2 * Copyright (C) 2003 James Klicman <james@klicman.org>
4 * This file is part of FFmpeg.
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.
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.
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
26 #include "libavutil/attributes.h"
27 #include "libavutil/cpu.h"
28 #include "libavutil/ppc/cpu.h"
29 #include "libavcodec/fdctdsp.h"
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))
40 #define C1 0.98078525066375732421875000 /* cos(1 * PI / 16) */
41 #define C2 0.92387950420379638671875000 /* cos(2 * PI / 16) */
42 #define C3 0.83146959543228149414062500 /* cos(3 * PI / 16) */
43 #define C4 0.70710676908493041992187500 /* cos(4 * PI / 16) */
44 #define C5 0.55557024478912353515625000 /* cos(5 * PI / 16) */
45 #define C6 0.38268342614173889160156250 /* cos(6 * PI / 16) */
46 #define C7 0.19509032368659973144531250 /* cos(7 * PI / 16) */
47 #define SQRT_2 1.41421353816986083984375000 /* sqrt(2) */
51 #define W2 (SQRT_2 * C6)
52 #define W3 (SQRT_2 * C3)
53 #define W4 (SQRT_2 * (-C1 + C3 + C5 - C7))
54 #define W5 (SQRT_2 * (C1 + C3 - C5 + C7))
55 #define W6 (SQRT_2 * (C1 + C3 + C5 - C7))
56 #define W7 (SQRT_2 * (C1 + C3 - C5 - C7))
57 #define W8 (SQRT_2 * (C7 - C3))
58 #define W9 (SQRT_2 * (-C1 - C3))
59 #define WA (SQRT_2 * (-C3 - C5))
60 #define WB (SQRT_2 * (C5 - C3))
62 static const vector float fdctconsts[3] = {
68 #define LD_W0 vec_splat(cnsts0, 0)
69 #define LD_W1 vec_splat(cnsts0, 1)
70 #define LD_W2 vec_splat(cnsts0, 2)
71 #define LD_W3 vec_splat(cnsts0, 3)
72 #define LD_W4 vec_splat(cnsts1, 0)
73 #define LD_W5 vec_splat(cnsts1, 1)
74 #define LD_W6 vec_splat(cnsts1, 2)
75 #define LD_W7 vec_splat(cnsts1, 3)
76 #define LD_W8 vec_splat(cnsts2, 0)
77 #define LD_W9 vec_splat(cnsts2, 1)
78 #define LD_WA vec_splat(cnsts2, 2)
79 #define LD_WB vec_splat(cnsts2, 3)
82 #define VEC_FMERGEH(a, b) vec_mergeh(a, b)
83 #define VEC_FMERGEL(a, b) vec_mergel(a, b)
85 #define VEC_FMERGEH(a, b) vec_mergel(b, a)
86 #define VEC_FMERGEL(a, b) vec_mergeh(b, a)
89 #define FDCTROW(b0, b1, b2, b3, b4, b5, b6, b7) /* {{{ */ \
90 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
91 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
92 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
93 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
94 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
95 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
96 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
97 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
99 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
100 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
101 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
102 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
104 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
105 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
106 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
108 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
110 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
112 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
114 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
115 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
116 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
117 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
118 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
120 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
123 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
125 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
127 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
129 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
132 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
134 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
136 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
138 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
140 b7 = vec_add(b7, x2); /* b7 = b7 + x2; */ \
141 b5 = vec_add(b5, x3); /* b5 = b5 + x3; */ \
142 b3 = vec_add(b3, x2); /* b3 = b3 + x2; */ \
143 b1 = vec_add(b1, x3) /* b1 = b1 + x3; */ \
146 #define FDCTCOL(b0, b1, b2, b3, b4, b5, b6, b7) /* {{{ */ \
147 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
148 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
149 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
150 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
151 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
152 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
153 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
154 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
156 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
157 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
158 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
159 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
161 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
162 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
163 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
165 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
167 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
169 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
171 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
172 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
173 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
174 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
175 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
177 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
180 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
182 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
184 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
186 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
189 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
191 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
193 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
195 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
197 b7 = vec_add(b7, x2); /* b7 += x2; */ \
198 b5 = vec_add(b5, x3); /* b5 += x3; */ \
199 b3 = vec_add(b3, x2); /* b3 += x2; */ \
200 b1 = vec_add(b1, x3) /* b1 += x3; */ \
203 /* two dimensional discrete cosine transform */
204 void ff_fdct_altivec(int16_t *block)
206 vector signed short *bp;
207 const vector float *cp = fdctconsts;
208 vector float b00, b10, b20, b30, b40, b50, b60, b70;
209 vector float b01, b11, b21, b31, b41, b51, b61, b71;
210 vector float mzero, cnst, cnsts0, cnsts1, cnsts2;
211 vector float x0, x1, x2, x3, x4, x5, x6, x7, x8;
213 /* setup constants {{{ */
215 mzero = ((vector float) vec_splat_u32(-1));
216 mzero = ((vector float) vec_sl(vu32(mzero), vu32(mzero)));
217 cnsts0 = vec_ld(0, cp);
219 cnsts1 = vec_ld(0, cp);
221 cnsts2 = vec_ld(0, cp);
224 /* 8x8 matrix transpose (vector short[8]) {{{ */
225 #define MERGE_S16(hl, a, b) vec_merge ## hl(vs16(a), vs16(b))
227 bp = (vector signed short *) block;
228 b00 = ((vector float) vec_ld(0, bp));
229 b40 = ((vector float) vec_ld(16 * 4, bp));
230 b01 = ((vector float) MERGE_S16(h, b00, b40));
231 b11 = ((vector float) MERGE_S16(l, b00, b40));
233 b10 = ((vector float) vec_ld(0, bp));
234 b50 = ((vector float) vec_ld(16 * 4, bp));
235 b21 = ((vector float) MERGE_S16(h, b10, b50));
236 b31 = ((vector float) MERGE_S16(l, b10, b50));
238 b20 = ((vector float) vec_ld(0, bp));
239 b60 = ((vector float) vec_ld(16 * 4, bp));
240 b41 = ((vector float) MERGE_S16(h, b20, b60));
241 b51 = ((vector float) MERGE_S16(l, b20, b60));
243 b30 = ((vector float) vec_ld(0, bp));
244 b70 = ((vector float) vec_ld(16 * 4, bp));
245 b61 = ((vector float) MERGE_S16(h, b30, b70));
246 b71 = ((vector float) MERGE_S16(l, b30, b70));
248 x0 = ((vector float) MERGE_S16(h, b01, b41));
249 x1 = ((vector float) MERGE_S16(l, b01, b41));
250 x2 = ((vector float) MERGE_S16(h, b11, b51));
251 x3 = ((vector float) MERGE_S16(l, b11, b51));
252 x4 = ((vector float) MERGE_S16(h, b21, b61));
253 x5 = ((vector float) MERGE_S16(l, b21, b61));
254 x6 = ((vector float) MERGE_S16(h, b31, b71));
255 x7 = ((vector float) MERGE_S16(l, b31, b71));
257 b00 = ((vector float) MERGE_S16(h, x0, x4));
258 b10 = ((vector float) MERGE_S16(l, x0, x4));
259 b20 = ((vector float) MERGE_S16(h, x1, x5));
260 b30 = ((vector float) MERGE_S16(l, x1, x5));
261 b40 = ((vector float) MERGE_S16(h, x2, x6));
262 b50 = ((vector float) MERGE_S16(l, x2, x6));
263 b60 = ((vector float) MERGE_S16(h, x3, x7));
264 b70 = ((vector float) MERGE_S16(l, x3, x7));
269 /* Some of the initial calculations can be done as vector short
270 * before conversion to vector float. The following code section
271 * takes advantage of this. */
274 x0 = ((vector float) vec_add(vs16(b00), vs16(b70)));
275 x7 = ((vector float) vec_sub(vs16(b00), vs16(b70)));
276 x1 = ((vector float) vec_add(vs16(b10), vs16(b60)));
277 x6 = ((vector float) vec_sub(vs16(b10), vs16(b60)));
278 x2 = ((vector float) vec_add(vs16(b20), vs16(b50)));
279 x5 = ((vector float) vec_sub(vs16(b20), vs16(b50)));
280 x3 = ((vector float) vec_add(vs16(b30), vs16(b40)));
281 x4 = ((vector float) vec_sub(vs16(b30), vs16(b40)));
283 b70 = ((vector float) vec_add(vs16(x0), vs16(x3)));
284 b10 = ((vector float) vec_add(vs16(x1), vs16(x2)));
286 b00 = ((vector float) vec_add(vs16(b70), vs16(b10)));
287 b40 = ((vector float) vec_sub(vs16(b70), vs16(b10)));
290 b ## n ## 1 = ((vector float) vec_unpackl(vs16(b ## n ## 0))); \
291 b ## n ## 0 = ((vector float) vec_unpackh(vs16(b ## n ## 0))); \
292 b ## n ## 1 = vec_ctf(vs32(b ## n ## 1), 0); \
293 b ## n ## 0 = vec_ctf(vs32(b ## n ## 0), 0)
298 b20 = ((vector float) vec_sub(vs16(x0), vs16(x3)));
299 b60 = ((vector float) vec_sub(vs16(x1), vs16(x2)));
306 x0 = vec_add(b60, b20);
307 x1 = vec_add(b61, b21);
310 x0 = vec_madd(cnst, x0, mzero);
311 x1 = vec_madd(cnst, x1, mzero);
313 b20 = vec_madd(cnst, b20, x0);
314 b21 = vec_madd(cnst, b21, x1);
316 b60 = vec_madd(cnst, b60, x0);
317 b61 = vec_madd(cnst, b61, x1);
320 b ## 0 = ((vector float) vec_unpackh(vs16(x))); \
321 b ## 1 = ((vector float) vec_unpackl(vs16(x))); \
322 b ## 0 = vec_ctf(vs32(b ## 0), 0); \
323 b ## 1 = vec_ctf(vs32(b ## 1), 0)
332 x0 = vec_add(b70, b10);
333 x1 = vec_add(b50, b30);
334 x2 = vec_add(b70, b30);
335 x3 = vec_add(b50, b10);
336 x8 = vec_add(x2, x3);
338 x8 = vec_madd(cnst, x8, mzero);
341 x0 = vec_madd(cnst, x0, mzero);
343 x1 = vec_madd(cnst, x1, mzero);
345 x2 = vec_madd(cnst, x2, x8);
347 x3 = vec_madd(cnst, x3, x8);
350 b70 = vec_madd(cnst, b70, x0);
352 b50 = vec_madd(cnst, b50, x1);
354 b30 = vec_madd(cnst, b30, x1);
356 b10 = vec_madd(cnst, b10, x0);
358 b70 = vec_add(b70, x2);
359 b50 = vec_add(b50, x3);
360 b30 = vec_add(b30, x2);
361 b10 = vec_add(b10, x3);
363 x0 = vec_add(b71, b11);
364 x1 = vec_add(b51, b31);
365 x2 = vec_add(b71, b31);
366 x3 = vec_add(b51, b11);
367 x8 = vec_add(x2, x3);
369 x8 = vec_madd(cnst, x8, mzero);
372 x0 = vec_madd(cnst, x0, mzero);
374 x1 = vec_madd(cnst, x1, mzero);
376 x2 = vec_madd(cnst, x2, x8);
378 x3 = vec_madd(cnst, x3, x8);
381 b71 = vec_madd(cnst, b71, x0);
383 b51 = vec_madd(cnst, b51, x1);
385 b31 = vec_madd(cnst, b31, x1);
387 b11 = vec_madd(cnst, b11, x0);
389 b71 = vec_add(b71, x2);
390 b51 = vec_add(b51, x3);
391 b31 = vec_add(b31, x2);
392 b11 = vec_add(b11, x3);
395 /* 8x8 matrix transpose (vector float[8][2]) {{{ */
396 x0 = VEC_FMERGEL(b00, b20);
397 x1 = VEC_FMERGEH(b00, b20);
398 x2 = VEC_FMERGEL(b10, b30);
399 x3 = VEC_FMERGEH(b10, b30);
401 b00 = VEC_FMERGEH(x1, x3);
402 b10 = VEC_FMERGEL(x1, x3);
403 b20 = VEC_FMERGEH(x0, x2);
404 b30 = VEC_FMERGEL(x0, x2);
406 x4 = VEC_FMERGEL(b41, b61);
407 x5 = VEC_FMERGEH(b41, b61);
408 x6 = VEC_FMERGEL(b51, b71);
409 x7 = VEC_FMERGEH(b51, b71);
411 b41 = VEC_FMERGEH(x5, x7);
412 b51 = VEC_FMERGEL(x5, x7);
413 b61 = VEC_FMERGEH(x4, x6);
414 b71 = VEC_FMERGEL(x4, x6);
416 x0 = VEC_FMERGEL(b01, b21);
417 x1 = VEC_FMERGEH(b01, b21);
418 x2 = VEC_FMERGEL(b11, b31);
419 x3 = VEC_FMERGEH(b11, b31);
421 x4 = VEC_FMERGEL(b40, b60);
422 x5 = VEC_FMERGEH(b40, b60);
423 x6 = VEC_FMERGEL(b50, b70);
424 x7 = VEC_FMERGEH(b50, b70);
426 b40 = VEC_FMERGEH(x1, x3);
427 b50 = VEC_FMERGEL(x1, x3);
428 b60 = VEC_FMERGEH(x0, x2);
429 b70 = VEC_FMERGEL(x0, x2);
431 b01 = VEC_FMERGEH(x5, x7);
432 b11 = VEC_FMERGEL(x5, x7);
433 b21 = VEC_FMERGEH(x4, x6);
434 b31 = VEC_FMERGEL(x4, x6);
437 FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
438 FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71);
440 /* round, convert back to short {{{ */
442 b ## n ## 0 = vec_round(b ## n ## 0); \
443 b ## n ## 1 = vec_round(b ## n ## 1); \
444 b ## n ## 0 = ((vector float) vec_cts(b ## n ## 0, 0)); \
445 b ## n ## 1 = ((vector float) vec_cts(b ## n ## 1, 0)); \
446 b ## n ## 0 = ((vector float) vec_pack(vs32(b ## n ## 0), \
447 vs32(b ## n ## 1))); \
448 vec_st(vs16(b ## n ## 0), 0, bp)
450 bp = (vector signed short *) block;
471 #endif /* HAVE_ALTIVEC */
473 av_cold void ff_fdctdsp_init_ppc(FDCTDSPContext *c, AVCodecContext *avctx,
474 unsigned high_bit_depth)
477 if (!PPC_ALTIVEC(av_get_cpu_flags()))
480 if (!high_bit_depth) {
481 if (avctx->dct_algo == FF_DCT_AUTO ||
482 avctx->dct_algo == FF_DCT_ALTIVEC) {
483 c->fdct = ff_fdct_altivec;
486 #endif /* HAVE_ALTIVEC */