2 * MMX optimized forward DCT
3 * The gcc porting is Copyright (c) 2001 Fabrice Bellard.
4 * cleanup/optimizations are Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
5 * SSE2 optimization is Copyright (c) 2004 Denes Balatoni.
7 * from fdctam32.c - AP922 MMX(3D-Now) forward-DCT
9 * Intel Application Note AP-922 - fast, precise implementation of DCT
10 * http://developer.intel.com/vtune/cbts/appnotes.htm
12 * Also of inspiration:
13 * a page about fdct at http://www.geocities.com/ssavekar/dct.htm
14 * Skal's fdct at http://skal.planet-d.net/coding/dct.html
16 * This file is part of FFmpeg.
18 * FFmpeg is free software; you can redistribute it and/or
19 * modify it under the terms of the GNU Lesser General Public
20 * License as published by the Free Software Foundation; either
21 * version 2.1 of the License, or (at your option) any later version.
23 * FFmpeg is distributed in the hope that it will be useful,
24 * but WITHOUT ANY WARRANTY; without even the implied warranty of
25 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
26 * Lesser General Public License for more details.
28 * You should have received a copy of the GNU Lesser General Public
29 * License along with FFmpeg; if not, write to the Free Software
30 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
36 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
38 //////////////////////////////////////////////////////////////////////
40 // constants for the forward DCT
41 // -----------------------------
43 // Be sure to check that your compiler is aligning all constants to QWORD
44 // (8-byte) memory boundaries! Otherwise the unaligned memory access will
45 // severely stall MMX execution.
47 //////////////////////////////////////////////////////////////////////
49 #define BITS_FRW_ACC 3 //; 2 or 3 for accuracy
50 #define SHIFT_FRW_COL BITS_FRW_ACC
51 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17 - 3)
52 #define RND_FRW_ROW (1 << (SHIFT_FRW_ROW-1))
53 //#define RND_FRW_COL (1 << (SHIFT_FRW_COL-1))
55 #define X8(x) x,x,x,x,x,x,x,x
57 //concatenated table, for forward DCT transformation
58 static const int16_t fdct_tg_all_16[24] ATTR_ALIGN(16) = {
59 X8(13036), // tg * (2<<16) + 0.5
60 X8(27146), // tg * (2<<16) + 0.5
61 X8(-21746) // tg * (2<<16) + 0.5
64 static const int16_t ocos_4_16[8] ATTR_ALIGN(16) = {
65 X8(23170) //cos * (2<<15) + 0.5
68 static const int16_t fdct_one_corr[8] ATTR_ALIGN(16) = { X8(1) };
70 static const int32_t fdct_r_row[2] ATTR_ALIGN(8) = {RND_FRW_ROW, RND_FRW_ROW };
74 const int32_t fdct_r_row_sse2[4] ATTR_ALIGN(16);
75 } fdct_r_row_sse2 ATTR_ALIGN(16)=
77 RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW
79 //static const long fdct_r_row_sse2[4] ATTR_ALIGN(16) = {RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW};
81 static const int16_t tab_frw_01234567[] ATTR_ALIGN(8) = { // forward_dct coeff table
82 16384, 16384, 22725, 19266,
83 16384, 16384, 12873, 4520,
84 21407, 8867, 19266, -4520,
85 -8867, -21407, -22725, -12873,
86 16384, -16384, 12873, -22725,
87 -16384, 16384, 4520, 19266,
88 8867, -21407, 4520, -12873,
89 21407, -8867, 19266, -22725,
91 22725, 22725, 31521, 26722,
92 22725, 22725, 17855, 6270,
93 29692, 12299, 26722, -6270,
94 -12299, -29692, -31521, -17855,
95 22725, -22725, 17855, -31521,
96 -22725, 22725, 6270, 26722,
97 12299, -29692, 6270, -17855,
98 29692, -12299, 26722, -31521,
100 21407, 21407, 29692, 25172,
101 21407, 21407, 16819, 5906,
102 27969, 11585, 25172, -5906,
103 -11585, -27969, -29692, -16819,
104 21407, -21407, 16819, -29692,
105 -21407, 21407, 5906, 25172,
106 11585, -27969, 5906, -16819,
107 27969, -11585, 25172, -29692,
109 19266, 19266, 26722, 22654,
110 19266, 19266, 15137, 5315,
111 25172, 10426, 22654, -5315,
112 -10426, -25172, -26722, -15137,
113 19266, -19266, 15137, -26722,
114 -19266, 19266, 5315, 22654,
115 10426, -25172, 5315, -15137,
116 25172, -10426, 22654, -26722,
118 16384, 16384, 22725, 19266,
119 16384, 16384, 12873, 4520,
120 21407, 8867, 19266, -4520,
121 -8867, -21407, -22725, -12873,
122 16384, -16384, 12873, -22725,
123 -16384, 16384, 4520, 19266,
124 8867, -21407, 4520, -12873,
125 21407, -8867, 19266, -22725,
127 19266, 19266, 26722, 22654,
128 19266, 19266, 15137, 5315,
129 25172, 10426, 22654, -5315,
130 -10426, -25172, -26722, -15137,
131 19266, -19266, 15137, -26722,
132 -19266, 19266, 5315, 22654,
133 10426, -25172, 5315, -15137,
134 25172, -10426, 22654, -26722,
136 21407, 21407, 29692, 25172,
137 21407, 21407, 16819, 5906,
138 27969, 11585, 25172, -5906,
139 -11585, -27969, -29692, -16819,
140 21407, -21407, 16819, -29692,
141 -21407, 21407, 5906, 25172,
142 11585, -27969, 5906, -16819,
143 27969, -11585, 25172, -29692,
145 22725, 22725, 31521, 26722,
146 22725, 22725, 17855, 6270,
147 29692, 12299, 26722, -6270,
148 -12299, -29692, -31521, -17855,
149 22725, -22725, 17855, -31521,
150 -22725, 22725, 6270, 26722,
151 12299, -29692, 6270, -17855,
152 29692, -12299, 26722, -31521,
157 const int16_t tab_frw_01234567_sse2[256] ATTR_ALIGN(16);
158 } tab_frw_01234567_sse2 ATTR_ALIGN(16) =
160 //static const int16_t tab_frw_01234567_sse2[] ATTR_ALIGN(16) = { // forward_dct coeff table
161 #define TABLE_SSE2 C4, C4, C1, C3, -C6, -C2, -C1, -C5, \
162 C4, C4, C5, C7, C2, C6, C3, -C7, \
163 -C4, C4, C7, C3, C6, -C2, C7, -C5, \
164 C4, -C4, C5, -C1, C2, -C6, C3, -C1,
165 // c1..c7 * cos(pi/4) * 2^15
288 #define FDCT_COL(cpu, mm, mov)\
289 static av_always_inline void fdct_col_##cpu(const int16_t *in, int16_t *out, int offset)\
291 mov##_m2r(*(in + offset + 1 * 8), mm##0);\
292 mov##_m2r(*(in + offset + 6 * 8), mm##1);\
293 mov##_r2r(mm##0, mm##2);\
294 mov##_m2r(*(in + offset + 2 * 8), mm##3);\
295 paddsw_r2r(mm##1, mm##0);\
296 mov##_m2r(*(in + offset + 5 * 8), mm##4);\
297 psllw_i2r(SHIFT_FRW_COL, mm##0);\
298 mov##_m2r(*(in + offset + 0 * 8), mm##5);\
299 paddsw_r2r(mm##3, mm##4);\
300 paddsw_m2r(*(in + offset + 7 * 8), mm##5);\
301 psllw_i2r(SHIFT_FRW_COL, mm##4);\
302 mov##_r2r(mm##0, mm##6);\
303 psubsw_r2r(mm##1, mm##2);\
304 mov##_m2r(*(fdct_tg_all_16 + 8), mm##1);\
305 psubsw_r2r(mm##4, mm##0);\
306 mov##_m2r(*(in + offset + 3 * 8), mm##7);\
307 pmulhw_r2r(mm##0, mm##1);\
308 paddsw_m2r(*(in + offset + 4 * 8), mm##7);\
309 psllw_i2r(SHIFT_FRW_COL, mm##5);\
310 paddsw_r2r(mm##4, mm##6);\
311 psllw_i2r(SHIFT_FRW_COL, mm##7);\
312 mov##_r2r(mm##5, mm##4);\
313 psubsw_r2r(mm##7, mm##5);\
314 paddsw_r2r(mm##5, mm##1);\
315 paddsw_r2r(mm##7, mm##4);\
316 por_m2r(*fdct_one_corr, mm##1);\
317 psllw_i2r(SHIFT_FRW_COL + 1, mm##2);\
318 pmulhw_m2r(*(fdct_tg_all_16 + 8), mm##5);\
319 mov##_r2r(mm##4, mm##7);\
320 psubsw_m2r(*(in + offset + 5 * 8), mm##3);\
321 psubsw_r2r(mm##6, mm##4);\
322 mov##_r2m(mm##1, *(out + offset + 2 * 8));\
323 paddsw_r2r(mm##6, mm##7);\
324 mov##_m2r(*(in + offset + 3 * 8), mm##1);\
325 psllw_i2r(SHIFT_FRW_COL + 1, mm##3);\
326 psubsw_m2r(*(in + offset + 4 * 8), mm##1);\
327 mov##_r2r(mm##2, mm##6);\
328 mov##_r2m(mm##4, *(out + offset + 4 * 8));\
329 paddsw_r2r(mm##3, mm##2);\
330 pmulhw_m2r(*ocos_4_16, mm##2);\
331 psubsw_r2r(mm##3, mm##6);\
332 pmulhw_m2r(*ocos_4_16, mm##6);\
333 psubsw_r2r(mm##0, mm##5);\
334 por_m2r(*fdct_one_corr, mm##5);\
335 psllw_i2r(SHIFT_FRW_COL, mm##1);\
336 por_m2r(*fdct_one_corr, mm##2);\
337 mov##_r2r(mm##1, mm##4);\
338 mov##_m2r(*(in + offset + 0 * 8), mm##3);\
339 paddsw_r2r(mm##6, mm##1);\
340 psubsw_m2r(*(in + offset + 7 * 8), mm##3);\
341 psubsw_r2r(mm##6, mm##4);\
342 mov##_m2r(*(fdct_tg_all_16 + 0), mm##0);\
343 psllw_i2r(SHIFT_FRW_COL, mm##3);\
344 mov##_m2r(*(fdct_tg_all_16 + 16), mm##6);\
345 pmulhw_r2r(mm##1, mm##0);\
346 mov##_r2m(mm##7, *(out + offset + 0 * 8));\
347 pmulhw_r2r(mm##4, mm##6);\
348 mov##_r2m(mm##5, *(out + offset + 6 * 8));\
349 mov##_r2r(mm##3, mm##7);\
350 mov##_m2r(*(fdct_tg_all_16 + 16), mm##5);\
351 psubsw_r2r(mm##2, mm##7);\
352 paddsw_r2r(mm##2, mm##3);\
353 pmulhw_r2r(mm##7, mm##5);\
354 paddsw_r2r(mm##3, mm##0);\
355 paddsw_r2r(mm##4, mm##6);\
356 pmulhw_m2r(*(fdct_tg_all_16 + 0), mm##3);\
357 por_m2r(*fdct_one_corr, mm##0);\
358 paddsw_r2r(mm##7, mm##5);\
359 psubsw_r2r(mm##6, mm##7);\
360 mov##_r2m(mm##0, *(out + offset + 1 * 8));\
361 paddsw_r2r(mm##4, mm##5);\
362 mov##_r2m(mm##7, *(out + offset + 3 * 8));\
363 psubsw_r2r(mm##1, mm##3);\
364 mov##_r2m(mm##5, *(out + offset + 5 * 8));\
365 mov##_r2m(mm##3, *(out + offset + 7 * 8));\
368 FDCT_COL(mmx, mm, movq)
369 FDCT_COL(sse2, xmm, movdqa)
371 static av_always_inline void fdct_row_sse2(const int16_t *in, int16_t *out)
374 #define FDCT_ROW_SSE2_H1(i,t) \
375 "movq " #i "(%0), %%xmm2 \n\t" \
376 "movq " #i "+8(%0), %%xmm0 \n\t" \
377 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
378 "movdqa " #t "+48(%1), %%xmm7 \n\t" \
379 "movdqa " #t "(%1), %%xmm4 \n\t" \
380 "movdqa " #t "+16(%1), %%xmm5 \n\t"
382 #define FDCT_ROW_SSE2_H2(i,t) \
383 "movq " #i "(%0), %%xmm2 \n\t" \
384 "movq " #i "+8(%0), %%xmm0 \n\t" \
385 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
386 "movdqa " #t "+48(%1), %%xmm7 \n\t"
388 #define FDCT_ROW_SSE2(i) \
389 "movq %%xmm2, %%xmm1 \n\t" \
390 "pshuflw $27, %%xmm0, %%xmm0 \n\t" \
391 "paddsw %%xmm0, %%xmm1 \n\t" \
392 "psubsw %%xmm0, %%xmm2 \n\t" \
393 "punpckldq %%xmm2, %%xmm1 \n\t" \
394 "pshufd $78, %%xmm1, %%xmm2 \n\t" \
395 "pmaddwd %%xmm2, %%xmm3 \n\t" \
396 "pmaddwd %%xmm1, %%xmm7 \n\t" \
397 "pmaddwd %%xmm5, %%xmm2 \n\t" \
398 "pmaddwd %%xmm4, %%xmm1 \n\t" \
399 "paddd %%xmm7, %%xmm3 \n\t" \
400 "paddd %%xmm2, %%xmm1 \n\t" \
401 "paddd %%xmm6, %%xmm3 \n\t" \
402 "paddd %%xmm6, %%xmm1 \n\t" \
403 "psrad %3, %%xmm3 \n\t" \
404 "psrad %3, %%xmm1 \n\t" \
405 "packssdw %%xmm3, %%xmm1 \n\t" \
406 "movdqa %%xmm1, " #i "(%4) \n\t"
408 "movdqa (%2), %%xmm6 \n\t"
409 FDCT_ROW_SSE2_H1(0,0)
411 FDCT_ROW_SSE2_H2(64,0)
414 FDCT_ROW_SSE2_H1(16,64)
416 FDCT_ROW_SSE2_H2(112,64)
419 FDCT_ROW_SSE2_H1(32,128)
421 FDCT_ROW_SSE2_H2(96,128)
424 FDCT_ROW_SSE2_H1(48,192)
426 FDCT_ROW_SSE2_H2(80,192)
429 : "r" (in), "r" (tab_frw_01234567_sse2.tab_frw_01234567_sse2), "r" (fdct_r_row_sse2.fdct_r_row_sse2), "i" (SHIFT_FRW_ROW), "r" (out)
433 static av_always_inline void fdct_row_mmx2(const int16_t *in, int16_t *out, const int16_t *table)
435 pshufw_m2r(*(in + 4), mm5, 0x1B);
436 movq_m2r(*(in + 0), mm0);
438 paddsw_r2r(mm5, mm0);
439 psubsw_r2r(mm5, mm1);
441 punpckldq_r2r(mm1, mm0);
442 punpckhdq_r2r(mm1, mm2);
443 movq_m2r(*(table + 0), mm1);
444 movq_m2r(*(table + 4), mm3);
445 movq_m2r(*(table + 8), mm4);
446 movq_m2r(*(table + 12), mm5);
447 movq_m2r(*(table + 16), mm6);
448 movq_m2r(*(table + 20), mm7);
449 pmaddwd_r2r(mm0, mm1);
450 pmaddwd_r2r(mm2, mm3);
451 pmaddwd_r2r(mm0, mm4);
452 pmaddwd_r2r(mm2, mm5);
453 pmaddwd_r2r(mm0, mm6);
454 pmaddwd_r2r(mm2, mm7);
455 pmaddwd_m2r(*(table + 24), mm0);
456 pmaddwd_m2r(*(table + 28), mm2);
461 movq_m2r(*fdct_r_row, mm0);
466 psrad_i2r(SHIFT_FRW_ROW, mm3);
467 psrad_i2r(SHIFT_FRW_ROW, mm5);
468 psrad_i2r(SHIFT_FRW_ROW, mm7);
469 psrad_i2r(SHIFT_FRW_ROW, mm2);
470 packssdw_r2r(mm5, mm3);
471 packssdw_r2r(mm2, mm7);
472 movq_r2m(mm3, *(out + 0));
473 movq_r2m(mm7, *(out + 4));
476 static av_always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table)
478 //FIXME reorder (I do not have an old MMX-only CPU here to benchmark ...)
479 movd_m2r(*(in + 6), mm1);
480 punpcklwd_m2r(*(in + 4), mm1);
482 psrlq_i2r(0x20, mm1);
483 movq_m2r(*(in + 0), mm0);
484 punpcklwd_r2r(mm2, mm1);
486 paddsw_r2r(mm1, mm0);
487 psubsw_r2r(mm1, mm5);
489 punpckldq_r2r(mm5, mm0);
490 punpckhdq_r2r(mm5, mm2);
491 movq_m2r(*(table + 0), mm1);
492 movq_m2r(*(table + 4), mm3);
493 movq_m2r(*(table + 8), mm4);
494 movq_m2r(*(table + 12), mm5);
495 movq_m2r(*(table + 16), mm6);
496 movq_m2r(*(table + 20), mm7);
497 pmaddwd_r2r(mm0, mm1);
498 pmaddwd_r2r(mm2, mm3);
499 pmaddwd_r2r(mm0, mm4);
500 pmaddwd_r2r(mm2, mm5);
501 pmaddwd_r2r(mm0, mm6);
502 pmaddwd_r2r(mm2, mm7);
503 pmaddwd_m2r(*(table + 24), mm0);
504 pmaddwd_m2r(*(table + 28), mm2);
509 movq_m2r(*fdct_r_row, mm0);
514 psrad_i2r(SHIFT_FRW_ROW, mm3);
515 psrad_i2r(SHIFT_FRW_ROW, mm5);
516 psrad_i2r(SHIFT_FRW_ROW, mm7);
517 psrad_i2r(SHIFT_FRW_ROW, mm2);
518 packssdw_r2r(mm5, mm3);
519 packssdw_r2r(mm2, mm7);
520 movq_r2m(mm3, *(out + 0));
521 movq_r2m(mm7, *(out + 4));
524 void ff_fdct_mmx(int16_t *block)
526 int64_t align_tmp[16] ATTR_ALIGN(8);
527 int16_t * block1= (int16_t*)align_tmp;
528 const int16_t *table= tab_frw_01234567;
531 fdct_col_mmx(block, block1, 0);
532 fdct_col_mmx(block, block1, 4);
535 fdct_row_mmx(block1, block, table);
542 void ff_fdct_mmx2(int16_t *block)
544 int64_t align_tmp[16] ATTR_ALIGN(8);
545 int16_t *block1= (int16_t*)align_tmp;
546 const int16_t *table= tab_frw_01234567;
549 fdct_col_mmx(block, block1, 0);
550 fdct_col_mmx(block, block1, 4);
553 fdct_row_mmx2(block1, block, table);
560 void ff_fdct_sse2(int16_t *block)
562 int64_t align_tmp[16] ATTR_ALIGN(16);
563 int16_t * const block1= (int16_t*)align_tmp;
565 fdct_col_sse2(block, block1, 0);
566 fdct_row_sse2(block1, block);