Mercurial > libavcodec.hg
annotate i386/snowdsp_mmx.c @ 5554:a6475d1a9ea0 libavcodec
fixme note
author | michael |
---|---|
date | Mon, 20 Aug 2007 22:34:42 +0000 |
parents | 3c0c9723ed3c |
children | 0790af6d0028 |
rev | line source |
---|---|
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
1 /* |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
2 * MMX and SSE2 optimized snow DSP utils |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
3 * Copyright (c) 2005-2006 Robert Edele <yartrebo@earthlink.net> |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
4 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3566
diff
changeset
|
5 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3566
diff
changeset
|
6 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3566
diff
changeset
|
7 * FFmpeg is free software; you can redistribute it and/or |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
8 * modify it under the terms of the GNU Lesser General Public |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
9 * 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:
3566
diff
changeset
|
10 * version 2.1 of the License, or (at your option) any later version. |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
11 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3566
diff
changeset
|
12 * FFmpeg is distributed in the hope that it will be useful, |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
15 * Lesser General Public License for more details. |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
16 * |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
17 * 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:
3566
diff
changeset
|
18 * License along with FFmpeg; if not, write to the Free Software |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
20 */ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
21 |
5010
d5ba514e3f4a
Add libavcodec to compiler include flags in order to simplify header
diego
parents:
4436
diff
changeset
|
22 #include "avcodec.h" |
d5ba514e3f4a
Add libavcodec to compiler include flags in order to simplify header
diego
parents:
4436
diff
changeset
|
23 #include "snow.h" |
3398
e0927bc44a10
Move REG_* macros from libavcodec/i386/mmx.h to libavutil/x86_cpu.h
lucabe
parents:
3265
diff
changeset
|
24 #include "x86_cpu.h" |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
25 |
3210 | 26 void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width){ |
27 const int w2= (width+1)>>1; | |
28 // SSE2 code runs faster with pointers aligned on a 32-byte boundary. | |
29 DWTELEM temp_buf[(width>>1) + 4]; | |
30 DWTELEM * const temp = temp_buf + 4 - (((int)temp_buf & 0xF) >> 2); | |
31 const int w_l= (width>>1); | |
32 const int w_r= w2 - 1; | |
33 int i; | |
34 | |
35 { // Lift 0 | |
36 DWTELEM * const ref = b + w2 - 1; | |
37 DWTELEM b_0 = b[0]; //By allowing the first entry in b[0] to be calculated twice | |
38 // (the first time erroneously), we allow the SSE2 code to run an extra pass. | |
39 // The savings in code and time are well worth having to store this value and | |
40 // calculate b[0] correctly afterwards. | |
41 | |
42 i = 0; | |
43 asm volatile( | |
44 "pcmpeqd %%xmm7, %%xmm7 \n\t" | |
45 "pslld $31, %%xmm7 \n\t" | |
46 "psrld $29, %%xmm7 \n\t" | |
47 ::); | |
48 for(; i<w_l-7; i+=8){ | |
49 asm volatile( | |
50 "movdqu (%1), %%xmm1 \n\t" | |
51 "movdqu 16(%1), %%xmm5 \n\t" | |
52 "movdqu 4(%1), %%xmm2 \n\t" | |
53 "movdqu 20(%1), %%xmm6 \n\t" | |
54 "paddd %%xmm1, %%xmm2 \n\t" | |
55 "paddd %%xmm5, %%xmm6 \n\t" | |
56 "movdqa %%xmm2, %%xmm0 \n\t" | |
57 "movdqa %%xmm6, %%xmm4 \n\t" | |
58 "paddd %%xmm2, %%xmm2 \n\t" | |
59 "paddd %%xmm6, %%xmm6 \n\t" | |
60 "paddd %%xmm0, %%xmm2 \n\t" | |
61 "paddd %%xmm4, %%xmm6 \n\t" | |
62 "paddd %%xmm7, %%xmm2 \n\t" | |
63 "paddd %%xmm7, %%xmm6 \n\t" | |
64 "psrad $3, %%xmm2 \n\t" | |
65 "psrad $3, %%xmm6 \n\t" | |
66 "movdqa (%0), %%xmm0 \n\t" | |
67 "movdqa 16(%0), %%xmm4 \n\t" | |
68 "psubd %%xmm2, %%xmm0 \n\t" | |
69 "psubd %%xmm6, %%xmm4 \n\t" | |
70 "movdqa %%xmm0, (%0) \n\t" | |
71 "movdqa %%xmm4, 16(%0) \n\t" | |
72 :: "r"(&b[i]), "r"(&ref[i]) | |
73 : "memory" | |
74 ); | |
75 } | |
76 snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS); | |
77 b[0] = b_0 - ((W_DM * 2 * ref[1]+W_DO)>>W_DS); | |
78 } | |
79 | |
80 { // Lift 1 | |
81 DWTELEM * const dst = b+w2; | |
82 | |
83 i = 0; | |
84 for(; (((long)&dst[i]) & 0xF) && i<w_r; i++){ | |
85 dst[i] = dst[i] - (b[i] + b[i + 1]); | |
86 } | |
87 for(; i<w_r-7; i+=8){ | |
88 asm volatile( | |
89 "movdqu (%1), %%xmm1 \n\t" | |
90 "movdqu 16(%1), %%xmm5 \n\t" | |
91 "movdqu 4(%1), %%xmm2 \n\t" | |
92 "movdqu 20(%1), %%xmm6 \n\t" | |
93 "paddd %%xmm1, %%xmm2 \n\t" | |
94 "paddd %%xmm5, %%xmm6 \n\t" | |
95 "movdqa (%0), %%xmm0 \n\t" | |
96 "movdqa 16(%0), %%xmm4 \n\t" | |
97 "psubd %%xmm2, %%xmm0 \n\t" | |
98 "psubd %%xmm6, %%xmm4 \n\t" | |
99 "movdqa %%xmm0, (%0) \n\t" | |
100 "movdqa %%xmm4, 16(%0) \n\t" | |
101 :: "r"(&dst[i]), "r"(&b[i]) | |
102 : "memory" | |
103 ); | |
104 } | |
105 snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS); | |
106 } | |
107 | |
108 { // Lift 2 | |
109 DWTELEM * const ref = b+w2 - 1; | |
110 DWTELEM b_0 = b[0]; | |
111 | |
112 i = 0; | |
113 asm volatile( | |
114 "pslld $1, %%xmm7 \n\t" /* xmm7 already holds a '4' from 2 lifts ago. */ | |
115 ::); | |
116 for(; i<w_l-7; i+=8){ | |
117 asm volatile( | |
118 "movdqu (%1), %%xmm1 \n\t" | |
119 "movdqu 16(%1), %%xmm5 \n\t" | |
120 "movdqu 4(%1), %%xmm0 \n\t" | |
5554 | 121 "movdqu 20(%1), %%xmm4 \n\t" //FIXME try aligned reads and shifts |
3210 | 122 "paddd %%xmm1, %%xmm0 \n\t" |
123 "paddd %%xmm5, %%xmm4 \n\t" | |
124 "movdqa %%xmm7, %%xmm1 \n\t" | |
125 "movdqa %%xmm7, %%xmm5 \n\t" | |
126 "psubd %%xmm0, %%xmm1 \n\t" | |
127 "psubd %%xmm4, %%xmm5 \n\t" | |
128 "movdqa (%0), %%xmm0 \n\t" | |
129 "movdqa 16(%0), %%xmm4 \n\t" | |
130 "pslld $2, %%xmm0 \n\t" | |
131 "pslld $2, %%xmm4 \n\t" | |
132 "psubd %%xmm0, %%xmm1 \n\t" | |
133 "psubd %%xmm4, %%xmm5 \n\t" | |
134 "psrad $4, %%xmm1 \n\t" | |
135 "psrad $4, %%xmm5 \n\t" | |
136 "movdqa (%0), %%xmm0 \n\t" | |
137 "movdqa 16(%0), %%xmm4 \n\t" | |
138 "psubd %%xmm1, %%xmm0 \n\t" | |
139 "psubd %%xmm5, %%xmm4 \n\t" | |
140 "movdqa %%xmm0, (%0) \n\t" | |
141 "movdqa %%xmm4, 16(%0) \n\t" | |
142 :: "r"(&b[i]), "r"(&ref[i]) | |
143 : "memory" | |
144 ); | |
145 } | |
146 snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l); | |
147 b[0] = b_0 - (((-2 * ref[1] + W_BO) - 4 * b_0) >> W_BS); | |
148 } | |
149 | |
150 { // Lift 3 | |
151 DWTELEM * const src = b+w2; | |
152 | |
153 i = 0; | |
154 for(; (((long)&temp[i]) & 0xF) && i<w_r; i++){ | |
155 temp[i] = src[i] - ((-W_AM*(b[i] + b[i+1]))>>W_AS); | |
156 } | |
157 for(; i<w_r-7; i+=8){ | |
158 asm volatile( | |
159 "movdqu 4(%1), %%xmm2 \n\t" | |
160 "movdqu 20(%1), %%xmm6 \n\t" | |
161 "paddd (%1), %%xmm2 \n\t" | |
162 "paddd 16(%1), %%xmm6 \n\t" | |
163 "movdqa %%xmm2, %%xmm0 \n\t" | |
164 "movdqa %%xmm6, %%xmm4 \n\t" | |
165 "pslld $2, %%xmm2 \n\t" | |
166 "pslld $2, %%xmm6 \n\t" | |
167 "psubd %%xmm2, %%xmm0 \n\t" | |
168 "psubd %%xmm6, %%xmm4 \n\t" | |
169 "psrad $1, %%xmm0 \n\t" | |
170 "psrad $1, %%xmm4 \n\t" | |
171 "movdqu (%0), %%xmm2 \n\t" | |
172 "movdqu 16(%0), %%xmm6 \n\t" | |
173 "psubd %%xmm0, %%xmm2 \n\t" | |
174 "psubd %%xmm4, %%xmm6 \n\t" | |
175 "movdqa %%xmm2, (%2) \n\t" | |
176 "movdqa %%xmm6, 16(%2) \n\t" | |
177 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i]) | |
178 : "memory" | |
179 ); | |
180 } | |
181 snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS); | |
182 } | |
183 | |
184 { | |
185 snow_interleave_line_header(&i, width, b, temp); | |
186 | |
187 for (; (i & 0x1E) != 0x1E; i-=2){ | |
188 b[i+1] = temp[i>>1]; | |
189 b[i] = b[i>>1]; | |
190 } | |
191 for (i-=30; i>=0; i-=32){ | |
192 asm volatile( | |
193 "movdqa (%1), %%xmm0 \n\t" | |
194 "movdqa 16(%1), %%xmm2 \n\t" | |
195 "movdqa 32(%1), %%xmm4 \n\t" | |
196 "movdqa 48(%1), %%xmm6 \n\t" | |
197 "movdqa (%1), %%xmm1 \n\t" | |
198 "movdqa 16(%1), %%xmm3 \n\t" | |
199 "movdqa 32(%1), %%xmm5 \n\t" | |
200 "movdqa 48(%1), %%xmm7 \n\t" | |
201 "punpckldq (%2), %%xmm0 \n\t" | |
202 "punpckldq 16(%2), %%xmm2 \n\t" | |
203 "punpckldq 32(%2), %%xmm4 \n\t" | |
204 "punpckldq 48(%2), %%xmm6 \n\t" | |
205 "movdqa %%xmm0, (%0) \n\t" | |
206 "movdqa %%xmm2, 32(%0) \n\t" | |
207 "movdqa %%xmm4, 64(%0) \n\t" | |
208 "movdqa %%xmm6, 96(%0) \n\t" | |
209 "punpckhdq (%2), %%xmm1 \n\t" | |
210 "punpckhdq 16(%2), %%xmm3 \n\t" | |
211 "punpckhdq 32(%2), %%xmm5 \n\t" | |
212 "punpckhdq 48(%2), %%xmm7 \n\t" | |
213 "movdqa %%xmm1, 16(%0) \n\t" | |
214 "movdqa %%xmm3, 48(%0) \n\t" | |
215 "movdqa %%xmm5, 80(%0) \n\t" | |
216 "movdqa %%xmm7, 112(%0) \n\t" | |
217 :: "r"(&(b)[i]), "r"(&(b)[i>>1]), "r"(&(temp)[i>>1]) | |
218 : "memory" | |
219 ); | |
220 } | |
221 } | |
222 } | |
223 | |
224 void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width){ | |
225 const int w2= (width+1)>>1; | |
226 DWTELEM temp[width >> 1]; | |
227 const int w_l= (width>>1); | |
228 const int w_r= w2 - 1; | |
229 int i; | |
230 | |
231 { // Lift 0 | |
232 DWTELEM * const ref = b + w2 - 1; | |
233 | |
234 i = 1; | |
235 b[0] = b[0] - ((W_DM * 2 * ref[1]+W_DO)>>W_DS); | |
236 asm volatile( | |
237 "pcmpeqd %%mm7, %%mm7 \n\t" | |
238 "pslld $31, %%mm7 \n\t" | |
239 "psrld $29, %%mm7 \n\t" | |
240 ::); | |
241 for(; i<w_l-3; i+=4){ | |
242 asm volatile( | |
243 "movq (%1), %%mm2 \n\t" | |
244 "movq 8(%1), %%mm6 \n\t" | |
245 "paddd 4(%1), %%mm2 \n\t" | |
246 "paddd 12(%1), %%mm6 \n\t" | |
247 "movq %%mm2, %%mm0 \n\t" | |
248 "movq %%mm6, %%mm4 \n\t" | |
249 "paddd %%mm2, %%mm2 \n\t" | |
250 "paddd %%mm6, %%mm6 \n\t" | |
251 "paddd %%mm0, %%mm2 \n\t" | |
252 "paddd %%mm4, %%mm6 \n\t" | |
253 "paddd %%mm7, %%mm2 \n\t" | |
254 "paddd %%mm7, %%mm6 \n\t" | |
255 "psrad $3, %%mm2 \n\t" | |
256 "psrad $3, %%mm6 \n\t" | |
257 "movq (%0), %%mm0 \n\t" | |
258 "movq 8(%0), %%mm4 \n\t" | |
259 "psubd %%mm2, %%mm0 \n\t" | |
260 "psubd %%mm6, %%mm4 \n\t" | |
261 "movq %%mm0, (%0) \n\t" | |
262 "movq %%mm4, 8(%0) \n\t" | |
263 :: "r"(&b[i]), "r"(&ref[i]) | |
264 : "memory" | |
265 ); | |
266 } | |
267 snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS); | |
268 } | |
269 | |
270 { // Lift 1 | |
271 DWTELEM * const dst = b+w2; | |
272 | |
273 i = 0; | |
274 for(; i<w_r-3; i+=4){ | |
275 asm volatile( | |
276 "movq (%1), %%mm2 \n\t" | |
277 "movq 8(%1), %%mm6 \n\t" | |
278 "paddd 4(%1), %%mm2 \n\t" | |
279 "paddd 12(%1), %%mm6 \n\t" | |
280 "movq (%0), %%mm0 \n\t" | |
281 "movq 8(%0), %%mm4 \n\t" | |
282 "psubd %%mm2, %%mm0 \n\t" | |
283 "psubd %%mm6, %%mm4 \n\t" | |
284 "movq %%mm0, (%0) \n\t" | |
285 "movq %%mm4, 8(%0) \n\t" | |
286 :: "r"(&dst[i]), "r"(&b[i]) | |
287 : "memory" | |
288 ); | |
289 } | |
290 snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS); | |
291 } | |
292 | |
293 { // Lift 2 | |
294 DWTELEM * const ref = b+w2 - 1; | |
295 | |
296 i = 1; | |
5553
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
297 b[0] = b[0] + (((2 * ref[1] + W_BO-1) + 4 * b[0]) >> W_BS); |
3210 | 298 asm volatile( |
5553
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
299 "pcmpeqd %%mm7, %%mm7 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
300 "psrld $29, %%mm7 \n\t" |
3210 | 301 ::); |
302 for(; i<w_l-3; i+=4){ | |
303 asm volatile( | |
304 "movq (%1), %%mm0 \n\t" | |
305 "movq 8(%1), %%mm4 \n\t" | |
306 "paddd 4(%1), %%mm0 \n\t" | |
307 "paddd 12(%1), %%mm4 \n\t" | |
5553
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
308 "paddd %%mm7, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
309 "paddd %%mm7, %%mm4 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
310 "psrad $2, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
311 "psrad $2, %%mm4 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
312 "movq (%0), %%mm1 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
313 "movq 8(%0), %%mm5 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
314 "paddd %%mm1, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
315 "paddd %%mm5, %%mm4 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
316 "psrad $2, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
317 "psrad $2, %%mm4 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
318 "paddd %%mm1, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
319 "paddd %%mm5, %%mm4 \n\t" |
3210 | 320 "movq %%mm0, (%0) \n\t" |
321 "movq %%mm4, 8(%0) \n\t" | |
322 :: "r"(&b[i]), "r"(&ref[i]) | |
323 : "memory" | |
324 ); | |
325 } | |
326 snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l); | |
327 } | |
328 | |
329 { // Lift 3 | |
330 DWTELEM * const src = b+w2; | |
331 i = 0; | |
332 | |
333 for(; i<w_r-3; i+=4){ | |
334 asm volatile( | |
335 "movq 4(%1), %%mm2 \n\t" | |
336 "movq 12(%1), %%mm6 \n\t" | |
337 "paddd (%1), %%mm2 \n\t" | |
338 "paddd 8(%1), %%mm6 \n\t" | |
339 "movq %%mm2, %%mm0 \n\t" | |
340 "movq %%mm6, %%mm4 \n\t" | |
341 "pslld $2, %%mm2 \n\t" | |
342 "pslld $2, %%mm6 \n\t" | |
343 "psubd %%mm2, %%mm0 \n\t" | |
344 "psubd %%mm6, %%mm4 \n\t" | |
345 "psrad $1, %%mm0 \n\t" | |
346 "psrad $1, %%mm4 \n\t" | |
347 "movq (%0), %%mm2 \n\t" | |
348 "movq 8(%0), %%mm6 \n\t" | |
349 "psubd %%mm0, %%mm2 \n\t" | |
350 "psubd %%mm4, %%mm6 \n\t" | |
351 "movq %%mm2, (%2) \n\t" | |
352 "movq %%mm6, 8(%2) \n\t" | |
353 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i]) | |
354 : "memory" | |
355 ); | |
356 } | |
357 snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS); | |
358 } | |
359 | |
360 { | |
361 snow_interleave_line_header(&i, width, b, temp); | |
362 | |
363 for (; (i & 0xE) != 0xE; i-=2){ | |
364 b[i+1] = temp[i>>1]; | |
365 b[i] = b[i>>1]; | |
366 } | |
367 for (i-=14; i>=0; i-=16){ | |
368 asm volatile( | |
369 "movq (%1), %%mm0 \n\t" | |
370 "movq 8(%1), %%mm2 \n\t" | |
371 "movq 16(%1), %%mm4 \n\t" | |
372 "movq 24(%1), %%mm6 \n\t" | |
373 "movq (%1), %%mm1 \n\t" | |
374 "movq 8(%1), %%mm3 \n\t" | |
375 "movq 16(%1), %%mm5 \n\t" | |
376 "movq 24(%1), %%mm7 \n\t" | |
377 "punpckldq (%2), %%mm0 \n\t" | |
378 "punpckldq 8(%2), %%mm2 \n\t" | |
379 "punpckldq 16(%2), %%mm4 \n\t" | |
380 "punpckldq 24(%2), %%mm6 \n\t" | |
381 "movq %%mm0, (%0) \n\t" | |
382 "movq %%mm2, 16(%0) \n\t" | |
383 "movq %%mm4, 32(%0) \n\t" | |
384 "movq %%mm6, 48(%0) \n\t" | |
385 "punpckhdq (%2), %%mm1 \n\t" | |
386 "punpckhdq 8(%2), %%mm3 \n\t" | |
387 "punpckhdq 16(%2), %%mm5 \n\t" | |
388 "punpckhdq 24(%2), %%mm7 \n\t" | |
389 "movq %%mm1, 8(%0) \n\t" | |
390 "movq %%mm3, 24(%0) \n\t" | |
391 "movq %%mm5, 40(%0) \n\t" | |
392 "movq %%mm7, 56(%0) \n\t" | |
393 :: "r"(&b[i]), "r"(&b[i>>1]), "r"(&temp[i>>1]) | |
394 : "memory" | |
395 ); | |
396 } | |
397 } | |
398 } | |
399 | |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
400 #define snow_vertical_compose_sse2_load_add(op,r,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
401 ""op" (%%"r",%%"REG_d",4), %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
402 ""op" 16(%%"r",%%"REG_d",4), %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
403 ""op" 32(%%"r",%%"REG_d",4), %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
404 ""op" 48(%%"r",%%"REG_d",4), %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
405 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
406 #define snow_vertical_compose_sse2_load(r,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
407 snow_vertical_compose_sse2_load_add("movdqa",r,t0,t1,t2,t3) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
408 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
409 #define snow_vertical_compose_sse2_add(r,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
410 snow_vertical_compose_sse2_load_add("paddd",r,t0,t1,t2,t3) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
411 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
412 #define snow_vertical_compose_sse2_sub(s0,s1,s2,s3,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
413 "psubd %%"s0", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
414 "psubd %%"s1", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
415 "psubd %%"s2", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
416 "psubd %%"s3", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
417 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
418 #define snow_vertical_compose_sse2_store(w,s0,s1,s2,s3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
419 "movdqa %%"s0", (%%"w",%%"REG_d",4) \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
420 "movdqa %%"s1", 16(%%"w",%%"REG_d",4) \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
421 "movdqa %%"s2", 32(%%"w",%%"REG_d",4) \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
422 "movdqa %%"s3", 48(%%"w",%%"REG_d",4) \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
423 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
424 #define snow_vertical_compose_sse2_sra(n,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
425 "psrad $"n", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
426 "psrad $"n", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
427 "psrad $"n", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
428 "psrad $"n", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
429 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
430 #define snow_vertical_compose_sse2_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
431 "paddd %%"s0", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
432 "paddd %%"s1", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
433 "paddd %%"s2", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
434 "paddd %%"s3", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
435 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
436 #define snow_vertical_compose_sse2_sll(n,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
437 "pslld $"n", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
438 "pslld $"n", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
439 "pslld $"n", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
440 "pslld $"n", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
441 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
442 #define snow_vertical_compose_sse2_move(s0,s1,s2,s3,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
443 "movdqa %%"s0", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
444 "movdqa %%"s1", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
445 "movdqa %%"s2", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
446 "movdqa %%"s3", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
447 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
448 void ff_snow_vertical_compose97i_sse2(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width){ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
449 long i = width; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
450 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
451 while(i & 0xF) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
452 { |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
453 i--; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
454 b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
455 b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
456 b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
457 b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
458 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
459 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
460 asm volatile ( |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
461 "jmp 2f \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
462 "1: \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
463 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
464 "mov %6, %%"REG_a" \n\t" |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
465 "mov %4, %%"REG_S" \n\t" |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
466 |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
467 snow_vertical_compose_sse2_load(REG_S,"xmm0","xmm2","xmm4","xmm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
468 snow_vertical_compose_sse2_add(REG_a,"xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
469 snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
470 snow_vertical_compose_sse2_sll("1","xmm0","xmm2","xmm4","xmm6")\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
471 snow_vertical_compose_sse2_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
472 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
473 "pcmpeqd %%xmm1, %%xmm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
474 "pslld $31, %%xmm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
475 "psrld $29, %%xmm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
476 "mov %5, %%"REG_a" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
477 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
478 snow_vertical_compose_sse2_r2r_add("xmm1","xmm1","xmm1","xmm1","xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
479 snow_vertical_compose_sse2_sra("3","xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
480 snow_vertical_compose_sse2_load(REG_a,"xmm1","xmm3","xmm5","xmm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
481 snow_vertical_compose_sse2_sub("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
482 snow_vertical_compose_sse2_store(REG_a,"xmm1","xmm3","xmm5","xmm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
483 "mov %3, %%"REG_c" \n\t" |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
484 snow_vertical_compose_sse2_load(REG_S,"xmm0","xmm2","xmm4","xmm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
485 snow_vertical_compose_sse2_add(REG_c,"xmm1","xmm3","xmm5","xmm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
486 snow_vertical_compose_sse2_sub("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6") |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
487 snow_vertical_compose_sse2_store(REG_S,"xmm0","xmm2","xmm4","xmm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
488 "mov %2, %%"REG_a" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
489 snow_vertical_compose_sse2_load(REG_c,"xmm1","xmm3","xmm5","xmm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
490 snow_vertical_compose_sse2_add(REG_a,"xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
491 snow_vertical_compose_sse2_sll("2","xmm1","xmm3","xmm5","xmm7")\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
492 snow_vertical_compose_sse2_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
493 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
494 "pcmpeqd %%xmm1, %%xmm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
495 "pslld $31, %%xmm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
496 "psrld $28, %%xmm1 \n\t" |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
497 "mov %1, %%"REG_S" \n\t" |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
498 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
499 snow_vertical_compose_sse2_r2r_add("xmm1","xmm1","xmm1","xmm1","xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
500 snow_vertical_compose_sse2_sra("4","xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
501 snow_vertical_compose_sse2_add(REG_c,"xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
502 snow_vertical_compose_sse2_store(REG_c,"xmm0","xmm2","xmm4","xmm6") |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
503 snow_vertical_compose_sse2_add(REG_S,"xmm0","xmm2","xmm4","xmm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
504 snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
505 snow_vertical_compose_sse2_sll("1","xmm0","xmm2","xmm4","xmm6")\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
506 snow_vertical_compose_sse2_r2r_add("xmm1","xmm3","xmm5","xmm7","xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
507 snow_vertical_compose_sse2_sra("1","xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
508 snow_vertical_compose_sse2_add(REG_a,"xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
509 snow_vertical_compose_sse2_store(REG_a,"xmm0","xmm2","xmm4","xmm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
510 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
511 "2: \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
512 "sub $16, %%"REG_d" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
513 "jge 1b \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
514 :"+d"(i) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
515 : |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
516 "m"(b0),"m"(b1),"m"(b2),"m"(b3),"m"(b4),"m"(b5): |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
517 "%"REG_a"","%"REG_S"","%"REG_c""); |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
518 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
519 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
520 #define snow_vertical_compose_mmx_load_add(op,r,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
521 ""op" (%%"r",%%"REG_d",4), %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
522 ""op" 8(%%"r",%%"REG_d",4), %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
523 ""op" 16(%%"r",%%"REG_d",4), %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
524 ""op" 24(%%"r",%%"REG_d",4), %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
525 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
526 #define snow_vertical_compose_mmx_load(r,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
527 snow_vertical_compose_mmx_load_add("movq",r,t0,t1,t2,t3) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
528 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
529 #define snow_vertical_compose_mmx_add(r,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
530 snow_vertical_compose_mmx_load_add("paddd",r,t0,t1,t2,t3) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
531 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
532 #define snow_vertical_compose_mmx_sub(s0,s1,s2,s3,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
533 snow_vertical_compose_sse2_sub(s0,s1,s2,s3,t0,t1,t2,t3) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
534 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
535 #define snow_vertical_compose_mmx_store(w,s0,s1,s2,s3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
536 "movq %%"s0", (%%"w",%%"REG_d",4) \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
537 "movq %%"s1", 8(%%"w",%%"REG_d",4) \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
538 "movq %%"s2", 16(%%"w",%%"REG_d",4) \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
539 "movq %%"s3", 24(%%"w",%%"REG_d",4) \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
540 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
541 #define snow_vertical_compose_mmx_sra(n,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
542 snow_vertical_compose_sse2_sra(n,t0,t1,t2,t3) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
543 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
544 #define snow_vertical_compose_mmx_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
545 snow_vertical_compose_sse2_r2r_add(s0,s1,s2,s3,t0,t1,t2,t3) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
546 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
547 #define snow_vertical_compose_mmx_sll(n,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
548 snow_vertical_compose_sse2_sll(n,t0,t1,t2,t3) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
549 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
550 #define snow_vertical_compose_mmx_move(s0,s1,s2,s3,t0,t1,t2,t3)\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
551 "movq %%"s0", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
552 "movq %%"s1", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
553 "movq %%"s2", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
554 "movq %%"s3", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
555 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
556 void ff_snow_vertical_compose97i_mmx(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width){ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
557 long i = width; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
558 while(i & 0x7) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
559 { |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
560 i--; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
561 b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
562 b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
563 b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
564 b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
565 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
566 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
567 asm volatile( |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
568 "jmp 2f \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
569 "1: \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
570 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
571 "mov %6, %%"REG_a" \n\t" |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
572 "mov %4, %%"REG_S" \n\t" |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
573 |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
574 snow_vertical_compose_mmx_load(REG_S,"mm0","mm2","mm4","mm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
575 snow_vertical_compose_mmx_add(REG_a,"mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
576 snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
577 snow_vertical_compose_mmx_sll("1","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
578 snow_vertical_compose_mmx_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
579 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
580 "pcmpeqd %%mm1, %%mm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
581 "pslld $31, %%mm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
582 "psrld $29, %%mm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
583 "mov %5, %%"REG_a" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
584 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
585 snow_vertical_compose_mmx_r2r_add("mm1","mm1","mm1","mm1","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
586 snow_vertical_compose_mmx_sra("3","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
587 snow_vertical_compose_mmx_load(REG_a,"mm1","mm3","mm5","mm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
588 snow_vertical_compose_mmx_sub("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
589 snow_vertical_compose_mmx_store(REG_a,"mm1","mm3","mm5","mm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
590 "mov %3, %%"REG_c" \n\t" |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
591 snow_vertical_compose_mmx_load(REG_S,"mm0","mm2","mm4","mm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
592 snow_vertical_compose_mmx_add(REG_c,"mm1","mm3","mm5","mm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
593 snow_vertical_compose_mmx_sub("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6") |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
594 snow_vertical_compose_mmx_store(REG_S,"mm0","mm2","mm4","mm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
595 "mov %2, %%"REG_a" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
596 snow_vertical_compose_mmx_load(REG_c,"mm1","mm3","mm5","mm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
597 snow_vertical_compose_mmx_add(REG_a,"mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
598 snow_vertical_compose_mmx_sll("2","mm1","mm3","mm5","mm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
599 snow_vertical_compose_mmx_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
600 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
601 "pcmpeqd %%mm1, %%mm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
602 "pslld $31, %%mm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
603 "psrld $28, %%mm1 \n\t" |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
604 "mov %1, %%"REG_S" \n\t" |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
605 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
606 snow_vertical_compose_mmx_r2r_add("mm1","mm1","mm1","mm1","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
607 snow_vertical_compose_mmx_sra("4","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
608 snow_vertical_compose_mmx_add(REG_c,"mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
609 snow_vertical_compose_mmx_store(REG_c,"mm0","mm2","mm4","mm6") |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
610 snow_vertical_compose_mmx_add(REG_S,"mm0","mm2","mm4","mm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
611 snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
612 snow_vertical_compose_mmx_sll("1","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
613 snow_vertical_compose_mmx_r2r_add("mm1","mm3","mm5","mm7","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
614 snow_vertical_compose_mmx_sra("1","mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
615 snow_vertical_compose_mmx_add(REG_a,"mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
616 snow_vertical_compose_mmx_store(REG_a,"mm0","mm2","mm4","mm6") |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
617 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
618 "2: \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
619 "sub $8, %%"REG_d" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
620 "jge 1b \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
621 :"+d"(i) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
622 : |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
623 "m"(b0),"m"(b1),"m"(b2),"m"(b3),"m"(b4),"m"(b5): |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
624 "%"REG_a"","%"REG_S"","%"REG_c""); |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
625 } |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
626 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
627 #define snow_inner_add_yblock_sse2_header \ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
628 DWTELEM * * dst_array = sb->line + src_y;\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
629 long tmp;\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
630 asm volatile(\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
631 "mov %7, %%"REG_c" \n\t"\ |
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
632 "mov %6, %2 \n\t"\ |
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
633 "mov %4, %%"REG_S" \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
634 "pxor %%xmm7, %%xmm7 \n\t" /* 0 */\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
635 "pcmpeqd %%xmm3, %%xmm3 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
636 "pslld $31, %%xmm3 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
637 "psrld $24, %%xmm3 \n\t" /* FRAC_BITS >> 1 */\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
638 "1: \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
639 "mov %1, %%"REG_D" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
640 "mov (%%"REG_D"), %%"REG_D" \n\t"\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
641 "add %3, %%"REG_D" \n\t" |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
642 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
643 #define snow_inner_add_yblock_sse2_start_8(out_reg1, out_reg2, ptr_offset, s_offset)\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
644 "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
645 "movq (%%"REG_d"), %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
646 "movq (%%"REG_d", %%"REG_c"), %%"out_reg2" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
647 "punpcklbw %%xmm7, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
648 "punpcklbw %%xmm7, %%"out_reg2" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
649 "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
650 "movq "s_offset"+16(%%"REG_S"), %%xmm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
651 "punpcklbw %%xmm7, %%xmm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
652 "punpcklbw %%xmm7, %%xmm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
653 "pmullw %%xmm0, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
654 "pmullw %%xmm4, %%"out_reg2" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
655 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
656 #define snow_inner_add_yblock_sse2_start_16(out_reg1, out_reg2, ptr_offset, s_offset)\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
657 "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
658 "movq (%%"REG_d"), %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
659 "movq 8(%%"REG_d"), %%"out_reg2" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
660 "punpcklbw %%xmm7, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
661 "punpcklbw %%xmm7, %%"out_reg2" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
662 "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
663 "movq "s_offset"+8(%%"REG_S"), %%xmm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
664 "punpcklbw %%xmm7, %%xmm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
665 "punpcklbw %%xmm7, %%xmm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
666 "pmullw %%xmm0, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
667 "pmullw %%xmm4, %%"out_reg2" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
668 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
669 #define snow_inner_add_yblock_sse2_accum_8(ptr_offset, s_offset) \ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
670 snow_inner_add_yblock_sse2_start_8("xmm2", "xmm6", ptr_offset, s_offset)\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
671 "paddusw %%xmm2, %%xmm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
672 "paddusw %%xmm6, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
673 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
674 #define snow_inner_add_yblock_sse2_accum_16(ptr_offset, s_offset) \ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
675 snow_inner_add_yblock_sse2_start_16("xmm2", "xmm6", ptr_offset, s_offset)\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
676 "paddusw %%xmm2, %%xmm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
677 "paddusw %%xmm6, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
678 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
679 #define snow_inner_add_yblock_sse2_end_common1\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
680 "add $32, %%"REG_S" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
681 "add %%"REG_c", %0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
682 "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
683 "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
684 "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
685 "add %%"REG_c", (%%"REG_a") \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
686 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
687 #define snow_inner_add_yblock_sse2_end_common2\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
688 "jnz 1b \n\t"\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
689 :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
690 :\ |
3265
ba26b176dcf8
gcc-2.95 compile fix, patch by Wolfram Gloger <wmglo A dent PIS med PIS uni-muenchen PIS de>
gpoirier
parents:
3223
diff
changeset
|
691 "rm"((long)(src_x<<2)),"m"(obmc),"a"(block),"m"((long)b_h),"m"((long)src_stride):\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
692 "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d""); |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
693 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
694 #define snow_inner_add_yblock_sse2_end_8\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
695 "sal $1, %%"REG_c" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
696 "add $"PTR_SIZE"*2, %1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
697 snow_inner_add_yblock_sse2_end_common1\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
698 "sar $1, %%"REG_c" \n\t"\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
699 "sub $2, %2 \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
700 snow_inner_add_yblock_sse2_end_common2 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
701 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
702 #define snow_inner_add_yblock_sse2_end_16\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
703 "add $"PTR_SIZE"*1, %1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
704 snow_inner_add_yblock_sse2_end_common1\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
705 "dec %2 \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
706 snow_inner_add_yblock_sse2_end_common2 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
707 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
708 static void inner_add_yblock_bw_8_obmc_16_bh_even_sse2(const uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h, |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
709 int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
710 snow_inner_add_yblock_sse2_header |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
711 snow_inner_add_yblock_sse2_start_8("xmm1", "xmm5", "3", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
712 snow_inner_add_yblock_sse2_accum_8("2", "8") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
713 snow_inner_add_yblock_sse2_accum_8("1", "128") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
714 snow_inner_add_yblock_sse2_accum_8("0", "136") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
715 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
716 "mov %0, %%"REG_d" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
717 "movdqa (%%"REG_D"), %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
718 "movdqa %%xmm1, %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
719 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
720 "punpckhwd %%xmm7, %%xmm1 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
721 "punpcklwd %%xmm7, %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
722 "paddd %%xmm2, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
723 "movdqa 16(%%"REG_D"), %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
724 "paddd %%xmm1, %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
725 "paddd %%xmm3, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
726 "paddd %%xmm3, %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
727 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
728 "mov %1, %%"REG_D" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
729 "mov "PTR_SIZE"(%%"REG_D"), %%"REG_D";\n\t" |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
730 "add %3, %%"REG_D" \n\t" |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
731 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
732 "movdqa (%%"REG_D"), %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
733 "movdqa %%xmm5, %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
734 "punpckhwd %%xmm7, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
735 "punpcklwd %%xmm7, %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
736 "paddd %%xmm6, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
737 "movdqa 16(%%"REG_D"), %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
738 "paddd %%xmm5, %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
739 "paddd %%xmm3, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
740 "paddd %%xmm3, %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
741 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
742 "psrad $8, %%xmm0 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
743 "psrad $8, %%xmm2 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
744 "packssdw %%xmm2, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
745 "packuswb %%xmm7, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
746 "movq %%xmm0, (%%"REG_d") \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
747 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
748 "psrad $8, %%xmm4 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
749 "psrad $8, %%xmm6 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
750 "packssdw %%xmm6, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
751 "packuswb %%xmm7, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
752 "movq %%xmm4, (%%"REG_d",%%"REG_c");\n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
753 snow_inner_add_yblock_sse2_end_8 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
754 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
755 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
756 static void inner_add_yblock_bw_16_obmc_32_sse2(const uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h, |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
757 int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
758 snow_inner_add_yblock_sse2_header |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
759 snow_inner_add_yblock_sse2_start_16("xmm1", "xmm5", "3", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
760 snow_inner_add_yblock_sse2_accum_16("2", "16") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
761 snow_inner_add_yblock_sse2_accum_16("1", "512") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
762 snow_inner_add_yblock_sse2_accum_16("0", "528") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
763 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
764 "mov %0, %%"REG_d" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
765 "movdqa %%xmm1, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
766 "movdqa %%xmm5, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
767 "punpcklwd %%xmm7, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
768 "paddd (%%"REG_D"), %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
769 "punpckhwd %%xmm7, %%xmm1 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
770 "paddd 16(%%"REG_D"), %%xmm1 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
771 "punpcklwd %%xmm7, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
772 "paddd 32(%%"REG_D"), %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
773 "punpckhwd %%xmm7, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
774 "paddd 48(%%"REG_D"), %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
775 "paddd %%xmm3, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
776 "paddd %%xmm3, %%xmm1 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
777 "paddd %%xmm3, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
778 "paddd %%xmm3, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
779 "psrad $8, %%xmm0 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
780 "psrad $8, %%xmm1 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
781 "psrad $8, %%xmm4 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
782 "psrad $8, %%xmm5 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
783 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
784 "packssdw %%xmm1, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
785 "packssdw %%xmm5, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
786 "packuswb %%xmm4, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
787 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
788 "movdqu %%xmm0, (%%"REG_d") \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
789 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
790 snow_inner_add_yblock_sse2_end_16 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
791 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
792 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
793 #define snow_inner_add_yblock_mmx_header \ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
794 DWTELEM * * dst_array = sb->line + src_y;\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
795 long tmp;\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
796 asm volatile(\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
797 "mov %7, %%"REG_c" \n\t"\ |
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
798 "mov %6, %2 \n\t"\ |
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
799 "mov %4, %%"REG_S" \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
800 "pxor %%mm7, %%mm7 \n\t" /* 0 */\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
801 "pcmpeqd %%mm3, %%mm3 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
802 "pslld $31, %%mm3 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
803 "psrld $24, %%mm3 \n\t" /* FRAC_BITS >> 1 */\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
804 "1: \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
805 "mov %1, %%"REG_D" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
806 "mov (%%"REG_D"), %%"REG_D" \n\t"\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
807 "add %3, %%"REG_D" \n\t" |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
808 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
809 #define snow_inner_add_yblock_mmx_start(out_reg1, out_reg2, ptr_offset, s_offset, d_offset)\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
810 "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
811 "movd "d_offset"(%%"REG_d"), %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
812 "movd "d_offset"+4(%%"REG_d"), %%"out_reg2" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
813 "punpcklbw %%mm7, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
814 "punpcklbw %%mm7, %%"out_reg2" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
815 "movd "s_offset"(%%"REG_S"), %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
816 "movd "s_offset"+4(%%"REG_S"), %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
817 "punpcklbw %%mm7, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
818 "punpcklbw %%mm7, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
819 "pmullw %%mm0, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
820 "pmullw %%mm4, %%"out_reg2" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
821 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
822 #define snow_inner_add_yblock_mmx_accum(ptr_offset, s_offset, d_offset) \ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
823 snow_inner_add_yblock_mmx_start("mm2", "mm6", ptr_offset, s_offset, d_offset)\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
824 "paddusw %%mm2, %%mm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
825 "paddusw %%mm6, %%mm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
826 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
827 #define snow_inner_add_yblock_mmx_mix(read_offset, write_offset)\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
828 "mov %0, %%"REG_d" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
829 "movq %%mm1, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
830 "movq %%mm5, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
831 "punpcklwd %%mm7, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
832 "paddd "read_offset"(%%"REG_D"), %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
833 "punpckhwd %%mm7, %%mm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
834 "paddd "read_offset"+8(%%"REG_D"), %%mm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
835 "punpcklwd %%mm7, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
836 "paddd "read_offset"+16(%%"REG_D"), %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
837 "punpckhwd %%mm7, %%mm5 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
838 "paddd "read_offset"+24(%%"REG_D"), %%mm5 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
839 "paddd %%mm3, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
840 "paddd %%mm3, %%mm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
841 "paddd %%mm3, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
842 "paddd %%mm3, %%mm5 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
843 "psrad $8, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
844 "psrad $8, %%mm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
845 "psrad $8, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
846 "psrad $8, %%mm5 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
847 \ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
848 "packssdw %%mm1, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
849 "packssdw %%mm5, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
850 "packuswb %%mm4, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
851 "movq %%mm0, "write_offset"(%%"REG_d") \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
852 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
853 #define snow_inner_add_yblock_mmx_end(s_step)\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
854 "add $"s_step", %%"REG_S" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
855 "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
856 "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
857 "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
858 "add %%"REG_c", (%%"REG_a") \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
859 "add $"PTR_SIZE"*1, %1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
860 "add %%"REG_c", %0 \n\t"\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
861 "dec %2 \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
862 "jnz 1b \n\t"\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
863 :"+m"(dst8),"+m"(dst_array),"=&r"(tmp)\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
864 :\ |
3265
ba26b176dcf8
gcc-2.95 compile fix, patch by Wolfram Gloger <wmglo A dent PIS med PIS uni-muenchen PIS de>
gpoirier
parents:
3223
diff
changeset
|
865 "rm"((long)(src_x<<2)),"m"(obmc),"a"(block),"m"((long)b_h),"m"((long)src_stride):\ |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
866 "%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d""); |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
867 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
868 static void inner_add_yblock_bw_8_obmc_16_mmx(const uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h, |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
869 int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
870 snow_inner_add_yblock_mmx_header |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
871 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
872 snow_inner_add_yblock_mmx_accum("2", "8", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
873 snow_inner_add_yblock_mmx_accum("1", "128", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
874 snow_inner_add_yblock_mmx_accum("0", "136", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
875 snow_inner_add_yblock_mmx_mix("0", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
876 snow_inner_add_yblock_mmx_end("16") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
877 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
878 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
879 static void inner_add_yblock_bw_16_obmc_32_mmx(const uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h, |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
880 int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
881 snow_inner_add_yblock_mmx_header |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
882 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
883 snow_inner_add_yblock_mmx_accum("2", "16", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
884 snow_inner_add_yblock_mmx_accum("1", "512", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
885 snow_inner_add_yblock_mmx_accum("0", "528", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
886 snow_inner_add_yblock_mmx_mix("0", "0") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
887 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
888 snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "8", "8") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
889 snow_inner_add_yblock_mmx_accum("2", "24", "8") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
890 snow_inner_add_yblock_mmx_accum("1", "520", "8") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
891 snow_inner_add_yblock_mmx_accum("0", "536", "8") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
892 snow_inner_add_yblock_mmx_mix("32", "8") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
893 snow_inner_add_yblock_mmx_end("32") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
894 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
895 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
896 void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
897 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
898 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
899 if (b_w == 16) |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
900 inner_add_yblock_bw_16_obmc_32_sse2(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8); |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
901 else if (b_w == 8 && obmc_stride == 16) { |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
902 if (!(b_h & 1)) |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
903 inner_add_yblock_bw_8_obmc_16_bh_even_sse2(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8); |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
904 else |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
905 inner_add_yblock_bw_8_obmc_16_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8); |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
906 } else |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
907 ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8); |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
908 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
909 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
910 void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
911 int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
912 if (b_w == 16) |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
913 inner_add_yblock_bw_16_obmc_32_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8); |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
914 else if (b_w == 8 && obmc_stride == 16) |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
915 inner_add_yblock_bw_8_obmc_16_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8); |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
916 else |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
917 ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8); |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
918 } |