Mercurial > libavcodec.hg
annotate i386/snowdsp_mmx.c @ 5563:538d55152d09 libavcodec
remove code which become unused by the previous changes
author | michael |
---|---|
date | Tue, 21 Aug 2007 00:05:30 +0000 |
parents | 0074dd43ae5e |
children | 93082c591c8b |
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( | |
5555
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
114 "pcmpeqd %%xmm7, %%xmm7 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
115 "psrad $29, %%xmm7 \n\t" |
3210 | 116 ::); |
117 for(; i<w_l-7; i+=8){ | |
118 asm volatile( | |
119 "movdqu (%1), %%xmm1 \n\t" | |
120 "movdqu 16(%1), %%xmm5 \n\t" | |
121 "movdqu 4(%1), %%xmm0 \n\t" | |
5554 | 122 "movdqu 20(%1), %%xmm4 \n\t" //FIXME try aligned reads and shifts |
3210 | 123 "paddd %%xmm1, %%xmm0 \n\t" |
124 "paddd %%xmm5, %%xmm4 \n\t" | |
5555
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
125 "paddd %%xmm7, %%xmm0 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
126 "paddd %%xmm7, %%xmm4 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
127 "movdqa (%0), %%xmm1 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
128 "movdqa 16(%0), %%xmm5 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
129 "psrad $2, %%xmm0 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
130 "psrad $2, %%xmm4 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
131 "paddd %%xmm1, %%xmm0 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
132 "paddd %%xmm5, %%xmm4 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
133 "psrad $2, %%xmm0 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
134 "psrad $2, %%xmm4 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
135 "paddd %%xmm1, %%xmm0 \n\t" |
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
136 "paddd %%xmm5, %%xmm4 \n\t" |
3210 | 137 "movdqa %%xmm0, (%0) \n\t" |
138 "movdqa %%xmm4, 16(%0) \n\t" | |
139 :: "r"(&b[i]), "r"(&ref[i]) | |
140 : "memory" | |
141 ); | |
142 } | |
143 snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l); | |
5555
0790af6d0028
remove idiotc double subtraction from the sse2 code (untested, no sse2 here)
michael
parents:
5554
diff
changeset
|
144 b[0] = b_0 + ((2 * ref[1] + W_BO-1 + 4 * b_0) >> W_BS); |
3210 | 145 } |
146 | |
147 { // Lift 3 | |
148 DWTELEM * const src = b+w2; | |
149 | |
150 i = 0; | |
151 for(; (((long)&temp[i]) & 0xF) && i<w_r; i++){ | |
152 temp[i] = src[i] - ((-W_AM*(b[i] + b[i+1]))>>W_AS); | |
153 } | |
154 for(; i<w_r-7; i+=8){ | |
155 asm volatile( | |
156 "movdqu 4(%1), %%xmm2 \n\t" | |
157 "movdqu 20(%1), %%xmm6 \n\t" | |
158 "paddd (%1), %%xmm2 \n\t" | |
159 "paddd 16(%1), %%xmm6 \n\t" | |
160 "movdqa %%xmm2, %%xmm0 \n\t" | |
161 "movdqa %%xmm6, %%xmm4 \n\t" | |
162 "pslld $2, %%xmm2 \n\t" | |
163 "pslld $2, %%xmm6 \n\t" | |
164 "psubd %%xmm2, %%xmm0 \n\t" | |
165 "psubd %%xmm6, %%xmm4 \n\t" | |
166 "psrad $1, %%xmm0 \n\t" | |
167 "psrad $1, %%xmm4 \n\t" | |
168 "movdqu (%0), %%xmm2 \n\t" | |
169 "movdqu 16(%0), %%xmm6 \n\t" | |
170 "psubd %%xmm0, %%xmm2 \n\t" | |
171 "psubd %%xmm4, %%xmm6 \n\t" | |
172 "movdqa %%xmm2, (%2) \n\t" | |
173 "movdqa %%xmm6, 16(%2) \n\t" | |
174 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i]) | |
175 : "memory" | |
176 ); | |
177 } | |
178 snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS); | |
179 } | |
180 | |
181 { | |
182 snow_interleave_line_header(&i, width, b, temp); | |
183 | |
184 for (; (i & 0x1E) != 0x1E; i-=2){ | |
185 b[i+1] = temp[i>>1]; | |
186 b[i] = b[i>>1]; | |
187 } | |
188 for (i-=30; i>=0; i-=32){ | |
189 asm volatile( | |
190 "movdqa (%1), %%xmm0 \n\t" | |
191 "movdqa 16(%1), %%xmm2 \n\t" | |
192 "movdqa 32(%1), %%xmm4 \n\t" | |
193 "movdqa 48(%1), %%xmm6 \n\t" | |
194 "movdqa (%1), %%xmm1 \n\t" | |
195 "movdqa 16(%1), %%xmm3 \n\t" | |
196 "movdqa 32(%1), %%xmm5 \n\t" | |
197 "movdqa 48(%1), %%xmm7 \n\t" | |
198 "punpckldq (%2), %%xmm0 \n\t" | |
199 "punpckldq 16(%2), %%xmm2 \n\t" | |
200 "punpckldq 32(%2), %%xmm4 \n\t" | |
201 "punpckldq 48(%2), %%xmm6 \n\t" | |
202 "movdqa %%xmm0, (%0) \n\t" | |
203 "movdqa %%xmm2, 32(%0) \n\t" | |
204 "movdqa %%xmm4, 64(%0) \n\t" | |
205 "movdqa %%xmm6, 96(%0) \n\t" | |
206 "punpckhdq (%2), %%xmm1 \n\t" | |
207 "punpckhdq 16(%2), %%xmm3 \n\t" | |
208 "punpckhdq 32(%2), %%xmm5 \n\t" | |
209 "punpckhdq 48(%2), %%xmm7 \n\t" | |
210 "movdqa %%xmm1, 16(%0) \n\t" | |
211 "movdqa %%xmm3, 48(%0) \n\t" | |
212 "movdqa %%xmm5, 80(%0) \n\t" | |
213 "movdqa %%xmm7, 112(%0) \n\t" | |
214 :: "r"(&(b)[i]), "r"(&(b)[i>>1]), "r"(&(temp)[i>>1]) | |
215 : "memory" | |
216 ); | |
217 } | |
218 } | |
219 } | |
220 | |
221 void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width){ | |
222 const int w2= (width+1)>>1; | |
223 DWTELEM temp[width >> 1]; | |
224 const int w_l= (width>>1); | |
225 const int w_r= w2 - 1; | |
226 int i; | |
227 | |
228 { // Lift 0 | |
229 DWTELEM * const ref = b + w2 - 1; | |
230 | |
231 i = 1; | |
232 b[0] = b[0] - ((W_DM * 2 * ref[1]+W_DO)>>W_DS); | |
233 asm volatile( | |
234 "pcmpeqd %%mm7, %%mm7 \n\t" | |
235 "pslld $31, %%mm7 \n\t" | |
236 "psrld $29, %%mm7 \n\t" | |
237 ::); | |
238 for(; i<w_l-3; i+=4){ | |
239 asm volatile( | |
240 "movq (%1), %%mm2 \n\t" | |
241 "movq 8(%1), %%mm6 \n\t" | |
242 "paddd 4(%1), %%mm2 \n\t" | |
243 "paddd 12(%1), %%mm6 \n\t" | |
244 "movq %%mm2, %%mm0 \n\t" | |
245 "movq %%mm6, %%mm4 \n\t" | |
246 "paddd %%mm2, %%mm2 \n\t" | |
247 "paddd %%mm6, %%mm6 \n\t" | |
248 "paddd %%mm0, %%mm2 \n\t" | |
249 "paddd %%mm4, %%mm6 \n\t" | |
250 "paddd %%mm7, %%mm2 \n\t" | |
251 "paddd %%mm7, %%mm6 \n\t" | |
252 "psrad $3, %%mm2 \n\t" | |
253 "psrad $3, %%mm6 \n\t" | |
254 "movq (%0), %%mm0 \n\t" | |
255 "movq 8(%0), %%mm4 \n\t" | |
256 "psubd %%mm2, %%mm0 \n\t" | |
257 "psubd %%mm6, %%mm4 \n\t" | |
258 "movq %%mm0, (%0) \n\t" | |
259 "movq %%mm4, 8(%0) \n\t" | |
260 :: "r"(&b[i]), "r"(&ref[i]) | |
261 : "memory" | |
262 ); | |
263 } | |
264 snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS); | |
265 } | |
266 | |
267 { // Lift 1 | |
268 DWTELEM * const dst = b+w2; | |
269 | |
270 i = 0; | |
271 for(; i<w_r-3; i+=4){ | |
272 asm volatile( | |
273 "movq (%1), %%mm2 \n\t" | |
274 "movq 8(%1), %%mm6 \n\t" | |
275 "paddd 4(%1), %%mm2 \n\t" | |
276 "paddd 12(%1), %%mm6 \n\t" | |
277 "movq (%0), %%mm0 \n\t" | |
278 "movq 8(%0), %%mm4 \n\t" | |
279 "psubd %%mm2, %%mm0 \n\t" | |
280 "psubd %%mm6, %%mm4 \n\t" | |
281 "movq %%mm0, (%0) \n\t" | |
282 "movq %%mm4, 8(%0) \n\t" | |
283 :: "r"(&dst[i]), "r"(&b[i]) | |
284 : "memory" | |
285 ); | |
286 } | |
287 snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS); | |
288 } | |
289 | |
290 { // Lift 2 | |
291 DWTELEM * const ref = b+w2 - 1; | |
292 | |
293 i = 1; | |
5553
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
294 b[0] = b[0] + (((2 * ref[1] + W_BO-1) + 4 * b[0]) >> W_BS); |
3210 | 295 asm volatile( |
5553
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
296 "pcmpeqd %%mm7, %%mm7 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
297 "psrld $29, %%mm7 \n\t" |
3210 | 298 ::); |
299 for(; i<w_l-3; i+=4){ | |
300 asm volatile( | |
301 "movq (%1), %%mm0 \n\t" | |
302 "movq 8(%1), %%mm4 \n\t" | |
303 "paddd 4(%1), %%mm0 \n\t" | |
304 "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
|
305 "paddd %%mm7, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
306 "paddd %%mm7, %%mm4 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
307 "psrad $2, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
308 "psrad $2, %%mm4 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
309 "movq (%0), %%mm1 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
310 "movq 8(%0), %%mm5 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
311 "paddd %%mm1, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
312 "paddd %%mm5, %%mm4 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
313 "psrad $2, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
314 "psrad $2, %%mm4 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
315 "paddd %%mm1, %%mm0 \n\t" |
3c0c9723ed3c
and of course the unneeded double subtractions were blindly put in the
michael
parents:
5010
diff
changeset
|
316 "paddd %%mm5, %%mm4 \n\t" |
3210 | 317 "movq %%mm0, (%0) \n\t" |
318 "movq %%mm4, 8(%0) \n\t" | |
319 :: "r"(&b[i]), "r"(&ref[i]) | |
320 : "memory" | |
321 ); | |
322 } | |
323 snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l); | |
324 } | |
325 | |
326 { // Lift 3 | |
327 DWTELEM * const src = b+w2; | |
328 i = 0; | |
329 | |
330 for(; i<w_r-3; i+=4){ | |
331 asm volatile( | |
332 "movq 4(%1), %%mm2 \n\t" | |
333 "movq 12(%1), %%mm6 \n\t" | |
334 "paddd (%1), %%mm2 \n\t" | |
335 "paddd 8(%1), %%mm6 \n\t" | |
5556
d2d3b185c8c8
slightly change horizontal lift3 so it needs 1 bit less headroom
michael
parents:
5555
diff
changeset
|
336 "pxor %%mm0, %%mm0 \n\t" //note: the 2 xor could be avoided if we would flip the rounding direction |
d2d3b185c8c8
slightly change horizontal lift3 so it needs 1 bit less headroom
michael
parents:
5555
diff
changeset
|
337 "pxor %%mm4, %%mm4 \n\t" |
3210 | 338 "psubd %%mm2, %%mm0 \n\t" |
339 "psubd %%mm6, %%mm4 \n\t" | |
340 "psrad $1, %%mm0 \n\t" | |
341 "psrad $1, %%mm4 \n\t" | |
342 "psubd %%mm0, %%mm2 \n\t" | |
343 "psubd %%mm4, %%mm6 \n\t" | |
5556
d2d3b185c8c8
slightly change horizontal lift3 so it needs 1 bit less headroom
michael
parents:
5555
diff
changeset
|
344 "movq (%0), %%mm0 \n\t" |
d2d3b185c8c8
slightly change horizontal lift3 so it needs 1 bit less headroom
michael
parents:
5555
diff
changeset
|
345 "movq 8(%0), %%mm4 \n\t" |
d2d3b185c8c8
slightly change horizontal lift3 so it needs 1 bit less headroom
michael
parents:
5555
diff
changeset
|
346 "paddd %%mm0, %%mm2 \n\t" |
d2d3b185c8c8
slightly change horizontal lift3 so it needs 1 bit less headroom
michael
parents:
5555
diff
changeset
|
347 "paddd %%mm4, %%mm6 \n\t" |
3210 | 348 "movq %%mm2, (%2) \n\t" |
349 "movq %%mm6, 8(%2) \n\t" | |
350 :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i]) | |
351 : "memory" | |
352 ); | |
353 } | |
354 snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS); | |
355 } | |
356 | |
357 { | |
358 snow_interleave_line_header(&i, width, b, temp); | |
359 | |
360 for (; (i & 0xE) != 0xE; i-=2){ | |
361 b[i+1] = temp[i>>1]; | |
362 b[i] = b[i>>1]; | |
363 } | |
364 for (i-=14; i>=0; i-=16){ | |
365 asm volatile( | |
366 "movq (%1), %%mm0 \n\t" | |
367 "movq 8(%1), %%mm2 \n\t" | |
368 "movq 16(%1), %%mm4 \n\t" | |
369 "movq 24(%1), %%mm6 \n\t" | |
370 "movq (%1), %%mm1 \n\t" | |
371 "movq 8(%1), %%mm3 \n\t" | |
372 "movq 16(%1), %%mm5 \n\t" | |
373 "movq 24(%1), %%mm7 \n\t" | |
374 "punpckldq (%2), %%mm0 \n\t" | |
375 "punpckldq 8(%2), %%mm2 \n\t" | |
376 "punpckldq 16(%2), %%mm4 \n\t" | |
377 "punpckldq 24(%2), %%mm6 \n\t" | |
378 "movq %%mm0, (%0) \n\t" | |
379 "movq %%mm2, 16(%0) \n\t" | |
380 "movq %%mm4, 32(%0) \n\t" | |
381 "movq %%mm6, 48(%0) \n\t" | |
382 "punpckhdq (%2), %%mm1 \n\t" | |
383 "punpckhdq 8(%2), %%mm3 \n\t" | |
384 "punpckhdq 16(%2), %%mm5 \n\t" | |
385 "punpckhdq 24(%2), %%mm7 \n\t" | |
386 "movq %%mm1, 8(%0) \n\t" | |
387 "movq %%mm3, 24(%0) \n\t" | |
388 "movq %%mm5, 40(%0) \n\t" | |
389 "movq %%mm7, 56(%0) \n\t" | |
390 :: "r"(&b[i]), "r"(&b[i>>1]), "r"(&temp[i>>1]) | |
391 : "memory" | |
392 ); | |
393 } | |
394 } | |
395 } | |
396 | |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
397 #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
|
398 ""op" (%%"r",%%"REG_d",4), %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
399 ""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
|
400 ""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
|
401 ""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
|
402 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
403 #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
|
404 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
|
405 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
406 #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
|
407 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
|
408 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
409 #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
|
410 "psubd %%"s0", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
411 "psubd %%"s1", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
412 "psubd %%"s2", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
413 "psubd %%"s3", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
414 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
415 #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
|
416 "movdqa %%"s0", (%%"w",%%"REG_d",4) \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
417 "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
|
418 "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
|
419 "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
|
420 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
421 #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
|
422 "psrad $"n", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
423 "psrad $"n", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
424 "psrad $"n", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
425 "psrad $"n", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
426 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
427 #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
|
428 "paddd %%"s0", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
429 "paddd %%"s1", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
430 "paddd %%"s2", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
431 "paddd %%"s3", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
432 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
433 #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
|
434 "movdqa %%"s0", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
435 "movdqa %%"s1", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
436 "movdqa %%"s2", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
437 "movdqa %%"s3", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
438 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
439 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
|
440 long i = width; |
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 while(i & 0xF) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
443 { |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
444 i--; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
445 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
|
446 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
|
447 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
|
448 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
|
449 } |
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 asm volatile ( |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
452 "jmp 2f \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
453 "1: \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
454 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
455 "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
|
456 "mov %4, %%"REG_S" \n\t" |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
457 |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
458 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
|
459 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
|
460 snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7") |
5558 | 461 snow_vertical_compose_sse2_r2r_add("xmm0","xmm2","xmm4","xmm6","xmm0","xmm2","xmm4","xmm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
462 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
|
463 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
464 "pcmpeqd %%xmm1, %%xmm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
465 "pslld $31, %%xmm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
466 "psrld $29, %%xmm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
467 "mov %5, %%"REG_a" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
468 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
469 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
|
470 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
|
471 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
|
472 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
|
473 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
|
474 "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
|
475 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
|
476 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
|
477 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
|
478 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
|
479 "mov %2, %%"REG_a" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
480 snow_vertical_compose_sse2_add(REG_a,"xmm0","xmm2","xmm4","xmm6") |
5560
2e63bd8e894d
Simplify and speedup code, reduce needed headroom by 2 bits in the 3rd
michael
parents:
5559
diff
changeset
|
481 snow_vertical_compose_sse2_sra("2","xmm0","xmm2","xmm4","xmm6") |
2e63bd8e894d
Simplify and speedup code, reduce needed headroom by 2 bits in the 3rd
michael
parents:
5559
diff
changeset
|
482 snow_vertical_compose_sse2_add(REG_c,"xmm0","xmm2","xmm4","xmm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
483 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
484 "pcmpeqd %%xmm1, %%xmm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
485 "pslld $31, %%xmm1 \n\t" |
5560
2e63bd8e894d
Simplify and speedup code, reduce needed headroom by 2 bits in the 3rd
michael
parents:
5559
diff
changeset
|
486 "psrld $30, %%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
|
487 "mov %1, %%"REG_S" \n\t" |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
488 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
489 snow_vertical_compose_sse2_r2r_add("xmm1","xmm1","xmm1","xmm1","xmm0","xmm2","xmm4","xmm6") |
5560
2e63bd8e894d
Simplify and speedup code, reduce needed headroom by 2 bits in the 3rd
michael
parents:
5559
diff
changeset
|
490 snow_vertical_compose_sse2_sra("2","xmm0","xmm2","xmm4","xmm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
491 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
|
492 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
|
493 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
|
494 snow_vertical_compose_sse2_move("xmm0","xmm2","xmm4","xmm6","xmm1","xmm3","xmm5","xmm7") |
5562
0074dd43ae5e
Simplify and optimize the 4th vertical lifting step of the SSE2 code (untested)
michael
parents:
5561
diff
changeset
|
495 snow_vertical_compose_sse2_sra("1","xmm0","xmm2","xmm4","xmm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
496 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
|
497 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
|
498 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
|
499 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
500 "2: \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
501 "sub $16, %%"REG_d" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
502 "jge 1b \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
503 :"+d"(i) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
504 : |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
505 "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
|
506 "%"REG_a"","%"REG_S"","%"REG_c""); |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
507 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
508 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
509 #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
|
510 ""op" (%%"r",%%"REG_d",4), %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
511 ""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
|
512 ""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
|
513 ""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
|
514 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
515 #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
|
516 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
|
517 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
518 #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
|
519 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
|
520 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
521 #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
|
522 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
|
523 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
524 #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
|
525 "movq %%"s0", (%%"w",%%"REG_d",4) \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
526 "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
|
527 "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
|
528 "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
|
529 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
530 #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
|
531 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
|
532 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
533 #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
|
534 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
|
535 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
536 #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
|
537 "movq %%"s0", %%"t0" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
538 "movq %%"s1", %%"t1" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
539 "movq %%"s2", %%"t2" \n\t"\ |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
540 "movq %%"s3", %%"t3" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
541 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
542 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
|
543 long i = width; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
544 while(i & 0x7) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
545 { |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
546 i--; |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
547 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
|
548 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
|
549 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
|
550 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
|
551 } |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
552 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
553 asm volatile( |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
554 "jmp 2f \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
555 "1: \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
556 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
557 "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
|
558 "mov %4, %%"REG_S" \n\t" |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
559 |
3566
b63ef6fcbc70
Fix x86 SIMD asm and pic, patch from Martin von Gagern <Martin.vGagern@gmx.net>
lu_zero
parents:
3398
diff
changeset
|
560 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
|
561 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
|
562 snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7") |
5557 | 563 snow_vertical_compose_mmx_r2r_add("mm0","mm2","mm4","mm6","mm0","mm2","mm4","mm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
564 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
|
565 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
566 "pcmpeqd %%mm1, %%mm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
567 "pslld $31, %%mm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
568 "psrld $29, %%mm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
569 "mov %5, %%"REG_a" \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 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
|
572 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
|
573 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
|
574 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
|
575 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
|
576 "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
|
577 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
|
578 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
|
579 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
|
580 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
|
581 "mov %2, %%"REG_a" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
582 snow_vertical_compose_mmx_add(REG_a,"mm0","mm2","mm4","mm6") |
5559
9993d38f59cf
simplify, speedup and reduce needed headroom by 2 bits in the 3rd
michael
parents:
5558
diff
changeset
|
583 snow_vertical_compose_mmx_sra("2","mm0","mm2","mm4","mm6") |
9993d38f59cf
simplify, speedup and reduce needed headroom by 2 bits in the 3rd
michael
parents:
5558
diff
changeset
|
584 snow_vertical_compose_mmx_add(REG_c,"mm0","mm2","mm4","mm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
585 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
586 "pcmpeqd %%mm1, %%mm1 \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
587 "pslld $31, %%mm1 \n\t" |
5559
9993d38f59cf
simplify, speedup and reduce needed headroom by 2 bits in the 3rd
michael
parents:
5558
diff
changeset
|
588 "psrld $30, %%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
|
589 "mov %1, %%"REG_S" \n\t" |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
590 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
591 snow_vertical_compose_mmx_r2r_add("mm1","mm1","mm1","mm1","mm0","mm2","mm4","mm6") |
5559
9993d38f59cf
simplify, speedup and reduce needed headroom by 2 bits in the 3rd
michael
parents:
5558
diff
changeset
|
592 snow_vertical_compose_mmx_sra("2","mm0","mm2","mm4","mm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
593 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
|
594 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
|
595 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
|
596 snow_vertical_compose_mmx_move("mm0","mm2","mm4","mm6","mm1","mm3","mm5","mm7") |
5561
225867be27f7
Simplify and optimize the 4th vertical lifting step of the MMX code
michael
parents:
5560
diff
changeset
|
597 snow_vertical_compose_mmx_sra("1","mm0","mm2","mm4","mm6") |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
598 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
|
599 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
|
600 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
|
601 |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
602 "2: \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
603 "sub $8, %%"REG_d" \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
604 "jge 1b \n\t" |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
605 :"+d"(i) |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
606 : |
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
607 "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
|
608 "%"REG_a"","%"REG_S"","%"REG_c""); |
3207
33110c1008a4
Add the mmx and sse2 implementations of ff_snow_vertical_compose().
gpoirier
parents:
diff
changeset
|
609 } |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
610 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
611 #define snow_inner_add_yblock_sse2_header \ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
612 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
|
613 long tmp;\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
614 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
|
615 "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
|
616 "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
|
617 "mov %4, %%"REG_S" \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
618 "pxor %%xmm7, %%xmm7 \n\t" /* 0 */\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
619 "pcmpeqd %%xmm3, %%xmm3 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
620 "pslld $31, %%xmm3 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
621 "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
|
622 "1: \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
623 "mov %1, %%"REG_D" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
624 "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
|
625 "add %3, %%"REG_D" \n\t" |
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_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
|
628 "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
|
629 "movq (%%"REG_d"), %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
630 "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
|
631 "punpcklbw %%xmm7, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
632 "punpcklbw %%xmm7, %%"out_reg2" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
633 "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
|
634 "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
|
635 "punpcklbw %%xmm7, %%xmm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
636 "punpcklbw %%xmm7, %%xmm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
637 "pmullw %%xmm0, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
638 "pmullw %%xmm4, %%"out_reg2" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
639 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
640 #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
|
641 "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
|
642 "movq (%%"REG_d"), %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
643 "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
|
644 "punpcklbw %%xmm7, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
645 "punpcklbw %%xmm7, %%"out_reg2" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
646 "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
|
647 "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
|
648 "punpcklbw %%xmm7, %%xmm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
649 "punpcklbw %%xmm7, %%xmm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
650 "pmullw %%xmm0, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
651 "pmullw %%xmm4, %%"out_reg2" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
652 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
653 #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
|
654 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
|
655 "paddusw %%xmm2, %%xmm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
656 "paddusw %%xmm6, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
657 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
658 #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
|
659 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
|
660 "paddusw %%xmm2, %%xmm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
661 "paddusw %%xmm6, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
662 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
663 #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
|
664 "add $32, %%"REG_S" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
665 "add %%"REG_c", %0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
666 "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
|
667 "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
|
668 "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
|
669 "add %%"REG_c", (%%"REG_a") \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
670 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
671 #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
|
672 "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
|
673 :"+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
|
674 :\ |
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
|
675 "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
|
676 "%"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
|
677 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
678 #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
|
679 "sal $1, %%"REG_c" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
680 "add $"PTR_SIZE"*2, %1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
681 snow_inner_add_yblock_sse2_end_common1\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
682 "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
|
683 "sub $2, %2 \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
684 snow_inner_add_yblock_sse2_end_common2 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
685 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
686 #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
|
687 "add $"PTR_SIZE"*1, %1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
688 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
|
689 "dec %2 \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
690 snow_inner_add_yblock_sse2_end_common2 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
691 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
692 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
|
693 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
|
694 snow_inner_add_yblock_sse2_header |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
695 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
|
696 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
|
697 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
|
698 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
|
699 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
700 "mov %0, %%"REG_d" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
701 "movdqa (%%"REG_D"), %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
702 "movdqa %%xmm1, %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
703 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
704 "punpckhwd %%xmm7, %%xmm1 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
705 "punpcklwd %%xmm7, %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
706 "paddd %%xmm2, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
707 "movdqa 16(%%"REG_D"), %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
708 "paddd %%xmm1, %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
709 "paddd %%xmm3, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
710 "paddd %%xmm3, %%xmm2 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
711 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
712 "mov %1, %%"REG_D" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
713 "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
|
714 "add %3, %%"REG_D" \n\t" |
3211
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 "movdqa (%%"REG_D"), %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
717 "movdqa %%xmm5, %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
718 "punpckhwd %%xmm7, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
719 "punpcklwd %%xmm7, %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
720 "paddd %%xmm6, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
721 "movdqa 16(%%"REG_D"), %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
722 "paddd %%xmm5, %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
723 "paddd %%xmm3, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
724 "paddd %%xmm3, %%xmm6 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
725 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
726 "psrad $8, %%xmm0 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
727 "psrad $8, %%xmm2 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
728 "packssdw %%xmm2, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
729 "packuswb %%xmm7, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
730 "movq %%xmm0, (%%"REG_d") \n\t" |
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 "psrad $8, %%xmm4 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
733 "psrad $8, %%xmm6 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
734 "packssdw %%xmm6, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
735 "packuswb %%xmm7, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
736 "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
|
737 snow_inner_add_yblock_sse2_end_8 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
738 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
739 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
740 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
|
741 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
|
742 snow_inner_add_yblock_sse2_header |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
743 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
|
744 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
|
745 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
|
746 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
|
747 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
748 "mov %0, %%"REG_d" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
749 "movdqa %%xmm1, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
750 "movdqa %%xmm5, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
751 "punpcklwd %%xmm7, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
752 "paddd (%%"REG_D"), %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
753 "punpckhwd %%xmm7, %%xmm1 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
754 "paddd 16(%%"REG_D"), %%xmm1 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
755 "punpcklwd %%xmm7, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
756 "paddd 32(%%"REG_D"), %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
757 "punpckhwd %%xmm7, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
758 "paddd 48(%%"REG_D"), %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
759 "paddd %%xmm3, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
760 "paddd %%xmm3, %%xmm1 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
761 "paddd %%xmm3, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
762 "paddd %%xmm3, %%xmm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
763 "psrad $8, %%xmm0 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
764 "psrad $8, %%xmm1 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
765 "psrad $8, %%xmm4 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
766 "psrad $8, %%xmm5 \n\t" /* FRAC_BITS. */ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
767 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
768 "packssdw %%xmm1, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
769 "packssdw %%xmm5, %%xmm4 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
770 "packuswb %%xmm4, %%xmm0 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
771 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
772 "movdqu %%xmm0, (%%"REG_d") \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
773 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
774 snow_inner_add_yblock_sse2_end_16 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
775 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
776 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
777 #define snow_inner_add_yblock_mmx_header \ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
778 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
|
779 long tmp;\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
780 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
|
781 "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
|
782 "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
|
783 "mov %4, %%"REG_S" \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
784 "pxor %%mm7, %%mm7 \n\t" /* 0 */\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
785 "pcmpeqd %%mm3, %%mm3 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
786 "pslld $31, %%mm3 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
787 "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
|
788 "1: \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
789 "mov %1, %%"REG_D" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
790 "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
|
791 "add %3, %%"REG_D" \n\t" |
3211
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_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
|
794 "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
|
795 "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
|
796 "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
|
797 "punpcklbw %%mm7, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
798 "punpcklbw %%mm7, %%"out_reg2" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
799 "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
|
800 "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
|
801 "punpcklbw %%mm7, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
802 "punpcklbw %%mm7, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
803 "pmullw %%mm0, %%"out_reg1" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
804 "pmullw %%mm4, %%"out_reg2" \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
805 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
806 #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
|
807 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
|
808 "paddusw %%mm2, %%mm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
809 "paddusw %%mm6, %%mm5 \n\t" |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
810 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
811 #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
|
812 "mov %0, %%"REG_d" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
813 "movq %%mm1, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
814 "movq %%mm5, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
815 "punpcklwd %%mm7, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
816 "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
|
817 "punpckhwd %%mm7, %%mm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
818 "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
|
819 "punpcklwd %%mm7, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
820 "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
|
821 "punpckhwd %%mm7, %%mm5 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
822 "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
|
823 "paddd %%mm3, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
824 "paddd %%mm3, %%mm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
825 "paddd %%mm3, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
826 "paddd %%mm3, %%mm5 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
827 "psrad $8, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
828 "psrad $8, %%mm1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
829 "psrad $8, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
830 "psrad $8, %%mm5 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
831 \ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
832 "packssdw %%mm1, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
833 "packssdw %%mm5, %%mm4 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
834 "packuswb %%mm4, %%mm0 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
835 "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
|
836 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
837 #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
|
838 "add $"s_step", %%"REG_S" \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
839 "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
|
840 "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
|
841 "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
|
842 "add %%"REG_c", (%%"REG_a") \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
843 "add $"PTR_SIZE"*1, %1 \n\t"\ |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
844 "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
|
845 "dec %2 \n\t"\ |
3211
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
846 "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
|
847 :"+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
|
848 :\ |
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
|
849 "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
|
850 "%"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
|
851 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
852 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
|
853 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
|
854 snow_inner_add_yblock_mmx_header |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
855 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
|
856 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
|
857 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
|
858 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
|
859 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
|
860 snow_inner_add_yblock_mmx_end("16") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
861 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
862 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
863 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
|
864 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
|
865 snow_inner_add_yblock_mmx_header |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
866 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
|
867 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
|
868 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
|
869 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
|
870 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
|
871 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
872 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
|
873 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
|
874 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
|
875 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
|
876 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
|
877 snow_inner_add_yblock_mmx_end("32") |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
878 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
879 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
880 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
|
881 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
|
882 |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
883 if (b_w == 16) |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
884 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
|
885 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
|
886 if (!(b_h & 1)) |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
887 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
|
888 else |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
889 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
|
890 } else |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
891 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
|
892 } |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
893 |
4436
d3e389536b0a
Add the const specifier as needed to reduce the number of warnings.
takis
parents:
3947
diff
changeset
|
894 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
|
895 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
|
896 if (b_w == 16) |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
897 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
|
898 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
|
899 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
|
900 else |
b77b5e7072d6
add MMX and SSE versions of ff_snow_inner_add_yblock
gpoirier
parents:
3210
diff
changeset
|
901 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
|
902 } |