Mercurial > libavcodec.hg

comparison ppc/fdct_altivec.c @ 1578:6a4cfc5f9f96 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
AltiVec optimized fdct patch by (James Klicman <james at klicman dot org>)
author michael
date Sun, 26 Oct 2003 10:14:05 +0000
parents
children 210cca8f5764
comparison
equal deleted inserted replaced
1577:dbd0ede6623d 1578:6a4cfc5f9f96
1 /* ffmpeg/libavcodec/ppc/fdct_altivec.c, this file is part of the
2 * AltiVec optimized library for the FFMPEG Multimedia System
3 * Copyright (C) 2003 James Klicman <james@klicman.org>
4 *
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 2.1 of the License, or (at your option) any later version.
9 *
10 * This library is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * Lesser General Public License for more details.
14 *
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
18 */
19
20
21 #include "../common.h"
22 #include "../dsputil.h"
23 #include "dsputil_altivec.h"
24 #include "gcc_fixes.h"
25
26
27 #define vs16(v) ((vector signed short)(v))
28 #define vs32(v) ((vector signed int)(v))
29 #define vu8(v) ((vector unsigned char)(v))
30 #define vu16(v) ((vector unsigned short)(v))
31 #define vu32(v) ((vector unsigned int)(v))
32
33
34 #define C1 0.98078525066375732421875000 /* cos(1*PI/16) */
35 #define C2 0.92387950420379638671875000 /* cos(2*PI/16) */
36 #define C3 0.83146959543228149414062500 /* cos(3*PI/16) */
37 #define C4 0.70710676908493041992187500 /* cos(4*PI/16) */
38 #define C5 0.55557024478912353515625000 /* cos(5*PI/16) */
39 #define C6 0.38268342614173889160156250 /* cos(6*PI/16) */
40 #define C7 0.19509032368659973144531250 /* cos(7*PI/16) */
41 #define SQRT_2 1.41421353816986083984375000 /* sqrt(2) */
42
43
44 #define W0 -(2 * C2)
45 #define W1 (2 * C6)
46 #define W2 (SQRT_2 * C6)
47 #define W3 (SQRT_2 * C3)
48 #define W4 (SQRT_2 * (-C1 + C3 + C5 - C7))
49 #define W5 (SQRT_2 * ( C1 + C3 - C5 + C7))
50 #define W6 (SQRT_2 * ( C1 + C3 + C5 - C7))
51 #define W7 (SQRT_2 * ( C1 + C3 - C5 - C7))
52 #define W8 (SQRT_2 * ( C7 - C3))
53 #define W9 (SQRT_2 * (-C1 - C3))
54 #define WA (SQRT_2 * (-C3 - C5))
55 #define WB (SQRT_2 * ( C5 - C3))
56
57
58 static vector float fdctconsts[3] = {
59 (vector float)AVV( W0, W1, W2, W3 ),
60 (vector float)AVV( W4, W5, W6, W7 ),
61 (vector float)AVV( W8, W9, WA, WB )
62 };
63
64 #define LD_W0 vec_splat(cnsts0, 0)
65 #define LD_W1 vec_splat(cnsts0, 1)
66 #define LD_W2 vec_splat(cnsts0, 2)
67 #define LD_W3 vec_splat(cnsts0, 3)
68 #define LD_W4 vec_splat(cnsts1, 0)
69 #define LD_W5 vec_splat(cnsts1, 1)
70 #define LD_W6 vec_splat(cnsts1, 2)
71 #define LD_W7 vec_splat(cnsts1, 3)
72 #define LD_W8 vec_splat(cnsts2, 0)
73 #define LD_W9 vec_splat(cnsts2, 1)
74 #define LD_WA vec_splat(cnsts2, 2)
75 #define LD_WB vec_splat(cnsts2, 3)
76
77
78 #define FDCTROW(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \
79 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
80 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
81 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
82 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
83 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
84 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
85 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
86 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
87 \
88 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
89 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
90 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
91 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
92 \
93 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
94 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
95 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
96 cnst = LD_W2; \
97 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
98 cnst = LD_W1; \
99 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
100 cnst = LD_W0; \
101 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
102 \
103 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
104 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
105 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
106 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
107 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
108 cnst = LD_W3; \
109 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
110 \
111 cnst = LD_W8; \
112 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
113 cnst = LD_W9; \
114 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
115 cnst = LD_WA; \
116 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
117 cnst = LD_WB; \
118 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
119 \
120 cnst = LD_W4; \
121 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
122 cnst = LD_W5; \
123 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
124 cnst = LD_W6; \
125 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
126 cnst = LD_W7; \
127 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
128 \
129 b7 = vec_add(b7, x2); /* b7 = b7 + x2; */ \
130 b5 = vec_add(b5, x3); /* b5 = b5 + x3; */ \
131 b3 = vec_add(b3, x2); /* b3 = b3 + x2; */ \
132 b1 = vec_add(b1, x3); /* b1 = b1 + x3; */ \
133 /* }}} */
134
135 #define FDCTCOL(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \
136 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
137 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
138 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
139 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
140 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
141 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
142 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
143 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
144 \
145 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
146 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
147 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
148 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
149 \
150 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
151 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
152 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
153 cnst = LD_W2; \
154 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
155 cnst = LD_W1; \
156 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
157 cnst = LD_W0; \
158 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
159 \
160 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
161 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
162 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
163 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
164 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
165 cnst = LD_W3; \
166 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
167 \
168 cnst = LD_W8; \
169 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
170 cnst = LD_W9; \
171 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
172 cnst = LD_WA; \
173 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
174 cnst = LD_WB; \
175 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
176 \
177 cnst = LD_W4; \
178 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
179 cnst = LD_W5; \
180 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
181 cnst = LD_W6; \
182 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
183 cnst = LD_W7; \
184 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
185 \
186 b7 = vec_add(b7, x2); /* b7 += x2; */ \
187 b5 = vec_add(b5, x3); /* b5 += x3; */ \
188 b3 = vec_add(b3, x2); /* b3 += x2; */ \
189 b1 = vec_add(b1, x3); /* b1 += x3; */ \
190 /* }}} */
191
192
193
194 /* two dimensional discrete cosine transform */
195
196 void fdct_altivec(int16_t *block)
197 {
198 POWERPC_PERF_DECLARE(altivec_fdct, 1);
199 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
200 POWERPC_PERF_START_COUNT(altivec_fdct, 1);
201 void ff_jpeg_fdct_islow(int16_t *block);
202 ff_jpeg_fdct_islow(block);
203 POWERPC_PERF_STOP_COUNT(altivec_fdct, 1);
204 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
205 vector signed short *bp;
206 vector float *cp;
207 vector float b00, b10, b20, b30, b40, b50, b60, b70;
208 vector float b01, b11, b21, b31, b41, b51, b61, b71;
209 vector float mzero, cnst, cnsts0, cnsts1, cnsts2;
210 vector float x0, x1, x2, x3, x4, x5, x6, x7, x8;
211
212 POWERPC_PERF_START_COUNT(altivec_fdct, 1);
213
214
215 /* setup constants {{{ */
216 /* mzero = -0.0 */
217 vu32(mzero) = vec_splat_u32(-1);
218 vu32(mzero) = vec_sl(vu32(mzero), vu32(mzero));
219 cp = fdctconsts;
220 cnsts0 = vec_ld(0, cp); cp++;
221 cnsts1 = vec_ld(0, cp); cp++;
222 cnsts2 = vec_ld(0, cp);
223 /* }}} */
224
225
226 /* 8x8 matrix transpose (vector short[8]) {{{ */
227 #define MERGE_S16(hl,a,b) vec_merge##hl(vs16(a), vs16(b))
228
229 bp = (vector signed short*)block;
230 vs16(b00) = vec_ld(0, bp);
231 vs16(b40) = vec_ld(16*4, bp);
232 vs16(b01) = MERGE_S16(h, b00, b40);
233 vs16(b11) = MERGE_S16(l, b00, b40);
234 bp++;
235 vs16(b10) = vec_ld(0, bp);
236 vs16(b50) = vec_ld(16*4, bp);
237 vs16(b21) = MERGE_S16(h, b10, b50);
238 vs16(b31) = MERGE_S16(l, b10, b50);
239 bp++;
240 vs16(b20) = vec_ld(0, bp);
241 vs16(b60) = vec_ld(16*4, bp);
242 vs16(b41) = MERGE_S16(h, b20, b60);
243 vs16(b51) = MERGE_S16(l, b20, b60);
244 bp++;
245 vs16(b30) = vec_ld(0, bp);
246 vs16(b70) = vec_ld(16*4, bp);
247 vs16(b61) = MERGE_S16(h, b30, b70);
248 vs16(b71) = MERGE_S16(l, b30, b70);
249
250 vs16(x0) = MERGE_S16(h, b01, b41);
251 vs16(x1) = MERGE_S16(l, b01, b41);
252 vs16(x2) = MERGE_S16(h, b11, b51);
253 vs16(x3) = MERGE_S16(l, b11, b51);
254 vs16(x4) = MERGE_S16(h, b21, b61);
255 vs16(x5) = MERGE_S16(l, b21, b61);
256 vs16(x6) = MERGE_S16(h, b31, b71);
257 vs16(x7) = MERGE_S16(l, b31, b71);
258
259 vs16(b00) = MERGE_S16(h, x0, x4);
260 vs16(b10) = MERGE_S16(l, x0, x4);
261 vs16(b20) = MERGE_S16(h, x1, x5);
262 vs16(b30) = MERGE_S16(l, x1, x5);
263 vs16(b40) = MERGE_S16(h, x2, x6);
264 vs16(b50) = MERGE_S16(l, x2, x6);
265 vs16(b60) = MERGE_S16(h, x3, x7);
266 vs16(b70) = MERGE_S16(l, x3, x7);
267
268 #undef MERGE_S16
269 /* }}} */
270
271
272 /* Some of the initial calculations can be done as vector short before
273 * conversion to vector float. The following code section takes advantage
274 * of this.
275 */
276 #if 1
277 /* fdct rows {{{ */
278 vs16(x0) = vec_add(vs16(b00), vs16(b70));
279 vs16(x7) = vec_sub(vs16(b00), vs16(b70));
280 vs16(x1) = vec_add(vs16(b10), vs16(b60));
281 vs16(x6) = vec_sub(vs16(b10), vs16(b60));
282 vs16(x2) = vec_add(vs16(b20), vs16(b50));
283 vs16(x5) = vec_sub(vs16(b20), vs16(b50));
284 vs16(x3) = vec_add(vs16(b30), vs16(b40));
285 vs16(x4) = vec_sub(vs16(b30), vs16(b40));
286
287 vs16(b70) = vec_add(vs16(x0), vs16(x3));
288 vs16(b10) = vec_add(vs16(x1), vs16(x2));
289
290 vs16(b00) = vec_add(vs16(b70), vs16(b10));
291 vs16(b40) = vec_sub(vs16(b70), vs16(b10));
292
293 #define CTF0(n) \
294 vs32(b##n##1) = vec_unpackl(vs16(b##n##0)); \
295 vs32(b##n##0) = vec_unpackh(vs16(b##n##0)); \
296 b##n##1 = vec_ctf(vs32(b##n##1), 0); \
297 b##n##0 = vec_ctf(vs32(b##n##0), 0);
298
299 CTF0(0);
300 CTF0(4);
301
302 vs16(b20) = vec_sub(vs16(x0), vs16(x3));
303 vs16(b60) = vec_sub(vs16(x1), vs16(x2));
304
305 CTF0(2);
306 CTF0(6);
307
308 #undef CTF0
309
310 x0 = vec_add(b60, b20);
311 x1 = vec_add(b61, b21);
312
313 cnst = LD_W2;
314 x0 = vec_madd(cnst, x0, mzero);
315 x1 = vec_madd(cnst, x1, mzero);
316 cnst = LD_W1;
317 b20 = vec_madd(cnst, b20, x0);
318 b21 = vec_madd(cnst, b21, x1);
319 cnst = LD_W0;
320 b60 = vec_madd(cnst, b60, x0);
321 b61 = vec_madd(cnst, b61, x1);
322
323 #define CTFX(x,b) \
324 vs32(b##0) = vec_unpackh(vs16(x)); \
325 vs32(b##1) = vec_unpackl(vs16(x)); \
326 b##0 = vec_ctf(vs32(b##0), 0); \
327 b##1 = vec_ctf(vs32(b##1), 0); \
328
329 CTFX(x4, b7);
330 CTFX(x5, b5);
331 CTFX(x6, b3);
332 CTFX(x7, b1);
333
334 #undef CTFX
335
336
337 x0 = vec_add(b70, b10);
338 x1 = vec_add(b50, b30);
339 x2 = vec_add(b70, b30);
340 x3 = vec_add(b50, b10);
341 x8 = vec_add(x2, x3);
342 cnst = LD_W3;
343 x8 = vec_madd(cnst, x8, mzero);
344
345 cnst = LD_W8;
346 x0 = vec_madd(cnst, x0, mzero);
347 cnst = LD_W9;
348 x1 = vec_madd(cnst, x1, mzero);
349 cnst = LD_WA;
350 x2 = vec_madd(cnst, x2, x8);
351 cnst = LD_WB;
352 x3 = vec_madd(cnst, x3, x8);
353
354 cnst = LD_W4;
355 b70 = vec_madd(cnst, b70, x0);
356 cnst = LD_W5;
357 b50 = vec_madd(cnst, b50, x1);
358 cnst = LD_W6;
359 b30 = vec_madd(cnst, b30, x1);
360 cnst = LD_W7;
361 b10 = vec_madd(cnst, b10, x0);
362
363 b70 = vec_add(b70, x2);
364 b50 = vec_add(b50, x3);
365 b30 = vec_add(b30, x2);
366 b10 = vec_add(b10, x3);
367
368
369 x0 = vec_add(b71, b11);
370 x1 = vec_add(b51, b31);
371 x2 = vec_add(b71, b31);
372 x3 = vec_add(b51, b11);
373 x8 = vec_add(x2, x3);
374 cnst = LD_W3;
375 x8 = vec_madd(cnst, x8, mzero);
376
377 cnst = LD_W8;
378 x0 = vec_madd(cnst, x0, mzero);
379 cnst = LD_W9;
380 x1 = vec_madd(cnst, x1, mzero);
381 cnst = LD_WA;
382 x2 = vec_madd(cnst, x2, x8);
383 cnst = LD_WB;
384 x3 = vec_madd(cnst, x3, x8);
385
386 cnst = LD_W4;
387 b71 = vec_madd(cnst, b71, x0);
388 cnst = LD_W5;
389 b51 = vec_madd(cnst, b51, x1);
390 cnst = LD_W6;
391 b31 = vec_madd(cnst, b31, x1);
392 cnst = LD_W7;
393 b11 = vec_madd(cnst, b11, x0);
394
395 b71 = vec_add(b71, x2);
396 b51 = vec_add(b51, x3);
397 b31 = vec_add(b31, x2);
398 b11 = vec_add(b11, x3);
399 /* }}} */
400 #else
401 /* convert to float {{{ */
402 #define CTF(n) \
403 vs32(b##n##1) = vec_unpackl(vs16(b##n##0)); \
404 vs32(b##n##0) = vec_unpackh(vs16(b##n##0)); \
405 b##n##1 = vec_ctf(vs32(b##n##1), 0); \
406 b##n##0 = vec_ctf(vs32(b##n##0), 0); \
407
408 CTF(0);
409 CTF(1);
410 CTF(2);
411 CTF(3);
412 CTF(4);
413 CTF(5);
414 CTF(6);
415 CTF(7);
416
417 #undef CTF
418 /* }}} */
419
420 FDCTROW(b00, b10, b20, b30, b40, b50, b60, b70);
421 FDCTROW(b01, b11, b21, b31, b41, b51, b61, b71);
422 #endif
423
424
425 /* 8x8 matrix transpose (vector float[8][2]) {{{ */
426 x0 = vec_mergel(b00, b20);
427 x1 = vec_mergeh(b00, b20);
428 x2 = vec_mergel(b10, b30);
429 x3 = vec_mergeh(b10, b30);
430
431 b00 = vec_mergeh(x1, x3);
432 b10 = vec_mergel(x1, x3);
433 b20 = vec_mergeh(x0, x2);
434 b30 = vec_mergel(x0, x2);
435
436 x4 = vec_mergel(b41, b61);
437 x5 = vec_mergeh(b41, b61);
438 x6 = vec_mergel(b51, b71);
439 x7 = vec_mergeh(b51, b71);
440
441 b41 = vec_mergeh(x5, x7);
442 b51 = vec_mergel(x5, x7);
443 b61 = vec_mergeh(x4, x6);
444 b71 = vec_mergel(x4, x6);
445
446 x0 = vec_mergel(b01, b21);
447 x1 = vec_mergeh(b01, b21);
448 x2 = vec_mergel(b11, b31);
449 x3 = vec_mergeh(b11, b31);
450
451 x4 = vec_mergel(b40, b60);
452 x5 = vec_mergeh(b40, b60);
453 x6 = vec_mergel(b50, b70);
454 x7 = vec_mergeh(b50, b70);
455
456 b40 = vec_mergeh(x1, x3);
457 b50 = vec_mergel(x1, x3);
458 b60 = vec_mergeh(x0, x2);
459 b70 = vec_mergel(x0, x2);
460
461 b01 = vec_mergeh(x5, x7);
462 b11 = vec_mergel(x5, x7);
463 b21 = vec_mergeh(x4, x6);
464 b31 = vec_mergel(x4, x6);
465 /* }}} */
466
467
468 FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
469 FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71);
470
471
472 /* round, convert back to short {{{ */
473 #define CTS(n) \
474 b##n##0 = vec_round(b##n##0); \
475 b##n##1 = vec_round(b##n##1); \
476 vs32(b##n##0) = vec_cts(b##n##0, 0); \
477 vs32(b##n##1) = vec_cts(b##n##1, 0); \
478 vs16(b##n##0) = vec_pack(vs32(b##n##0), vs32(b##n##1)); \
479 vec_st(vs16(b##n##0), 0, bp);
480
481 bp = (vector signed short*)block;
482 CTS(0); bp++;
483 CTS(1); bp++;
484 CTS(2); bp++;
485 CTS(3); bp++;
486 CTS(4); bp++;
487 CTS(5); bp++;
488 CTS(6); bp++;
489 CTS(7);
490
491 #undef CTS
492 /* }}} */
493
494 POWERPC_PERF_STOP_COUNT(altivec_fdct, 1);
495 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
496 }
497
498 /* vim:set foldmethod=marker foldlevel=0: */