Mercurial > libavcodec.hg
annotate imgresample.c @ 1103:b05aa3f49cc8 libavcodec
doxygen config file, so ppl just have to call doxygen in the libavcodec dir and a beatifull doxy is generated ;)
author | michaelni |
---|---|
date | Thu, 06 Mar 2003 10:21:57 +0000 |
parents | b32afefe7d33 |
children | 1e39f273ecd6 |
rev | line source |
---|---|
0 | 1 /* |
2 * High quality image resampling with polyphase filters | |
429 | 3 * Copyright (c) 2001 Fabrice Bellard. |
0 | 4 * |
429 | 5 * This library is free software; you can redistribute it and/or |
6 * modify it under the terms of the GNU Lesser General Public | |
7 * License as published by the Free Software Foundation; either | |
8 * version 2 of the License, or (at your option) any later version. | |
0 | 9 * |
429 | 10 * This library is distributed in the hope that it will be useful, |
0 | 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
429 | 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
13 * Lesser General Public License for more details. | |
0 | 14 * |
429 | 15 * You should have received a copy of the GNU Lesser General Public |
16 * License along with this library; if not, write to the Free Software | |
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
0 | 18 */ |
396
fce0a2520551
removed useless header includes - use av memory functions
glantau
parents:
18
diff
changeset
|
19 #include "avcodec.h" |
0 | 20 #include "dsputil.h" |
21 | |
17 | 22 #ifdef USE_FASTMEMCPY |
23 #include "fastmemcpy.h" | |
24 #endif | |
25 | |
0 | 26 #define NB_COMPONENTS 3 |
27 | |
28 #define PHASE_BITS 4 | |
29 #define NB_PHASES (1 << PHASE_BITS) | |
30 #define NB_TAPS 4 | |
31 #define FCENTER 1 /* index of the center of the filter */ | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
32 //#define TEST 1 /* Test it */ |
0 | 33 |
34 #define POS_FRAC_BITS 16 | |
35 #define POS_FRAC (1 << POS_FRAC_BITS) | |
36 /* 6 bits precision is needed for MMX */ | |
37 #define FILTER_BITS 8 | |
38 | |
39 #define LINE_BUF_HEIGHT (NB_TAPS * 4) | |
40 | |
41 struct ImgReSampleContext { | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
42 int iwidth, iheight, owidth, oheight, topBand, bottomBand, leftBand, rightBand; |
0 | 43 int h_incr, v_incr; |
1064 | 44 int16_t h_filters[NB_PHASES][NB_TAPS] __align8; /* horizontal filters */ |
45 int16_t v_filters[NB_PHASES][NB_TAPS] __align8; /* vertical filters */ | |
46 uint8_t *line_buf; | |
0 | 47 }; |
48 | |
49 static inline int get_phase(int pos) | |
50 { | |
51 return ((pos) >> (POS_FRAC_BITS - PHASE_BITS)) & ((1 << PHASE_BITS) - 1); | |
52 } | |
53 | |
54 /* This function must be optimized */ | |
1064 | 55 static void h_resample_fast(uint8_t *dst, int dst_width, uint8_t *src, int src_width, |
56 int src_start, int src_incr, int16_t *filters) | |
0 | 57 { |
58 int src_pos, phase, sum, i; | |
1064 | 59 uint8_t *s; |
60 int16_t *filter; | |
0 | 61 |
62 src_pos = src_start; | |
63 for(i=0;i<dst_width;i++) { | |
64 #ifdef TEST | |
65 /* test */ | |
66 if ((src_pos >> POS_FRAC_BITS) < 0 || | |
67 (src_pos >> POS_FRAC_BITS) > (src_width - NB_TAPS)) | |
653 | 68 av_abort(); |
0 | 69 #endif |
70 s = src + (src_pos >> POS_FRAC_BITS); | |
71 phase = get_phase(src_pos); | |
72 filter = filters + phase * NB_TAPS; | |
73 #if NB_TAPS == 4 | |
74 sum = s[0] * filter[0] + | |
75 s[1] * filter[1] + | |
76 s[2] * filter[2] + | |
77 s[3] * filter[3]; | |
78 #else | |
79 { | |
80 int j; | |
81 sum = 0; | |
82 for(j=0;j<NB_TAPS;j++) | |
83 sum += s[j] * filter[j]; | |
84 } | |
85 #endif | |
86 sum = sum >> FILTER_BITS; | |
87 if (sum < 0) | |
88 sum = 0; | |
89 else if (sum > 255) | |
90 sum = 255; | |
91 dst[0] = sum; | |
92 src_pos += src_incr; | |
93 dst++; | |
94 } | |
95 } | |
96 | |
97 /* This function must be optimized */ | |
1064 | 98 static void v_resample(uint8_t *dst, int dst_width, uint8_t *src, int wrap, |
99 int16_t *filter) | |
0 | 100 { |
101 int sum, i; | |
1064 | 102 uint8_t *s; |
0 | 103 |
104 s = src; | |
105 for(i=0;i<dst_width;i++) { | |
106 #if NB_TAPS == 4 | |
107 sum = s[0 * wrap] * filter[0] + | |
108 s[1 * wrap] * filter[1] + | |
109 s[2 * wrap] * filter[2] + | |
110 s[3 * wrap] * filter[3]; | |
111 #else | |
112 { | |
113 int j; | |
1064 | 114 uint8_t *s1 = s; |
0 | 115 |
116 sum = 0; | |
117 for(j=0;j<NB_TAPS;j++) { | |
118 sum += s1[0] * filter[j]; | |
119 s1 += wrap; | |
120 } | |
121 } | |
122 #endif | |
123 sum = sum >> FILTER_BITS; | |
124 if (sum < 0) | |
125 sum = 0; | |
126 else if (sum > 255) | |
127 sum = 255; | |
128 dst[0] = sum; | |
129 dst++; | |
130 s++; | |
131 } | |
132 } | |
133 | |
2 | 134 #ifdef HAVE_MMX |
0 | 135 |
136 #include "i386/mmx.h" | |
137 | |
138 #define FILTER4(reg) \ | |
139 {\ | |
140 s = src + (src_pos >> POS_FRAC_BITS);\ | |
141 phase = get_phase(src_pos);\ | |
142 filter = filters + phase * NB_TAPS;\ | |
143 movq_m2r(*s, reg);\ | |
144 punpcklbw_r2r(mm7, reg);\ | |
145 movq_m2r(*filter, mm6);\ | |
146 pmaddwd_r2r(reg, mm6);\ | |
147 movq_r2r(mm6, reg);\ | |
148 psrlq_i2r(32, reg);\ | |
149 paddd_r2r(mm6, reg);\ | |
150 psrad_i2r(FILTER_BITS, reg);\ | |
151 src_pos += src_incr;\ | |
152 } | |
153 | |
154 #define DUMP(reg) movq_r2m(reg, tmp); printf(#reg "=%016Lx\n", tmp.uq); | |
155 | |
156 /* XXX: do four pixels at a time */ | |
1064 | 157 static void h_resample_fast4_mmx(uint8_t *dst, int dst_width, uint8_t *src, int src_width, |
158 int src_start, int src_incr, int16_t *filters) | |
0 | 159 { |
160 int src_pos, phase; | |
1064 | 161 uint8_t *s; |
162 int16_t *filter; | |
0 | 163 mmx_t tmp; |
164 | |
165 src_pos = src_start; | |
166 pxor_r2r(mm7, mm7); | |
167 | |
168 while (dst_width >= 4) { | |
169 | |
170 FILTER4(mm0); | |
171 FILTER4(mm1); | |
172 FILTER4(mm2); | |
173 FILTER4(mm3); | |
174 | |
175 packuswb_r2r(mm7, mm0); | |
176 packuswb_r2r(mm7, mm1); | |
177 packuswb_r2r(mm7, mm3); | |
178 packuswb_r2r(mm7, mm2); | |
179 movq_r2m(mm0, tmp); | |
180 dst[0] = tmp.ub[0]; | |
181 movq_r2m(mm1, tmp); | |
182 dst[1] = tmp.ub[0]; | |
183 movq_r2m(mm2, tmp); | |
184 dst[2] = tmp.ub[0]; | |
185 movq_r2m(mm3, tmp); | |
186 dst[3] = tmp.ub[0]; | |
187 dst += 4; | |
188 dst_width -= 4; | |
189 } | |
190 while (dst_width > 0) { | |
191 FILTER4(mm0); | |
192 packuswb_r2r(mm7, mm0); | |
193 movq_r2m(mm0, tmp); | |
194 dst[0] = tmp.ub[0]; | |
195 dst++; | |
196 dst_width--; | |
197 } | |
198 emms(); | |
199 } | |
200 | |
1064 | 201 static void v_resample4_mmx(uint8_t *dst, int dst_width, uint8_t *src, int wrap, |
202 int16_t *filter) | |
0 | 203 { |
204 int sum, i, v; | |
1064 | 205 uint8_t *s; |
0 | 206 mmx_t tmp; |
207 mmx_t coefs[4]; | |
208 | |
209 for(i=0;i<4;i++) { | |
210 v = filter[i]; | |
211 coefs[i].uw[0] = v; | |
212 coefs[i].uw[1] = v; | |
213 coefs[i].uw[2] = v; | |
214 coefs[i].uw[3] = v; | |
215 } | |
216 | |
217 pxor_r2r(mm7, mm7); | |
218 s = src; | |
219 while (dst_width >= 4) { | |
220 movq_m2r(s[0 * wrap], mm0); | |
221 punpcklbw_r2r(mm7, mm0); | |
222 movq_m2r(s[1 * wrap], mm1); | |
223 punpcklbw_r2r(mm7, mm1); | |
224 movq_m2r(s[2 * wrap], mm2); | |
225 punpcklbw_r2r(mm7, mm2); | |
226 movq_m2r(s[3 * wrap], mm3); | |
227 punpcklbw_r2r(mm7, mm3); | |
228 | |
229 pmullw_m2r(coefs[0], mm0); | |
230 pmullw_m2r(coefs[1], mm1); | |
231 pmullw_m2r(coefs[2], mm2); | |
232 pmullw_m2r(coefs[3], mm3); | |
233 | |
234 paddw_r2r(mm1, mm0); | |
235 paddw_r2r(mm3, mm2); | |
236 paddw_r2r(mm2, mm0); | |
237 psraw_i2r(FILTER_BITS, mm0); | |
238 | |
239 packuswb_r2r(mm7, mm0); | |
240 movq_r2m(mm0, tmp); | |
241 | |
1064 | 242 *(uint32_t *)dst = tmp.ud[0]; |
0 | 243 dst += 4; |
244 s += 4; | |
245 dst_width -= 4; | |
246 } | |
247 while (dst_width > 0) { | |
248 sum = s[0 * wrap] * filter[0] + | |
249 s[1 * wrap] * filter[1] + | |
250 s[2 * wrap] * filter[2] + | |
251 s[3 * wrap] * filter[3]; | |
252 sum = sum >> FILTER_BITS; | |
253 if (sum < 0) | |
254 sum = 0; | |
255 else if (sum > 255) | |
256 sum = 255; | |
257 dst[0] = sum; | |
258 dst++; | |
259 s++; | |
260 dst_width--; | |
261 } | |
262 emms(); | |
263 } | |
264 #endif | |
265 | |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
266 #ifdef HAVE_ALTIVEC |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
267 typedef union { |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
268 vector unsigned char v; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
269 unsigned char c[16]; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
270 } vec_uc_t; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
271 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
272 typedef union { |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
273 vector signed short v; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
274 signed short s[8]; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
275 } vec_ss_t; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
276 |
1064 | 277 void v_resample16_altivec(uint8_t *dst, int dst_width, uint8_t *src, int wrap, |
278 int16_t *filter) | |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
279 { |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
280 int sum, i; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
281 uint8_t *s; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
282 vector unsigned char *tv, tmp, dstv, zero; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
283 vec_ss_t srchv[4], srclv[4], fv[4]; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
284 vector signed short zeros, sumhv, sumlv; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
285 s = src; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
286 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
287 for(i=0;i<4;i++) |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
288 { |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
289 /* |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
290 The vec_madds later on does an implicit >>15 on the result. |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
291 Since FILTER_BITS is 8, and we have 15 bits of magnitude in |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
292 a signed short, we have just enough bits to pre-shift our |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
293 filter constants <<7 to compensate for vec_madds. |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
294 */ |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
295 fv[i].s[0] = filter[i] << (15-FILTER_BITS); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
296 fv[i].v = vec_splat(fv[i].v, 0); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
297 } |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
298 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
299 zero = vec_splat_u8(0); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
300 zeros = vec_splat_s16(0); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
301 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
302 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
303 /* |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
304 When we're resampling, we'd ideally like both our input buffers, |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
305 and output buffers to be 16-byte aligned, so we can do both aligned |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
306 reads and writes. Sadly we can't always have this at the moment, so |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
307 we opt for aligned writes, as unaligned writes have a huge overhead. |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
308 To do this, do enough scalar resamples to get dst 16-byte aligned. |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
309 */ |
898
6d5e3fe7aea1
Simplify an expression and eliminate a compile warning
philipjsg
parents:
894
diff
changeset
|
310 i = (-(int)dst) & 0xf; |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
311 while(i>0) { |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
312 sum = s[0 * wrap] * filter[0] + |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
313 s[1 * wrap] * filter[1] + |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
314 s[2 * wrap] * filter[2] + |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
315 s[3 * wrap] * filter[3]; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
316 sum = sum >> FILTER_BITS; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
317 if (sum<0) sum = 0; else if (sum>255) sum=255; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
318 dst[0] = sum; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
319 dst++; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
320 s++; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
321 dst_width--; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
322 i--; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
323 } |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
324 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
325 /* Do our altivec resampling on 16 pixels at once. */ |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
326 while(dst_width>=16) { |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
327 /* |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
328 Read 16 (potentially unaligned) bytes from each of |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
329 4 lines into 4 vectors, and split them into shorts. |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
330 Interleave the multipy/accumulate for the resample |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
331 filter with the loads to hide the 3 cycle latency |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
332 the vec_madds have. |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
333 */ |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
334 tv = (vector unsigned char *) &s[0 * wrap]; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
335 tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[i * wrap])); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
336 srchv[0].v = (vector signed short) vec_mergeh(zero, tmp); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
337 srclv[0].v = (vector signed short) vec_mergel(zero, tmp); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
338 sumhv = vec_madds(srchv[0].v, fv[0].v, zeros); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
339 sumlv = vec_madds(srclv[0].v, fv[0].v, zeros); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
340 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
341 tv = (vector unsigned char *) &s[1 * wrap]; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
342 tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[1 * wrap])); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
343 srchv[1].v = (vector signed short) vec_mergeh(zero, tmp); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
344 srclv[1].v = (vector signed short) vec_mergel(zero, tmp); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
345 sumhv = vec_madds(srchv[1].v, fv[1].v, sumhv); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
346 sumlv = vec_madds(srclv[1].v, fv[1].v, sumlv); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
347 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
348 tv = (vector unsigned char *) &s[2 * wrap]; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
349 tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[2 * wrap])); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
350 srchv[2].v = (vector signed short) vec_mergeh(zero, tmp); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
351 srclv[2].v = (vector signed short) vec_mergel(zero, tmp); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
352 sumhv = vec_madds(srchv[2].v, fv[2].v, sumhv); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
353 sumlv = vec_madds(srclv[2].v, fv[2].v, sumlv); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
354 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
355 tv = (vector unsigned char *) &s[3 * wrap]; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
356 tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[3 * wrap])); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
357 srchv[3].v = (vector signed short) vec_mergeh(zero, tmp); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
358 srclv[3].v = (vector signed short) vec_mergel(zero, tmp); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
359 sumhv = vec_madds(srchv[3].v, fv[3].v, sumhv); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
360 sumlv = vec_madds(srclv[3].v, fv[3].v, sumlv); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
361 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
362 /* |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
363 Pack the results into our destination vector, |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
364 and do an aligned write of that back to memory. |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
365 */ |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
366 dstv = vec_packsu(sumhv, sumlv) ; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
367 vec_st(dstv, 0, (vector unsigned char *) dst); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
368 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
369 dst+=16; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
370 s+=16; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
371 dst_width-=16; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
372 } |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
373 |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
374 /* |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
375 If there are any leftover pixels, resample them |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
376 with the slow scalar method. |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
377 */ |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
378 while(dst_width>0) { |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
379 sum = s[0 * wrap] * filter[0] + |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
380 s[1 * wrap] * filter[1] + |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
381 s[2 * wrap] * filter[2] + |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
382 s[3 * wrap] * filter[3]; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
383 sum = sum >> FILTER_BITS; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
384 if (sum<0) sum = 0; else if (sum>255) sum=255; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
385 dst[0] = sum; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
386 dst++; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
387 s++; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
388 dst_width--; |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
389 } |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
390 } |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
391 #endif |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
392 |
0 | 393 /* slow version to handle limit cases. Does not need optimisation */ |
1064 | 394 static void h_resample_slow(uint8_t *dst, int dst_width, uint8_t *src, int src_width, |
395 int src_start, int src_incr, int16_t *filters) | |
0 | 396 { |
397 int src_pos, phase, sum, j, v, i; | |
1064 | 398 uint8_t *s, *src_end; |
399 int16_t *filter; | |
0 | 400 |
401 src_end = src + src_width; | |
402 src_pos = src_start; | |
403 for(i=0;i<dst_width;i++) { | |
404 s = src + (src_pos >> POS_FRAC_BITS); | |
405 phase = get_phase(src_pos); | |
406 filter = filters + phase * NB_TAPS; | |
407 sum = 0; | |
408 for(j=0;j<NB_TAPS;j++) { | |
409 if (s < src) | |
410 v = src[0]; | |
411 else if (s >= src_end) | |
412 v = src_end[-1]; | |
413 else | |
414 v = s[0]; | |
415 sum += v * filter[j]; | |
416 s++; | |
417 } | |
418 sum = sum >> FILTER_BITS; | |
419 if (sum < 0) | |
420 sum = 0; | |
421 else if (sum > 255) | |
422 sum = 255; | |
423 dst[0] = sum; | |
424 src_pos += src_incr; | |
425 dst++; | |
426 } | |
427 } | |
428 | |
1064 | 429 static void h_resample(uint8_t *dst, int dst_width, uint8_t *src, int src_width, |
430 int src_start, int src_incr, int16_t *filters) | |
0 | 431 { |
432 int n, src_end; | |
433 | |
434 if (src_start < 0) { | |
435 n = (0 - src_start + src_incr - 1) / src_incr; | |
436 h_resample_slow(dst, n, src, src_width, src_start, src_incr, filters); | |
437 dst += n; | |
438 dst_width -= n; | |
439 src_start += n * src_incr; | |
440 } | |
441 src_end = src_start + dst_width * src_incr; | |
442 if (src_end > ((src_width - NB_TAPS) << POS_FRAC_BITS)) { | |
443 n = (((src_width - NB_TAPS + 1) << POS_FRAC_BITS) - 1 - src_start) / | |
444 src_incr; | |
445 } else { | |
446 n = dst_width; | |
447 } | |
2 | 448 #ifdef HAVE_MMX |
0 | 449 if ((mm_flags & MM_MMX) && NB_TAPS == 4) |
450 h_resample_fast4_mmx(dst, n, | |
451 src, src_width, src_start, src_incr, filters); | |
452 else | |
453 #endif | |
454 h_resample_fast(dst, n, | |
455 src, src_width, src_start, src_incr, filters); | |
456 if (n < dst_width) { | |
457 dst += n; | |
458 dst_width -= n; | |
459 src_start += n * src_incr; | |
460 h_resample_slow(dst, dst_width, | |
461 src, src_width, src_start, src_incr, filters); | |
462 } | |
463 } | |
464 | |
465 static void component_resample(ImgReSampleContext *s, | |
1064 | 466 uint8_t *output, int owrap, int owidth, int oheight, |
467 uint8_t *input, int iwrap, int iwidth, int iheight) | |
0 | 468 { |
469 int src_y, src_y1, last_src_y, ring_y, phase_y, y1, y; | |
1064 | 470 uint8_t *new_line, *src_line; |
0 | 471 |
472 last_src_y = - FCENTER - 1; | |
473 /* position of the bottom of the filter in the source image */ | |
474 src_y = (last_src_y + NB_TAPS) * POS_FRAC; | |
475 ring_y = NB_TAPS; /* position in ring buffer */ | |
476 for(y=0;y<oheight;y++) { | |
477 /* apply horizontal filter on new lines from input if needed */ | |
478 src_y1 = src_y >> POS_FRAC_BITS; | |
479 while (last_src_y < src_y1) { | |
480 if (++ring_y >= LINE_BUF_HEIGHT + NB_TAPS) | |
481 ring_y = NB_TAPS; | |
482 last_src_y++; | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
483 /* handle limit conditions : replicate line (slightly |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
484 inefficient because we filter multiple times) */ |
0 | 485 y1 = last_src_y; |
486 if (y1 < 0) { | |
487 y1 = 0; | |
488 } else if (y1 >= iheight) { | |
489 y1 = iheight - 1; | |
490 } | |
491 src_line = input + y1 * iwrap; | |
492 new_line = s->line_buf + ring_y * owidth; | |
493 /* apply filter and handle limit cases correctly */ | |
494 h_resample(new_line, owidth, | |
495 src_line, iwidth, - FCENTER * POS_FRAC, s->h_incr, | |
496 &s->h_filters[0][0]); | |
497 /* handle ring buffer wraping */ | |
498 if (ring_y >= LINE_BUF_HEIGHT) { | |
499 memcpy(s->line_buf + (ring_y - LINE_BUF_HEIGHT) * owidth, | |
500 new_line, owidth); | |
501 } | |
502 } | |
503 /* apply vertical filter */ | |
504 phase_y = get_phase(src_y); | |
2 | 505 #ifdef HAVE_MMX |
0 | 506 /* desactivated MMX because loss of precision */ |
507 if ((mm_flags & MM_MMX) && NB_TAPS == 4 && 0) | |
508 v_resample4_mmx(output, owidth, | |
509 s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth, | |
510 &s->v_filters[phase_y][0]); | |
511 else | |
512 #endif | |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
513 #ifdef HAVE_ALTIVEC |
920
a0ad8e3452f2
practically disabling altivec resampling code (some ppl said its broken) patch by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
898
diff
changeset
|
514 if ((mm_flags & MM_ALTIVEC) && NB_TAPS == 4 && FILTER_BITS <= 6) |
894
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
515 v_resample16_altivec(output, owidth, |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
516 s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth, |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
517 &s->v_filters[phase_y][0]); |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
518 else |
a408778eff87
altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents:
653
diff
changeset
|
519 #endif |
0 | 520 v_resample(output, owidth, |
521 s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth, | |
522 &s->v_filters[phase_y][0]); | |
523 | |
524 src_y += s->v_incr; | |
525 output += owrap; | |
526 } | |
527 } | |
528 | |
529 /* XXX: the following filter is quite naive, but it seems to suffice | |
530 for 4 taps */ | |
1064 | 531 static void build_filter(int16_t *filter, float factor) |
0 | 532 { |
533 int ph, i, v; | |
534 float x, y, tab[NB_TAPS], norm, mult; | |
535 | |
536 /* if upsampling, only need to interpolate, no filter */ | |
537 if (factor > 1.0) | |
538 factor = 1.0; | |
539 | |
540 for(ph=0;ph<NB_PHASES;ph++) { | |
541 norm = 0; | |
542 for(i=0;i<NB_TAPS;i++) { | |
543 | |
544 x = M_PI * ((float)(i - FCENTER) - (float)ph / NB_PHASES) * factor; | |
545 if (x == 0) | |
546 y = 1.0; | |
547 else | |
548 y = sin(x) / x; | |
549 tab[i] = y; | |
550 norm += y; | |
551 } | |
552 | |
553 /* normalize so that an uniform color remains the same */ | |
554 mult = (float)(1 << FILTER_BITS) / norm; | |
555 for(i=0;i<NB_TAPS;i++) { | |
556 v = (int)(tab[i] * mult); | |
557 filter[ph * NB_TAPS + i] = v; | |
558 } | |
559 } | |
560 } | |
561 | |
562 ImgReSampleContext *img_resample_init(int owidth, int oheight, | |
563 int iwidth, int iheight) | |
564 { | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
565 return img_resample_full_init(owidth, oheight, iwidth, iheight, 0, 0, 0, 0); |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
566 } |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
567 |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
568 ImgReSampleContext *img_resample_full_init(int owidth, int oheight, |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
569 int iwidth, int iheight, |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
570 int topBand, int bottomBand, |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
571 int leftBand, int rightBand) |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
572 { |
0 | 573 ImgReSampleContext *s; |
574 | |
575 s = av_mallocz(sizeof(ImgReSampleContext)); | |
576 if (!s) | |
577 return NULL; | |
578 s->line_buf = av_mallocz(owidth * (LINE_BUF_HEIGHT + NB_TAPS)); | |
579 if (!s->line_buf) | |
580 goto fail; | |
581 | |
582 s->owidth = owidth; | |
583 s->oheight = oheight; | |
584 s->iwidth = iwidth; | |
585 s->iheight = iheight; | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
586 s->topBand = topBand; |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
587 s->bottomBand = bottomBand; |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
588 s->leftBand = leftBand; |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
589 s->rightBand = rightBand; |
0 | 590 |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
591 s->h_incr = ((iwidth - leftBand - rightBand) * POS_FRAC) / owidth; |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
592 s->v_incr = ((iheight - topBand - bottomBand) * POS_FRAC) / oheight; |
0 | 593 |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
594 build_filter(&s->h_filters[0][0], (float) owidth / (float) (iwidth - leftBand - rightBand)); |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
595 build_filter(&s->v_filters[0][0], (float) oheight / (float) (iheight - topBand - bottomBand)); |
0 | 596 |
597 return s; | |
598 fail: | |
396
fce0a2520551
removed useless header includes - use av memory functions
glantau
parents:
18
diff
changeset
|
599 av_free(s); |
0 | 600 return NULL; |
601 } | |
602 | |
603 void img_resample(ImgReSampleContext *s, | |
604 AVPicture *output, AVPicture *input) | |
605 { | |
606 int i, shift; | |
607 | |
608 for(i=0;i<3;i++) { | |
609 shift = (i == 0) ? 0 : 1; | |
610 component_resample(s, output->data[i], output->linesize[i], | |
611 s->owidth >> shift, s->oheight >> shift, | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
612 input->data[i] + (input->linesize[i] * (s->topBand >> shift)) + (s->leftBand >> shift), |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
613 input->linesize[i], ((s->iwidth - s->leftBand - s->rightBand) >> shift), |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
614 (s->iheight - s->topBand - s->bottomBand) >> shift); |
0 | 615 } |
616 } | |
617 | |
618 void img_resample_close(ImgReSampleContext *s) | |
619 { | |
396
fce0a2520551
removed useless header includes - use av memory functions
glantau
parents:
18
diff
changeset
|
620 av_free(s->line_buf); |
fce0a2520551
removed useless header includes - use av memory functions
glantau
parents:
18
diff
changeset
|
621 av_free(s); |
0 | 622 } |
623 | |
624 #ifdef TEST | |
625 | |
626 void *av_mallocz(int size) | |
627 { | |
628 void *ptr; | |
629 ptr = malloc(size); | |
630 memset(ptr, 0, size); | |
631 return ptr; | |
632 } | |
633 | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
634 void av_free(void *ptr) |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
635 { |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
636 /* XXX: this test should not be needed on most libcs */ |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
637 if (ptr) |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
638 free(ptr); |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
639 } |
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
640 |
0 | 641 /* input */ |
642 #define XSIZE 256 | |
643 #define YSIZE 256 | |
1064 | 644 uint8_t img[XSIZE * YSIZE]; |
0 | 645 |
646 /* output */ | |
647 #define XSIZE1 512 | |
648 #define YSIZE1 512 | |
1064 | 649 uint8_t img1[XSIZE1 * YSIZE1]; |
650 uint8_t img2[XSIZE1 * YSIZE1]; | |
0 | 651 |
1064 | 652 void save_pgm(const char *filename, uint8_t *img, int xsize, int ysize) |
0 | 653 { |
654 FILE *f; | |
655 f=fopen(filename,"w"); | |
656 fprintf(f,"P5\n%d %d\n%d\n", xsize, ysize, 255); | |
657 fwrite(img,1, xsize * ysize,f); | |
658 fclose(f); | |
659 } | |
660 | |
1064 | 661 static void dump_filter(int16_t *filter) |
0 | 662 { |
663 int i, ph; | |
664 | |
665 for(ph=0;ph<NB_PHASES;ph++) { | |
666 printf("%2d: ", ph); | |
667 for(i=0;i<NB_TAPS;i++) { | |
668 printf(" %5.2f", filter[ph * NB_TAPS + i] / 256.0); | |
669 } | |
670 printf("\n"); | |
671 } | |
672 } | |
673 | |
2 | 674 #ifdef HAVE_MMX |
644 | 675 int mm_flags; |
0 | 676 #endif |
677 | |
678 int main(int argc, char **argv) | |
679 { | |
680 int x, y, v, i, xsize, ysize; | |
681 ImgReSampleContext *s; | |
682 float fact, factors[] = { 1/2.0, 3.0/4.0, 1.0, 4.0/3.0, 16.0/9.0, 2.0 }; | |
683 char buf[256]; | |
684 | |
685 /* build test image */ | |
686 for(y=0;y<YSIZE;y++) { | |
687 for(x=0;x<XSIZE;x++) { | |
688 if (x < XSIZE/2 && y < YSIZE/2) { | |
689 if (x < XSIZE/4 && y < YSIZE/4) { | |
690 if ((x % 10) <= 6 && | |
691 (y % 10) <= 6) | |
692 v = 0xff; | |
693 else | |
694 v = 0x00; | |
695 } else if (x < XSIZE/4) { | |
696 if (x & 1) | |
697 v = 0xff; | |
698 else | |
699 v = 0; | |
700 } else if (y < XSIZE/4) { | |
701 if (y & 1) | |
702 v = 0xff; | |
703 else | |
704 v = 0; | |
705 } else { | |
706 if (y < YSIZE*3/8) { | |
707 if ((y+x) & 1) | |
708 v = 0xff; | |
709 else | |
710 v = 0; | |
711 } else { | |
712 if (((x+3) % 4) <= 1 && | |
713 ((y+3) % 4) <= 1) | |
714 v = 0xff; | |
715 else | |
716 v = 0x00; | |
717 } | |
718 } | |
719 } else if (x < XSIZE/2) { | |
720 v = ((x - (XSIZE/2)) * 255) / (XSIZE/2); | |
721 } else if (y < XSIZE/2) { | |
722 v = ((y - (XSIZE/2)) * 255) / (XSIZE/2); | |
723 } else { | |
724 v = ((x + y - XSIZE) * 255) / XSIZE; | |
725 } | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
726 img[(YSIZE - y) * XSIZE + (XSIZE - x)] = v; |
0 | 727 } |
728 } | |
729 save_pgm("/tmp/in.pgm", img, XSIZE, YSIZE); | |
730 for(i=0;i<sizeof(factors)/sizeof(float);i++) { | |
731 fact = factors[i]; | |
732 xsize = (int)(XSIZE * fact); | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
733 ysize = (int)((YSIZE - 100) * fact); |
644 | 734 s = img_resample_full_init(xsize, ysize, XSIZE, YSIZE, 50 ,50, 0, 0); |
0 | 735 printf("Factor=%0.2f\n", fact); |
736 dump_filter(&s->h_filters[0][0]); | |
737 component_resample(s, img1, xsize, xsize, ysize, | |
630
b4ee42142ad1
croping patch by (talus25 at speakeasy dot net) with fixes from atmos & me
michaelni
parents:
429
diff
changeset
|
738 img + 50 * XSIZE, XSIZE, XSIZE, YSIZE - 100); |
0 | 739 img_resample_close(s); |
740 | |
741 sprintf(buf, "/tmp/out%d.pgm", i); | |
742 save_pgm(buf, img1, xsize, ysize); | |
743 } | |
744 | |
745 /* mmx test */ | |
2 | 746 #ifdef HAVE_MMX |
0 | 747 printf("MMX test\n"); |
748 fact = 0.72; | |
749 xsize = (int)(XSIZE * fact); | |
750 ysize = (int)(YSIZE * fact); | |
751 mm_flags = MM_MMX; | |
752 s = img_resample_init(xsize, ysize, XSIZE, YSIZE); | |
753 component_resample(s, img1, xsize, xsize, ysize, | |
754 img, XSIZE, XSIZE, YSIZE); | |
755 | |
756 mm_flags = 0; | |
757 s = img_resample_init(xsize, ysize, XSIZE, YSIZE); | |
758 component_resample(s, img2, xsize, xsize, ysize, | |
759 img, XSIZE, XSIZE, YSIZE); | |
760 if (memcmp(img1, img2, xsize * ysize) != 0) { | |
761 fprintf(stderr, "mmx error\n"); | |
762 exit(1); | |
763 } | |
764 printf("MMX OK\n"); | |
765 #endif | |
766 return 0; | |
767 } | |
768 | |
769 #endif |