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 "libavutil/x86_cpu.h"
35 #include "libavcodec/dsputil.h"
37 //////////////////////////////////////////////////////////////////////
39 // constants for the forward DCT
40 // -----------------------------
42 // Be sure to check that your compiler is aligning all constants to QWORD
43 // (8-byte) memory boundaries! Otherwise the unaligned memory access will
44 // severely stall MMX execution.
46 //////////////////////////////////////////////////////////////////////
48 #define BITS_FRW_ACC 3 //; 2 or 3 for accuracy
49 #define SHIFT_FRW_COL BITS_FRW_ACC
50 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17 - 3)
51 #define RND_FRW_ROW (1 << (SHIFT_FRW_ROW-1))
52 //#define RND_FRW_COL (1 << (SHIFT_FRW_COL-1))
54 #define X8(x) x,x,x,x,x,x,x,x
56 //concatenated table, for forward DCT transformation
57 DECLARE_ALIGNED(16, static const int16_t, fdct_tg_all_16)[24] = {
58 X8(13036), // tg * (2<<16) + 0.5
59 X8(27146), // tg * (2<<16) + 0.5
60 X8(-21746) // tg * (2<<16) + 0.5
63 DECLARE_ALIGNED(16, static const int16_t, ocos_4_16)[8] = {
64 X8(23170) //cos * (2<<15) + 0.5
67 DECLARE_ALIGNED(16, static const int16_t, fdct_one_corr)[8] = { X8(1) };
69 DECLARE_ALIGNED(8, static const int32_t, fdct_r_row)[2] = {RND_FRW_ROW, RND_FRW_ROW };
73 DECLARE_ALIGNED(16, const int32_t, fdct_r_row_sse2)[4];
76 RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW
78 //DECLARE_ALIGNED(16, static const long, fdct_r_row_sse2)[4] = {RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW};
80 DECLARE_ALIGNED(8, static const int16_t, tab_frw_01234567)[] = { // forward_dct coeff table
81 16384, 16384, 22725, 19266,
82 16384, 16384, 12873, 4520,
83 21407, 8867, 19266, -4520,
84 -8867, -21407, -22725, -12873,
85 16384, -16384, 12873, -22725,
86 -16384, 16384, 4520, 19266,
87 8867, -21407, 4520, -12873,
88 21407, -8867, 19266, -22725,
90 22725, 22725, 31521, 26722,
91 22725, 22725, 17855, 6270,
92 29692, 12299, 26722, -6270,
93 -12299, -29692, -31521, -17855,
94 22725, -22725, 17855, -31521,
95 -22725, 22725, 6270, 26722,
96 12299, -29692, 6270, -17855,
97 29692, -12299, 26722, -31521,
99 21407, 21407, 29692, 25172,
100 21407, 21407, 16819, 5906,
101 27969, 11585, 25172, -5906,
102 -11585, -27969, -29692, -16819,
103 21407, -21407, 16819, -29692,
104 -21407, 21407, 5906, 25172,
105 11585, -27969, 5906, -16819,
106 27969, -11585, 25172, -29692,
108 19266, 19266, 26722, 22654,
109 19266, 19266, 15137, 5315,
110 25172, 10426, 22654, -5315,
111 -10426, -25172, -26722, -15137,
112 19266, -19266, 15137, -26722,
113 -19266, 19266, 5315, 22654,
114 10426, -25172, 5315, -15137,
115 25172, -10426, 22654, -26722,
117 16384, 16384, 22725, 19266,
118 16384, 16384, 12873, 4520,
119 21407, 8867, 19266, -4520,
120 -8867, -21407, -22725, -12873,
121 16384, -16384, 12873, -22725,
122 -16384, 16384, 4520, 19266,
123 8867, -21407, 4520, -12873,
124 21407, -8867, 19266, -22725,
126 19266, 19266, 26722, 22654,
127 19266, 19266, 15137, 5315,
128 25172, 10426, 22654, -5315,
129 -10426, -25172, -26722, -15137,
130 19266, -19266, 15137, -26722,
131 -19266, 19266, 5315, 22654,
132 10426, -25172, 5315, -15137,
133 25172, -10426, 22654, -26722,
135 21407, 21407, 29692, 25172,
136 21407, 21407, 16819, 5906,
137 27969, 11585, 25172, -5906,
138 -11585, -27969, -29692, -16819,
139 21407, -21407, 16819, -29692,
140 -21407, 21407, 5906, 25172,
141 11585, -27969, 5906, -16819,
142 27969, -11585, 25172, -29692,
144 22725, 22725, 31521, 26722,
145 22725, 22725, 17855, 6270,
146 29692, 12299, 26722, -6270,
147 -12299, -29692, -31521, -17855,
148 22725, -22725, 17855, -31521,
149 -22725, 22725, 6270, 26722,
150 12299, -29692, 6270, -17855,
151 29692, -12299, 26722, -31521,
156 DECLARE_ALIGNED(16, const int16_t, tab_frw_01234567_sse2)[256];
157 } tab_frw_01234567_sse2 =
159 //DECLARE_ALIGNED(16, static const int16_t, tab_frw_01234567_sse2)[] = { // forward_dct coeff table
160 #define TABLE_SSE2 C4, C4, C1, C3, -C6, -C2, -C1, -C5, \
161 C4, C4, C5, C7, C2, C6, C3, -C7, \
162 -C4, C4, C7, C3, C6, -C2, C7, -C5, \
163 C4, -C4, C5, -C1, C2, -C6, C3, -C1,
164 // c1..c7 * cos(pi/4) * 2^15
287 #define S(s) AV_TOSTRING(s) //AV_STRINGIFY is too long
289 #define FDCT_COL(cpu, mm, mov)\
290 static av_always_inline void fdct_col_##cpu(const int16_t *in, int16_t *out, int offset)\
293 #mov" 16(%0), %%"#mm"0 \n\t" \
294 #mov" 96(%0), %%"#mm"1 \n\t" \
295 #mov" %%"#mm"0, %%"#mm"2 \n\t" \
296 #mov" 32(%0), %%"#mm"3 \n\t" \
297 "paddsw %%"#mm"1, %%"#mm"0 \n\t" \
298 #mov" 80(%0), %%"#mm"4 \n\t" \
299 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"0 \n\t" \
300 #mov" (%0), %%"#mm"5 \n\t" \
301 "paddsw %%"#mm"3, %%"#mm"4 \n\t" \
302 "paddsw 112(%0), %%"#mm"5 \n\t" \
303 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"4 \n\t" \
304 #mov" %%"#mm"0, %%"#mm"6 \n\t" \
305 "psubsw %%"#mm"1, %%"#mm"2 \n\t" \
306 #mov" 16(%1), %%"#mm"1 \n\t" \
307 "psubsw %%"#mm"4, %%"#mm"0 \n\t" \
308 #mov" 48(%0), %%"#mm"7 \n\t" \
309 "pmulhw %%"#mm"0, %%"#mm"1 \n\t" \
310 "paddsw 64(%0), %%"#mm"7 \n\t" \
311 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"5 \n\t" \
312 "paddsw %%"#mm"4, %%"#mm"6 \n\t" \
313 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"7 \n\t" \
314 #mov" %%"#mm"5, %%"#mm"4 \n\t" \
315 "psubsw %%"#mm"7, %%"#mm"5 \n\t" \
316 "paddsw %%"#mm"5, %%"#mm"1 \n\t" \
317 "paddsw %%"#mm"7, %%"#mm"4 \n\t" \
318 "por (%2), %%"#mm"1 \n\t" \
319 "psllw $"S(SHIFT_FRW_COL)"+1, %%"#mm"2 \n\t" \
320 "pmulhw 16(%1), %%"#mm"5 \n\t" \
321 #mov" %%"#mm"4, %%"#mm"7 \n\t" \
322 "psubsw 80(%0), %%"#mm"3 \n\t" \
323 "psubsw %%"#mm"6, %%"#mm"4 \n\t" \
324 #mov" %%"#mm"1, 32(%3) \n\t" \
325 "paddsw %%"#mm"6, %%"#mm"7 \n\t" \
326 #mov" 48(%0), %%"#mm"1 \n\t" \
327 "psllw $"S(SHIFT_FRW_COL)"+1, %%"#mm"3 \n\t" \
328 "psubsw 64(%0), %%"#mm"1 \n\t" \
329 #mov" %%"#mm"2, %%"#mm"6 \n\t" \
330 #mov" %%"#mm"4, 64(%3) \n\t" \
331 "paddsw %%"#mm"3, %%"#mm"2 \n\t" \
332 "pmulhw (%4), %%"#mm"2 \n\t" \
333 "psubsw %%"#mm"3, %%"#mm"6 \n\t" \
334 "pmulhw (%4), %%"#mm"6 \n\t" \
335 "psubsw %%"#mm"0, %%"#mm"5 \n\t" \
336 "por (%2), %%"#mm"5 \n\t" \
337 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"1 \n\t" \
338 "por (%2), %%"#mm"2 \n\t" \
339 #mov" %%"#mm"1, %%"#mm"4 \n\t" \
340 #mov" (%0), %%"#mm"3 \n\t" \
341 "paddsw %%"#mm"6, %%"#mm"1 \n\t" \
342 "psubsw 112(%0), %%"#mm"3 \n\t" \
343 "psubsw %%"#mm"6, %%"#mm"4 \n\t" \
344 #mov" (%1), %%"#mm"0 \n\t" \
345 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"3 \n\t" \
346 #mov" 32(%1), %%"#mm"6 \n\t" \
347 "pmulhw %%"#mm"1, %%"#mm"0 \n\t" \
348 #mov" %%"#mm"7, (%3) \n\t" \
349 "pmulhw %%"#mm"4, %%"#mm"6 \n\t" \
350 #mov" %%"#mm"5, 96(%3) \n\t" \
351 #mov" %%"#mm"3, %%"#mm"7 \n\t" \
352 #mov" 32(%1), %%"#mm"5 \n\t" \
353 "psubsw %%"#mm"2, %%"#mm"7 \n\t" \
354 "paddsw %%"#mm"2, %%"#mm"3 \n\t" \
355 "pmulhw %%"#mm"7, %%"#mm"5 \n\t" \
356 "paddsw %%"#mm"3, %%"#mm"0 \n\t" \
357 "paddsw %%"#mm"4, %%"#mm"6 \n\t" \
358 "pmulhw (%1), %%"#mm"3 \n\t" \
359 "por (%2), %%"#mm"0 \n\t" \
360 "paddsw %%"#mm"7, %%"#mm"5 \n\t" \
361 "psubsw %%"#mm"6, %%"#mm"7 \n\t" \
362 #mov" %%"#mm"0, 16(%3) \n\t" \
363 "paddsw %%"#mm"4, %%"#mm"5 \n\t" \
364 #mov" %%"#mm"7, 48(%3) \n\t" \
365 "psubsw %%"#mm"1, %%"#mm"3 \n\t" \
366 #mov" %%"#mm"5, 80(%3) \n\t" \
367 #mov" %%"#mm"3, 112(%3) \n\t" \
369 : "r" (in + offset), "r" (fdct_tg_all_16), "r" (fdct_one_corr), \
370 "r" (out + offset), "r" (ocos_4_16)); \
373 FDCT_COL(mmx, mm, movq)
374 FDCT_COL(sse2, xmm, movdqa)
376 static av_always_inline void fdct_row_sse2(const int16_t *in, int16_t *out)
379 #define FDCT_ROW_SSE2_H1(i,t) \
380 "movq " #i "(%0), %%xmm2 \n\t" \
381 "movq " #i "+8(%0), %%xmm0 \n\t" \
382 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
383 "movdqa " #t "+48(%1), %%xmm7 \n\t" \
384 "movdqa " #t "(%1), %%xmm4 \n\t" \
385 "movdqa " #t "+16(%1), %%xmm5 \n\t"
387 #define FDCT_ROW_SSE2_H2(i,t) \
388 "movq " #i "(%0), %%xmm2 \n\t" \
389 "movq " #i "+8(%0), %%xmm0 \n\t" \
390 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
391 "movdqa " #t "+48(%1), %%xmm7 \n\t"
393 #define FDCT_ROW_SSE2(i) \
394 "movq %%xmm2, %%xmm1 \n\t" \
395 "pshuflw $27, %%xmm0, %%xmm0 \n\t" \
396 "paddsw %%xmm0, %%xmm1 \n\t" \
397 "psubsw %%xmm0, %%xmm2 \n\t" \
398 "punpckldq %%xmm2, %%xmm1 \n\t" \
399 "pshufd $78, %%xmm1, %%xmm2 \n\t" \
400 "pmaddwd %%xmm2, %%xmm3 \n\t" \
401 "pmaddwd %%xmm1, %%xmm7 \n\t" \
402 "pmaddwd %%xmm5, %%xmm2 \n\t" \
403 "pmaddwd %%xmm4, %%xmm1 \n\t" \
404 "paddd %%xmm7, %%xmm3 \n\t" \
405 "paddd %%xmm2, %%xmm1 \n\t" \
406 "paddd %%xmm6, %%xmm3 \n\t" \
407 "paddd %%xmm6, %%xmm1 \n\t" \
408 "psrad %3, %%xmm3 \n\t" \
409 "psrad %3, %%xmm1 \n\t" \
410 "packssdw %%xmm3, %%xmm1 \n\t" \
411 "movdqa %%xmm1, " #i "(%4) \n\t"
413 "movdqa (%2), %%xmm6 \n\t"
414 FDCT_ROW_SSE2_H1(0,0)
416 FDCT_ROW_SSE2_H2(64,0)
419 FDCT_ROW_SSE2_H1(16,64)
421 FDCT_ROW_SSE2_H2(112,64)
424 FDCT_ROW_SSE2_H1(32,128)
426 FDCT_ROW_SSE2_H2(96,128)
429 FDCT_ROW_SSE2_H1(48,192)
431 FDCT_ROW_SSE2_H2(80,192)
434 : "r" (in), "r" (tab_frw_01234567_sse2.tab_frw_01234567_sse2),
435 "r" (fdct_r_row_sse2.fdct_r_row_sse2), "i" (SHIFT_FRW_ROW), "r" (out)
436 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
437 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
441 static av_always_inline void fdct_row_mmx2(const int16_t *in, int16_t *out, const int16_t *table)
444 "pshufw $0x1B, 8(%0), %%mm5 \n\t"
445 "movq (%0), %%mm0 \n\t"
446 "movq %%mm0, %%mm1 \n\t"
447 "paddsw %%mm5, %%mm0 \n\t"
448 "psubsw %%mm5, %%mm1 \n\t"
449 "movq %%mm0, %%mm2 \n\t"
450 "punpckldq %%mm1, %%mm0 \n\t"
451 "punpckhdq %%mm1, %%mm2 \n\t"
452 "movq (%1), %%mm1 \n\t"
453 "movq 8(%1), %%mm3 \n\t"
454 "movq 16(%1), %%mm4 \n\t"
455 "movq 24(%1), %%mm5 \n\t"
456 "movq 32(%1), %%mm6 \n\t"
457 "movq 40(%1), %%mm7 \n\t"
458 "pmaddwd %%mm0, %%mm1 \n\t"
459 "pmaddwd %%mm2, %%mm3 \n\t"
460 "pmaddwd %%mm0, %%mm4 \n\t"
461 "pmaddwd %%mm2, %%mm5 \n\t"
462 "pmaddwd %%mm0, %%mm6 \n\t"
463 "pmaddwd %%mm2, %%mm7 \n\t"
464 "pmaddwd 48(%1), %%mm0 \n\t"
465 "pmaddwd 56(%1), %%mm2 \n\t"
466 "paddd %%mm1, %%mm3 \n\t"
467 "paddd %%mm4, %%mm5 \n\t"
468 "paddd %%mm6, %%mm7 \n\t"
469 "paddd %%mm0, %%mm2 \n\t"
470 "movq (%2), %%mm0 \n\t"
471 "paddd %%mm0, %%mm3 \n\t"
472 "paddd %%mm0, %%mm5 \n\t"
473 "paddd %%mm0, %%mm7 \n\t"
474 "paddd %%mm0, %%mm2 \n\t"
475 "psrad $"S(SHIFT_FRW_ROW)", %%mm3 \n\t"
476 "psrad $"S(SHIFT_FRW_ROW)", %%mm5 \n\t"
477 "psrad $"S(SHIFT_FRW_ROW)", %%mm7 \n\t"
478 "psrad $"S(SHIFT_FRW_ROW)", %%mm2 \n\t"
479 "packssdw %%mm5, %%mm3 \n\t"
480 "packssdw %%mm2, %%mm7 \n\t"
481 "movq %%mm3, (%3) \n\t"
482 "movq %%mm7, 8(%3) \n\t"
484 : "r" (in), "r" (table), "r" (fdct_r_row), "r" (out));
487 static av_always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table)
489 //FIXME reorder (I do not have an old MMX-only CPU here to benchmark ...)
491 "movd 12(%0), %%mm1 \n\t"
492 "punpcklwd 8(%0), %%mm1 \n\t"
493 "movq %%mm1, %%mm2 \n\t"
494 "psrlq $0x20, %%mm1 \n\t"
495 "movq 0(%0), %%mm0 \n\t"
496 "punpcklwd %%mm2, %%mm1 \n\t"
497 "movq %%mm0, %%mm5 \n\t"
498 "paddsw %%mm1, %%mm0 \n\t"
499 "psubsw %%mm1, %%mm5 \n\t"
500 "movq %%mm0, %%mm2 \n\t"
501 "punpckldq %%mm5, %%mm0 \n\t"
502 "punpckhdq %%mm5, %%mm2 \n\t"
503 "movq 0(%1), %%mm1 \n\t"
504 "movq 8(%1), %%mm3 \n\t"
505 "movq 16(%1), %%mm4 \n\t"
506 "movq 24(%1), %%mm5 \n\t"
507 "movq 32(%1), %%mm6 \n\t"
508 "movq 40(%1), %%mm7 \n\t"
509 "pmaddwd %%mm0, %%mm1 \n\t"
510 "pmaddwd %%mm2, %%mm3 \n\t"
511 "pmaddwd %%mm0, %%mm4 \n\t"
512 "pmaddwd %%mm2, %%mm5 \n\t"
513 "pmaddwd %%mm0, %%mm6 \n\t"
514 "pmaddwd %%mm2, %%mm7 \n\t"
515 "pmaddwd 48(%1), %%mm0 \n\t"
516 "pmaddwd 56(%1), %%mm2 \n\t"
517 "paddd %%mm1, %%mm3 \n\t"
518 "paddd %%mm4, %%mm5 \n\t"
519 "paddd %%mm6, %%mm7 \n\t"
520 "paddd %%mm0, %%mm2 \n\t"
521 "movq (%2), %%mm0 \n\t"
522 "paddd %%mm0, %%mm3 \n\t"
523 "paddd %%mm0, %%mm5 \n\t"
524 "paddd %%mm0, %%mm7 \n\t"
525 "paddd %%mm0, %%mm2 \n\t"
526 "psrad $"S(SHIFT_FRW_ROW)", %%mm3 \n\t"
527 "psrad $"S(SHIFT_FRW_ROW)", %%mm5 \n\t"
528 "psrad $"S(SHIFT_FRW_ROW)", %%mm7 \n\t"
529 "psrad $"S(SHIFT_FRW_ROW)", %%mm2 \n\t"
530 "packssdw %%mm5, %%mm3 \n\t"
531 "packssdw %%mm2, %%mm7 \n\t"
532 "movq %%mm3, 0(%3) \n\t"
533 "movq %%mm7, 8(%3) \n\t"
535 : "r" (in), "r" (table), "r" (fdct_r_row), "r" (out));
538 void ff_fdct_mmx(int16_t *block)
540 DECLARE_ALIGNED(8, int64_t, align_tmp)[16];
541 int16_t * block1= (int16_t*)align_tmp;
542 const int16_t *table= tab_frw_01234567;
545 fdct_col_mmx(block, block1, 0);
546 fdct_col_mmx(block, block1, 4);
549 fdct_row_mmx(block1, block, table);
556 void ff_fdct_mmx2(int16_t *block)
558 DECLARE_ALIGNED(8, int64_t, align_tmp)[16];
559 int16_t *block1= (int16_t*)align_tmp;
560 const int16_t *table= tab_frw_01234567;
563 fdct_col_mmx(block, block1, 0);
564 fdct_col_mmx(block, block1, 4);
567 fdct_row_mmx2(block1, block, table);
574 void ff_fdct_sse2(int16_t *block)
576 DECLARE_ALIGNED(16, int64_t, align_tmp)[16];
577 int16_t * const block1= (int16_t*)align_tmp;
579 fdct_col_sse2(block, block1, 0);
580 fdct_row_sse2(block1, block);