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
25 #include "libavutil/common.h"
26 #include "libavcodec/dsputil.h"
27 #include "dsputil_altivec.h"
29 #define vs16(v) ((vector signed short)(v))
30 #define vs32(v) ((vector signed int)(v))
31 #define vu8(v) ((vector unsigned char)(v))
32 #define vu16(v) ((vector unsigned short)(v))
33 #define vu32(v) ((vector unsigned int)(v))
36 #define C1 0.98078525066375732421875000 /* cos(1*PI/16) */
37 #define C2 0.92387950420379638671875000 /* cos(2*PI/16) */
38 #define C3 0.83146959543228149414062500 /* cos(3*PI/16) */
39 #define C4 0.70710676908493041992187500 /* cos(4*PI/16) */
40 #define C5 0.55557024478912353515625000 /* cos(5*PI/16) */
41 #define C6 0.38268342614173889160156250 /* cos(6*PI/16) */
42 #define C7 0.19509032368659973144531250 /* cos(7*PI/16) */
43 #define SQRT_2 1.41421353816986083984375000 /* sqrt(2) */
48 #define W2 (SQRT_2 * C6)
49 #define W3 (SQRT_2 * C3)
50 #define W4 (SQRT_2 * (-C1 + C3 + C5 - C7))
51 #define W5 (SQRT_2 * ( C1 + C3 - C5 + C7))
52 #define W6 (SQRT_2 * ( C1 + C3 + C5 - C7))
53 #define W7 (SQRT_2 * ( C1 + C3 - C5 - C7))
54 #define W8 (SQRT_2 * ( C7 - C3))
55 #define W9 (SQRT_2 * (-C1 - C3))
56 #define WA (SQRT_2 * (-C3 - C5))
57 #define WB (SQRT_2 * ( C5 - C3))
60 static vector float fdctconsts[3] = {
66 #define LD_W0 vec_splat(cnsts0, 0)
67 #define LD_W1 vec_splat(cnsts0, 1)
68 #define LD_W2 vec_splat(cnsts0, 2)
69 #define LD_W3 vec_splat(cnsts0, 3)
70 #define LD_W4 vec_splat(cnsts1, 0)
71 #define LD_W5 vec_splat(cnsts1, 1)
72 #define LD_W6 vec_splat(cnsts1, 2)
73 #define LD_W7 vec_splat(cnsts1, 3)
74 #define LD_W8 vec_splat(cnsts2, 0)
75 #define LD_W9 vec_splat(cnsts2, 1)
76 #define LD_WA vec_splat(cnsts2, 2)
77 #define LD_WB vec_splat(cnsts2, 3)
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; */ \
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; */ \
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; */ \
99 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
101 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
103 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
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; */ \
111 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
114 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
116 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
118 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
120 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
123 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
125 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
127 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
129 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
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; */ \
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; */ \
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; */ \
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; */ \
156 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
158 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
160 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
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; */ \
168 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
171 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
173 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
175 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
177 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
180 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
182 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
184 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
186 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
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; */ \
196 /* two dimensional discrete cosine transform */
198 void ff_fdct_altivec(int16_t *block)
200 vector signed short *bp;
202 vector float b00, b10, b20, b30, b40, b50, b60, b70;
203 vector float b01, b11, b21, b31, b41, b51, b61, b71;
204 vector float mzero, cnst, cnsts0, cnsts1, cnsts2;
205 vector float x0, x1, x2, x3, x4, x5, x6, x7, x8;
207 /* setup constants {{{ */
209 mzero = ((vector float)vec_splat_u32(-1));
210 mzero = ((vector float)vec_sl(vu32(mzero), vu32(mzero)));
212 cnsts0 = vec_ld(0, cp); cp++;
213 cnsts1 = vec_ld(0, cp); cp++;
214 cnsts2 = vec_ld(0, cp);
218 /* 8x8 matrix transpose (vector short[8]) {{{ */
219 #define MERGE_S16(hl,a,b) vec_merge##hl(vs16(a), vs16(b))
221 bp = (vector signed short*)block;
222 b00 = ((vector float)vec_ld(0, bp));
223 b40 = ((vector float)vec_ld(16*4, bp));
224 b01 = ((vector float)MERGE_S16(h, b00, b40));
225 b11 = ((vector float)MERGE_S16(l, b00, b40));
227 b10 = ((vector float)vec_ld(0, bp));
228 b50 = ((vector float)vec_ld(16*4, bp));
229 b21 = ((vector float)MERGE_S16(h, b10, b50));
230 b31 = ((vector float)MERGE_S16(l, b10, b50));
232 b20 = ((vector float)vec_ld(0, bp));
233 b60 = ((vector float)vec_ld(16*4, bp));
234 b41 = ((vector float)MERGE_S16(h, b20, b60));
235 b51 = ((vector float)MERGE_S16(l, b20, b60));
237 b30 = ((vector float)vec_ld(0, bp));
238 b70 = ((vector float)vec_ld(16*4, bp));
239 b61 = ((vector float)MERGE_S16(h, b30, b70));
240 b71 = ((vector float)MERGE_S16(l, b30, b70));
242 x0 = ((vector float)MERGE_S16(h, b01, b41));
243 x1 = ((vector float)MERGE_S16(l, b01, b41));
244 x2 = ((vector float)MERGE_S16(h, b11, b51));
245 x3 = ((vector float)MERGE_S16(l, b11, b51));
246 x4 = ((vector float)MERGE_S16(h, b21, b61));
247 x5 = ((vector float)MERGE_S16(l, b21, b61));
248 x6 = ((vector float)MERGE_S16(h, b31, b71));
249 x7 = ((vector float)MERGE_S16(l, b31, b71));
251 b00 = ((vector float)MERGE_S16(h, x0, x4));
252 b10 = ((vector float)MERGE_S16(l, x0, x4));
253 b20 = ((vector float)MERGE_S16(h, x1, x5));
254 b30 = ((vector float)MERGE_S16(l, x1, x5));
255 b40 = ((vector float)MERGE_S16(h, x2, x6));
256 b50 = ((vector float)MERGE_S16(l, x2, x6));
257 b60 = ((vector float)MERGE_S16(h, x3, x7));
258 b70 = ((vector float)MERGE_S16(l, x3, x7));
264 /* Some of the initial calculations can be done as vector short before
265 * conversion to vector float. The following code section takes advantage
269 x0 = ((vector float)vec_add(vs16(b00), vs16(b70)));
270 x7 = ((vector float)vec_sub(vs16(b00), vs16(b70)));
271 x1 = ((vector float)vec_add(vs16(b10), vs16(b60)));
272 x6 = ((vector float)vec_sub(vs16(b10), vs16(b60)));
273 x2 = ((vector float)vec_add(vs16(b20), vs16(b50)));
274 x5 = ((vector float)vec_sub(vs16(b20), vs16(b50)));
275 x3 = ((vector float)vec_add(vs16(b30), vs16(b40)));
276 x4 = ((vector float)vec_sub(vs16(b30), vs16(b40)));
278 b70 = ((vector float)vec_add(vs16(x0), vs16(x3)));
279 b10 = ((vector float)vec_add(vs16(x1), vs16(x2)));
281 b00 = ((vector float)vec_add(vs16(b70), vs16(b10)));
282 b40 = ((vector float)vec_sub(vs16(b70), vs16(b10)));
285 b##n##1 = ((vector float)vec_unpackl(vs16(b##n##0))); \
286 b##n##0 = ((vector float)vec_unpackh(vs16(b##n##0))); \
287 b##n##1 = vec_ctf(vs32(b##n##1), 0); \
288 b##n##0 = vec_ctf(vs32(b##n##0), 0);
293 b20 = ((vector float)vec_sub(vs16(x0), vs16(x3)));
294 b60 = ((vector float)vec_sub(vs16(x1), vs16(x2)));
301 x0 = vec_add(b60, b20);
302 x1 = vec_add(b61, b21);
305 x0 = vec_madd(cnst, x0, mzero);
306 x1 = vec_madd(cnst, x1, mzero);
308 b20 = vec_madd(cnst, b20, x0);
309 b21 = vec_madd(cnst, b21, x1);
311 b60 = vec_madd(cnst, b60, x0);
312 b61 = vec_madd(cnst, b61, x1);
315 b##0 = ((vector float)vec_unpackh(vs16(x))); \
316 b##1 = ((vector float)vec_unpackl(vs16(x))); \
317 b##0 = vec_ctf(vs32(b##0), 0); \
318 b##1 = vec_ctf(vs32(b##1), 0); \
328 x0 = vec_add(b70, b10);
329 x1 = vec_add(b50, b30);
330 x2 = vec_add(b70, b30);
331 x3 = vec_add(b50, b10);
332 x8 = vec_add(x2, x3);
334 x8 = vec_madd(cnst, x8, mzero);
337 x0 = vec_madd(cnst, x0, mzero);
339 x1 = vec_madd(cnst, x1, mzero);
341 x2 = vec_madd(cnst, x2, x8);
343 x3 = vec_madd(cnst, x3, x8);
346 b70 = vec_madd(cnst, b70, x0);
348 b50 = vec_madd(cnst, b50, x1);
350 b30 = vec_madd(cnst, b30, x1);
352 b10 = vec_madd(cnst, b10, x0);
354 b70 = vec_add(b70, x2);
355 b50 = vec_add(b50, x3);
356 b30 = vec_add(b30, x2);
357 b10 = vec_add(b10, x3);
360 x0 = vec_add(b71, b11);
361 x1 = vec_add(b51, b31);
362 x2 = vec_add(b71, b31);
363 x3 = vec_add(b51, b11);
364 x8 = vec_add(x2, x3);
366 x8 = vec_madd(cnst, x8, mzero);
369 x0 = vec_madd(cnst, x0, mzero);
371 x1 = vec_madd(cnst, x1, mzero);
373 x2 = vec_madd(cnst, x2, x8);
375 x3 = vec_madd(cnst, x3, x8);
378 b71 = vec_madd(cnst, b71, x0);
380 b51 = vec_madd(cnst, b51, x1);
382 b31 = vec_madd(cnst, b31, x1);
384 b11 = vec_madd(cnst, b11, x0);
386 b71 = vec_add(b71, x2);
387 b51 = vec_add(b51, x3);
388 b31 = vec_add(b31, x2);
389 b11 = vec_add(b11, x3);
393 /* 8x8 matrix transpose (vector float[8][2]) {{{ */
394 x0 = vec_mergel(b00, b20);
395 x1 = vec_mergeh(b00, b20);
396 x2 = vec_mergel(b10, b30);
397 x3 = vec_mergeh(b10, b30);
399 b00 = vec_mergeh(x1, x3);
400 b10 = vec_mergel(x1, x3);
401 b20 = vec_mergeh(x0, x2);
402 b30 = vec_mergel(x0, x2);
404 x4 = vec_mergel(b41, b61);
405 x5 = vec_mergeh(b41, b61);
406 x6 = vec_mergel(b51, b71);
407 x7 = vec_mergeh(b51, b71);
409 b41 = vec_mergeh(x5, x7);
410 b51 = vec_mergel(x5, x7);
411 b61 = vec_mergeh(x4, x6);
412 b71 = vec_mergel(x4, x6);
414 x0 = vec_mergel(b01, b21);
415 x1 = vec_mergeh(b01, b21);
416 x2 = vec_mergel(b11, b31);
417 x3 = vec_mergeh(b11, b31);
419 x4 = vec_mergel(b40, b60);
420 x5 = vec_mergeh(b40, b60);
421 x6 = vec_mergel(b50, b70);
422 x7 = vec_mergeh(b50, b70);
424 b40 = vec_mergeh(x1, x3);
425 b50 = vec_mergel(x1, x3);
426 b60 = vec_mergeh(x0, x2);
427 b70 = vec_mergel(x0, x2);
429 b01 = vec_mergeh(x5, x7);
430 b11 = vec_mergel(x5, x7);
431 b21 = vec_mergeh(x4, x6);
432 b31 = vec_mergel(x4, x6);
436 FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
437 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), vs32(b##n##1))); \
447 vec_st(vs16(b##n##0), 0, bp);
449 bp = (vector signed short*)block;
463 /* vim:set foldmethod=marker foldlevel=0: */