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
33 #include "libavutil/common.h"
34 #include "libavcodec/dsputil.h"
36 //////////////////////////////////////////////////////////////////////
38 // constants for the forward DCT
39 // -----------------------------
41 // Be sure to check that your compiler is aligning all constants to QWORD
42 // (8-byte) memory boundaries! Otherwise the unaligned memory access will
43 // severely stall MMX execution.
45 //////////////////////////////////////////////////////////////////////
47 #define BITS_FRW_ACC 3 //; 2 or 3 for accuracy
48 #define SHIFT_FRW_COL BITS_FRW_ACC
49 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17 - 3)
50 #define RND_FRW_ROW (1 << (SHIFT_FRW_ROW-1))
51 //#define RND_FRW_COL (1 << (SHIFT_FRW_COL-1))
53 #define X8(x) x,x,x,x,x,x,x,x
55 //concatenated table, for forward DCT transformation
56 DECLARE_ALIGNED(16, static const int16_t, fdct_tg_all_16)[24] = {
57 X8(13036), // tg * (2<<16) + 0.5
58 X8(27146), // tg * (2<<16) + 0.5
59 X8(-21746) // tg * (2<<16) + 0.5
62 DECLARE_ALIGNED(16, static const int16_t, ocos_4_16)[8] = {
63 X8(23170) //cos * (2<<15) + 0.5
66 DECLARE_ALIGNED(16, static const int16_t, fdct_one_corr)[8] = { X8(1) };
68 DECLARE_ALIGNED(8, static const int32_t, fdct_r_row)[2] = {RND_FRW_ROW, RND_FRW_ROW };
72 DECLARE_ALIGNED(16, const int32_t, fdct_r_row_sse2)[4];
75 RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW
77 //DECLARE_ALIGNED(16, static const long, fdct_r_row_sse2)[4] = {RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW};
79 DECLARE_ALIGNED(8, static const int16_t, tab_frw_01234567)[] = { // forward_dct coeff table
80 16384, 16384, 22725, 19266,
81 16384, 16384, 12873, 4520,
82 21407, 8867, 19266, -4520,
83 -8867, -21407, -22725, -12873,
84 16384, -16384, 12873, -22725,
85 -16384, 16384, 4520, 19266,
86 8867, -21407, 4520, -12873,
87 21407, -8867, 19266, -22725,
89 22725, 22725, 31521, 26722,
90 22725, 22725, 17855, 6270,
91 29692, 12299, 26722, -6270,
92 -12299, -29692, -31521, -17855,
93 22725, -22725, 17855, -31521,
94 -22725, 22725, 6270, 26722,
95 12299, -29692, 6270, -17855,
96 29692, -12299, 26722, -31521,
98 21407, 21407, 29692, 25172,
99 21407, 21407, 16819, 5906,
100 27969, 11585, 25172, -5906,
101 -11585, -27969, -29692, -16819,
102 21407, -21407, 16819, -29692,
103 -21407, 21407, 5906, 25172,
104 11585, -27969, 5906, -16819,
105 27969, -11585, 25172, -29692,
107 19266, 19266, 26722, 22654,
108 19266, 19266, 15137, 5315,
109 25172, 10426, 22654, -5315,
110 -10426, -25172, -26722, -15137,
111 19266, -19266, 15137, -26722,
112 -19266, 19266, 5315, 22654,
113 10426, -25172, 5315, -15137,
114 25172, -10426, 22654, -26722,
116 16384, 16384, 22725, 19266,
117 16384, 16384, 12873, 4520,
118 21407, 8867, 19266, -4520,
119 -8867, -21407, -22725, -12873,
120 16384, -16384, 12873, -22725,
121 -16384, 16384, 4520, 19266,
122 8867, -21407, 4520, -12873,
123 21407, -8867, 19266, -22725,
125 19266, 19266, 26722, 22654,
126 19266, 19266, 15137, 5315,
127 25172, 10426, 22654, -5315,
128 -10426, -25172, -26722, -15137,
129 19266, -19266, 15137, -26722,
130 -19266, 19266, 5315, 22654,
131 10426, -25172, 5315, -15137,
132 25172, -10426, 22654, -26722,
134 21407, 21407, 29692, 25172,
135 21407, 21407, 16819, 5906,
136 27969, 11585, 25172, -5906,
137 -11585, -27969, -29692, -16819,
138 21407, -21407, 16819, -29692,
139 -21407, 21407, 5906, 25172,
140 11585, -27969, 5906, -16819,
141 27969, -11585, 25172, -29692,
143 22725, 22725, 31521, 26722,
144 22725, 22725, 17855, 6270,
145 29692, 12299, 26722, -6270,
146 -12299, -29692, -31521, -17855,
147 22725, -22725, 17855, -31521,
148 -22725, 22725, 6270, 26722,
149 12299, -29692, 6270, -17855,
150 29692, -12299, 26722, -31521,
155 DECLARE_ALIGNED(16, const int16_t, tab_frw_01234567_sse2)[256];
156 } tab_frw_01234567_sse2 =
158 //DECLARE_ALIGNED(16, static const int16_t, tab_frw_01234567_sse2)[] = { // forward_dct coeff table
159 #define TABLE_SSE2 C4, C4, C1, C3, -C6, -C2, -C1, -C5, \
160 C4, C4, C5, C7, C2, C6, C3, -C7, \
161 -C4, C4, C7, C3, C6, -C2, C7, -C5, \
162 C4, -C4, C5, -C1, C2, -C6, C3, -C1,
163 // c1..c7 * cos(pi/4) * 2^15
286 #define S(s) AV_TOSTRING(s) //AV_STRINGIFY is too long
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)\
292 #mov" 16(%0), %%"#mm"0 \n\t" \
293 #mov" 96(%0), %%"#mm"1 \n\t" \
294 #mov" %%"#mm"0, %%"#mm"2 \n\t" \
295 #mov" 32(%0), %%"#mm"3 \n\t" \
296 "paddsw %%"#mm"1, %%"#mm"0 \n\t" \
297 #mov" 80(%0), %%"#mm"4 \n\t" \
298 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"0 \n\t" \
299 #mov" (%0), %%"#mm"5 \n\t" \
300 "paddsw %%"#mm"3, %%"#mm"4 \n\t" \
301 "paddsw 112(%0), %%"#mm"5 \n\t" \
302 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"4 \n\t" \
303 #mov" %%"#mm"0, %%"#mm"6 \n\t" \
304 "psubsw %%"#mm"1, %%"#mm"2 \n\t" \
305 #mov" 16(%1), %%"#mm"1 \n\t" \
306 "psubsw %%"#mm"4, %%"#mm"0 \n\t" \
307 #mov" 48(%0), %%"#mm"7 \n\t" \
308 "pmulhw %%"#mm"0, %%"#mm"1 \n\t" \
309 "paddsw 64(%0), %%"#mm"7 \n\t" \
310 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"5 \n\t" \
311 "paddsw %%"#mm"4, %%"#mm"6 \n\t" \
312 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"7 \n\t" \
313 #mov" %%"#mm"5, %%"#mm"4 \n\t" \
314 "psubsw %%"#mm"7, %%"#mm"5 \n\t" \
315 "paddsw %%"#mm"5, %%"#mm"1 \n\t" \
316 "paddsw %%"#mm"7, %%"#mm"4 \n\t" \
317 "por (%2), %%"#mm"1 \n\t" \
318 "psllw $"S(SHIFT_FRW_COL)"+1, %%"#mm"2 \n\t" \
319 "pmulhw 16(%1), %%"#mm"5 \n\t" \
320 #mov" %%"#mm"4, %%"#mm"7 \n\t" \
321 "psubsw 80(%0), %%"#mm"3 \n\t" \
322 "psubsw %%"#mm"6, %%"#mm"4 \n\t" \
323 #mov" %%"#mm"1, 32(%3) \n\t" \
324 "paddsw %%"#mm"6, %%"#mm"7 \n\t" \
325 #mov" 48(%0), %%"#mm"1 \n\t" \
326 "psllw $"S(SHIFT_FRW_COL)"+1, %%"#mm"3 \n\t" \
327 "psubsw 64(%0), %%"#mm"1 \n\t" \
328 #mov" %%"#mm"2, %%"#mm"6 \n\t" \
329 #mov" %%"#mm"4, 64(%3) \n\t" \
330 "paddsw %%"#mm"3, %%"#mm"2 \n\t" \
331 "pmulhw (%4), %%"#mm"2 \n\t" \
332 "psubsw %%"#mm"3, %%"#mm"6 \n\t" \
333 "pmulhw (%4), %%"#mm"6 \n\t" \
334 "psubsw %%"#mm"0, %%"#mm"5 \n\t" \
335 "por (%2), %%"#mm"5 \n\t" \
336 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"1 \n\t" \
337 "por (%2), %%"#mm"2 \n\t" \
338 #mov" %%"#mm"1, %%"#mm"4 \n\t" \
339 #mov" (%0), %%"#mm"3 \n\t" \
340 "paddsw %%"#mm"6, %%"#mm"1 \n\t" \
341 "psubsw 112(%0), %%"#mm"3 \n\t" \
342 "psubsw %%"#mm"6, %%"#mm"4 \n\t" \
343 #mov" (%1), %%"#mm"0 \n\t" \
344 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"3 \n\t" \
345 #mov" 32(%1), %%"#mm"6 \n\t" \
346 "pmulhw %%"#mm"1, %%"#mm"0 \n\t" \
347 #mov" %%"#mm"7, (%3) \n\t" \
348 "pmulhw %%"#mm"4, %%"#mm"6 \n\t" \
349 #mov" %%"#mm"5, 96(%3) \n\t" \
350 #mov" %%"#mm"3, %%"#mm"7 \n\t" \
351 #mov" 32(%1), %%"#mm"5 \n\t" \
352 "psubsw %%"#mm"2, %%"#mm"7 \n\t" \
353 "paddsw %%"#mm"2, %%"#mm"3 \n\t" \
354 "pmulhw %%"#mm"7, %%"#mm"5 \n\t" \
355 "paddsw %%"#mm"3, %%"#mm"0 \n\t" \
356 "paddsw %%"#mm"4, %%"#mm"6 \n\t" \
357 "pmulhw (%1), %%"#mm"3 \n\t" \
358 "por (%2), %%"#mm"0 \n\t" \
359 "paddsw %%"#mm"7, %%"#mm"5 \n\t" \
360 "psubsw %%"#mm"6, %%"#mm"7 \n\t" \
361 #mov" %%"#mm"0, 16(%3) \n\t" \
362 "paddsw %%"#mm"4, %%"#mm"5 \n\t" \
363 #mov" %%"#mm"7, 48(%3) \n\t" \
364 "psubsw %%"#mm"1, %%"#mm"3 \n\t" \
365 #mov" %%"#mm"5, 80(%3) \n\t" \
366 #mov" %%"#mm"3, 112(%3) \n\t" \
368 : "r" (in + offset), "r" (fdct_tg_all_16), "r" (fdct_one_corr), \
369 "r" (out + offset), "r" (ocos_4_16)); \
372 FDCT_COL(mmx, mm, movq)
373 FDCT_COL(sse2, xmm, movdqa)
375 static av_always_inline void fdct_row_sse2(const int16_t *in, int16_t *out)
378 #define FDCT_ROW_SSE2_H1(i,t) \
379 "movq " #i "(%0), %%xmm2 \n\t" \
380 "movq " #i "+8(%0), %%xmm0 \n\t" \
381 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
382 "movdqa " #t "+48(%1), %%xmm7 \n\t" \
383 "movdqa " #t "(%1), %%xmm4 \n\t" \
384 "movdqa " #t "+16(%1), %%xmm5 \n\t"
386 #define FDCT_ROW_SSE2_H2(i,t) \
387 "movq " #i "(%0), %%xmm2 \n\t" \
388 "movq " #i "+8(%0), %%xmm0 \n\t" \
389 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
390 "movdqa " #t "+48(%1), %%xmm7 \n\t"
392 #define FDCT_ROW_SSE2(i) \
393 "movq %%xmm2, %%xmm1 \n\t" \
394 "pshuflw $27, %%xmm0, %%xmm0 \n\t" \
395 "paddsw %%xmm0, %%xmm1 \n\t" \
396 "psubsw %%xmm0, %%xmm2 \n\t" \
397 "punpckldq %%xmm2, %%xmm1 \n\t" \
398 "pshufd $78, %%xmm1, %%xmm2 \n\t" \
399 "pmaddwd %%xmm2, %%xmm3 \n\t" \
400 "pmaddwd %%xmm1, %%xmm7 \n\t" \
401 "pmaddwd %%xmm5, %%xmm2 \n\t" \
402 "pmaddwd %%xmm4, %%xmm1 \n\t" \
403 "paddd %%xmm7, %%xmm3 \n\t" \
404 "paddd %%xmm2, %%xmm1 \n\t" \
405 "paddd %%xmm6, %%xmm3 \n\t" \
406 "paddd %%xmm6, %%xmm1 \n\t" \
407 "psrad %3, %%xmm3 \n\t" \
408 "psrad %3, %%xmm1 \n\t" \
409 "packssdw %%xmm3, %%xmm1 \n\t" \
410 "movdqa %%xmm1, " #i "(%4) \n\t"
412 "movdqa (%2), %%xmm6 \n\t"
413 FDCT_ROW_SSE2_H1(0,0)
415 FDCT_ROW_SSE2_H2(64,0)
418 FDCT_ROW_SSE2_H1(16,64)
420 FDCT_ROW_SSE2_H2(112,64)
423 FDCT_ROW_SSE2_H1(32,128)
425 FDCT_ROW_SSE2_H2(96,128)
428 FDCT_ROW_SSE2_H1(48,192)
430 FDCT_ROW_SSE2_H2(80,192)
433 : "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)
437 static av_always_inline void fdct_row_mmx2(const int16_t *in, int16_t *out, const int16_t *table)
440 "pshufw $0x1B, 8(%0), %%mm5 \n\t"
441 "movq (%0), %%mm0 \n\t"
442 "movq %%mm0, %%mm1 \n\t"
443 "paddsw %%mm5, %%mm0 \n\t"
444 "psubsw %%mm5, %%mm1 \n\t"
445 "movq %%mm0, %%mm2 \n\t"
446 "punpckldq %%mm1, %%mm0 \n\t"
447 "punpckhdq %%mm1, %%mm2 \n\t"
448 "movq (%1), %%mm1 \n\t"
449 "movq 8(%1), %%mm3 \n\t"
450 "movq 16(%1), %%mm4 \n\t"
451 "movq 24(%1), %%mm5 \n\t"
452 "movq 32(%1), %%mm6 \n\t"
453 "movq 40(%1), %%mm7 \n\t"
454 "pmaddwd %%mm0, %%mm1 \n\t"
455 "pmaddwd %%mm2, %%mm3 \n\t"
456 "pmaddwd %%mm0, %%mm4 \n\t"
457 "pmaddwd %%mm2, %%mm5 \n\t"
458 "pmaddwd %%mm0, %%mm6 \n\t"
459 "pmaddwd %%mm2, %%mm7 \n\t"
460 "pmaddwd 48(%1), %%mm0 \n\t"
461 "pmaddwd 56(%1), %%mm2 \n\t"
462 "paddd %%mm1, %%mm3 \n\t"
463 "paddd %%mm4, %%mm5 \n\t"
464 "paddd %%mm6, %%mm7 \n\t"
465 "paddd %%mm0, %%mm2 \n\t"
466 "movq (%2), %%mm0 \n\t"
467 "paddd %%mm0, %%mm3 \n\t"
468 "paddd %%mm0, %%mm5 \n\t"
469 "paddd %%mm0, %%mm7 \n\t"
470 "paddd %%mm0, %%mm2 \n\t"
471 "psrad $"S(SHIFT_FRW_ROW)", %%mm3 \n\t"
472 "psrad $"S(SHIFT_FRW_ROW)", %%mm5 \n\t"
473 "psrad $"S(SHIFT_FRW_ROW)", %%mm7 \n\t"
474 "psrad $"S(SHIFT_FRW_ROW)", %%mm2 \n\t"
475 "packssdw %%mm5, %%mm3 \n\t"
476 "packssdw %%mm2, %%mm7 \n\t"
477 "movq %%mm3, (%3) \n\t"
478 "movq %%mm7, 8(%3) \n\t"
480 : "r" (in), "r" (table), "r" (fdct_r_row), "r" (out));
483 static av_always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table)
485 //FIXME reorder (I do not have an old MMX-only CPU here to benchmark ...)
487 "movd 12(%0), %%mm1 \n\t"
488 "punpcklwd 8(%0), %%mm1 \n\t"
489 "movq %%mm1, %%mm2 \n\t"
490 "psrlq $0x20, %%mm1 \n\t"
491 "movq 0(%0), %%mm0 \n\t"
492 "punpcklwd %%mm2, %%mm1 \n\t"
493 "movq %%mm0, %%mm5 \n\t"
494 "paddsw %%mm1, %%mm0 \n\t"
495 "psubsw %%mm1, %%mm5 \n\t"
496 "movq %%mm0, %%mm2 \n\t"
497 "punpckldq %%mm5, %%mm0 \n\t"
498 "punpckhdq %%mm5, %%mm2 \n\t"
499 "movq 0(%1), %%mm1 \n\t"
500 "movq 8(%1), %%mm3 \n\t"
501 "movq 16(%1), %%mm4 \n\t"
502 "movq 24(%1), %%mm5 \n\t"
503 "movq 32(%1), %%mm6 \n\t"
504 "movq 40(%1), %%mm7 \n\t"
505 "pmaddwd %%mm0, %%mm1 \n\t"
506 "pmaddwd %%mm2, %%mm3 \n\t"
507 "pmaddwd %%mm0, %%mm4 \n\t"
508 "pmaddwd %%mm2, %%mm5 \n\t"
509 "pmaddwd %%mm0, %%mm6 \n\t"
510 "pmaddwd %%mm2, %%mm7 \n\t"
511 "pmaddwd 48(%1), %%mm0 \n\t"
512 "pmaddwd 56(%1), %%mm2 \n\t"
513 "paddd %%mm1, %%mm3 \n\t"
514 "paddd %%mm4, %%mm5 \n\t"
515 "paddd %%mm6, %%mm7 \n\t"
516 "paddd %%mm0, %%mm2 \n\t"
517 "movq (%2), %%mm0 \n\t"
518 "paddd %%mm0, %%mm3 \n\t"
519 "paddd %%mm0, %%mm5 \n\t"
520 "paddd %%mm0, %%mm7 \n\t"
521 "paddd %%mm0, %%mm2 \n\t"
522 "psrad $"S(SHIFT_FRW_ROW)", %%mm3 \n\t"
523 "psrad $"S(SHIFT_FRW_ROW)", %%mm5 \n\t"
524 "psrad $"S(SHIFT_FRW_ROW)", %%mm7 \n\t"
525 "psrad $"S(SHIFT_FRW_ROW)", %%mm2 \n\t"
526 "packssdw %%mm5, %%mm3 \n\t"
527 "packssdw %%mm2, %%mm7 \n\t"
528 "movq %%mm3, 0(%3) \n\t"
529 "movq %%mm7, 8(%3) \n\t"
531 : "r" (in), "r" (table), "r" (fdct_r_row), "r" (out));
534 void ff_fdct_mmx(int16_t *block)
536 DECLARE_ALIGNED(8, int64_t, align_tmp)[16];
537 int16_t * block1= (int16_t*)align_tmp;
538 const int16_t *table= tab_frw_01234567;
541 fdct_col_mmx(block, block1, 0);
542 fdct_col_mmx(block, block1, 4);
545 fdct_row_mmx(block1, block, table);
552 void ff_fdct_mmx2(int16_t *block)
554 DECLARE_ALIGNED(8, int64_t, align_tmp)[16];
555 int16_t *block1= (int16_t*)align_tmp;
556 const int16_t *table= tab_frw_01234567;
559 fdct_col_mmx(block, block1, 0);
560 fdct_col_mmx(block, block1, 4);
563 fdct_row_mmx2(block1, block, table);
570 void ff_fdct_sse2(int16_t *block)
572 DECLARE_ALIGNED(16, int64_t, align_tmp)[16];
573 int16_t * const block1= (int16_t*)align_tmp;
575 fdct_col_sse2(block, block1, 0);
576 fdct_row_sse2(block1, block);