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 "dsputil_altivec.h"
28 #define vs16(v) ((vector signed short)(v))
29 #define vs32(v) ((vector signed int)(v))
30 #define vu8(v) ((vector unsigned char)(v))
31 #define vu16(v) ((vector unsigned short)(v))
32 #define vu32(v) ((vector unsigned int)(v))
35 #define C1 0.98078525066375732421875000 /* cos(1*PI/16) */
36 #define C2 0.92387950420379638671875000 /* cos(2*PI/16) */
37 #define C3 0.83146959543228149414062500 /* cos(3*PI/16) */
38 #define C4 0.70710676908493041992187500 /* cos(4*PI/16) */
39 #define C5 0.55557024478912353515625000 /* cos(5*PI/16) */
40 #define C6 0.38268342614173889160156250 /* cos(6*PI/16) */
41 #define C7 0.19509032368659973144531250 /* cos(7*PI/16) */
42 #define SQRT_2 1.41421353816986083984375000 /* sqrt(2) */
47 #define W2 (SQRT_2 * C6)
48 #define W3 (SQRT_2 * C3)
49 #define W4 (SQRT_2 * (-C1 + C3 + C5 - C7))
50 #define W5 (SQRT_2 * ( C1 + C3 - C5 + C7))
51 #define W6 (SQRT_2 * ( C1 + C3 + C5 - C7))
52 #define W7 (SQRT_2 * ( C1 + C3 - C5 - C7))
53 #define W8 (SQRT_2 * ( C7 - C3))
54 #define W9 (SQRT_2 * (-C1 - C3))
55 #define WA (SQRT_2 * (-C3 - C5))
56 #define WB (SQRT_2 * ( C5 - C3))
59 static vector float fdctconsts[3] = {
65 #define LD_W0 vec_splat(cnsts0, 0)
66 #define LD_W1 vec_splat(cnsts0, 1)
67 #define LD_W2 vec_splat(cnsts0, 2)
68 #define LD_W3 vec_splat(cnsts0, 3)
69 #define LD_W4 vec_splat(cnsts1, 0)
70 #define LD_W5 vec_splat(cnsts1, 1)
71 #define LD_W6 vec_splat(cnsts1, 2)
72 #define LD_W7 vec_splat(cnsts1, 3)
73 #define LD_W8 vec_splat(cnsts2, 0)
74 #define LD_W9 vec_splat(cnsts2, 1)
75 #define LD_WA vec_splat(cnsts2, 2)
76 #define LD_WB vec_splat(cnsts2, 3)
79 #define FDCTROW(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \
80 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
81 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
82 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
83 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
84 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
85 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
86 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
87 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
89 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
90 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
91 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
92 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
94 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
95 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
96 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
98 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
100 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
102 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
104 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
105 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
106 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
107 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
108 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
110 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
113 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
115 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
117 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
119 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
122 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
124 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
126 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
128 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
130 b7 = vec_add(b7, x2); /* b7 = b7 + x2; */ \
131 b5 = vec_add(b5, x3); /* b5 = b5 + x3; */ \
132 b3 = vec_add(b3, x2); /* b3 = b3 + x2; */ \
133 b1 = vec_add(b1, x3); /* b1 = b1 + x3; */ \
136 #define FDCTCOL(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \
137 x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
138 x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
139 x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
140 x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
141 x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
142 x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
143 x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
144 x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
146 b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
147 b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
148 b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
149 b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
151 b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
152 b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
153 b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
155 b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
157 b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
159 b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
161 x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
162 x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
163 x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
164 x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
165 x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
167 x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
170 x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
172 x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
174 x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
176 x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
179 b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
181 b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
183 b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
185 b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
187 b7 = vec_add(b7, x2); /* b7 += x2; */ \
188 b5 = vec_add(b5, x3); /* b5 += x3; */ \
189 b3 = vec_add(b3, x2); /* b3 += x2; */ \
190 b1 = vec_add(b1, x3); /* b1 += x3; */ \
195 /* two dimensional discrete cosine transform */
197 void ff_fdct_altivec(int16_t *block)
199 vector signed short *bp;
201 vector float b00, b10, b20, b30, b40, b50, b60, b70;
202 vector float b01, b11, b21, b31, b41, b51, b61, b71;
203 vector float mzero, cnst, cnsts0, cnsts1, cnsts2;
204 vector float x0, x1, x2, x3, x4, x5, x6, x7, x8;
206 /* setup constants {{{ */
208 mzero = ((vector float)vec_splat_u32(-1));
209 mzero = ((vector float)vec_sl(vu32(mzero), vu32(mzero)));
211 cnsts0 = vec_ld(0, cp); cp++;
212 cnsts1 = vec_ld(0, cp); cp++;
213 cnsts2 = vec_ld(0, cp);
217 /* 8x8 matrix transpose (vector short[8]) {{{ */
218 #define MERGE_S16(hl,a,b) vec_merge##hl(vs16(a), vs16(b))
220 bp = (vector signed short*)block;
221 b00 = ((vector float)vec_ld(0, bp));
222 b40 = ((vector float)vec_ld(16*4, bp));
223 b01 = ((vector float)MERGE_S16(h, b00, b40));
224 b11 = ((vector float)MERGE_S16(l, b00, b40));
226 b10 = ((vector float)vec_ld(0, bp));
227 b50 = ((vector float)vec_ld(16*4, bp));
228 b21 = ((vector float)MERGE_S16(h, b10, b50));
229 b31 = ((vector float)MERGE_S16(l, b10, b50));
231 b20 = ((vector float)vec_ld(0, bp));
232 b60 = ((vector float)vec_ld(16*4, bp));
233 b41 = ((vector float)MERGE_S16(h, b20, b60));
234 b51 = ((vector float)MERGE_S16(l, b20, b60));
236 b30 = ((vector float)vec_ld(0, bp));
237 b70 = ((vector float)vec_ld(16*4, bp));
238 b61 = ((vector float)MERGE_S16(h, b30, b70));
239 b71 = ((vector float)MERGE_S16(l, b30, b70));
241 x0 = ((vector float)MERGE_S16(h, b01, b41));
242 x1 = ((vector float)MERGE_S16(l, b01, b41));
243 x2 = ((vector float)MERGE_S16(h, b11, b51));
244 x3 = ((vector float)MERGE_S16(l, b11, b51));
245 x4 = ((vector float)MERGE_S16(h, b21, b61));
246 x5 = ((vector float)MERGE_S16(l, b21, b61));
247 x6 = ((vector float)MERGE_S16(h, b31, b71));
248 x7 = ((vector float)MERGE_S16(l, b31, b71));
250 b00 = ((vector float)MERGE_S16(h, x0, x4));
251 b10 = ((vector float)MERGE_S16(l, x0, x4));
252 b20 = ((vector float)MERGE_S16(h, x1, x5));
253 b30 = ((vector float)MERGE_S16(l, x1, x5));
254 b40 = ((vector float)MERGE_S16(h, x2, x6));
255 b50 = ((vector float)MERGE_S16(l, x2, x6));
256 b60 = ((vector float)MERGE_S16(h, x3, x7));
257 b70 = ((vector float)MERGE_S16(l, x3, x7));
263 /* Some of the initial calculations can be done as vector short before
264 * conversion to vector float. The following code section takes advantage
268 x0 = ((vector float)vec_add(vs16(b00), vs16(b70)));
269 x7 = ((vector float)vec_sub(vs16(b00), vs16(b70)));
270 x1 = ((vector float)vec_add(vs16(b10), vs16(b60)));
271 x6 = ((vector float)vec_sub(vs16(b10), vs16(b60)));
272 x2 = ((vector float)vec_add(vs16(b20), vs16(b50)));
273 x5 = ((vector float)vec_sub(vs16(b20), vs16(b50)));
274 x3 = ((vector float)vec_add(vs16(b30), vs16(b40)));
275 x4 = ((vector float)vec_sub(vs16(b30), vs16(b40)));
277 b70 = ((vector float)vec_add(vs16(x0), vs16(x3)));
278 b10 = ((vector float)vec_add(vs16(x1), vs16(x2)));
280 b00 = ((vector float)vec_add(vs16(b70), vs16(b10)));
281 b40 = ((vector float)vec_sub(vs16(b70), vs16(b10)));
284 b##n##1 = ((vector float)vec_unpackl(vs16(b##n##0))); \
285 b##n##0 = ((vector float)vec_unpackh(vs16(b##n##0))); \
286 b##n##1 = vec_ctf(vs32(b##n##1), 0); \
287 b##n##0 = vec_ctf(vs32(b##n##0), 0);
292 b20 = ((vector float)vec_sub(vs16(x0), vs16(x3)));
293 b60 = ((vector float)vec_sub(vs16(x1), vs16(x2)));
300 x0 = vec_add(b60, b20);
301 x1 = vec_add(b61, b21);
304 x0 = vec_madd(cnst, x0, mzero);
305 x1 = vec_madd(cnst, x1, mzero);
307 b20 = vec_madd(cnst, b20, x0);
308 b21 = vec_madd(cnst, b21, x1);
310 b60 = vec_madd(cnst, b60, x0);
311 b61 = vec_madd(cnst, b61, x1);
314 b##0 = ((vector float)vec_unpackh(vs16(x))); \
315 b##1 = ((vector float)vec_unpackl(vs16(x))); \
316 b##0 = vec_ctf(vs32(b##0), 0); \
317 b##1 = vec_ctf(vs32(b##1), 0); \
327 x0 = vec_add(b70, b10);
328 x1 = vec_add(b50, b30);
329 x2 = vec_add(b70, b30);
330 x3 = vec_add(b50, b10);
331 x8 = vec_add(x2, x3);
333 x8 = vec_madd(cnst, x8, mzero);
336 x0 = vec_madd(cnst, x0, mzero);
338 x1 = vec_madd(cnst, x1, mzero);
340 x2 = vec_madd(cnst, x2, x8);
342 x3 = vec_madd(cnst, x3, x8);
345 b70 = vec_madd(cnst, b70, x0);
347 b50 = vec_madd(cnst, b50, x1);
349 b30 = vec_madd(cnst, b30, x1);
351 b10 = vec_madd(cnst, b10, x0);
353 b70 = vec_add(b70, x2);
354 b50 = vec_add(b50, x3);
355 b30 = vec_add(b30, x2);
356 b10 = vec_add(b10, x3);
359 x0 = vec_add(b71, b11);
360 x1 = vec_add(b51, b31);
361 x2 = vec_add(b71, b31);
362 x3 = vec_add(b51, b11);
363 x8 = vec_add(x2, x3);
365 x8 = vec_madd(cnst, x8, mzero);
368 x0 = vec_madd(cnst, x0, mzero);
370 x1 = vec_madd(cnst, x1, mzero);
372 x2 = vec_madd(cnst, x2, x8);
374 x3 = vec_madd(cnst, x3, x8);
377 b71 = vec_madd(cnst, b71, x0);
379 b51 = vec_madd(cnst, b51, x1);
381 b31 = vec_madd(cnst, b31, x1);
383 b11 = vec_madd(cnst, b11, x0);
385 b71 = vec_add(b71, x2);
386 b51 = vec_add(b51, x3);
387 b31 = vec_add(b31, x2);
388 b11 = vec_add(b11, x3);
392 /* 8x8 matrix transpose (vector float[8][2]) {{{ */
393 x0 = vec_mergel(b00, b20);
394 x1 = vec_mergeh(b00, b20);
395 x2 = vec_mergel(b10, b30);
396 x3 = vec_mergeh(b10, b30);
398 b00 = vec_mergeh(x1, x3);
399 b10 = vec_mergel(x1, x3);
400 b20 = vec_mergeh(x0, x2);
401 b30 = vec_mergel(x0, x2);
403 x4 = vec_mergel(b41, b61);
404 x5 = vec_mergeh(b41, b61);
405 x6 = vec_mergel(b51, b71);
406 x7 = vec_mergeh(b51, b71);
408 b41 = vec_mergeh(x5, x7);
409 b51 = vec_mergel(x5, x7);
410 b61 = vec_mergeh(x4, x6);
411 b71 = vec_mergel(x4, x6);
413 x0 = vec_mergel(b01, b21);
414 x1 = vec_mergeh(b01, b21);
415 x2 = vec_mergel(b11, b31);
416 x3 = vec_mergeh(b11, b31);
418 x4 = vec_mergel(b40, b60);
419 x5 = vec_mergeh(b40, b60);
420 x6 = vec_mergel(b50, b70);
421 x7 = vec_mergeh(b50, b70);
423 b40 = vec_mergeh(x1, x3);
424 b50 = vec_mergel(x1, x3);
425 b60 = vec_mergeh(x0, x2);
426 b70 = vec_mergel(x0, x2);
428 b01 = vec_mergeh(x5, x7);
429 b11 = vec_mergel(x5, x7);
430 b21 = vec_mergeh(x4, x6);
431 b31 = vec_mergel(x4, x6);
435 FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
436 FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71);
439 /* round, convert back to short {{{ */
441 b##n##0 = vec_round(b##n##0); \
442 b##n##1 = vec_round(b##n##1); \
443 b##n##0 = ((vector float)vec_cts(b##n##0, 0)); \
444 b##n##1 = ((vector float)vec_cts(b##n##1, 0)); \
445 b##n##0 = ((vector float)vec_pack(vs32(b##n##0), vs32(b##n##1))); \
446 vec_st(vs16(b##n##0), 0, bp);
448 bp = (vector signed short*)block;