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
17 #include "../dsputil.h"
20 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
22 //////////////////////////////////////////////////////////////////////
24 // constants for the forward DCT
25 // -----------------------------
27 // Be sure to check that your compiler is aligning all constants to QWORD
28 // (8-byte) memory boundaries! Otherwise the unaligned memory access will
29 // severely stall MMX execution.
31 //////////////////////////////////////////////////////////////////////
33 #define BITS_FRW_ACC 3 //; 2 or 3 for accuracy
34 #define SHIFT_FRW_COL BITS_FRW_ACC
35 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17 - 3)
36 #define RND_FRW_ROW (1 << (SHIFT_FRW_ROW-1))
37 //#define RND_FRW_COL (1 << (SHIFT_FRW_COL-1))
39 //concatenated table, for forward DCT transformation
40 static const int16_t fdct_tg_all_16[] ATTR_ALIGN(8) = {
41 13036, 13036, 13036, 13036, // tg * (2<<16) + 0.5
42 27146, 27146, 27146, 27146, // tg * (2<<16) + 0.5
43 -21746, -21746, -21746, -21746, // tg * (2<<16) + 0.5
46 static const int16_t ocos_4_16[4] ATTR_ALIGN(8) = {
47 23170, 23170, 23170, 23170, //cos * (2<<15) + 0.5
50 static const int64_t fdct_one_corr ATTR_ALIGN(8) = 0x0001000100010001LL;
52 static const int32_t fdct_r_row[2] ATTR_ALIGN(8) = {RND_FRW_ROW, RND_FRW_ROW };
56 const int32_t fdct_r_row_sse2[4] ATTR_ALIGN(16);
57 } fdct_r_row_sse2 ATTR_ALIGN(16)=
59 RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW
61 //static const long fdct_r_row_sse2[4] ATTR_ALIGN(16) = {RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW};
63 static const int16_t tab_frw_01234567[] ATTR_ALIGN(8) = { // forward_dct coeff table
64 16384, 16384, 22725, 19266,
65 16384, 16384, 12873, 4520,
66 21407, 8867, 19266, -4520,
67 -8867, -21407, -22725, -12873,
68 16384, -16384, 12873, -22725,
69 -16384, 16384, 4520, 19266,
70 8867, -21407, 4520, -12873,
71 21407, -8867, 19266, -22725,
73 22725, 22725, 31521, 26722,
74 22725, 22725, 17855, 6270,
75 29692, 12299, 26722, -6270,
76 -12299, -29692, -31521, -17855,
77 22725, -22725, 17855, -31521,
78 -22725, 22725, 6270, 26722,
79 12299, -29692, 6270, -17855,
80 29692, -12299, 26722, -31521,
82 21407, 21407, 29692, 25172,
83 21407, 21407, 16819, 5906,
84 27969, 11585, 25172, -5906,
85 -11585, -27969, -29692, -16819,
86 21407, -21407, 16819, -29692,
87 -21407, 21407, 5906, 25172,
88 11585, -27969, 5906, -16819,
89 27969, -11585, 25172, -29692,
91 19266, 19266, 26722, 22654,
92 19266, 19266, 15137, 5315,
93 25172, 10426, 22654, -5315,
94 -10426, -25172, -26722, -15137,
95 19266, -19266, 15137, -26722,
96 -19266, 19266, 5315, 22654,
97 10426, -25172, 5315, -15137,
98 25172, -10426, 22654, -26722,
100 16384, 16384, 22725, 19266,
101 16384, 16384, 12873, 4520,
102 21407, 8867, 19266, -4520,
103 -8867, -21407, -22725, -12873,
104 16384, -16384, 12873, -22725,
105 -16384, 16384, 4520, 19266,
106 8867, -21407, 4520, -12873,
107 21407, -8867, 19266, -22725,
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 21407, 21407, 29692, 25172,
119 21407, 21407, 16819, 5906,
120 27969, 11585, 25172, -5906,
121 -11585, -27969, -29692, -16819,
122 21407, -21407, 16819, -29692,
123 -21407, 21407, 5906, 25172,
124 11585, -27969, 5906, -16819,
125 27969, -11585, 25172, -29692,
127 22725, 22725, 31521, 26722,
128 22725, 22725, 17855, 6270,
129 29692, 12299, 26722, -6270,
130 -12299, -29692, -31521, -17855,
131 22725, -22725, 17855, -31521,
132 -22725, 22725, 6270, 26722,
133 12299, -29692, 6270, -17855,
134 29692, -12299, 26722, -31521,
139 const int16_t tab_frw_01234567_sse2[256] ATTR_ALIGN(16);
140 } tab_frw_01234567_sse2 ATTR_ALIGN(16) =
142 //static const int16_t tab_frw_01234567_sse2[] ATTR_ALIGN(16) = { // forward_dct coeff table
143 #define TABLE_SSE2 C4, C4, C1, C3, -C6, -C2, -C1, -C5, \
144 C4, C4, C5, C7, C2, C6, C3, -C7, \
145 -C4, C4, C7, C3, C6, -C2, C7, -C5, \
146 C4, -C4, C5, -C1, C2, -C6, C3, -C1,
147 // c1..c7 * cos(pi/4) * 2^15
271 static always_inline void fdct_col(const int16_t *in, int16_t *out, int offset)
273 movq_m2r(*(in + offset + 1 * 8), mm0);
274 movq_m2r(*(in + offset + 6 * 8), mm1);
276 movq_m2r(*(in + offset + 2 * 8), mm3);
277 paddsw_r2r(mm1, mm0);
278 movq_m2r(*(in + offset + 5 * 8), mm4);
279 psllw_i2r(SHIFT_FRW_COL, mm0);
280 movq_m2r(*(in + offset + 0 * 8), mm5);
281 paddsw_r2r(mm3, mm4);
282 paddsw_m2r(*(in + offset + 7 * 8), mm5);
283 psllw_i2r(SHIFT_FRW_COL, mm4);
285 psubsw_r2r(mm1, mm2);
286 movq_m2r(*(fdct_tg_all_16 + 4), mm1);
287 psubsw_r2r(mm4, mm0);
288 movq_m2r(*(in + offset + 3 * 8), mm7);
289 pmulhw_r2r(mm0, mm1);
290 paddsw_m2r(*(in + offset + 4 * 8), mm7);
291 psllw_i2r(SHIFT_FRW_COL, mm5);
292 paddsw_r2r(mm4, mm6);
293 psllw_i2r(SHIFT_FRW_COL, mm7);
295 psubsw_r2r(mm7, mm5);
296 paddsw_r2r(mm5, mm1);
297 paddsw_r2r(mm7, mm4);
298 por_m2r(fdct_one_corr, mm1);
299 psllw_i2r(SHIFT_FRW_COL + 1, mm2);
300 pmulhw_m2r(*(fdct_tg_all_16 + 4), mm5);
302 psubsw_m2r(*(in + offset + 5 * 8), mm3);
303 psubsw_r2r(mm6, mm4);
304 movq_r2m(mm1, *(out + offset + 2 * 8));
305 paddsw_r2r(mm6, mm7);
306 movq_m2r(*(in + offset + 3 * 8), mm1);
307 psllw_i2r(SHIFT_FRW_COL + 1, mm3);
308 psubsw_m2r(*(in + offset + 4 * 8), mm1);
310 movq_r2m(mm4, *(out + offset + 4 * 8));
311 paddsw_r2r(mm3, mm2);
312 pmulhw_m2r(*ocos_4_16, mm2);
313 psubsw_r2r(mm3, mm6);
314 pmulhw_m2r(*ocos_4_16, mm6);
315 psubsw_r2r(mm0, mm5);
316 por_m2r(fdct_one_corr, mm5);
317 psllw_i2r(SHIFT_FRW_COL, mm1);
318 por_m2r(fdct_one_corr, mm2);
320 movq_m2r(*(in + offset + 0 * 8), mm3);
321 paddsw_r2r(mm6, mm1);
322 psubsw_m2r(*(in + offset + 7 * 8), mm3);
323 psubsw_r2r(mm6, mm4);
324 movq_m2r(*(fdct_tg_all_16 + 0), mm0);
325 psllw_i2r(SHIFT_FRW_COL, mm3);
326 movq_m2r(*(fdct_tg_all_16 + 8), mm6);
327 pmulhw_r2r(mm1, mm0);
328 movq_r2m(mm7, *(out + offset + 0 * 8));
329 pmulhw_r2r(mm4, mm6);
330 movq_r2m(mm5, *(out + offset + 6 * 8));
332 movq_m2r(*(fdct_tg_all_16 + 8), mm5);
333 psubsw_r2r(mm2, mm7);
334 paddsw_r2r(mm2, mm3);
335 pmulhw_r2r(mm7, mm5);
336 paddsw_r2r(mm3, mm0);
337 paddsw_r2r(mm4, mm6);
338 pmulhw_m2r(*(fdct_tg_all_16 + 0), mm3);
339 por_m2r(fdct_one_corr, mm0);
340 paddsw_r2r(mm7, mm5);
341 psubsw_r2r(mm6, mm7);
342 movq_r2m(mm0, *(out + offset + 1 * 8));
343 paddsw_r2r(mm4, mm5);
344 movq_r2m(mm7, *(out + offset + 3 * 8));
345 psubsw_r2r(mm1, mm3);
346 movq_r2m(mm5, *(out + offset + 5 * 8));
347 movq_r2m(mm3, *(out + offset + 7 * 8));
351 static always_inline void fdct_row_sse2(const int16_t *in, int16_t *out)
354 #define FDCT_ROW_SSE2_H1(i,t) \
355 "movq " #i "(%0), %%xmm2 \n\t" \
356 "movq " #i "+8(%0), %%xmm0 \n\t" \
357 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
358 "movdqa " #t "+48(%1), %%xmm7 \n\t" \
359 "movdqa " #t "(%1), %%xmm4 \n\t" \
360 "movdqa " #t "+16(%1), %%xmm5 \n\t"
362 #define FDCT_ROW_SSE2_H2(i,t) \
363 "movq " #i "(%0), %%xmm2 \n\t" \
364 "movq " #i "+8(%0), %%xmm0 \n\t" \
365 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
366 "movdqa " #t "+48(%1), %%xmm7 \n\t"
368 #define FDCT_ROW_SSE2(i) \
369 "movq %%xmm2, %%xmm1 \n\t" \
370 "pshuflw $27, %%xmm0, %%xmm0 \n\t" \
371 "paddsw %%xmm0, %%xmm1 \n\t" \
372 "psubsw %%xmm0, %%xmm2 \n\t" \
373 "punpckldq %%xmm2, %%xmm1 \n\t" \
374 "pshufd $78, %%xmm1, %%xmm2 \n\t" \
375 "pmaddwd %%xmm2, %%xmm3 \n\t" \
376 "pmaddwd %%xmm1, %%xmm7 \n\t" \
377 "pmaddwd %%xmm5, %%xmm2 \n\t" \
378 "pmaddwd %%xmm4, %%xmm1 \n\t" \
379 "paddd %%xmm7, %%xmm3 \n\t" \
380 "paddd %%xmm2, %%xmm1 \n\t" \
381 "paddd %%xmm6, %%xmm3 \n\t" \
382 "paddd %%xmm6, %%xmm1 \n\t" \
383 "psrad %3, %%xmm3 \n\t" \
384 "psrad %3, %%xmm1 \n\t" \
385 "packssdw %%xmm3, %%xmm1 \n\t" \
386 "movdqa %%xmm1, " #i "(%4) \n\t"
388 "movdqa (%2), %%xmm6 \n\t"
389 FDCT_ROW_SSE2_H1(0,0)
391 FDCT_ROW_SSE2_H2(64,0)
394 FDCT_ROW_SSE2_H1(16,64)
396 FDCT_ROW_SSE2_H2(112,64)
399 FDCT_ROW_SSE2_H1(32,128)
401 FDCT_ROW_SSE2_H2(96,128)
404 FDCT_ROW_SSE2_H1(48,192)
406 FDCT_ROW_SSE2_H2(80,192)
409 : "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)
413 static always_inline void fdct_row_mmx2(const int16_t *in, int16_t *out, const int16_t *table)
415 pshufw_m2r(*(in + 4), mm5, 0x1B);
416 movq_m2r(*(in + 0), mm0);
418 paddsw_r2r(mm5, mm0);
419 psubsw_r2r(mm5, mm1);
421 punpckldq_r2r(mm1, mm0);
422 punpckhdq_r2r(mm1, mm2);
423 movq_m2r(*(table + 0), mm1);
424 movq_m2r(*(table + 4), mm3);
425 movq_m2r(*(table + 8), mm4);
426 movq_m2r(*(table + 12), mm5);
427 movq_m2r(*(table + 16), mm6);
428 movq_m2r(*(table + 20), mm7);
429 pmaddwd_r2r(mm0, mm1);
430 pmaddwd_r2r(mm2, mm3);
431 pmaddwd_r2r(mm0, mm4);
432 pmaddwd_r2r(mm2, mm5);
433 pmaddwd_r2r(mm0, mm6);
434 pmaddwd_r2r(mm2, mm7);
435 pmaddwd_m2r(*(table + 24), mm0);
436 pmaddwd_m2r(*(table + 28), mm2);
441 movq_m2r(*fdct_r_row, mm0);
446 psrad_i2r(SHIFT_FRW_ROW, mm3);
447 psrad_i2r(SHIFT_FRW_ROW, mm5);
448 psrad_i2r(SHIFT_FRW_ROW, mm7);
449 psrad_i2r(SHIFT_FRW_ROW, mm2);
450 packssdw_r2r(mm5, mm3);
451 packssdw_r2r(mm2, mm7);
452 movq_r2m(mm3, *(out + 0));
453 movq_r2m(mm7, *(out + 4));
456 static always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table)
458 //FIXME reorder (i dont have a old mmx only cpu here to benchmark ...)
459 movd_m2r(*(in + 6), mm1);
460 punpcklwd_m2r(*(in + 4), mm1);
462 psrlq_i2r(0x20, mm1);
463 movq_m2r(*(in + 0), mm0);
464 punpcklwd_r2r(mm2, mm1);
466 paddsw_r2r(mm1, mm0);
467 psubsw_r2r(mm1, mm5);
469 punpckldq_r2r(mm5, mm0);
470 punpckhdq_r2r(mm5, mm2);
471 movq_m2r(*(table + 0), mm1);
472 movq_m2r(*(table + 4), mm3);
473 movq_m2r(*(table + 8), mm4);
474 movq_m2r(*(table + 12), mm5);
475 movq_m2r(*(table + 16), mm6);
476 movq_m2r(*(table + 20), mm7);
477 pmaddwd_r2r(mm0, mm1);
478 pmaddwd_r2r(mm2, mm3);
479 pmaddwd_r2r(mm0, mm4);
480 pmaddwd_r2r(mm2, mm5);
481 pmaddwd_r2r(mm0, mm6);
482 pmaddwd_r2r(mm2, mm7);
483 pmaddwd_m2r(*(table + 24), mm0);
484 pmaddwd_m2r(*(table + 28), mm2);
489 movq_m2r(*fdct_r_row, mm0);
494 psrad_i2r(SHIFT_FRW_ROW, mm3);
495 psrad_i2r(SHIFT_FRW_ROW, mm5);
496 psrad_i2r(SHIFT_FRW_ROW, mm7);
497 psrad_i2r(SHIFT_FRW_ROW, mm2);
498 packssdw_r2r(mm5, mm3);
499 packssdw_r2r(mm2, mm7);
500 movq_r2m(mm3, *(out + 0));
501 movq_r2m(mm7, *(out + 4));
504 void ff_fdct_mmx(int16_t *block)
506 int64_t align_tmp[16] ATTR_ALIGN(8);
507 int16_t * block1= (int16_t*)align_tmp;
508 const int16_t *table= tab_frw_01234567;
511 fdct_col(block, block1, 0);
512 fdct_col(block, block1, 4);
515 fdct_row_mmx(block1, block, table);
522 void ff_fdct_mmx2(int16_t *block)
524 int64_t align_tmp[16] ATTR_ALIGN(8);
525 int16_t *block1= (int16_t*)align_tmp;
526 const int16_t *table= tab_frw_01234567;
529 fdct_col(block, block1, 0);
530 fdct_col(block, block1, 4);
533 fdct_row_mmx2(block1, block, table);
540 void ff_fdct_sse2(int16_t *block)
542 int64_t align_tmp[16] ATTR_ALIGN(16);
543 int16_t * const block1= (int16_t*)align_tmp;
545 fdct_col(block, block1, 0);
546 fdct_col(block, block1, 4);
548 fdct_row_sse2(block1, block);