1 /*
2 * Copyright (C) 2003 James Klicman <james@klicman.org>
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg 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 2.1 of the License, or (at your option) any later version.
10 *
11 * FFmpeg 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 FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21 #include "config.h"
22
23 #include "libavutil/attributes.h"
24 #include "libavutil/cpu.h"
25 #include "libavutil/ppc/cpu.h"
26 #include "libavutil/ppc/util_altivec.h"
27
28 #include "libavcodec/fdctdsp.h"
29
30 #include "fdct.h"
31
32 #if HAVE_ALTIVEC
33
34 #define vs16(v) ((vector signed short) (v))
35 #define vs32(v) ((vector signed int) (v))
36 #define vu8(v) ((vector unsigned char) (v))
37 #define vu16(v) ((vector unsigned short) (v))
38 #define vu32(v) ((vector unsigned int) (v))
39
40 #define C1 0.98078528040323044912618224 /* cos(1 * PI / 16) */
41 #define C2 0.92387953251128675612818319 /* cos(2 * PI / 16) */
42 #define C3 0.83146961230254523707878838 /* cos(3 * PI / 16) */
43 #define C4 0.70710678118654752440084436 /* cos(4 * PI / 16) */
44 #define C5 0.55557023301960222474283081 /* cos(5 * PI / 16) */
45 #define C6 0.38268343236508977172845998 /* cos(6 * PI / 16) */
46 #define C7 0.19509032201612826784828487 /* cos(7 * PI / 16) */
47
48 #define W0 -(2 * C2)
49 #define W1 (2 * C6)
50 #define W2 (M_SQRT2 * C6)
51 #define W3 (M_SQRT2 * C3)
52 #define W4 (M_SQRT2 * (-C1 + C3 + C5 - C7))
53 #define W5 (M_SQRT2 * (C1 + C3 - C5 + C7))
54 #define W6 (M_SQRT2 * (C1 + C3 + C5 - C7))
55 #define W7 (M_SQRT2 * (C1 + C3 - C5 - C7))
56 #define W8 (M_SQRT2 * (C7 - C3))
57 #define W9 (M_SQRT2 * (-C1 - C3))
58 #define WA (M_SQRT2 * (-C3 - C5))
59 #define WB (M_SQRT2 * (C5 - C3))
60
61 static const vector float fdctconsts[3] = {
62 { W0, W1, W2, W3 },
63 { W4, W5, W6, W7 },
64 { W8, W9, WA, WB }
65 };
66
67 #define LD_W0 vec_splat(cnsts0, 0)
68 #define LD_W1 vec_splat(cnsts0, 1)
69 #define LD_W2 vec_splat(cnsts0, 2)
70 #define LD_W3 vec_splat(cnsts0, 3)
71 #define LD_W4 vec_splat(cnsts1, 0)
72 #define LD_W5 vec_splat(cnsts1, 1)
73 #define LD_W6 vec_splat(cnsts1, 2)
74 #define LD_W7 vec_splat(cnsts1, 3)
75 #define LD_W8 vec_splat(cnsts2, 0)
76 #define LD_W9 vec_splat(cnsts2, 1)
77 #define LD_WA vec_splat(cnsts2, 2)
78 #define LD_WB vec_splat(cnsts2, 3)
79
80 #define FDCTROW(b0, b1, b2, b3, b4, b5, b6, b7) /* {{{ */ \
81 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
82 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
83 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
84 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
85 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
86 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
87 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
88 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
89 \
90 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
91 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
92 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
93 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
94 \
95 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
96 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
97 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
98 cnst = LD_W2; \
99 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
100 cnst = LD_W1; \
101 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
102 cnst = LD_W0; \
103 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
104 \
105 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
106 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
107 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
108 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
109 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
110 cnst = LD_W3; \
111 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
112 \
113 cnst = LD_W8; \
114 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
115 cnst = LD_W9; \
116 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
117 cnst = LD_WA; \
118 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
119 cnst = LD_WB; \
120 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
121 \
122 cnst = LD_W4; \
123 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
124 cnst = LD_W5; \
125 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
126 cnst = LD_W6; \
127 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
128 cnst = LD_W7; \
129 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
130 \
131 b7 = vec_add(b7, x2); /* b7 = b7 + x2; */ \
132 b5 = vec_add(b5, x3); /* b5 = b5 + x3; */ \
133 b3 = vec_add(b3, x2); /* b3 = b3 + x2; */ \
134 b1 = vec_add(b1, x3) /* b1 = b1 + x3; */ \
135 /* }}} */
136
137 #define FDCTCOL(b0, b1, b2, b3, b4, b5, b6, b7) /* {{{ */ \
138 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
139 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
140 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
141 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
142 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
143 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
144 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
145 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
146 \
147 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
148 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
149 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
150 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
151 \
152 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
153 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
154 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
155 cnst = LD_W2; \
156 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
157 cnst = LD_W1; \
158 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
159 cnst = LD_W0; \
160 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
161 \
162 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
163 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
164 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
165 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
166 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
167 cnst = LD_W3; \
168 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
169 \
170 cnst = LD_W8; \
171 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
172 cnst = LD_W9; \
173 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
174 cnst = LD_WA; \
175 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
176 cnst = LD_WB; \
177 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
178 \
179 cnst = LD_W4; \
180 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
181 cnst = LD_W5; \
182 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
183 cnst = LD_W6; \
184 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
185 cnst = LD_W7; \
186 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
187 \
188 b7 = vec_add(b7, x2); /* b7 += x2; */ \
189 b5 = vec_add(b5, x3); /* b5 += x3; */ \
190 b3 = vec_add(b3, x2); /* b3 += x2; */ \
191 b1 = vec_add(b1, x3) /* b1 += x3; */ \
192 /* }}} */
193
194 /* two dimensional discrete cosine transform */
ff_fdct_altivec(int16_t * block)195 void ff_fdct_altivec(int16_t *block)
196 {
197 vector signed short *bp;
198 const vector float *cp = fdctconsts;
199 vector float b00, b10, b20, b30, b40, b50, b60, b70;
200 vector float b01, b11, b21, b31, b41, b51, b61, b71;
201 vector float mzero, cnst, cnsts0, cnsts1, cnsts2;
202 vector float x0, x1, x2, x3, x4, x5, x6, x7, x8;
203
204 /* setup constants {{{ */
205 /* mzero = -0.0 */
206 mzero = ((vector float) vec_splat_u32(-1));
207 mzero = ((vector float) vec_sl(vu32(mzero), vu32(mzero)));
208 cnsts0 = vec_ld(0, cp);
209 cp++;
210 cnsts1 = vec_ld(0, cp);
211 cp++;
212 cnsts2 = vec_ld(0, cp);
213 /* }}} */
214
215 /* 8x8 matrix transpose (vector short[8]) {{{ */
216 #define MERGE_S16(hl, a, b) vec_merge ## hl(vs16(a), vs16(b))
217
218 bp = (vector signed short *) block;
219 b00 = ((vector float) vec_ld(0, bp));
220 b40 = ((vector float) vec_ld(16 * 4, bp));
221 b01 = ((vector float) MERGE_S16(h, b00, b40));
222 b11 = ((vector float) MERGE_S16(l, b00, b40));
223 bp++;
224 b10 = ((vector float) vec_ld(0, bp));
225 b50 = ((vector float) vec_ld(16 * 4, bp));
226 b21 = ((vector float) MERGE_S16(h, b10, b50));
227 b31 = ((vector float) MERGE_S16(l, b10, b50));
228 bp++;
229 b20 = ((vector float) vec_ld(0, bp));
230 b60 = ((vector float) vec_ld(16 * 4, bp));
231 b41 = ((vector float) MERGE_S16(h, b20, b60));
232 b51 = ((vector float) MERGE_S16(l, b20, b60));
233 bp++;
234 b30 = ((vector float) vec_ld(0, bp));
235 b70 = ((vector float) vec_ld(16 * 4, bp));
236 b61 = ((vector float) MERGE_S16(h, b30, b70));
237 b71 = ((vector float) MERGE_S16(l, b30, b70));
238
239 x0 = ((vector float) MERGE_S16(h, b01, b41));
240 x1 = ((vector float) MERGE_S16(l, b01, b41));
241 x2 = ((vector float) MERGE_S16(h, b11, b51));
242 x3 = ((vector float) MERGE_S16(l, b11, b51));
243 x4 = ((vector float) MERGE_S16(h, b21, b61));
244 x5 = ((vector float) MERGE_S16(l, b21, b61));
245 x6 = ((vector float) MERGE_S16(h, b31, b71));
246 x7 = ((vector float) MERGE_S16(l, b31, b71));
247
248 b00 = ((vector float) MERGE_S16(h, x0, x4));
249 b10 = ((vector float) MERGE_S16(l, x0, x4));
250 b20 = ((vector float) MERGE_S16(h, x1, x5));
251 b30 = ((vector float) MERGE_S16(l, x1, x5));
252 b40 = ((vector float) MERGE_S16(h, x2, x6));
253 b50 = ((vector float) MERGE_S16(l, x2, x6));
254 b60 = ((vector float) MERGE_S16(h, x3, x7));
255 b70 = ((vector float) MERGE_S16(l, x3, x7));
256
257 #undef MERGE_S16
258 /* }}} */
259
260 /* Some of the initial calculations can be done as vector short
261 * before conversion to vector float. The following code section
262 * takes advantage of this. */
263
264 /* fdct rows {{{ */
265 x0 = ((vector float) vec_add(vs16(b00), vs16(b70)));
266 x7 = ((vector float) vec_sub(vs16(b00), vs16(b70)));
267 x1 = ((vector float) vec_add(vs16(b10), vs16(b60)));
268 x6 = ((vector float) vec_sub(vs16(b10), vs16(b60)));
269 x2 = ((vector float) vec_add(vs16(b20), vs16(b50)));
270 x5 = ((vector float) vec_sub(vs16(b20), vs16(b50)));
271 x3 = ((vector float) vec_add(vs16(b30), vs16(b40)));
272 x4 = ((vector float) vec_sub(vs16(b30), vs16(b40)));
273
274 b70 = ((vector float) vec_add(vs16(x0), vs16(x3)));
275 b10 = ((vector float) vec_add(vs16(x1), vs16(x2)));
276
277 b00 = ((vector float) vec_add(vs16(b70), vs16(b10)));
278 b40 = ((vector float) vec_sub(vs16(b70), vs16(b10)));
279
280 #define CTF0(n) \
281 b ## n ## 1 = ((vector float) vec_unpackl(vs16(b ## n ## 0))); \
282 b ## n ## 0 = ((vector float) vec_unpackh(vs16(b ## n ## 0))); \
283 b ## n ## 1 = vec_ctf(vs32(b ## n ## 1), 0); \
284 b ## n ## 0 = vec_ctf(vs32(b ## n ## 0), 0)
285
286 CTF0(0);
287 CTF0(4);
288
289 b20 = ((vector float) vec_sub(vs16(x0), vs16(x3)));
290 b60 = ((vector float) vec_sub(vs16(x1), vs16(x2)));
291
292 CTF0(2);
293 CTF0(6);
294
295 #undef CTF0
296
297 x0 = vec_add(b60, b20);
298 x1 = vec_add(b61, b21);
299
300 cnst = LD_W2;
301 x0 = vec_madd(cnst, x0, mzero);
302 x1 = vec_madd(cnst, x1, mzero);
303 cnst = LD_W1;
304 b20 = vec_madd(cnst, b20, x0);
305 b21 = vec_madd(cnst, b21, x1);
306 cnst = LD_W0;
307 b60 = vec_madd(cnst, b60, x0);
308 b61 = vec_madd(cnst, b61, x1);
309
310 #define CTFX(x, b) \
311 b ## 0 = ((vector float) vec_unpackh(vs16(x))); \
312 b ## 1 = ((vector float) vec_unpackl(vs16(x))); \
313 b ## 0 = vec_ctf(vs32(b ## 0), 0); \
314 b ## 1 = vec_ctf(vs32(b ## 1), 0)
315
316 CTFX(x4, b7);
317 CTFX(x5, b5);
318 CTFX(x6, b3);
319 CTFX(x7, b1);
320
321 #undef CTFX
322
323 x0 = vec_add(b70, b10);
324 x1 = vec_add(b50, b30);
325 x2 = vec_add(b70, b30);
326 x3 = vec_add(b50, b10);
327 x8 = vec_add(x2, x3);
328 cnst = LD_W3;
329 x8 = vec_madd(cnst, x8, mzero);
330
331 cnst = LD_W8;
332 x0 = vec_madd(cnst, x0, mzero);
333 cnst = LD_W9;
334 x1 = vec_madd(cnst, x1, mzero);
335 cnst = LD_WA;
336 x2 = vec_madd(cnst, x2, x8);
337 cnst = LD_WB;
338 x3 = vec_madd(cnst, x3, x8);
339
340 cnst = LD_W4;
341 b70 = vec_madd(cnst, b70, x0);
342 cnst = LD_W5;
343 b50 = vec_madd(cnst, b50, x1);
344 cnst = LD_W6;
345 b30 = vec_madd(cnst, b30, x1);
346 cnst = LD_W7;
347 b10 = vec_madd(cnst, b10, x0);
348
349 b70 = vec_add(b70, x2);
350 b50 = vec_add(b50, x3);
351 b30 = vec_add(b30, x2);
352 b10 = vec_add(b10, x3);
353
354 x0 = vec_add(b71, b11);
355 x1 = vec_add(b51, b31);
356 x2 = vec_add(b71, b31);
357 x3 = vec_add(b51, b11);
358 x8 = vec_add(x2, x3);
359 cnst = LD_W3;
360 x8 = vec_madd(cnst, x8, mzero);
361
362 cnst = LD_W8;
363 x0 = vec_madd(cnst, x0, mzero);
364 cnst = LD_W9;
365 x1 = vec_madd(cnst, x1, mzero);
366 cnst = LD_WA;
367 x2 = vec_madd(cnst, x2, x8);
368 cnst = LD_WB;
369 x3 = vec_madd(cnst, x3, x8);
370
371 cnst = LD_W4;
372 b71 = vec_madd(cnst, b71, x0);
373 cnst = LD_W5;
374 b51 = vec_madd(cnst, b51, x1);
375 cnst = LD_W6;
376 b31 = vec_madd(cnst, b31, x1);
377 cnst = LD_W7;
378 b11 = vec_madd(cnst, b11, x0);
379
380 b71 = vec_add(b71, x2);
381 b51 = vec_add(b51, x3);
382 b31 = vec_add(b31, x2);
383 b11 = vec_add(b11, x3);
384 /* }}} */
385
386 /* 8x8 matrix transpose (vector float[8][2]) {{{ */
387 x0 = vec_mergel(b00, b20);
388 x1 = vec_mergeh(b00, b20);
389 x2 = vec_mergel(b10, b30);
390 x3 = vec_mergeh(b10, b30);
391
392 b00 = vec_mergeh(x1, x3);
393 b10 = vec_mergel(x1, x3);
394 b20 = vec_mergeh(x0, x2);
395 b30 = vec_mergel(x0, x2);
396
397 x4 = vec_mergel(b41, b61);
398 x5 = vec_mergeh(b41, b61);
399 x6 = vec_mergel(b51, b71);
400 x7 = vec_mergeh(b51, b71);
401
402 b41 = vec_mergeh(x5, x7);
403 b51 = vec_mergel(x5, x7);
404 b61 = vec_mergeh(x4, x6);
405 b71 = vec_mergel(x4, x6);
406
407 x0 = vec_mergel(b01, b21);
408 x1 = vec_mergeh(b01, b21);
409 x2 = vec_mergel(b11, b31);
410 x3 = vec_mergeh(b11, b31);
411
412 x4 = vec_mergel(b40, b60);
413 x5 = vec_mergeh(b40, b60);
414 x6 = vec_mergel(b50, b70);
415 x7 = vec_mergeh(b50, b70);
416
417 b40 = vec_mergeh(x1, x3);
418 b50 = vec_mergel(x1, x3);
419 b60 = vec_mergeh(x0, x2);
420 b70 = vec_mergel(x0, x2);
421
422 b01 = vec_mergeh(x5, x7);
423 b11 = vec_mergel(x5, x7);
424 b21 = vec_mergeh(x4, x6);
425 b31 = vec_mergel(x4, x6);
426 /* }}} */
427
428 FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
429 FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71);
430
431 /* round, convert back to short {{{ */
432 #define CTS(n) \
433 b ## n ## 0 = vec_round(b ## n ## 0); \
434 b ## n ## 1 = vec_round(b ## n ## 1); \
435 b ## n ## 0 = ((vector float) vec_cts(b ## n ## 0, 0)); \
436 b ## n ## 1 = ((vector float) vec_cts(b ## n ## 1, 0)); \
437 b ## n ## 0 = ((vector float) vec_pack(vs32(b ## n ## 0), \
438 vs32(b ## n ## 1))); \
439 vec_st(vs16(b ## n ## 0), 0, bp)
440
441 bp = (vector signed short *) block;
442 CTS(0);
443 bp++;
444 CTS(1);
445 bp++;
446 CTS(2);
447 bp++;
448 CTS(3);
449 bp++;
450 CTS(4);
451 bp++;
452 CTS(5);
453 bp++;
454 CTS(6);
455 bp++;
456 CTS(7);
457
458 #undef CTS
459 /* }}} */
460 }
461
462 #endif /* HAVE_ALTIVEC */
463
ff_fdctdsp_init_ppc(FDCTDSPContext * c,AVCodecContext * avctx,unsigned high_bit_depth)464 av_cold void ff_fdctdsp_init_ppc(FDCTDSPContext *c, AVCodecContext *avctx,
465 unsigned high_bit_depth)
466 {
467 #if HAVE_ALTIVEC
468 if (!PPC_ALTIVEC(av_get_cpu_flags()))
469 return;
470
471 if (!high_bit_depth) {
472 if (avctx->dct_algo == FF_DCT_AUTO ||
473 avctx->dct_algo == FF_DCT_ALTIVEC) {
474 c->fdct = ff_fdct_altivec;
475 }
476 }
477 #endif /* HAVE_ALTIVEC */
478 }
479