tmpStride, srcStride, 8);\ |
6331 | 1809 }\ |
1810 static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1811 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst, tmp, src, dstStride, tmpStride, srcStride, 16);\ |
6331 | 1812 }\ |
1813 | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1814 #define put_pixels8_l2_sse2 put_pixels8_l2_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1815 #define avg_pixels8_l2_sse2 avg_pixels8_l2_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1816 #define put_pixels16_l2_sse2 put_pixels16_l2_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1817 #define avg_pixels16_l2_sse2 avg_pixels16_l2_mmx2 |
6331 | 1818 #define put_pixels8_l2_ssse3 put_pixels8_l2_mmx2 |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1819 #define avg_pixels8_l2_ssse3 avg_pixels8_l2_mmx2 |
6331 | 1820 #define put_pixels16_l2_ssse3 put_pixels16_l2_mmx2 |
1821 #define avg_pixels16_l2_ssse3 avg_pixels16_l2_mmx2 | |
1822 | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1823 #define put_pixels8_l2_shift5_sse2 put_pixels8_l2_shift5_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1824 #define avg_pixels8_l2_shift5_sse2 avg_pixels8_l2_shift5_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1825 #define put_pixels16_l2_shift5_sse2 put_pixels16_l2_shift5_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1826 #define avg_pixels16_l2_shift5_sse2 avg_pixels16_l2_shift5_mmx2 |
6331 | 1827 #define put_pixels8_l2_shift5_ssse3 put_pixels8_l2_shift5_mmx2 |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1828 #define avg_pixels8_l2_shift5_ssse3 avg_pixels8_l2_shift5_mmx2 |
6331 | 1829 #define put_pixels16_l2_shift5_ssse3 put_pixels16_l2_shift5_mmx2 |
1830 #define avg_pixels16_l2_shift5_ssse3 avg_pixels16_l2_shift5_mmx2 | |
1831 | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1832 #define put_h264_qpel8_h_lowpass_l2_sse2 put_h264_qpel8_h_lowpass_l2_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1833 #define avg_h264_qpel8_h_lowpass_l2_sse2 avg_h264_qpel8_h_lowpass_l2_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1834 #define put_h264_qpel16_h_lowpass_l2_sse2 put_h264_qpel16_h_lowpass_l2_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1835 #define avg_h264_qpel16_h_lowpass_l2_sse2 avg_h264_qpel16_h_lowpass_l2_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1836 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1837 #define put_h264_qpel8_v_lowpass_ssse3 put_h264_qpel8_v_lowpass_sse2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1838 #define avg_h264_qpel8_v_lowpass_ssse3 avg_h264_qpel8_v_lowpass_sse2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1839 #define put_h264_qpel16_v_lowpass_ssse3 put_h264_qpel16_v_lowpass_sse2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1840 #define avg_h264_qpel16_v_lowpass_ssse3 avg_h264_qpel16_v_lowpass_sse2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1841 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1842 #define put_h264_qpel8or16_hv2_lowpass_sse2 put_h264_qpel8or16_hv2_lowpass_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1843 #define avg_h264_qpel8or16_hv2_lowpass_sse2 avg_h264_qpel8or16_hv2_lowpass_mmx2 |
6331 | 1844 |
1845 #define H264_MC(OPNAME, SIZE, MMX, ALIGN) \ | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1846 H264_MC_C(OPNAME, SIZE, MMX, ALIGN)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1847 H264_MC_V(OPNAME, SIZE, MMX, ALIGN)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1848 H264_MC_H(OPNAME, SIZE, MMX, ALIGN)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1849 H264_MC_HV(OPNAME, SIZE, MMX, ALIGN)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1850 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1851 static void put_h264_qpel16_mc00_sse2 (uint8_t *dst, uint8_t *src, int stride){ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1852 put_pixels16_sse2(dst, src, stride, 16); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1853 } |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1854 static void avg_h264_qpel16_mc00_sse2 (uint8_t *dst, uint8_t *src, int stride){ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1855 avg_pixels16_sse2(dst, src, stride, 16); |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1856 } |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1857 #define put_h264_qpel8_mc00_sse2 put_h264_qpel8_mc00_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1858 #define avg_h264_qpel8_mc00_sse2 avg_h264_qpel8_mc00_mmx2 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1859 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1860 #define H264_MC_C(OPNAME, SIZE, MMX, ALIGN) \ |
2754 | 1861 static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\ |
6321 | 1862 OPNAME ## pixels ## SIZE ## _ ## MMX(dst, src, stride, SIZE);\ |
2754 | 1863 }\ |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1864 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1865 #define H264_MC_H(OPNAME, SIZE, MMX, ALIGN) \ |
2754 | 1866 static void OPNAME ## h264_qpel ## SIZE ## _mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
3156 | 1867 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src, stride, stride);\ |
2754 | 1868 }\ |
1869 \ | |
1870 static void OPNAME ## h264_qpel ## SIZE ## _mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1871 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1872 }\ | |
1873 \ | |
1874 static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
3156 | 1875 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src+1, stride, stride);\ |
2754 | 1876 }\ |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1877 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1878 #define H264_MC_V(OPNAME, SIZE, MMX, ALIGN) \ |
2754 | 1879 static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
6334 | 1880 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*SIZE]);\ |
1881 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\ | |
1882 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src, temp, stride, stride, SIZE);\ | |
2754 | 1883 }\ |
1884 \ | |
1885 static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
1886 OPNAME ## h264_qpel ## SIZE ## _v_lowpass_ ## MMX(dst, src, stride, stride);\ | |
1887 }\ | |
1888 \ | |
1889 static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
6334 | 1890 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*SIZE]);\ |
1891 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\ | |
1892 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src+stride, temp, stride, stride, SIZE);\ | |
2754 | 1893 }\ |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1894 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1895 #define H264_MC_HV(OPNAME, SIZE, MMX, ALIGN) \ |
2754 | 1896 static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ |
6334 | 1897 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*SIZE]);\ |
1898 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\ | |
1899 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, temp, stride, SIZE);\ | |
2754 | 1900 }\ |
1901 \ | |
1902 static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
6334 | 1903 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*SIZE]);\ |
1904 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src+1, SIZE, stride);\ | |
1905 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, temp, stride, SIZE);\ | |
2754 | 1906 }\ |
1907 \ | |
1908 static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
6334 | 1909 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*SIZE]);\ |
1910 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\ | |
1911 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, temp, stride, SIZE);\ | |
2754 | 1912 }\ |
1913 \ | |
1914 static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
6334 | 1915 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*SIZE]);\ |
1916 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src+1, SIZE, stride);\ | |
1917 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, temp, stride, SIZE);\ | |
2754 | 1918 }\ |
1919 \ | |
1920 static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
6334 | 1921 DECLARE_ALIGNED(ALIGN, uint16_t, temp[SIZE*(SIZE<8?12:24)]);\ |
1922 OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(dst, temp, src, stride, SIZE, stride);\ | |
2754 | 1923 }\ |
1924 \ | |
1925 static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
6334 | 1926 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE]);\ |
1927 uint8_t * const halfHV= temp;\ | |
1928 int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\ | |
5809
7301ea0ae221
Fix intended order of operations for 4 assert() checks.
cehoyos
parents:
4939
diff
changeset
|
1929 assert(((int)temp & 7) == 0);\ |
6334 | 1930 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\ |
3156 | 1931 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfHV, stride, SIZE);\ |
2754 | 1932 }\ |
1933 \ | |
1934 static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
6334 | 1935 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE]);\ |
1936 uint8_t * const halfHV= temp;\ | |
1937 int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\ | |
5809
7301ea0ae221
Fix intended order of operations for 4 assert() checks.
cehoyos
parents:
4939
diff
changeset
|
1938 assert(((int)temp & 7) == 0);\ |
6334 | 1939 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\ |
3156 | 1940 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfHV, stride, SIZE);\ |
2754 | 1941 }\ |
1942 \ | |
1943 static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
6334 | 1944 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE]);\ |
1945 uint8_t * const halfHV= temp;\ | |
1946 int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\ | |
5809
7301ea0ae221
Fix intended order of operations for 4 assert() checks.
cehoyos
parents:
4939
diff
changeset
|
1947 assert(((int)temp & 7) == 0);\ |
3095 | 1948 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\ |
1949 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+2, halfHV, stride, SIZE, SIZE);\ | |
2754 | 1950 }\ |
1951 \ | |
1952 static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\ | |
6334 | 1953 DECLARE_ALIGNED(ALIGN, uint8_t, temp[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE]);\ |
1954 uint8_t * const halfHV= temp;\ | |
1955 int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\ | |
5809
7301ea0ae221
Fix intended order of operations for 4 assert() checks.
cehoyos
parents:
4939
diff
changeset
|
1956 assert(((int)temp & 7) == 0);\ |
3095 | 1957 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\ |
1958 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+3, halfHV, stride, SIZE, SIZE);\ | |
2754 | 1959 }\ |
1960 | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1961 #define H264_MC_4816(MMX)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1962 H264_MC(put_, 4, MMX, 8)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1963 H264_MC(put_, 8, MMX, 8)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1964 H264_MC(put_, 16,MMX, 8)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1965 H264_MC(avg_, 4, MMX, 8)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1966 H264_MC(avg_, 8, MMX, 8)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1967 H264_MC(avg_, 16,MMX, 8)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1968 |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1969 #define H264_MC_816(QPEL, XMM)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1970 QPEL(put_, 8, XMM, 16)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1971 QPEL(put_, 16,XMM, 16)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1972 QPEL(avg_, 8, XMM, 16)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1973 QPEL(avg_, 16,XMM, 16)\ |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1974 |
2754 | 1975 |
1976 #define AVG_3DNOW_OP(a,b,temp, size) \ | |
2979 | 1977 "mov" #size " " #b ", " #temp " \n\t"\ |
1978 "pavgusb " #temp ", " #a " \n\t"\ | |
1979 "mov" #size " " #a ", " #b " \n\t" | |
2754 | 1980 #define AVG_MMX2_OP(a,b,temp, size) \ |
2979 | 1981 "mov" #size " " #b ", " #temp " \n\t"\ |
1982 "pavgb " #temp ", " #a " \n\t"\ | |
1983 "mov" #size " " #a ", " #b " \n\t" | |
2754 | 1984 |
3095 | 1985 #define PAVGB "pavgusb" |
2754 | 1986 QPEL_H264(put_, PUT_OP, 3dnow) |
1987 QPEL_H264(avg_, AVG_3DNOW_OP, 3dnow) | |
3095 | 1988 #undef PAVGB |
1989 #define PAVGB "pavgb" | |
2754 | 1990 QPEL_H264(put_, PUT_OP, mmx2) |
1991 QPEL_H264(avg_, AVG_MMX2_OP, mmx2) | |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1992 QPEL_H264_V_XMM(put_, PUT_OP, sse2) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1993 QPEL_H264_V_XMM(avg_, AVG_MMX2_OP, sse2) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1994 QPEL_H264_HV_XMM(put_, PUT_OP, sse2) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1995 QPEL_H264_HV_XMM(avg_, AVG_MMX2_OP, sse2) |
6331 | 1996 #ifdef HAVE_SSSE3 |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1997 QPEL_H264_H_XMM(put_, PUT_OP, ssse3) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1998 QPEL_H264_H_XMM(avg_, AVG_MMX2_OP, ssse3) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
1999 QPEL_H264_HV2_XMM(put_, PUT_OP, ssse3) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2000 QPEL_H264_HV2_XMM(avg_, AVG_MMX2_OP, ssse3) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2001 QPEL_H264_HV_XMM(put_, PUT_OP, ssse3) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2002 QPEL_H264_HV_XMM(avg_, AVG_MMX2_OP, ssse3) |
6331 | 2003 #endif |
3095 | 2004 #undef PAVGB |
2754 | 2005 |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2006 H264_MC_4816(3dnow) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2007 H264_MC_4816(mmx2) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2008 H264_MC_816(H264_MC_V, sse2) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2009 H264_MC_816(H264_MC_HV, sse2) |
6331 | 2010 #ifdef HAVE_SSSE3 |
6336
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2011 H264_MC_816(H264_MC_H, ssse3) |
ef3fb5a7e275
sse2 h264 motion compensation. not new code, just separate out the cases that didn't need ssse3.
lorenm
parents:
6335
diff
changeset
|
2012 H264_MC_816(H264_MC_HV, ssse3) |
6331 | 2013 #endif |
2014 | |
2754 | 2015 |
2016 #define H264_CHROMA_OP(S,D) | |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2017 #define H264_CHROMA_OP4(S,D,T) |
2754 | 2018 #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
|
2019 #define H264_CHROMA_MC4_TMPL put_h264_chroma_mc4_mmx |
3213 | 2020 #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
|
2021 #define H264_CHROMA_MC8_MV0 put_pixels8_mmx |
2754 | 2022 #include "dsputil_h264_template_mmx.c" |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2023 |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2024 static void put_h264_chroma_mc8_mmx_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y) |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2025 { |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2026 put_h264_chroma_mc8_mmx(dst, src, stride, h, x, y, 1); |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2027 } |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2028 static void put_h264_chroma_mc8_mmx_nornd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y) |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2029 { |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2030 put_h264_chroma_mc8_mmx(dst, src, stride, h, x, y, 0); |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2031 } |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2032 |
2754 | 2033 #undef H264_CHROMA_OP |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2034 #undef H264_CHROMA_OP4 |
2754 | 2035 #undef H264_CHROMA_MC8_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2036 #undef H264_CHROMA_MC4_TMPL |
3213 | 2037 #undef H264_CHROMA_MC2_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2038 #undef H264_CHROMA_MC8_MV0 |
2754 | 2039 |
2040 #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
|
2041 #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
|
2042 "pavgb " #T ", " #D " \n\t" |
2754 | 2043 #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
|
2044 #define H264_CHROMA_MC4_TMPL avg_h264_chroma_mc4_mmx2 |
3213 | 2045 #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
|
2046 #define H264_CHROMA_MC8_MV0 avg_pixels8_mmx2 |
2754 | 2047 #include "dsputil_h264_template_mmx.c" |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2048 static void avg_h264_chroma_mc8_mmx2_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y) |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2049 { |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2050 avg_h264_chroma_mc8_mmx2(dst, src, stride, h, x, y, 1); |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2051 } |
2754 | 2052 #undef H264_CHROMA_OP |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2053 #undef H264_CHROMA_OP4 |
2754 | 2054 #undef H264_CHROMA_MC8_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2055 #undef H264_CHROMA_MC4_TMPL |
3213 | 2056 #undef H264_CHROMA_MC2_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2057 #undef H264_CHROMA_MC8_MV0 |
2754 | 2058 |
2059 #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
|
2060 #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
|
2061 "pavgusb " #T ", " #D " \n\t" |
2754 | 2062 #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
|
2063 #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
|
2064 #define H264_CHROMA_MC8_MV0 avg_pixels8_3dnow |
2754 | 2065 #include "dsputil_h264_template_mmx.c" |
6057
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2066 static void avg_h264_chroma_mc8_3dnow_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y) |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2067 { |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2068 avg_h264_chroma_mc8_3dnow(dst, src, stride, h, x, y, 1); |
03febc8f506f
add MMX version for put_no_rnd_h264_chroma_mc8_c, used in VC-1 decoding.
gpoirier
parents:
5947
diff
changeset
|
2069 } |
2754 | 2070 #undef H264_CHROMA_OP |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2071 #undef H264_CHROMA_OP4 |
2754 | 2072 #undef H264_CHROMA_MC8_TMPL |
2922
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2073 #undef H264_CHROMA_MC4_TMPL |
d772011258ec
faster h264_chroma_mc8_mmx, added h264_chroma_mc4_mmx.
lorenm
parents:
2902
diff
changeset
|
2074 #undef H264_CHROMA_MC8_MV0 |
2754 | 2075 |
6557 | 2076 #ifdef HAVE_SSSE3 |
2077 #define AVG_OP(X) | |
2078 #undef H264_CHROMA_MC8_TMPL | |
2079 #undef H264_CHROMA_MC4_TMPL | |
2080 #define H264_CHROMA_MC8_TMPL put_h264_chroma_mc8_ssse3 | |
2081 #define H264_CHROMA_MC4_TMPL put_h264_chroma_mc4_ssse3 | |
2082 #define H264_CHROMA_MC8_MV0 put_pixels8_mmx | |
2083 #include "dsputil_h264_template_ssse3.c" | |
2084 static void put_h264_chroma_mc8_ssse3_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y) | |
2085 { | |
2086 put_h264_chroma_mc8_ssse3(dst, src, stride, h, x, y, 1); | |
2087 } | |
2088 static void put_h264_chroma_mc8_ssse3_nornd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y) | |
2089 { | |
2090 put_h264_chroma_mc8_ssse3(dst, src, stride, h, x, y, 0); | |
2091 } | |
2092 | |
2093 #undef AVG_OP | |
2094 #undef H264_CHROMA_MC8_TMPL | |
2095 #undef H264_CHROMA_MC4_TMPL | |
2096 #undef H264_CHROMA_MC8_MV0 | |
2097 #define AVG_OP(X) X | |
2098 #define H264_CHROMA_MC8_TMPL avg_h264_chroma_mc8_ssse3 | |
2099 #define H264_CHROMA_MC4_TMPL avg_h264_chroma_mc4_ssse3 | |
2100 #define H264_CHROMA_MC8_MV0 avg_pixels8_mmx2 | |
2101 #include "dsputil_h264_template_ssse3.c" | |
2102 static void avg_h264_chroma_mc8_ssse3_rnd(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y) | |
2103 { | |
2104 avg_h264_chroma_mc8_ssse3(dst, src, stride, h, x, y, 1); | |
2105 } | |
2106 #undef AVG_OP | |
2107 #undef H264_CHROMA_MC8_TMPL | |
2108 #undef H264_CHROMA_MC4_TMPL | |
2109 #undef H264_CHROMA_MC8_MV0 | |
2110 #endif | |
2111 | |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2112 /***********************************/ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2113 /* weighted prediction */ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2114 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2115 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
|
2116 { |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2117 int x, y; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2118 offset <<= log2_denom; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2119 offset += (1 << log2_denom) >> 1; |
8031 | 2120 __asm__ volatile( |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2121 "movd %0, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2122 "movd %1, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2123 "movd %2, %%mm6 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2124 "pshufw $0, %%mm4, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2125 "pshufw $0, %%mm5, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2126 "pxor %%mm7, %%mm7 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2127 :: "g"(weight), "g"(offset), "g"(log2_denom) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2128 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2129 for(y=0; y<h; y+=2){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2130 for(x=0; x<w; x+=4){ |
8031 | 2131 __asm__ volatile( |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2132 "movd %0, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2133 "movd %1, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2134 "punpcklbw %%mm7, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2135 "punpcklbw %%mm7, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2136 "pmullw %%mm4, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2137 "pmullw %%mm4, %%mm1 \n\t" |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
2138 "paddsw %%mm5, %%mm0 \n\t" |
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
2139 "paddsw %%mm5, %%mm1 \n\t" |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2140 "psraw %%mm6, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2141 "psraw %%mm6, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2142 "packuswb %%mm7, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2143 "packuswb %%mm7, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2144 "movd %%mm0, %0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2145 "movd %%mm1, %1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2146 : "+m"(*(uint32_t*)(dst+x)), |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2147 "+m"(*(uint32_t*)(dst+x+stride)) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2148 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2149 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2150 dst += 2*stride; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2151 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2152 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2153 |
3029 | 2154 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
|
2155 { |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2156 int x, y; |
3029 | 2157 offset = ((offset + 1) | 1) << log2_denom; |
8031 | 2158 __asm__ volatile( |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2159 "movd %0, %%mm3 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2160 "movd %1, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2161 "movd %2, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2162 "movd %3, %%mm6 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2163 "pshufw $0, %%mm3, %%mm3 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2164 "pshufw $0, %%mm4, %%mm4 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2165 "pshufw $0, %%mm5, %%mm5 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2166 "pxor %%mm7, %%mm7 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2167 :: "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
|
2168 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2169 for(y=0; y<h; y++){ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2170 for(x=0; x<w; x+=4){ |
8031 | 2171 __asm__ volatile( |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2172 "movd %0, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2173 "movd %1, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2174 "punpcklbw %%mm7, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2175 "punpcklbw %%mm7, %%mm1 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2176 "pmullw %%mm3, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2177 "pmullw %%mm4, %%mm1 \n\t" |
3001
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
2178 "paddsw %%mm1, %%mm0 \n\t" |
b52d8ee430f6
fix some potential arithmetic overflows in pred_direct_motion() and
lorenm
parents:
2979
diff
changeset
|
2179 "paddsw %%mm5, %%mm0 \n\t" |
2902
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2180 "psraw %%mm6, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2181 "packuswb %%mm0, %%mm0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2182 "movd %%mm0, %0 \n\t" |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2183 : "+m"(*(uint32_t*)(dst+x)) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2184 : "m"(*(uint32_t*)(src+x)) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2185 ); |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2186 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2187 src += stride; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2188 dst += stride; |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2189 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2190 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2191 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2192 #define H264_WEIGHT(W,H) \ |
3029 | 2193 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){ \ |
2194 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
|
2195 } \ |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2196 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
|
2197 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
|
2198 } |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2199 |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2200 H264_WEIGHT(16,16) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2201 H264_WEIGHT(16, 8) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2202 H264_WEIGHT( 8,16) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2203 H264_WEIGHT( 8, 8) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2204 H264_WEIGHT( 8, 4) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2205 H264_WEIGHT( 4, 8) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2206 H264_WEIGHT( 4, 4) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2207 H264_WEIGHT( 4, 2) |
3c79bc9f3aa9
h264 mmx weighted prediction. up to 3% overall speedup.
lorenm
parents:
2855
diff
changeset
|
2208 |