sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
[libav.git] / libavcodec / sh4 / idct_sh4.c
1 /*
2 * idct for sh4
3 *
4 * Copyright (c) 2001-2003 BERO <bero@geocities.co.jp>
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 */
20
21 #include "../dsputil.h"
22 #define c1 1.38703984532214752434 /* sqrt(2)*cos(1*pi/16) */
23 #define c2 1.30656296487637657577 /* sqrt(2)*cos(2*pi/16) */
24 #define c3 1.17587560241935884520 /* sqrt(2)*cos(3*pi/16) */
25 #define c4 1.00000000000000000000 /* sqrt(2)*cos(4*pi/16) */
26 #define c5 0.78569495838710234903 /* sqrt(2)*cos(5*pi/16) */
27 #define c6 0.54119610014619712324 /* sqrt(2)*cos(6*pi/16) */
28 #define c7 0.27589937928294311353 /* sqrt(2)*cos(7*pi/16) */
29
30 const static float even_table[] __attribute__ ((aligned(8))) = {
31 c4, c4, c4, c4,
32 c2, c6,-c6,-c2,
33 c4,-c4,-c4, c4,
34 c6,-c2, c2,-c6
35 };
36
37 const static float odd_table[] __attribute__ ((aligned(8))) = {
38 c1, c3, c5, c7,
39 c3,-c7,-c1,-c5,
40 c5,-c1, c7, c3,
41 c7,-c5, c3,-c1
42 };
43
44 #undef c1
45 #undef c2
46 #undef c3
47 #undef c4
48 #undef c5
49 #undef c6
50 #undef c7
51
52 #if defined(__SH4_SINGLE__) || defined(__SH4_SINGLE_ONLY__)
53
54 #define load_matrix(table) \
55 __asm__ volatile( \
56 " fschg\n" \
57 " fmov @%0+,xd0\n" \
58 " fmov @%0+,xd2\n" \
59 " fmov @%0+,xd4\n" \
60 " fmov @%0+,xd6\n" \
61 " fmov @%0+,xd8\n" \
62 " fmov @%0+,xd10\n" \
63 " fmov @%0+,xd12\n" \
64 " fmov @%0+,xd14\n" \
65 " fschg\n" \
66 :\
67 : "r"(table)\
68 : "0" \
69 )
70
71 #define ftrv() \
72 __asm__ volatile("ftrv xmtrx,fv0" \
73 : "=f"(fr0),"=f"(fr1),"=f"(fr2),"=f"(fr3) \
74 : "0"(fr0), "1"(fr1), "2"(fr2), "3"(fr3) );
75
76 #define DEFREG \
77 register float fr0 __asm__("fr0"); \
78 register float fr1 __asm__("fr1"); \
79 register float fr2 __asm__("fr2"); \
80 register float fr3 __asm__("fr3")
81
82 #else
83
84 /* generic C code for check */
85
86 static void ftrv_(const float xf[],float fv[])
87 {
88 float f0,f1,f2,f3;
89 f0 = fv[0];
90 f1 = fv[1];
91 f2 = fv[2];
92 f3 = fv[3];
93 fv[0] = xf[0]*f0 + xf[4]*f1 + xf[ 8]*f2 + xf[12]*f3;
94 fv[1] = xf[1]*f0 + xf[5]*f1 + xf[ 9]*f2 + xf[13]*f3;
95 fv[2] = xf[2]*f0 + xf[6]*f1 + xf[10]*f2 + xf[14]*f3;
96 fv[3] = xf[3]*f0 + xf[7]*f1 + xf[11]*f2 + xf[15]*f3;
97 }
98
99 static void load_matrix_(float xf[],const float table[])
100 {
101 int i;
102 for(i=0;i<16;i++) xf[i]=table[i];
103 }
104
105 #define ftrv() ftrv_(xf,fv)
106 #define load_matrix(table) load_matrix_(xf,table)
107
108 #define DEFREG \
109 float fv[4],xf[16]
110
111 #define fr0 fv[0]
112 #define fr1 fv[1]
113 #define fr2 fv[2]
114 #define fr3 fv[3]
115
116 #endif
117
118 #if 1
119 #define DESCALE(x,n) (x)*(1.0f/(1<<(n)))
120 #else
121 #define DESCALE(x,n) (((int)(x)+(1<<(n-1)))>>(n))
122 #endif
123
124 /* this code work worse on gcc cvs. 3.2.3 work fine */
125
126
127 #if 1
128 //optimized
129
130 void idct_sh4(DCTELEM *block)
131 {
132 DEFREG;
133
134 int i;
135 float tblock[8*8],*fblock;
136 int ofs1,ofs2,ofs3;
137
138 #if defined(__SH4__)
139 #error "FIXME!! change to single float"
140 #endif
141
142 /* row */
143
144 /* even part */
145 load_matrix(even_table);
146
147 fblock = tblock+4;
148 i = 8;
149 do {
150 fr0 = block[0];
151 fr1 = block[2];
152 fr2 = block[4];
153 fr3 = block[6];
154 block+=8;
155 ftrv();
156 *--fblock = fr3;
157 *--fblock = fr2;
158 *--fblock = fr1;
159 *--fblock = fr0;
160 fblock+=8+4;
161 } while(--i);
162 block-=8*8;
163 fblock-=8*8+4;
164
165 load_matrix(odd_table);
166
167 i = 8;
168
169 // ofs1 = sizeof(float)*1;
170 // ofs2 = sizeof(float)*2;
171 // ofs3 = sizeof(float)*3;
172
173 do {
174 float t0,t1,t2,t3;
175 fr0 = block[1];
176 fr1 = block[3];
177 fr2 = block[5];
178 fr3 = block[7];
179 block+=8;
180 ftrv();
181 t0 = *fblock++;
182 t1 = *fblock++;
183 t2 = *fblock++;
184 t3 = *fblock++;
185 fblock+=4;
186 *--fblock = t0 - fr0;
187 *--fblock = t1 - fr1;
188 *--fblock = t2 - fr2;
189 *--fblock = t3 - fr3;
190 *--fblock = t3 + fr3;
191 *--fblock = t2 + fr2;
192 *--fblock = t1 + fr1;
193 *--fblock = t0 + fr0;
194 fblock+=8;
195 } while(--i);
196 block-=8*8;
197 fblock-=8*8;
198
199 /* col */
200
201 /* even part */
202 load_matrix(even_table);
203
204 ofs1 = sizeof(float)*2*8;
205 ofs2 = sizeof(float)*4*8;
206 ofs3 = sizeof(float)*6*8;
207
208 i = 8;
209
210 #define OA(fblock,ofs) *(float*)((char*)fblock + ofs)
211
212 do {
213 fr0 = OA(fblock, 0);
214 fr1 = OA(fblock,ofs1);
215 fr2 = OA(fblock,ofs2);
216 fr3 = OA(fblock,ofs3);
217 ftrv();
218 OA(fblock,0 ) = fr0;
219 OA(fblock,ofs1) = fr1;
220 OA(fblock,ofs2) = fr2;
221 OA(fblock,ofs3) = fr3;
222 fblock++;
223 } while(--i);
224 fblock-=8;
225
226 load_matrix(odd_table);
227
228 i=8;
229 do {
230 float t0,t1,t2,t3;
231 t0 = OA(fblock, 0); /* [8*0] */
232 t1 = OA(fblock,ofs1); /* [8*2] */
233 t2 = OA(fblock,ofs2); /* [8*4] */
234 t3 = OA(fblock,ofs3); /* [8*6] */
235 fblock+=8;
236 fr0 = OA(fblock, 0); /* [8*1] */
237 fr1 = OA(fblock,ofs1); /* [8*3] */
238 fr2 = OA(fblock,ofs2); /* [8*5] */
239 fr3 = OA(fblock,ofs3); /* [8*7] */
240 fblock+=-8+1;
241 ftrv();
242 block[8*0] = DESCALE(t0 + fr0,3);
243 block[8*7] = DESCALE(t0 - fr0,3);
244 block[8*1] = DESCALE(t1 + fr1,3);
245 block[8*6] = DESCALE(t1 - fr1,3);
246 block[8*2] = DESCALE(t2 + fr2,3);
247 block[8*5] = DESCALE(t2 - fr2,3);
248 block[8*3] = DESCALE(t3 + fr3,3);
249 block[8*4] = DESCALE(t3 - fr3,3);
250 block++;
251 } while(--i);
252
253 #if defined(__SH4__)
254 #error "FIXME!! change to double"
255 #endif
256 }
257 #else
258 void idct_sh4(DCTELEM *block)
259 {
260 DEFREG;
261
262 int i;
263 float tblock[8*8],*fblock;
264
265 /* row */
266
267 /* even part */
268 load_matrix(even_table);
269
270 fblock = tblock;
271 i = 8;
272 do {
273 fr0 = block[0];
274 fr1 = block[2];
275 fr2 = block[4];
276 fr3 = block[6];
277 block+=8;
278 ftrv();
279 fblock[0] = fr0;
280 fblock[2] = fr1;
281 fblock[4] = fr2;
282 fblock[6] = fr3;
283 fblock+=8;
284 } while(--i);
285 block-=8*8;
286 fblock-=8*8;
287
288 load_matrix(odd_table);
289
290 i = 8;
291
292 do {
293 float t0,t1,t2,t3;
294 fr0 = block[1];
295 fr1 = block[3];
296 fr2 = block[5];
297 fr3 = block[7];
298 block+=8;
299 ftrv();
300 t0 = fblock[0];
301 t1 = fblock[2];
302 t2 = fblock[4];
303 t3 = fblock[6];
304 fblock[0] = t0 + fr0;
305 fblock[7] = t0 - fr0;
306 fblock[1] = t1 + fr1;
307 fblock[6] = t1 - fr1;
308 fblock[2] = t2 + fr2;
309 fblock[5] = t2 - fr2;
310 fblock[3] = t3 + fr3;
311 fblock[4] = t3 - fr3;
312 fblock+=8;
313 } while(--i);
314 block-=8*8;
315 fblock-=8*8;
316
317 /* col */
318
319 /* even part */
320 load_matrix(even_table);
321
322 i = 8;
323
324 do {
325 fr0 = fblock[8*0];
326 fr1 = fblock[8*2];
327 fr2 = fblock[8*4];
328 fr3 = fblock[8*6];
329 ftrv();
330 fblock[8*0] = fr0;
331 fblock[8*2] = fr1;
332 fblock[8*4] = fr2;
333 fblock[8*6] = fr3;
334 fblock++;
335 } while(--i);
336 fblock-=8;
337
338 load_matrix(odd_table);
339
340 i=8;
341 do {
342 float t0,t1,t2,t3;
343 fr0 = fblock[8*1];
344 fr1 = fblock[8*3];
345 fr2 = fblock[8*5];
346 fr3 = fblock[8*7];
347 ftrv();
348 t0 = fblock[8*0];
349 t1 = fblock[8*2];
350 t2 = fblock[8*4];
351 t3 = fblock[8*6];
352 fblock++;
353 block[8*0] = DESCALE(t0 + fr0,3);
354 block[8*7] = DESCALE(t0 - fr0,3);
355 block[8*1] = DESCALE(t1 + fr1,3);
356 block[8*6] = DESCALE(t1 - fr1,3);
357 block[8*2] = DESCALE(t2 + fr2,3);
358 block[8*5] = DESCALE(t2 - fr2,3);
359 block[8*3] = DESCALE(t3 + fr3,3);
360 block[8*4] = DESCALE(t3 - fr3,3);
361 block++;
362 } while(--i);
363 }
364 #endif