?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Remove commented-out and unused fmt_list array.
[mplayer/greg.git] / libmpcodecs / vf_fspp.c
blob52c0556b306379b62b97a1cb072997941d3f457a
1 /*
2 Copyright (C) 2003 Michael Niedermayer <michaelni@gmx.at>
3 Copyright (C) 2005 Nikolaj Poroshin <porosh3@psu.ru>
4
5 This program is free software; you can redistribute it and/or modify
6 it under the terms of the GNU General Public License as published by
7 the Free Software Foundation; either version 2 of the License, or
8 (at your option) any later version.
9
10 This program is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 GNU General Public License for more details.
14
15 You should have received a copy of the GNU General Public License
16 along with this program; if not, write to the Free Software
17 Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 */
19
20 /*
21 * This implementation is based on an algorithm described in
22 * "Aria Nosratinia Embedded Post-Processing for
23 * Enhancement of Compressed Images (1999)"
24 * (http://citeseer.nj.nec.com/nosratinia99embedded.html)
25 * Futher, with splitting (i)dct into hor/ver passes, one of them can be
26 * performed once per block, not pixel. This allows for much better speed.
27 */
28
29 /*
30 Heavily optimized version of SPP filter by Nikolaj
31 */
32
33 #include <stdio.h>
34 #include <stdlib.h>
35 #include <string.h>
36 #include <inttypes.h>
37 #include <math.h>
38
39 #include "config.h"
40
41 #include "mp_msg.h"
42 #include "cpudetect.h"
43
44 #include "libavcodec/avcodec.h"
45 #include "libavcodec/dsputil.h"
46
47 #ifdef HAVE_MALLOC_H
48 #include <malloc.h>
49 #endif
50
51 #include "img_format.h"
52 #include "mp_image.h"
53 #include "vf.h"
54 #include "libvo/fastmemcpy.h"
55
56 //===========================================================================//
57 #define BLOCKSZ 12
58
59 static const short custom_threshold[64]=
60 // values (296) can't be too high
61 // -it causes too big quant dependence
62 // or maybe overflow(check), which results in some flashing
63 { 71, 296, 295, 237, 71, 40, 38, 19,
64 245, 193, 185, 121, 102, 73, 53, 27,
65 158, 129, 141, 107, 97, 73, 50, 26,
66 102, 116, 109, 98, 82, 66, 45, 23,
67 71, 94, 95, 81, 70, 56, 38, 20,
68 56, 77, 74, 66, 56, 44, 30, 15,
69 38, 53, 50, 45, 38, 30, 21, 11,
70 20, 27, 26, 23, 20, 15, 11, 5
71 };
72
73 static const uint8_t __attribute__((aligned(32))) dither[8][8]={
74 { 0, 48, 12, 60, 3, 51, 15, 63, },
75 { 32, 16, 44, 28, 35, 19, 47, 31, },
76 { 8, 56, 4, 52, 11, 59, 7, 55, },
77 { 40, 24, 36, 20, 43, 27, 39, 23, },
78 { 2, 50, 14, 62, 1, 49, 13, 61, },
79 { 34, 18, 46, 30, 33, 17, 45, 29, },
80 { 10, 58, 6, 54, 9, 57, 5, 53, },
81 { 42, 26, 38, 22, 41, 25, 37, 21, },
82 };
83
84 struct vf_priv_s { //align 16 !
85 uint64_t threshold_mtx_noq[8*2];
86 uint64_t threshold_mtx[8*2];//used in both C & MMX (& later SSE2) versions
87
88 int log2_count;
89 int temp_stride;
90 int qp;
91 int mpeg2;
92 int prev_q;
93 uint8_t *src;
94 int16_t *temp;
95 int bframes;
96 char *non_b_qp;
97 };
98
99
100 #ifndef HAVE_MMX
101
102 //This func reads from 1 slice, 1 and clears 0 & 1
103 static void store_slice_c(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale)
104 {int y, x;
105 #define STORE(pos) \
106 temp= (src[x + pos] + (d[pos]>>log2_scale))>>(6-log2_scale); \
107 src[x + pos]=src[x + pos - 8*src_stride]=0; \
108 if(temp & 0x100) temp= ~(temp>>31); \
109 dst[x + pos]= temp;
110
111 for(y=0; y<height; y++){
112 const uint8_t *d= dither[y];
113 for(x=0; x<width; x+=8){
114 int temp;
115 STORE(0);
116 STORE(1);
117 STORE(2);
118 STORE(3);
119 STORE(4);
120 STORE(5);
121 STORE(6);
122 STORE(7);
123 }
124 src+=src_stride;
125 dst+=dst_stride;
126 }
127 }
128
129 //This func reads from 2 slices, 0 & 2 and clears 2-nd
130 static void store_slice2_c(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale)
131 {int y, x;
132 #define STORE2(pos) \
133 temp= (src[x + pos] + src[x + pos + 16*src_stride] + (d[pos]>>log2_scale))>>(6-log2_scale); \
134 src[x + pos + 16*src_stride]=0; \
135 if(temp & 0x100) temp= ~(temp>>31); \
136 dst[x + pos]= temp;
137
138 for(y=0; y<height; y++){
139 const uint8_t *d= dither[y];
140 for(x=0; x<width; x+=8){
141 int temp;
142 STORE2(0);
143 STORE2(1);
144 STORE2(2);
145 STORE2(3);
146 STORE2(4);
147 STORE2(5);
148 STORE2(6);
149 STORE2(7);
150 }
151 src+=src_stride;
152 dst+=dst_stride;
153 }
154 }
155
156 static void mul_thrmat_c(struct vf_priv_s *p,int q)
157 {
158 int a;
159 for(a=0;a<64;a++)
160 ((short*)p->threshold_mtx)[a]=q * ((short*)p->threshold_mtx_noq)[a];//ints faster in C
161 }
162
163 static void column_fidct_c(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt);
164 static void row_idct_c(DCTELEM* workspace,
165 int16_t* output_adr, int output_stride, int cnt);
166 static void row_fdct_c(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt);
167
168 //this is rather ugly, but there is no need for function pointers
169 #define store_slice_s store_slice_c
170 #define store_slice2_s store_slice2_c
171 #define mul_thrmat_s mul_thrmat_c
172 #define column_fidct_s column_fidct_c
173 #define row_idct_s row_idct_c
174 #define row_fdct_s row_fdct_c
175
176 #else /* HAVE_MMX */
177
178 //This func reads from 1 slice, 1 and clears 0 & 1
179 static void store_slice_mmx(uint8_t *dst, int16_t *src, long dst_stride, long src_stride, long width, long height, long log2_scale)
180 {
181 const uint8_t *od=&dither[0][0];
182 const uint8_t *end=&dither[height][0];
183 width = (width+7)&~7;
184 dst_stride-=width;
185 //src_stride=(src_stride-width)*2;
186 asm volatile(
187 "mov %5, %%"REG_d" \n\t"
188 "mov %6, %%"REG_S" \n\t"
189 "mov %7, %%"REG_D" \n\t"
190 "mov %1, %%"REG_a" \n\t"
191 "movd %%"REG_d", %%mm5 \n\t"
192 "xor $-1, %%"REG_d" \n\t"
193 "mov %%"REG_a", %%"REG_c" \n\t"
194 "add $7, %%"REG_d" \n\t"
195 "neg %%"REG_a" \n\t"
196 "sub %0, %%"REG_c" \n\t"
197 "add %%"REG_c", %%"REG_c" \n\t"
198 "movd %%"REG_d", %%mm2 \n\t"
199 "mov %%"REG_c", %1 \n\t"
200 "mov %2, %%"REG_d" \n\t"
201 "shl $4, %%"REG_a" \n\t"
202
203 "2: \n\t"
204 "movq (%%"REG_d"), %%mm3 \n\t"
205 "movq %%mm3, %%mm4 \n\t"
206 "pxor %%mm7, %%mm7 \n\t"
207 "punpcklbw %%mm7, %%mm3 \n\t"
208 "punpckhbw %%mm7, %%mm4 \n\t"
209 "mov %0, %%"REG_c" \n\t"
210 "psraw %%mm5, %%mm3 \n\t"
211 "psraw %%mm5, %%mm4 \n\t"
212 "1: \n\t"
213 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t"
214 "movq (%%"REG_S"), %%mm0 \n\t"
215 "movq 8(%%"REG_S"), %%mm1 \n\t"
216
217 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t"
218 "paddw %%mm3, %%mm0 \n\t"
219 "paddw %%mm4, %%mm1 \n\t"
220
221 "movq %%mm7, (%%"REG_S") \n\t"
222 "psraw %%mm2, %%mm0 \n\t"
223 "psraw %%mm2, %%mm1 \n\t"
224
225 "movq %%mm7, 8(%%"REG_S") \n\t"
226 "packuswb %%mm1, %%mm0 \n\t"
227 "add $16, %%"REG_S" \n\t"
228
229 "movq %%mm0, (%%"REG_D") \n\t"
230 "add $8, %%"REG_D" \n\t"
231 "sub $8, %%"REG_c" \n\t"
232 "jg 1b \n\t"
233 "add %1, %%"REG_S" \n\t"
234 "add $8, %%"REG_d" \n\t"
235 "add %3, %%"REG_D" \n\t"
236 "cmp %4, %%"REG_d" \n\t"
237 "jl 2b \n\t"
238
239 :
240 : "m" (width), "m" (src_stride), "g" (od), "m" (dst_stride), "g" (end),
241 "m" (log2_scale), "m" (src), "m" (dst) //input
242 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
243 );
244 }
245
246 //This func reads from 2 slices, 0 & 2 and clears 2-nd
247 static void store_slice2_mmx(uint8_t *dst, int16_t *src, long dst_stride, long src_stride, long width, long height, long log2_scale)
248 {
249 const uint8_t *od=&dither[0][0];
250 const uint8_t *end=&dither[height][0];
251 width = (width+7)&~7;
252 dst_stride-=width;
253 //src_stride=(src_stride-width)*2;
254 asm volatile(
255 "mov %5, %%"REG_d" \n\t"
256 "mov %6, %%"REG_S" \n\t"
257 "mov %7, %%"REG_D" \n\t"
258 "mov %1, %%"REG_a" \n\t"
259 "movd %%"REG_d", %%mm5 \n\t"
260 "xor $-1, %%"REG_d" \n\t"
261 "mov %%"REG_a", %%"REG_c" \n\t"
262 "add $7, %%"REG_d" \n\t"
263 "sub %0, %%"REG_c" \n\t"
264 "add %%"REG_c", %%"REG_c" \n\t"
265 "movd %%"REG_d", %%mm2 \n\t"
266 "mov %%"REG_c", %1 \n\t"
267 "mov %2, %%"REG_d" \n\t"
268 "shl $5, %%"REG_a" \n\t"
269
270 "2: \n\t"
271 "movq (%%"REG_d"), %%mm3 \n\t"
272 "movq %%mm3, %%mm4 \n\t"
273 "pxor %%mm7, %%mm7 \n\t"
274 "punpcklbw %%mm7, %%mm3 \n\t"
275 "punpckhbw %%mm7, %%mm4 \n\t"
276 "mov %0, %%"REG_c" \n\t"
277 "psraw %%mm5, %%mm3 \n\t"
278 "psraw %%mm5, %%mm4 \n\t"
279 "1: \n\t"
280 "movq (%%"REG_S"), %%mm0 \n\t"
281 "movq 8(%%"REG_S"), %%mm1 \n\t"
282 "paddw %%mm3, %%mm0 \n\t"
283
284 "paddw (%%"REG_S",%%"REG_a",), %%mm0 \n\t"
285 "paddw %%mm4, %%mm1 \n\t"
286 "movq 8(%%"REG_S",%%"REG_a",), %%mm6 \n\t"
287
288 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t"
289 "psraw %%mm2, %%mm0 \n\t"
290 "paddw %%mm6, %%mm1 \n\t"
291
292 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t"
293 "psraw %%mm2, %%mm1 \n\t"
294 "packuswb %%mm1, %%mm0 \n\t"
295
296 "movq %%mm0, (%%"REG_D") \n\t"
297 "add $16, %%"REG_S" \n\t"
298 "add $8, %%"REG_D" \n\t"
299 "sub $8, %%"REG_c" \n\t"
300 "jg 1b \n\t"
301 "add %1, %%"REG_S" \n\t"
302 "add $8, %%"REG_d" \n\t"
303 "add %3, %%"REG_D" \n\t"
304 "cmp %4, %%"REG_d" \n\t"
305 "jl 2b \n\t"
306
307 :
308 : "m" (width), "m" (src_stride), "g" (od), "m" (dst_stride), "g" (end),
309 "m" (log2_scale), "m" (src), "m" (dst) //input
310 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_D, "%"REG_S
311 );
312 }
313
314 static void mul_thrmat_mmx(struct vf_priv_s *p, int q)
315 {
316 uint64_t *adr=&p->threshold_mtx_noq[0];
317 asm volatile(
318 "movd %0, %%mm7 \n\t"
319 "add $8*8*2, %%"REG_D" \n\t"
320 "movq 0*8(%%"REG_S"), %%mm0 \n\t"
321 "punpcklwd %%mm7, %%mm7 \n\t"
322 "movq 1*8(%%"REG_S"), %%mm1 \n\t"
323 "punpckldq %%mm7, %%mm7 \n\t"
324 "pmullw %%mm7, %%mm0 \n\t"
325
326 "movq 2*8(%%"REG_S"), %%mm2 \n\t"
327 "pmullw %%mm7, %%mm1 \n\t"
328
329 "movq 3*8(%%"REG_S"), %%mm3 \n\t"
330 "pmullw %%mm7, %%mm2 \n\t"
331
332 "movq %%mm0, 0*8(%%"REG_D") \n\t"
333 "movq 4*8(%%"REG_S"), %%mm4 \n\t"
334 "pmullw %%mm7, %%mm3 \n\t"
335
336 "movq %%mm1, 1*8(%%"REG_D") \n\t"
337 "movq 5*8(%%"REG_S"), %%mm5 \n\t"
338 "pmullw %%mm7, %%mm4 \n\t"
339
340 "movq %%mm2, 2*8(%%"REG_D") \n\t"
341 "movq 6*8(%%"REG_S"), %%mm6 \n\t"
342 "pmullw %%mm7, %%mm5 \n\t"
343
344 "movq %%mm3, 3*8(%%"REG_D") \n\t"
345 "movq 7*8+0*8(%%"REG_S"), %%mm0 \n\t"
346 "pmullw %%mm7, %%mm6 \n\t"
347
348 "movq %%mm4, 4*8(%%"REG_D") \n\t"
349 "movq 7*8+1*8(%%"REG_S"), %%mm1 \n\t"
350 "pmullw %%mm7, %%mm0 \n\t"
351
352 "movq %%mm5, 5*8(%%"REG_D") \n\t"
353 "movq 7*8+2*8(%%"REG_S"), %%mm2 \n\t"
354 "pmullw %%mm7, %%mm1 \n\t"
355
356 "movq %%mm6, 6*8(%%"REG_D") \n\t"
357 "movq 7*8+3*8(%%"REG_S"), %%mm3 \n\t"
358 "pmullw %%mm7, %%mm2 \n\t"
359
360 "movq %%mm0, 7*8+0*8(%%"REG_D") \n\t"
361 "movq 7*8+4*8(%%"REG_S"), %%mm4 \n\t"
362 "pmullw %%mm7, %%mm3 \n\t"
363
364 "movq %%mm1, 7*8+1*8(%%"REG_D") \n\t"
365 "movq 7*8+5*8(%%"REG_S"), %%mm5 \n\t"
366 "pmullw %%mm7, %%mm4 \n\t"
367
368 "movq %%mm2, 7*8+2*8(%%"REG_D") \n\t"
369 "movq 7*8+6*8(%%"REG_S"), %%mm6 \n\t"
370 "pmullw %%mm7, %%mm5 \n\t"
371
372 "movq %%mm3, 7*8+3*8(%%"REG_D") \n\t"
373 "movq 14*8+0*8(%%"REG_S"), %%mm0 \n\t"
374 "pmullw %%mm7, %%mm6 \n\t"
375
376 "movq %%mm4, 7*8+4*8(%%"REG_D") \n\t"
377 "movq 14*8+1*8(%%"REG_S"), %%mm1 \n\t"
378 "pmullw %%mm7, %%mm0 \n\t"
379
380 "movq %%mm5, 7*8+5*8(%%"REG_D") \n\t"
381 "pmullw %%mm7, %%mm1 \n\t"
382
383 "movq %%mm6, 7*8+6*8(%%"REG_D") \n\t"
384 "movq %%mm0, 14*8+0*8(%%"REG_D") \n\t"
385 "movq %%mm1, 14*8+1*8(%%"REG_D") \n\t"
386
387 : "+g" (q), "+S" (adr), "+D" (adr)
388 :
389 );
390 }
391
392 static void column_fidct_mmx(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt);
393 static void row_idct_mmx(DCTELEM* workspace,
394 int16_t* output_adr, int output_stride, int cnt);
395 static void row_fdct_mmx(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt);
396
397 #define store_slice_s store_slice_mmx
398 #define store_slice2_s store_slice2_mmx
399 #define mul_thrmat_s mul_thrmat_mmx
400 #define column_fidct_s column_fidct_mmx
401 #define row_idct_s row_idct_mmx
402 #define row_fdct_s row_fdct_mmx
403 #endif // HAVE_MMX
404
405 static void filter(struct vf_priv_s *p, uint8_t *dst, uint8_t *src,
406 int dst_stride, int src_stride,
407 int width, int height,
408 uint8_t *qp_store, int qp_stride, int is_luma)
409 {
410 int x, x0, y, es, qy, t;
411 const int stride= is_luma ? p->temp_stride : (width+16);//((width+16+15)&(~15))
412 const int step=6-p->log2_count;
413 const int qps= 3 + is_luma;
414 int32_t __attribute__((aligned(32))) block_align[4*8*BLOCKSZ+ 4*8*BLOCKSZ];
415 DCTELEM *block= (DCTELEM *)block_align;
416 DCTELEM *block3=(DCTELEM *)(block_align+4*8*BLOCKSZ);
417
418 memset(block3, 0, 4*8*BLOCKSZ);
419
420 //p->src=src-src_stride*8-8;//!
421 if (!src || !dst) return; // HACK avoid crash for Y8 colourspace
422 for(y=0; y<height; y++){
423 int index= 8 + 8*stride + y*stride;
424 fast_memcpy(p->src + index, src + y*src_stride, width);//this line can be avoided by using DR & user fr.buffers
425 for(x=0; x<8; x++){
426 p->src[index - x - 1]= p->src[index + x ];
427 p->src[index + width + x ]= p->src[index + width - x - 1];
428 }
429 }
430 for(y=0; y<8; y++){
431 fast_memcpy(p->src + ( 7-y)*stride, p->src + ( y+8)*stride, stride);
432 fast_memcpy(p->src + (height+8+y)*stride, p->src + (height-y+7)*stride, stride);
433 }
434 //FIXME (try edge emu)
435
436 for(y=8; y<24; y++)
437 memset(p->temp+ 8 +y*stride, 0,width*sizeof(int16_t));
438
439 for(y=step; y<height+8; y+=step){ //step= 1,2
440 qy=y-4;
441 if (qy>height-1) qy=height-1;
442 if (qy<0) qy=0;
443 qy=(qy>>qps)*qp_stride;
444 row_fdct_s(block, p->src + y*stride +2-(y&1), stride, 2);
445 for(x0=0; x0<width+8-8*(BLOCKSZ-1); x0+=8*(BLOCKSZ-1)){
446 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, 2*(BLOCKSZ-1));
447 if(p->qp)
448 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+0*8, block3+0*8, 8*(BLOCKSZ-1)); //yes, this is a HOTSPOT
449 else
450 for (x=0; x<8*(BLOCKSZ-1); x+=8) {
451 t=x+x0-2; //correct t=x+x0-2-(y&1), but its the same
452 if (t<0) t=0;//t always < width-2
453 t=qp_store[qy+(t>>qps)];
454 if(p->mpeg2) t>>=1; //copy p->mpeg2,prev_q to locals?
455 if (t!=p->prev_q) p->prev_q=t, mul_thrmat_s(p, t);
456 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+x*8, block3+x*8, 8); //yes, this is a HOTSPOT
457 }
458 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, 2*(BLOCKSZ-1));
459 memcpy(block, block+(BLOCKSZ-1)*64, 8*8*sizeof(DCTELEM)); //cycling
460 memcpy(block3, block3+(BLOCKSZ-1)*64, 6*8*sizeof(DCTELEM));
461 }
462 //
463 es=width+8-x0; // 8, ...
464 if (es>8)
465 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, (es-4)>>2);
466 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block, block3, es&(~1));
467 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, es>>2);
468 {const int y1=y-8+step;//l5-7 l4-6
469 if (!(y1&7) && y1) {
470 if (y1&8) store_slice_s(dst + (y1-8)*dst_stride, p->temp+ 8 +8*stride,
471 dst_stride, stride, width, 8, 5-p->log2_count);
472 else store_slice2_s(dst + (y1-8)*dst_stride, p->temp+ 8 +0*stride,
473 dst_stride, stride, width, 8, 5-p->log2_count);
474 } }
475 }
476
477 if (y&7) { // == height & 7
478 if (y&8) store_slice_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +8*stride,
479 dst_stride, stride, width, y&7, 5-p->log2_count);
480 else store_slice2_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +0*stride,
481 dst_stride, stride, width, y&7, 5-p->log2_count);
482 }
483 }
484
485 static int config(struct vf_instance_s* vf,
486 int width, int height, int d_width, int d_height,
487 unsigned int flags, unsigned int outfmt)
488 {
489 int h= (height+16+15)&(~15);
490
491 vf->priv->temp_stride= (width+16+15)&(~15);
492 vf->priv->temp= (int16_t*)av_mallocz(vf->priv->temp_stride*3*8*sizeof(int16_t));
493 //this can also be avoided, see above
494 vf->priv->src = (uint8_t*)av_malloc(vf->priv->temp_stride*h*sizeof(uint8_t));
495
496 return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt);
497 }
498
499 static void get_image(struct vf_instance_s* vf, mp_image_t *mpi)
500 {
501 if(mpi->flags&MP_IMGFLAG_PRESERVE) return; // don't change
502 // ok, we can do pp in-place (or pp disabled):
503 vf->dmpi=vf_get_image(vf->next,mpi->imgfmt,
504 mpi->type, mpi->flags, mpi->width, mpi->height);
505 mpi->planes[0]=vf->dmpi->planes[0];
506 mpi->stride[0]=vf->dmpi->stride[0];
507 mpi->width=vf->dmpi->width;
508 if(mpi->flags&MP_IMGFLAG_PLANAR){
509 mpi->planes[1]=vf->dmpi->planes[1];
510 mpi->planes[2]=vf->dmpi->planes[2];
511 mpi->stride[1]=vf->dmpi->stride[1];
512 mpi->stride[2]=vf->dmpi->stride[2];
513 }
514 mpi->flags|=MP_IMGFLAG_DIRECT;
515 }
516
517 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts)
518 {
519 mp_image_t *dmpi;
520 if(!(mpi->flags&MP_IMGFLAG_DIRECT)){
521 // no DR, so get a new image! hope we'll get DR buffer:
522 dmpi=vf_get_image(vf->next,mpi->imgfmt,
523 MP_IMGTYPE_TEMP,
524 MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE,
525 mpi->width,mpi->height);
526 vf_clone_mpi_attributes(dmpi, mpi);
527 }else{
528 dmpi=vf->dmpi;
529 }
530
531 vf->priv->mpeg2= mpi->qscale_type;
532 if(mpi->pict_type != 3 && mpi->qscale && !vf->priv->qp){
533 if(!vf->priv->non_b_qp)
534 vf->priv->non_b_qp= malloc(mpi->qstride * ((mpi->h + 15) >> 4));
535 fast_memcpy(vf->priv->non_b_qp, mpi->qscale, mpi->qstride * ((mpi->h + 15) >> 4));
536 }
537 if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){
538 char *qp_tab= vf->priv->non_b_qp;
539 if(vf->priv->bframes || !qp_tab)
540 qp_tab= mpi->qscale;
541
542 if(qp_tab || vf->priv->qp){
543 filter(vf->priv, dmpi->planes[0], mpi->planes[0], dmpi->stride[0], mpi->stride[0],
544 mpi->w, mpi->h, qp_tab, mpi->qstride, 1);
545 filter(vf->priv, dmpi->planes[1], mpi->planes[1], dmpi->stride[1], mpi->stride[1],
546 mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0);
547 filter(vf->priv, dmpi->planes[2], mpi->planes[2], dmpi->stride[2], mpi->stride[2],
548 mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0);
549 }else{
550 memcpy_pic(dmpi->planes[0], mpi->planes[0], mpi->w, mpi->h, dmpi->stride[0], mpi->stride[0]);
551 memcpy_pic(dmpi->planes[1], mpi->planes[1], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[1], mpi->stride[1]);
552 memcpy_pic(dmpi->planes[2], mpi->planes[2], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[2], mpi->stride[2]);
553 }
554 }
555
556 #ifdef HAVE_MMX
557 if(gCpuCaps.hasMMX) asm volatile ("emms\n\t");
558 #endif
559 #ifdef HAVE_MMX2
560 if(gCpuCaps.hasMMX2) asm volatile ("sfence\n\t");
561 #endif
562 return vf_next_put_image(vf,dmpi, pts);
563 }
564
565 static void uninit(struct vf_instance_s* vf)
566 {
567 if(!vf->priv) return;
568
569 if(vf->priv->temp) av_free(vf->priv->temp);
570 vf->priv->temp= NULL;
571 if(vf->priv->src) av_free(vf->priv->src);
572 vf->priv->src= NULL;
573 //if(vf->priv->avctx) free(vf->priv->avctx);
574 //vf->priv->avctx= NULL;
575 if(vf->priv->non_b_qp) free(vf->priv->non_b_qp);
576 vf->priv->non_b_qp= NULL;
577
578 av_free(vf->priv);
579 vf->priv=NULL;
580 }
581
582 //===========================================================================//
583
584 static int query_format(struct vf_instance_s* vf, unsigned int fmt)
585 {
586 switch(fmt){
587 case IMGFMT_YVU9:
588 case IMGFMT_IF09:
589 case IMGFMT_YV12:
590 case IMGFMT_I420:
591 case IMGFMT_IYUV:
592 case IMGFMT_CLPL:
593 case IMGFMT_Y800:
594 case IMGFMT_Y8:
595 case IMGFMT_444P:
596 case IMGFMT_422P:
597 case IMGFMT_411P:
598 return vf_next_query_format(vf,fmt);
599 }
600 return 0;
601 }
602
603 static int control(struct vf_instance_s* vf, int request, void* data)
604 {
605 switch(request){
606 case VFCTRL_QUERY_MAX_PP_LEVEL:
607 return 5;
608 case VFCTRL_SET_PP_LEVEL:
609 vf->priv->log2_count= *((unsigned int*)data);
610 if (vf->priv->log2_count < 4) vf->priv->log2_count=4;
611 return CONTROL_TRUE;
612 }
613 return vf_next_control(vf,request,data);
614 }
615
616 static int open(vf_instance_t *vf, char* args)
617 {
618 int i=0, bias;
619 int custom_threshold_m[64];
620 int log2c=-1;
621
622 vf->config=config;
623 vf->put_image=put_image;
624 vf->get_image=get_image;
625 vf->query_format=query_format;
626 vf->uninit=uninit;
627 vf->control= control;
628 vf->priv=av_mallocz(sizeof(struct vf_priv_s));//assumes align 16 !
629
630 avcodec_init();
631
632 //vf->priv->avctx= avcodec_alloc_context();
633 //dsputil_init(&vf->priv->dsp, vf->priv->avctx);
634
635 vf->priv->log2_count= 4;
636 vf->priv->bframes = 0;
637
638 if (args) sscanf(args, "%d:%d:%d:%d", &log2c, &vf->priv->qp, &i, &vf->priv->bframes);
639
640 if( log2c >=4 && log2c <=5 )
641 vf->priv->log2_count = log2c;
642 else if( log2c >= 6 )
643 vf->priv->log2_count = 5;
644
645 if(vf->priv->qp < 0)
646 vf->priv->qp = 0;
647
648 if (i < -15) i = -15;
649 if (i > 32) i = 32;
650
651 bias= (1<<4)+i; //regulable
652 vf->priv->prev_q=0;
653 //
654 for(i=0;i<64;i++) //FIXME: tune custom_threshold[] and remove this !
655 custom_threshold_m[i]=(int)(custom_threshold[i]*(bias/71.)+ 0.5);
656 for(i=0;i<8;i++){
657 vf->priv->threshold_mtx_noq[2*i]=(uint64_t)custom_threshold_m[i*8+2]
658 |(((uint64_t)custom_threshold_m[i*8+6])<<16)
659 |(((uint64_t)custom_threshold_m[i*8+0])<<32)
660 |(((uint64_t)custom_threshold_m[i*8+4])<<48);
661 vf->priv->threshold_mtx_noq[2*i+1]=(uint64_t)custom_threshold_m[i*8+5]
662 |(((uint64_t)custom_threshold_m[i*8+3])<<16)
663 |(((uint64_t)custom_threshold_m[i*8+1])<<32)
664 |(((uint64_t)custom_threshold_m[i*8+7])<<48);
665 }
666
667 if (vf->priv->qp) vf->priv->prev_q=vf->priv->qp, mul_thrmat_s(vf->priv, vf->priv->qp);
668
669 return 1;
670 }
671
672 vf_info_t vf_info_fspp = {
673 "fast simple postprocess",
674 "fspp",
675 "Michael Niedermayer, Nikolaj Poroshin",
676 "",
677 open,
678 NULL
679 };
680
681 //====================================================================
682 //Specific spp's dct, idct and threshold functions
683 //I'd prefer to have them in the separate file.
684
685 #include "mangle.h"
686 //#define MANGLE(a) #a
687
688 //typedef int16_t DCTELEM; //! only int16_t
689
690 #define DCTSIZE 8
691 #define DCTSIZE_S "8"
692
693 #define FIX(x,s) ((int) ((x) * (1<<s) + 0.5)&0xffff)
694 #define C64(x) ((uint64_t)((x)|(x)<<16))<<32 | (uint64_t)(x) | (uint64_t)(x)<<16
695 #define FIX64(x,s) C64(FIX(x,s))
696
697 #define MULTIPLY16H(x,k) (((x)*(k))>>16)
698 #define THRESHOLD(r,x,t) if(((unsigned)((x)+t))>t*2) r=(x);else r=0;
699 #define DESCALE(x,n) (((x) + (1 << ((n)-1))) >> n)
700
701 #ifdef HAVE_MMX
702
703 static uint64_t attribute_used __attribute__((aligned(8))) temps[4];//!!
704
705 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_382683433=FIX64(0.382683433, 14);
706 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_541196100=FIX64(0.541196100, 14);
707 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_707106781=FIX64(0.707106781, 14);
708 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_306562965=FIX64(1.306562965, 14);
709
710 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_414213562_A=FIX64(1.414213562, 14);
711
712 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_847759065=FIX64(1.847759065, 13);
713 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_2_613125930=FIX64(-2.613125930, 13); //-
714 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_414213562=FIX64(1.414213562, 13);
715 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_082392200=FIX64(1.082392200, 13);
716 //for t3,t5,t7 == 0 shortcut
717 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_847759065=FIX64(0.847759065, 14);
718 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_566454497=FIX64(0.566454497, 14);
719 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_198912367=FIX64(0.198912367, 14);
720
721 static uint64_t attribute_used __attribute__((aligned(8))) MM_DESCALE_RND=C64(4);
722 static uint64_t attribute_used __attribute__((aligned(8))) MM_2=C64(2);
723
724 #else /* !HAVE_MMX */
725
726 typedef int32_t int_simd16_t;
727 static int16_t FIX_0_382683433=FIX(0.382683433, 14);
728 static int16_t FIX_0_541196100=FIX(0.541196100, 14);
729 static int16_t FIX_0_707106781=FIX(0.707106781, 14);
730 static int16_t FIX_1_306562965=FIX(1.306562965, 14);
731 static int16_t FIX_1_414213562_A=FIX(1.414213562, 14);
732 static int16_t FIX_1_847759065=FIX(1.847759065, 13);
733 static int16_t FIX_2_613125930=FIX(-2.613125930, 13); //-
734 static int16_t FIX_1_414213562=FIX(1.414213562, 13);
735 static int16_t FIX_1_082392200=FIX(1.082392200, 13);
736
737 #endif
738
739 #ifndef HAVE_MMX
740
741 static void column_fidct_c(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt)
742 {
743 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
744 int_simd16_t tmp10, tmp11, tmp12, tmp13;
745 int_simd16_t z1,z2,z3,z4,z5, z10, z11, z12, z13;
746 int_simd16_t d0, d1, d2, d3, d4, d5, d6, d7;
747
748 DCTELEM* dataptr;
749 DCTELEM* wsptr;
750 int16_t *threshold;
751 int ctr;
752
753 dataptr = data;
754 wsptr = output;
755
756 for (; cnt > 0; cnt-=2) { //start positions
757 threshold=(int16_t*)thr_adr;//threshold_mtx
758 for (ctr = DCTSIZE; ctr > 0; ctr--) {
759 // Process columns from input, add to output.
760 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
761 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
762
763 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
764 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
765
766 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
767 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
768
769 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
770 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
771
772 // Even part of FDCT
773
774 tmp10 = tmp0 + tmp3;
775 tmp13 = tmp0 - tmp3;
776 tmp11 = tmp1 + tmp2;
777 tmp12 = tmp1 - tmp2;
778
779 d0 = tmp10 + tmp11;
780 d4 = tmp10 - tmp11;
781
782 z1 = MULTIPLY16H((tmp12 + tmp13) <<2, FIX_0_707106781);
783 d2 = tmp13 + z1;
784 d6 = tmp13 - z1;
785
786 // Even part of IDCT
787
788 THRESHOLD(tmp0, d0, threshold[0*8]);
789 THRESHOLD(tmp1, d2, threshold[2*8]);
790 THRESHOLD(tmp2, d4, threshold[4*8]);
791 THRESHOLD(tmp3, d6, threshold[6*8]);
792 tmp0+=2;
793 tmp10 = (tmp0 + tmp2)>>2;
794 tmp11 = (tmp0 - tmp2)>>2;
795
796 tmp13 = (tmp1 + tmp3)>>2; //+2 ! (psnr decides)
797 tmp12 = MULTIPLY16H((tmp1 - tmp3), FIX_1_414213562_A) - tmp13; //<<2
798
799 tmp0 = tmp10 + tmp13; //->temps
800 tmp3 = tmp10 - tmp13; //->temps
801 tmp1 = tmp11 + tmp12; //->temps
802 tmp2 = tmp11 - tmp12; //->temps
803
804 // Odd part of FDCT
805
806 tmp10 = tmp4 + tmp5;
807 tmp11 = tmp5 + tmp6;
808 tmp12 = tmp6 + tmp7;
809
810 z5 = MULTIPLY16H((tmp10 - tmp12)<<2, FIX_0_382683433);
811 z2 = MULTIPLY16H(tmp10 <<2, FIX_0_541196100) + z5;
812 z4 = MULTIPLY16H(tmp12 <<2, FIX_1_306562965) + z5;
813 z3 = MULTIPLY16H(tmp11 <<2, FIX_0_707106781);
814
815 z11 = tmp7 + z3;
816 z13 = tmp7 - z3;
817
818 d5 = z13 + z2;
819 d3 = z13 - z2;
820 d1 = z11 + z4;
821 d7 = z11 - z4;
822
823 // Odd part of IDCT
824
825 THRESHOLD(tmp4, d1, threshold[1*8]);
826 THRESHOLD(tmp5, d3, threshold[3*8]);
827 THRESHOLD(tmp6, d5, threshold[5*8]);
828 THRESHOLD(tmp7, d7, threshold[7*8]);
829
830 //Simd version uses here a shortcut for the tmp5,tmp6,tmp7 == 0
831 z13 = tmp6 + tmp5;
832 z10 = (tmp6 - tmp5)<<1;
833 z11 = tmp4 + tmp7;
834 z12 = (tmp4 - tmp7)<<1;
835
836 tmp7 = (z11 + z13)>>2; //+2 !
837 tmp11 = MULTIPLY16H((z11 - z13)<<1, FIX_1_414213562);
838 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
839 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
840 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - !!
841
842 tmp6 = tmp12 - tmp7;
843 tmp5 = tmp11 - tmp6;
844 tmp4 = tmp10 + tmp5;
845
846 wsptr[DCTSIZE*0]+= (tmp0 + tmp7);
847 wsptr[DCTSIZE*1]+= (tmp1 + tmp6);
848 wsptr[DCTSIZE*2]+= (tmp2 + tmp5);
849 wsptr[DCTSIZE*3]+= (tmp3 - tmp4);
850 wsptr[DCTSIZE*4]+= (tmp3 + tmp4);
851 wsptr[DCTSIZE*5]+= (tmp2 - tmp5);
852 wsptr[DCTSIZE*6]= (tmp1 - tmp6);
853 wsptr[DCTSIZE*7]= (tmp0 - tmp7);
854 //
855 dataptr++; //next column
856 wsptr++;
857 threshold++;
858 }
859 dataptr+=8; //skip each second start pos
860 wsptr +=8;
861 }
862 }
863
864 #else /* HAVE_MMX */
865
866 static void column_fidct_mmx(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt)
867 {
868 asm volatile(
869 ASMALIGN(4)
870 "1: \n\t"
871 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
872 //
873 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
874 "movq %%mm1, %%mm0 \n\t"
875
876 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0
877 "movq %%mm7, %%mm3 \n\t"
878
879 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
880 "movq %%mm1, %%mm5 \n\t"
881
882 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
883 "psubw %%mm7, %%mm1 \n\t" //t13
884
885 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
886 "movq %%mm6, %%mm4 \n\t"
887
888 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
889 "paddw %%mm7, %%mm5 \n\t" //t10
890
891 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
892 "movq %%mm6, %%mm7 \n\t"
893
894 "paddw %%mm2, %%mm6 \n\t" //t11
895 "psubw %%mm2, %%mm7 \n\t" //t12
896
897 "movq %%mm5, %%mm2 \n\t"
898 "paddw %%mm6, %%mm5 \n\t" //d0
899 // i0 t13 t12 i3 i1 d0 - d4
900 "psubw %%mm6, %%mm2 \n\t" //d4
901 "paddw %%mm1, %%mm7 \n\t"
902
903 "movq 4*16(%%"REG_d"), %%mm6 \n\t"
904 "psllw $2, %%mm7 \n\t"
905
906 "psubw 0*16(%%"REG_d"), %%mm5 \n\t"
907 "psubw %%mm6, %%mm2 \n\t"
908
909 "paddusw 0*16(%%"REG_d"), %%mm5 \n\t"
910 "paddusw %%mm6, %%mm2 \n\t"
911
912 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t"
913 //
914 "paddw 0*16(%%"REG_d"), %%mm5 \n\t"
915 "paddw %%mm6, %%mm2 \n\t"
916
917 "psubusw 0*16(%%"REG_d"), %%mm5 \n\t"
918 "psubusw %%mm6, %%mm2 \n\t"
919
920 //This func is totally compute-bound, operates at huge speed. So, DC shortcut
921 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
922 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
923 "paddw "MANGLE(MM_2)", %%mm5 \n\t"
924 "movq %%mm2, %%mm6 \n\t"
925
926 "paddw %%mm5, %%mm2 \n\t"
927 "psubw %%mm6, %%mm5 \n\t"
928
929 "movq %%mm1, %%mm6 \n\t"
930 "paddw %%mm7, %%mm1 \n\t" //d2
931
932 "psubw 2*16(%%"REG_d"), %%mm1 \n\t"
933 "psubw %%mm7, %%mm6 \n\t" //d6
934
935 "movq 6*16(%%"REG_d"), %%mm7 \n\t"
936 "psraw $2, %%mm5 \n\t"
937
938 "paddusw 2*16(%%"REG_d"), %%mm1 \n\t"
939 "psubw %%mm7, %%mm6 \n\t"
940 // t7 d2 /t11 t4 t6 - d6 /t10
941
942 "paddw 2*16(%%"REG_d"), %%mm1 \n\t"
943 "paddusw %%mm7, %%mm6 \n\t"
944
945 "psubusw 2*16(%%"REG_d"), %%mm1 \n\t"
946 "paddw %%mm7, %%mm6 \n\t"
947
948 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t"
949 "psubusw %%mm7, %%mm6 \n\t"
950
951 //movq [edi+"DCTSIZE_S"*2*2], mm1
952 //movq [edi+"DCTSIZE_S"*6*2], mm6
953 "movq %%mm1, %%mm7 \n\t"
954 "psraw $2, %%mm2 \n\t"
955
956 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t"
957 "psubw %%mm6, %%mm1 \n\t"
958
959 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t"
960 "paddw %%mm7, %%mm6 \n\t" //'t13
961
962 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
963 "movq %%mm2, %%mm7 \n\t"
964
965 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
966 "paddw %%mm6, %%mm2 \n\t" //'t0
967
968 "movq %%mm2, "MANGLE(temps)"+0*8 \n\t" //!
969 "psubw %%mm6, %%mm7 \n\t" //'t3
970
971 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
972 "psubw %%mm6, %%mm1 \n\t" //'t12
973
974 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5
975 "movq %%mm5, %%mm6 \n\t"
976
977 "movq %%mm7, "MANGLE(temps)"+3*8 \n\t"
978 "paddw %%mm2, %%mm3 \n\t" //t10
979
980 "paddw %%mm4, %%mm2 \n\t" //t11
981 "paddw %%mm0, %%mm4 \n\t" //t12
982
983 "movq %%mm3, %%mm7 \n\t"
984 "psubw %%mm4, %%mm3 \n\t"
985
986 "psllw $2, %%mm3 \n\t"
987 "psllw $2, %%mm7 \n\t" //opt for P6
988
989 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
990 "psllw $2, %%mm4 \n\t"
991
992 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t"
993 "psllw $2, %%mm2 \n\t"
994
995 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
996 "paddw %%mm1, %%mm5 \n\t" //'t1
997
998 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t"
999 "psubw %%mm1, %%mm6 \n\t" //'t2
1000 // t7 't12 't11 t4 t6 - 't13 't10 ---
1001
1002 "paddw %%mm3, %%mm7 \n\t" //z2
1003
1004 "movq %%mm5, "MANGLE(temps)"+1*8 \n\t"
1005 "paddw %%mm3, %%mm4 \n\t" //z4
1006
1007 "movq 3*16(%%"REG_d"), %%mm3 \n\t"
1008 "movq %%mm0, %%mm1 \n\t"
1009
1010 "movq %%mm6, "MANGLE(temps)"+2*8 \n\t"
1011 "psubw %%mm2, %%mm1 \n\t" //z13
1012
1013 //===
1014 "paddw %%mm2, %%mm0 \n\t" //z11
1015 "movq %%mm1, %%mm5 \n\t"
1016
1017 "movq 5*16(%%"REG_d"), %%mm2 \n\t"
1018 "psubw %%mm7, %%mm1 \n\t" //d3
1019
1020 "paddw %%mm7, %%mm5 \n\t" //d5
1021 "psubw %%mm3, %%mm1 \n\t"
1022
1023 "movq 1*16(%%"REG_d"), %%mm7 \n\t"
1024 "psubw %%mm2, %%mm5 \n\t"
1025
1026 "movq %%mm0, %%mm6 \n\t"
1027 "paddw %%mm4, %%mm0 \n\t" //d1
1028
1029 "paddusw %%mm3, %%mm1 \n\t"
1030 "psubw %%mm4, %%mm6 \n\t" //d7
1031
1032 // d1 d3 - - - d5 d7 -
1033 "movq 7*16(%%"REG_d"), %%mm4 \n\t"
1034 "psubw %%mm7, %%mm0 \n\t"
1035
1036 "psubw %%mm4, %%mm6 \n\t"
1037 "paddusw %%mm2, %%mm5 \n\t"
1038
1039 "paddusw %%mm4, %%mm6 \n\t"
1040 "paddw %%mm3, %%mm1 \n\t"
1041
1042 "paddw %%mm2, %%mm5 \n\t"
1043 "paddw %%mm4, %%mm6 \n\t"
1044
1045 "psubusw %%mm3, %%mm1 \n\t"
1046 "psubusw %%mm2, %%mm5 \n\t"
1047
1048 "psubusw %%mm4, %%mm6 \n\t"
1049 "movq %%mm1, %%mm4 \n\t"
1050
1051 "por %%mm5, %%mm4 \n\t"
1052 "paddusw %%mm7, %%mm0 \n\t"
1053
1054 "por %%mm6, %%mm4 \n\t"
1055 "paddw %%mm7, %%mm0 \n\t"
1056
1057 "packssdw %%mm4, %%mm4 \n\t"
1058 "psubusw %%mm7, %%mm0 \n\t"
1059
1060 "movd %%mm4, %%"REG_a" \n\t"
1061 "or %%"REG_a", %%"REG_a" \n\t"
1062 "jnz 2f \n\t"
1063 //movq [edi+"DCTSIZE_S"*3*2], mm1
1064 //movq [edi+"DCTSIZE_S"*5*2], mm5
1065 //movq [edi+"DCTSIZE_S"*1*2], mm0
1066 //movq [edi+"DCTSIZE_S"*7*2], mm6
1067 // t4 t5 - - - t6 t7 -
1068 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
1069 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
1070 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t"
1071 "movq %%mm0, %%mm1 \n\t"
1072
1073 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
1074 "movq %%mm1, %%mm2 \n\t"
1075
1076 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
1077 "movq %%mm2, %%mm3 \n\t"
1078
1079 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
1080 "paddw %%mm4, %%mm5 \n\t"
1081
1082 "movq "MANGLE(temps)"+1*8, %%mm6 \n\t"
1083 //paddw mm3, MM_2
1084 "psraw $2, %%mm3 \n\t" //tmp7
1085
1086 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
1087 "psubw %%mm3, %%mm4 \n\t"
1088
1089 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
1090 "paddw %%mm3, %%mm5 \n\t"
1091
1092 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
1093 "paddw %%mm6, %%mm7 \n\t"
1094
1095 "movq "MANGLE(temps)"+2*8, %%mm3 \n\t"
1096 "psubw %%mm0, %%mm6 \n\t"
1097
1098 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
1099 "paddw %%mm0, %%mm7 \n\t"
1100
1101 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
1102 "paddw %%mm3, %%mm4 \n\t"
1103
1104 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
1105 "psubw %%mm1, %%mm3 \n\t"
1106
1107 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
1108 "paddw %%mm1, %%mm4 \n\t"
1109
1110 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
1111 "paddw %%mm3, %%mm5 \n\t"
1112
1113 "movq "MANGLE(temps)"+3*8, %%mm0 \n\t"
1114 "add $8, %%"REG_S" \n\t"
1115
1116 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
1117 "paddw %%mm0, %%mm6 \n\t"
1118
1119 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
1120 "psubw %%mm2, %%mm0 \n\t"
1121
1122 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
1123 "paddw %%mm2, %%mm6 \n\t"
1124
1125 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
1126 "paddw %%mm0, %%mm7 \n\t"
1127
1128 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
1129
1130 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
1131 "add $8, %%"REG_D" \n\t"
1132 "jmp 4f \n\t"
1133
1134 "2: \n\t"
1135 //--- non DC2
1136 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
1137 //psraw mm5, 2
1138 //psraw mm0, 2
1139 //psraw mm6, 2
1140 "movq %%mm5, %%mm3 \n\t"
1141 "psubw %%mm1, %%mm5 \n\t"
1142
1143 "psllw $1, %%mm5 \n\t" //'z10
1144 "paddw %%mm1, %%mm3 \n\t" //'z13
1145
1146 "movq %%mm0, %%mm2 \n\t"
1147 "psubw %%mm6, %%mm0 \n\t"
1148
1149 "movq %%mm5, %%mm1 \n\t"
1150 "psllw $1, %%mm0 \n\t" //'z12
1151
1152 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
1153 "paddw %%mm0, %%mm5 \n\t"
1154
1155 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
1156 "paddw %%mm6, %%mm2 \n\t" //'z11
1157
1158 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
1159 "movq %%mm2, %%mm7 \n\t"
1160
1161 //---
1162 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t"
1163 "psubw %%mm3, %%mm2 \n\t"
1164
1165 "psllw $1, %%mm2 \n\t"
1166 "paddw %%mm3, %%mm7 \n\t" //'t7
1167
1168 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
1169 "movq %%mm4, %%mm6 \n\t"
1170 //paddw mm7, MM_2
1171 "psraw $2, %%mm7 \n\t"
1172
1173 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
1174 "psubw %%mm7, %%mm6 \n\t"
1175
1176 "movq "MANGLE(temps)"+1*8, %%mm3 \n\t"
1177 "paddw %%mm7, %%mm4 \n\t"
1178
1179 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
1180 "paddw %%mm5, %%mm1 \n\t" //'t12
1181
1182 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
1183 "psubw %%mm7, %%mm1 \n\t" //'t6
1184
1185 "movq "MANGLE(temps)"+2*8, %%mm7 \n\t"
1186 "psubw %%mm5, %%mm0 \n\t" //'t10
1187
1188 "movq "MANGLE(temps)"+3*8, %%mm6 \n\t"
1189 "movq %%mm3, %%mm5 \n\t"
1190
1191 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
1192 "psubw %%mm1, %%mm5 \n\t"
1193
1194 "psubw %%mm1, %%mm2 \n\t" //'t5
1195 "paddw %%mm1, %%mm3 \n\t"
1196
1197 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
1198 "movq %%mm7, %%mm4 \n\t"
1199
1200 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
1201 "psubw %%mm2, %%mm4 \n\t"
1202
1203 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
1204 "paddw %%mm2, %%mm7 \n\t"
1205
1206 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
1207 "paddw %%mm2, %%mm0 \n\t" //'t4
1208
1209 // 't4 't6 't5 - - - - 't7
1210 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
1211 "movq %%mm6, %%mm1 \n\t"
1212
1213 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
1214 "psubw %%mm0, %%mm1 \n\t"
1215
1216 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
1217 "paddw %%mm0, %%mm6 \n\t"
1218
1219 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
1220 "add $8, %%"REG_S" \n\t"
1221
1222 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
1223
1224 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
1225 "add $8, %%"REG_D" \n\t"
1226
1227 "4: \n\t"
1228 //=part 2 (the same)===========================================================
1229 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
1230 //
1231 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
1232 "movq %%mm1, %%mm0 \n\t"
1233
1234 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0
1235 "movq %%mm7, %%mm3 \n\t"
1236
1237 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
1238 "movq %%mm1, %%mm5 \n\t"
1239
1240 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
1241 "psubw %%mm7, %%mm1 \n\t" //t13
1242
1243 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
1244 "movq %%mm6, %%mm4 \n\t"
1245
1246 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
1247 "paddw %%mm7, %%mm5 \n\t" //t10
1248
1249 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
1250 "movq %%mm6, %%mm7 \n\t"
1251
1252 "paddw %%mm2, %%mm6 \n\t" //t11
1253 "psubw %%mm2, %%mm7 \n\t" //t12
1254
1255 "movq %%mm5, %%mm2 \n\t"
1256 "paddw %%mm6, %%mm5 \n\t" //d0
1257 // i0 t13 t12 i3 i1 d0 - d4
1258 "psubw %%mm6, %%mm2 \n\t" //d4
1259 "paddw %%mm1, %%mm7 \n\t"
1260
1261 "movq 1*8+4*16(%%"REG_d"), %%mm6 \n\t"
1262 "psllw $2, %%mm7 \n\t"
1263
1264 "psubw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
1265 "psubw %%mm6, %%mm2 \n\t"
1266
1267 "paddusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
1268 "paddusw %%mm6, %%mm2 \n\t"
1269
1270 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t"
1271 //
1272 "paddw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
1273 "paddw %%mm6, %%mm2 \n\t"
1274
1275 "psubusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
1276 "psubusw %%mm6, %%mm2 \n\t"
1277
1278 //This func is totally compute-bound, operates at huge speed. So, DC shortcut
1279 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
1280 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
1281 "paddw "MANGLE(MM_2)", %%mm5 \n\t"
1282 "movq %%mm2, %%mm6 \n\t"
1283
1284 "paddw %%mm5, %%mm2 \n\t"
1285 "psubw %%mm6, %%mm5 \n\t"
1286
1287 "movq %%mm1, %%mm6 \n\t"
1288 "paddw %%mm7, %%mm1 \n\t" //d2
1289
1290 "psubw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
1291 "psubw %%mm7, %%mm6 \n\t" //d6
1292
1293 "movq 1*8+6*16(%%"REG_d"), %%mm7 \n\t"
1294 "psraw $2, %%mm5 \n\t"
1295
1296 "paddusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
1297 "psubw %%mm7, %%mm6 \n\t"
1298 // t7 d2 /t11 t4 t6 - d6 /t10
1299
1300 "paddw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
1301 "paddusw %%mm7, %%mm6 \n\t"
1302
1303 "psubusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
1304 "paddw %%mm7, %%mm6 \n\t"
1305
1306 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t"
1307 "psubusw %%mm7, %%mm6 \n\t"
1308
1309 //movq [edi+"DCTSIZE_S"*2*2], mm1
1310 //movq [edi+"DCTSIZE_S"*6*2], mm6
1311 "movq %%mm1, %%mm7 \n\t"
1312 "psraw $2, %%mm2 \n\t"
1313
1314 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t"
1315 "psubw %%mm6, %%mm1 \n\t"
1316
1317 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t"
1318 "paddw %%mm7, %%mm6 \n\t" //'t13
1319
1320 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
1321 "movq %%mm2, %%mm7 \n\t"
1322
1323 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
1324 "paddw %%mm6, %%mm2 \n\t" //'t0
1325
1326 "movq %%mm2, "MANGLE(temps)"+0*8 \n\t" //!
1327 "psubw %%mm6, %%mm7 \n\t" //'t3
1328
1329 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
1330 "psubw %%mm6, %%mm1 \n\t" //'t12
1331
1332 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5
1333 "movq %%mm5, %%mm6 \n\t"
1334
1335 "movq %%mm7, "MANGLE(temps)"+3*8 \n\t"
1336 "paddw %%mm2, %%mm3 \n\t" //t10
1337
1338 "paddw %%mm4, %%mm2 \n\t" //t11
1339 "paddw %%mm0, %%mm4 \n\t" //t12
1340
1341 "movq %%mm3, %%mm7 \n\t"
1342 "psubw %%mm4, %%mm3 \n\t"
1343
1344 "psllw $2, %%mm3 \n\t"
1345 "psllw $2, %%mm7 \n\t" //opt for P6
1346
1347 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
1348 "psllw $2, %%mm4 \n\t"
1349
1350 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t"
1351 "psllw $2, %%mm2 \n\t"
1352
1353 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
1354 "paddw %%mm1, %%mm5 \n\t" //'t1
1355
1356 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t"
1357 "psubw %%mm1, %%mm6 \n\t" //'t2
1358 // t7 't12 't11 t4 t6 - 't13 't10 ---
1359
1360 "paddw %%mm3, %%mm7 \n\t" //z2
1361
1362 "movq %%mm5, "MANGLE(temps)"+1*8 \n\t"
1363 "paddw %%mm3, %%mm4 \n\t" //z4
1364
1365 "movq 1*8+3*16(%%"REG_d"), %%mm3 \n\t"
1366 "movq %%mm0, %%mm1 \n\t"
1367
1368 "movq %%mm6, "MANGLE(temps)"+2*8 \n\t"
1369 "psubw %%mm2, %%mm1 \n\t" //z13
1370
1371 //===
1372 "paddw %%mm2, %%mm0 \n\t" //z11
1373 "movq %%mm1, %%mm5 \n\t"
1374
1375 "movq 1*8+5*16(%%"REG_d"), %%mm2 \n\t"
1376 "psubw %%mm7, %%mm1 \n\t" //d3
1377
1378 "paddw %%mm7, %%mm5 \n\t" //d5
1379 "psubw %%mm3, %%mm1 \n\t"
1380
1381 "movq 1*8+1*16(%%"REG_d"), %%mm7 \n\t"
1382 "psubw %%mm2, %%mm5 \n\t"
1383
1384 "movq %%mm0, %%mm6 \n\t"
1385 "paddw %%mm4, %%mm0 \n\t" //d1
1386
1387 "paddusw %%mm3, %%mm1 \n\t"
1388 "psubw %%mm4, %%mm6 \n\t" //d7
1389
1390 // d1 d3 - - - d5 d7 -
1391 "movq 1*8+7*16(%%"REG_d"), %%mm4 \n\t"
1392 "psubw %%mm7, %%mm0 \n\t"
1393
1394 "psubw %%mm4, %%mm6 \n\t"
1395 "paddusw %%mm2, %%mm5 \n\t"
1396
1397 "paddusw %%mm4, %%mm6 \n\t"
1398 "paddw %%mm3, %%mm1 \n\t"
1399
1400 "paddw %%mm2, %%mm5 \n\t"
1401 "paddw %%mm4, %%mm6 \n\t"
1402
1403 "psubusw %%mm3, %%mm1 \n\t"
1404 "psubusw %%mm2, %%mm5 \n\t"
1405
1406 "psubusw %%mm4, %%mm6 \n\t"
1407 "movq %%mm1, %%mm4 \n\t"
1408
1409 "por %%mm5, %%mm4 \n\t"
1410 "paddusw %%mm7, %%mm0 \n\t"
1411
1412 "por %%mm6, %%mm4 \n\t"
1413 "paddw %%mm7, %%mm0 \n\t"
1414
1415 "packssdw %%mm4, %%mm4 \n\t"
1416 "psubusw %%mm7, %%mm0 \n\t"
1417
1418 "movd %%mm4, %%"REG_a" \n\t"
1419 "or %%"REG_a", %%"REG_a" \n\t"
1420 "jnz 3f \n\t"
1421 //movq [edi+"DCTSIZE_S"*3*2], mm1
1422 //movq [edi+"DCTSIZE_S"*5*2], mm5
1423 //movq [edi+"DCTSIZE_S"*1*2], mm0
1424 //movq [edi+"DCTSIZE_S"*7*2], mm6
1425 // t4 t5 - - - t6 t7 -
1426 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
1427 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
1428 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t"
1429 "movq %%mm0, %%mm1 \n\t"
1430
1431 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
1432 "movq %%mm1, %%mm2 \n\t"
1433
1434 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
1435 "movq %%mm2, %%mm3 \n\t"
1436
1437 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
1438 "paddw %%mm4, %%mm5 \n\t"
1439
1440 "movq "MANGLE(temps)"+1*8, %%mm6 \n\t"
1441 //paddw mm3, MM_2
1442 "psraw $2, %%mm3 \n\t" //tmp7
1443
1444 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
1445 "psubw %%mm3, %%mm4 \n\t"
1446
1447 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
1448 "paddw %%mm3, %%mm5 \n\t"
1449
1450 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
1451 "paddw %%mm6, %%mm7 \n\t"
1452
1453 "movq "MANGLE(temps)"+2*8, %%mm3 \n\t"
1454 "psubw %%mm0, %%mm6 \n\t"
1455
1456 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
1457 "paddw %%mm0, %%mm7 \n\t"
1458
1459 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
1460 "paddw %%mm3, %%mm4 \n\t"
1461
1462 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
1463 "psubw %%mm1, %%mm3 \n\t"
1464
1465 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
1466 "paddw %%mm1, %%mm4 \n\t"
1467
1468 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
1469 "paddw %%mm3, %%mm5 \n\t"
1470
1471 "movq "MANGLE(temps)"+3*8, %%mm0 \n\t"
1472 "add $24, %%"REG_S" \n\t"
1473
1474 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
1475 "paddw %%mm0, %%mm6 \n\t"
1476
1477 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
1478 "psubw %%mm2, %%mm0 \n\t"
1479
1480 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
1481 "paddw %%mm2, %%mm6 \n\t"
1482
1483 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
1484 "paddw %%mm0, %%mm7 \n\t"
1485
1486 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
1487
1488 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
1489 "add $24, %%"REG_D" \n\t"
1490 "sub $2, %%"REG_c" \n\t"
1491 "jnz 1b \n\t"
1492 "jmp 5f \n\t"
1493
1494 "3: \n\t"
1495 //--- non DC2
1496 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
1497 //psraw mm5, 2
1498 //psraw mm0, 2
1499 //psraw mm6, 2
1500 "movq %%mm5, %%mm3 \n\t"
1501 "psubw %%mm1, %%mm5 \n\t"
1502
1503 "psllw $1, %%mm5 \n\t" //'z10
1504 "paddw %%mm1, %%mm3 \n\t" //'z13
1505
1506 "movq %%mm0, %%mm2 \n\t"
1507 "psubw %%mm6, %%mm0 \n\t"
1508
1509 "movq %%mm5, %%mm1 \n\t"
1510 "psllw $1, %%mm0 \n\t" //'z12
1511
1512 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
1513 "paddw %%mm0, %%mm5 \n\t"
1514
1515 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
1516 "paddw %%mm6, %%mm2 \n\t" //'z11
1517
1518 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
1519 "movq %%mm2, %%mm7 \n\t"
1520
1521 //---
1522 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t"
1523 "psubw %%mm3, %%mm2 \n\t"
1524
1525 "psllw $1, %%mm2 \n\t"
1526 "paddw %%mm3, %%mm7 \n\t" //'t7
1527
1528 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
1529 "movq %%mm4, %%mm6 \n\t"
1530 //paddw mm7, MM_2
1531 "psraw $2, %%mm7 \n\t"
1532
1533 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
1534 "psubw %%mm7, %%mm6 \n\t"
1535
1536 "movq "MANGLE(temps)"+1*8, %%mm3 \n\t"
1537 "paddw %%mm7, %%mm4 \n\t"
1538
1539 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
1540 "paddw %%mm5, %%mm1 \n\t" //'t12
1541
1542 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
1543 "psubw %%mm7, %%mm1 \n\t" //'t6
1544
1545 "movq "MANGLE(temps)"+2*8, %%mm7 \n\t"
1546 "psubw %%mm5, %%mm0 \n\t" //'t10
1547
1548 "movq "MANGLE(temps)"+3*8, %%mm6 \n\t"
1549 "movq %%mm3, %%mm5 \n\t"
1550
1551 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
1552 "psubw %%mm1, %%mm5 \n\t"
1553
1554 "psubw %%mm1, %%mm2 \n\t" //'t5
1555 "paddw %%mm1, %%mm3 \n\t"
1556
1557 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
1558 "movq %%mm7, %%mm4 \n\t"
1559
1560 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
1561 "psubw %%mm2, %%mm4 \n\t"
1562
1563 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
1564 "paddw %%mm2, %%mm7 \n\t"
1565
1566 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
1567 "paddw %%mm2, %%mm0 \n\t" //'t4
1568
1569 // 't4 't6 't5 - - - - 't7
1570 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
1571 "movq %%mm6, %%mm1 \n\t"
1572
1573 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
1574 "psubw %%mm0, %%mm1 \n\t"
1575
1576 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
1577 "paddw %%mm0, %%mm6 \n\t"
1578
1579 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
1580 "add $24, %%"REG_S" \n\t"
1581
1582 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
1583
1584 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
1585 "add $24, %%"REG_D" \n\t"
1586 "sub $2, %%"REG_c" \n\t"
1587 "jnz 1b \n\t"
1588 "5: \n\t"
1589
1590 : "+S"(data), "+D"(output), "+c"(cnt)// input regs
1591 : "d"(thr_adr)
1592 : "%"REG_a
1593 );
1594 }
1595
1596 #endif // HAVE_MMX
1597
1598 #ifndef HAVE_MMX
1599
1600 static void row_idct_c(DCTELEM* workspace,
1601 int16_t* output_adr, int output_stride, int cnt)
1602 {
1603 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
1604 int_simd16_t tmp10, tmp11, tmp12, tmp13;
1605 int_simd16_t z5, z10, z11, z12, z13;
1606 int16_t* outptr;
1607 DCTELEM* wsptr;
1608
1609 cnt*=4;
1610 wsptr = workspace;
1611 outptr = output_adr;
1612 for (; cnt > 0; cnt--) {
1613 // Even part
1614 //Simd version reads 4x4 block and transposes it
1615 tmp10 = ( wsptr[2] + wsptr[3]);
1616 tmp11 = ( wsptr[2] - wsptr[3]);
1617
1618 tmp13 = ( wsptr[0] + wsptr[1]);
1619 tmp12 = (MULTIPLY16H( wsptr[0] - wsptr[1], FIX_1_414213562_A)<<2) - tmp13;//this shift order to avoid overflow
1620
1621 tmp0 = tmp10 + tmp13; //->temps
1622 tmp3 = tmp10 - tmp13; //->temps
1623 tmp1 = tmp11 + tmp12;
1624 tmp2 = tmp11 - tmp12;
1625
1626 // Odd part
1627 //Also transpose, with previous:
1628 // ---- ---- ||||
1629 // ---- ---- idct ||||
1630 // ---- ---- ---> ||||
1631 // ---- ---- ||||
1632 z13 = wsptr[4] + wsptr[5];
1633 z10 = wsptr[4] - wsptr[5];
1634 z11 = wsptr[6] + wsptr[7];
1635 z12 = wsptr[6] - wsptr[7];
1636
1637 tmp7 = z11 + z13;
1638 tmp11 = MULTIPLY16H(z11 - z13, FIX_1_414213562);
1639
1640 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
1641 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
1642 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - FIX_
1643
1644 tmp6 = (tmp12<<3) - tmp7;
1645 tmp5 = (tmp11<<3) - tmp6;
1646 tmp4 = (tmp10<<3) + tmp5;
1647
1648 // Final output stage: descale and write column
1649 outptr[0*output_stride]+= DESCALE(tmp0 + tmp7, 3);
1650 outptr[1*output_stride]+= DESCALE(tmp1 + tmp6, 3);
1651 outptr[2*output_stride]+= DESCALE(tmp2 + tmp5, 3);
1652 outptr[3*output_stride]+= DESCALE(tmp3 - tmp4, 3);
1653 outptr[4*output_stride]+= DESCALE(tmp3 + tmp4, 3);
1654 outptr[5*output_stride]+= DESCALE(tmp2 - tmp5, 3);
1655 outptr[6*output_stride]+= DESCALE(tmp1 - tmp6, 3); //no += ?
1656 outptr[7*output_stride]+= DESCALE(tmp0 - tmp7, 3); //no += ?
1657 outptr++;
1658
1659 wsptr += DCTSIZE; // advance pointer to next row
1660 }
1661 }
1662
1663 #else /* HAVE_MMX */
1664
1665 static void row_idct_mmx (DCTELEM* workspace,
1666 int16_t* output_adr, int output_stride, int cnt)
1667 {
1668 asm volatile(
1669 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t"
1670 "1: \n\t"
1671 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm0 \n\t"
1672 //
1673
1674 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm1 \n\t"
1675 "movq %%mm0, %%mm4 \n\t"
1676
1677 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
1678 "punpcklwd %%mm1, %%mm0 \n\t"
1679
1680 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm3 \n\t"
1681 "punpckhwd %%mm1, %%mm4 \n\t"
1682
1683 //transpose 4x4
1684 "movq %%mm2, %%mm7 \n\t"
1685 "punpcklwd %%mm3, %%mm2 \n\t"
1686
1687 "movq %%mm0, %%mm6 \n\t"
1688 "punpckldq %%mm2, %%mm0 \n\t" //0
1689
1690 "punpckhdq %%mm2, %%mm6 \n\t" //1
1691 "movq %%mm0, %%mm5 \n\t"
1692
1693 "punpckhwd %%mm3, %%mm7 \n\t"
1694 "psubw %%mm6, %%mm0 \n\t"
1695
1696 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm0 \n\t"
1697 "movq %%mm4, %%mm2 \n\t"
1698
1699 "punpckldq %%mm7, %%mm4 \n\t" //2
1700 "paddw %%mm6, %%mm5 \n\t"
1701
1702 "punpckhdq %%mm7, %%mm2 \n\t" //3
1703 "movq %%mm4, %%mm1 \n\t"
1704
1705 "psllw $2, %%mm0 \n\t"
1706 "paddw %%mm2, %%mm4 \n\t" //t10
1707
1708 "movq "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_S"), %%mm3 \n\t"
1709 "psubw %%mm2, %%mm1 \n\t" //t11
1710
1711 "movq "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_S"), %%mm2 \n\t"
1712 "psubw %%mm5, %%mm0 \n\t"
1713
1714 "movq %%mm4, %%mm6 \n\t"
1715 "paddw %%mm5, %%mm4 \n\t" //t0
1716
1717 "psubw %%mm5, %%mm6 \n\t" //t3
1718 "movq %%mm1, %%mm7 \n\t"
1719
1720 "movq "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_S"), %%mm5 \n\t"
1721 "paddw %%mm0, %%mm1 \n\t" //t1
1722
1723 "movq %%mm4, "MANGLE(temps)"+0*8 \n\t" //t0
1724 "movq %%mm3, %%mm4 \n\t"
1725
1726 "movq %%mm6, "MANGLE(temps)"+1*8 \n\t" //t3
1727 "punpcklwd %%mm2, %%mm3 \n\t"
1728
1729 //transpose 4x4
1730 "movq "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_S"), %%mm6 \n\t"
1731 "punpckhwd %%mm2, %%mm4 \n\t"
1732
1733 "movq %%mm5, %%mm2 \n\t"
1734 "punpcklwd %%mm6, %%mm5 \n\t"
1735
1736 "psubw %%mm0, %%mm7 \n\t" //t2
1737 "punpckhwd %%mm6, %%mm2 \n\t"
1738
1739 "movq %%mm3, %%mm0 \n\t"
1740 "punpckldq %%mm5, %%mm3 \n\t" //4
1741
1742 "punpckhdq %%mm5, %%mm0 \n\t" //5
1743 "movq %%mm4, %%mm5 \n\t"
1744
1745 //
1746 "movq %%mm3, %%mm6 \n\t"
1747 "punpckldq %%mm2, %%mm4 \n\t" //6
1748
1749 "psubw %%mm0, %%mm3 \n\t" //z10
1750 "punpckhdq %%mm2, %%mm5 \n\t" //7
1751
1752 "paddw %%mm0, %%mm6 \n\t" //z13
1753 "movq %%mm4, %%mm2 \n\t"
1754
1755 "movq %%mm3, %%mm0 \n\t"
1756 "psubw %%mm5, %%mm4 \n\t" //z12
1757
1758 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm0 \n\t" //-
1759 "paddw %%mm4, %%mm3 \n\t"
1760
1761 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm3 \n\t" //z5
1762 "paddw %%mm5, %%mm2 \n\t" //z11 >
1763
1764 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm4 \n\t"
1765 "movq %%mm2, %%mm5 \n\t"
1766
1767 "psubw %%mm6, %%mm2 \n\t"
1768 "paddw %%mm6, %%mm5 \n\t" //t7
1769
1770 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //t11
1771 "paddw %%mm3, %%mm0 \n\t" //t12
1772
1773 "psllw $3, %%mm0 \n\t"
1774 "psubw %%mm3, %%mm4 \n\t" //t10
1775
1776 "movq "MANGLE(temps)"+0*8, %%mm6 \n\t"
1777 "movq %%mm1, %%mm3 \n\t"
1778
1779 "psllw $3, %%mm4 \n\t"
1780 "psubw %%mm5, %%mm0 \n\t" //t6
1781
1782 "psllw $3, %%mm2 \n\t"
1783 "paddw %%mm0, %%mm1 \n\t" //d1
1784
1785 "psubw %%mm0, %%mm2 \n\t" //t5
1786 "psubw %%mm0, %%mm3 \n\t" //d6
1787
1788 "paddw %%mm2, %%mm4 \n\t" //t4
1789 "movq %%mm7, %%mm0 \n\t"
1790
1791 "paddw %%mm2, %%mm7 \n\t" //d2
1792 "psubw %%mm2, %%mm0 \n\t" //d5
1793
1794 "movq "MANGLE(MM_DESCALE_RND)", %%mm2 \n\t" //4
1795 "psubw %%mm5, %%mm6 \n\t" //d7
1796
1797 "paddw "MANGLE(temps)"+0*8, %%mm5 \n\t" //d0
1798 "paddw %%mm2, %%mm1 \n\t"
1799
1800 "paddw %%mm2, %%mm5 \n\t"
1801 "psraw $3, %%mm1 \n\t"
1802
1803 "paddw %%mm2, %%mm7 \n\t"
1804 "psraw $3, %%mm5 \n\t"
1805
1806 "paddw (%%"REG_D"), %%mm5 \n\t"
1807 "psraw $3, %%mm7 \n\t"
1808
1809 "paddw (%%"REG_D",%%"REG_a",), %%mm1 \n\t"
1810 "paddw %%mm2, %%mm0 \n\t"
1811
1812 "paddw (%%"REG_D",%%"REG_a",2), %%mm7 \n\t"
1813 "paddw %%mm2, %%mm3 \n\t"
1814
1815 "movq %%mm5, (%%"REG_D") \n\t"
1816 "paddw %%mm2, %%mm6 \n\t"
1817
1818 "movq %%mm1, (%%"REG_D",%%"REG_a",) \n\t"
1819 "psraw $3, %%mm0 \n\t"
1820
1821 "movq %%mm7, (%%"REG_D",%%"REG_a",2) \n\t"
1822 "add %%"REG_d", %%"REG_D" \n\t" //3*ls
1823
1824 "movq "MANGLE(temps)"+1*8, %%mm5 \n\t" //t3
1825 "psraw $3, %%mm3 \n\t"
1826
1827 "paddw (%%"REG_D",%%"REG_a",2), %%mm0 \n\t"
1828 "psubw %%mm4, %%mm5 \n\t" //d3
1829
1830 "paddw (%%"REG_D",%%"REG_d",), %%mm3 \n\t"
1831 "psraw $3, %%mm6 \n\t"
1832
1833 "paddw "MANGLE(temps)"+1*8, %%mm4 \n\t" //d4
1834 "paddw %%mm2, %%mm5 \n\t"
1835
1836 "paddw (%%"REG_D",%%"REG_a",4), %%mm6 \n\t"
1837 "paddw %%mm2, %%mm4 \n\t"
1838
1839 "movq %%mm0, (%%"REG_D",%%"REG_a",2) \n\t"
1840 "psraw $3, %%mm5 \n\t"
1841
1842 "paddw (%%"REG_D"), %%mm5 \n\t"
1843 "psraw $3, %%mm4 \n\t"
1844
1845 "paddw (%%"REG_D",%%"REG_a",), %%mm4 \n\t"
1846 "add $"DCTSIZE_S"*2*4, %%"REG_S" \n\t" //4 rows
1847
1848 "movq %%mm3, (%%"REG_D",%%"REG_d",) \n\t"
1849 "movq %%mm6, (%%"REG_D",%%"REG_a",4) \n\t"
1850 "movq %%mm5, (%%"REG_D") \n\t"
1851 "movq %%mm4, (%%"REG_D",%%"REG_a",) \n\t"
1852
1853 "sub %%"REG_d", %%"REG_D" \n\t"
1854 "add $8, %%"REG_D" \n\t"
1855 "dec %%"REG_c" \n\t"
1856 "jnz 1b \n\t"
1857
1858 : "+S"(workspace), "+D"(output_adr), "+c"(cnt) //input regs
1859 : "a"(output_stride*sizeof(short))
1860 : "%"REG_d
1861 );
1862 }
1863
1864 #endif // HAVE_MMX
1865
1866 #ifndef HAVE_MMX
1867
1868 static void row_fdct_c(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt)
1869 {
1870 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
1871 int_simd16_t tmp10, tmp11, tmp12, tmp13;
1872 int_simd16_t z1, z2, z3, z4, z5, z11, z13;
1873 DCTELEM *dataptr;
1874
1875 cnt*=4;
1876 // Pass 1: process rows.
1877
1878 dataptr = data;
1879 for (; cnt > 0; cnt--) {
1880 tmp0 = pixels[line_size*0] + pixels[line_size*7];
1881 tmp7 = pixels[line_size*0] - pixels[line_size*7];
1882 tmp1 = pixels[line_size*1] + pixels[line_size*6];
1883 tmp6 = pixels[line_size*1] - pixels[line_size*6];
1884 tmp2 = pixels[line_size*2] + pixels[line_size*5];
1885 tmp5 = pixels[line_size*2] - pixels[line_size*5];
1886 tmp3 = pixels[line_size*3] + pixels[line_size*4];
1887 tmp4 = pixels[line_size*3] - pixels[line_size*4];
1888
1889 // Even part
1890
1891 tmp10 = tmp0 + tmp3;
1892 tmp13 = tmp0 - tmp3;
1893 tmp11 = tmp1 + tmp2;
1894 tmp12 = tmp1 - tmp2;
1895 //Even columns are written first, this leads to different order of columns
1896 //in column_fidct(), but they are processed independently, so all ok.
1897 //Later in the row_idct() columns readed at the same order.
1898 dataptr[2] = tmp10 + tmp11;
1899 dataptr[3] = tmp10 - tmp11;
1900
1901 z1 = MULTIPLY16H((tmp12 + tmp13)<<2, FIX_0_707106781);
1902 dataptr[0] = tmp13 + z1;
1903 dataptr[1] = tmp13 - z1;
1904
1905 // Odd part
1906
1907 tmp10 = (tmp4 + tmp5) <<2;
1908 tmp11 = (tmp5 + tmp6) <<2;
1909 tmp12 = (tmp6 + tmp7) <<2;
1910
1911 z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433);
1912 z2 = MULTIPLY16H(tmp10, FIX_0_541196100) + z5;
1913 z4 = MULTIPLY16H(tmp12, FIX_1_306562965) + z5;
1914 z3 = MULTIPLY16H(tmp11, FIX_0_707106781);
1915
1916 z11 = tmp7 + z3;
1917 z13 = tmp7 - z3;
1918
1919 dataptr[4] = z13 + z2;
1920 dataptr[5] = z13 - z2;
1921 dataptr[6] = z11 + z4;
1922 dataptr[7] = z11 - z4;
1923
1924 pixels++; // advance pointer to next column
1925 dataptr += DCTSIZE;
1926 }
1927 }
1928
1929 #else /* HAVE_MMX */
1930
1931 static void row_fdct_mmx(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt)
1932 {
1933 asm volatile(
1934 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t"
1935 "6: \n\t"
1936 "movd (%%"REG_S"), %%mm0 \n\t"
1937 "pxor %%mm7, %%mm7 \n\t"
1938
1939 "movd (%%"REG_S",%%"REG_a",), %%mm1 \n\t"
1940 "punpcklbw %%mm7, %%mm0 \n\t"
1941
1942 "movd (%%"REG_S",%%"REG_a",2), %%mm2 \n\t"
1943 "punpcklbw %%mm7, %%mm1 \n\t"
1944
1945 "punpcklbw %%mm7, %%mm2 \n\t"
1946 "add %%"REG_d", %%"REG_S" \n\t"
1947
1948 "movq %%mm0, %%mm5 \n\t"
1949 //
1950
1951 "movd (%%"REG_S",%%"REG_a",4), %%mm3 \n\t" //7 ;prefetch!
1952 "movq %%mm1, %%mm6 \n\t"
1953
1954 "movd (%%"REG_S",%%"REG_d",), %%mm4 \n\t" //6
1955 "punpcklbw %%mm7, %%mm3 \n\t"
1956
1957 "psubw %%mm3, %%mm5 \n\t"
1958 "punpcklbw %%mm7, %%mm4 \n\t"
1959
1960 "paddw %%mm3, %%mm0 \n\t"
1961 "psubw %%mm4, %%mm6 \n\t"
1962
1963 "movd (%%"REG_S",%%"REG_a",2), %%mm3 \n\t" //5
1964 "paddw %%mm4, %%mm1 \n\t"
1965
1966 "movq %%mm5, "MANGLE(temps)"+0*8 \n\t" //t7
1967 "punpcklbw %%mm7, %%mm3 \n\t"
1968
1969 "movq %%mm6, "MANGLE(temps)"+1*8 \n\t" //t6
1970 "movq %%mm2, %%mm4 \n\t"
1971
1972 "movd (%%"REG_S"), %%mm5 \n\t" //3
1973 "paddw %%mm3, %%mm2 \n\t"
1974
1975 "movd (%%"REG_S",%%"REG_a",), %%mm6 \n\t" //4
1976 "punpcklbw %%mm7, %%mm5 \n\t"
1977
1978 "psubw %%mm3, %%mm4 \n\t"
1979 "punpcklbw %%mm7, %%mm6 \n\t"
1980
1981 "movq %%mm5, %%mm3 \n\t"
1982 "paddw %%mm6, %%mm5 \n\t" //t3
1983
1984 "psubw %%mm6, %%mm3 \n\t" //t4 ; t0 t1 t2 t4 t5 t3 - -
1985 "movq %%mm0, %%mm6 \n\t"
1986
1987 "movq %%mm1, %%mm7 \n\t"
1988 "psubw %%mm5, %%mm0 \n\t" //t13
1989
1990 "psubw %%mm2, %%mm1 \n\t"
1991 "paddw %%mm2, %%mm7 \n\t" //t11
1992
1993 "paddw %%mm0, %%mm1 \n\t"
1994 "movq %%mm7, %%mm2 \n\t"
1995
1996 "psllw $2, %%mm1 \n\t"
1997 "paddw %%mm5, %%mm6 \n\t" //t10
1998
1999 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm1 \n\t"
2000 "paddw %%mm6, %%mm7 \n\t" //d2
2001
2002 "psubw %%mm2, %%mm6 \n\t" //d3
2003 "movq %%mm0, %%mm5 \n\t"
2004
2005 //transpose 4x4
2006 "movq %%mm7, %%mm2 \n\t"
2007 "punpcklwd %%mm6, %%mm7 \n\t"
2008
2009 "paddw %%mm1, %%mm0 \n\t" //d0
2010 "punpckhwd %%mm6, %%mm2 \n\t"
2011
2012 "psubw %%mm1, %%mm5 \n\t" //d1
2013 "movq %%mm0, %%mm6 \n\t"
2014
2015 "movq "MANGLE(temps)"+1*8, %%mm1 \n\t"
2016 "punpcklwd %%mm5, %%mm0 \n\t"
2017
2018 "punpckhwd %%mm5, %%mm6 \n\t"
2019 "movq %%mm0, %%mm5 \n\t"
2020
2021 "punpckldq %%mm7, %%mm0 \n\t" //0
2022 "paddw %%mm4, %%mm3 \n\t"
2023
2024 "punpckhdq %%mm7, %%mm5 \n\t" //1
2025 "movq %%mm6, %%mm7 \n\t"
2026
2027 "movq %%mm0, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
2028 "punpckldq %%mm2, %%mm6 \n\t" //2
2029
2030 "movq %%mm5, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
2031 "punpckhdq %%mm2, %%mm7 \n\t" //3
2032
2033 "movq %%mm6, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
2034 "paddw %%mm1, %%mm4 \n\t"
2035
2036 "movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
2037 "psllw $2, %%mm3 \n\t" //t10
2038
2039 "movq "MANGLE(temps)"+0*8, %%mm2 \n\t"
2040 "psllw $2, %%mm4 \n\t" //t11
2041
2042 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm4 \n\t" //z3
2043 "paddw %%mm2, %%mm1 \n\t"
2044
2045 "psllw $2, %%mm1 \n\t" //t12
2046 "movq %%mm3, %%mm0 \n\t"
2047
2048 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm0 \n\t"
2049 "psubw %%mm1, %%mm3 \n\t"
2050
2051 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" //z5
2052 "movq %%mm2, %%mm5 \n\t"
2053
2054 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm1 \n\t"
2055 "psubw %%mm4, %%mm2 \n\t" //z13
2056
2057 "paddw %%mm4, %%mm5 \n\t" //z11
2058 "movq %%mm2, %%mm6 \n\t"
2059
2060 "paddw %%mm3, %%mm0 \n\t" //z2
2061 "movq %%mm5, %%mm7 \n\t"
2062
2063 "paddw %%mm0, %%mm2 \n\t" //d4
2064 "psubw %%mm0, %%mm6 \n\t" //d5
2065
2066 "movq %%mm2, %%mm4 \n\t"
2067 "paddw %%mm3, %%mm1 \n\t" //z4
2068
2069 //transpose 4x4
2070 "punpcklwd %%mm6, %%mm2 \n\t"
2071 "paddw %%mm1, %%mm5 \n\t" //d6
2072
2073 "punpckhwd %%mm6, %%mm4 \n\t"
2074 "psubw %%mm1, %%mm7 \n\t" //d7
2075
2076 "movq %%mm5, %%mm6 \n\t"
2077 "punpcklwd %%mm7, %%mm5 \n\t"
2078
2079 "punpckhwd %%mm7, %%mm6 \n\t"
2080 "movq %%mm2, %%mm7 \n\t"
2081
2082 "punpckldq %%mm5, %%mm2 \n\t" //4
2083 "sub %%"REG_d", %%"REG_S" \n\t"
2084
2085 "punpckhdq %%mm5, %%mm7 \n\t" //5
2086 "movq %%mm4, %%mm5 \n\t"
2087
2088 "movq %%mm2, "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_D") \n\t"
2089 "punpckldq %%mm6, %%mm4 \n\t" //6
2090
2091 "movq %%mm7, "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_D") \n\t"
2092 "punpckhdq %%mm6, %%mm5 \n\t" //7
2093
2094 "movq %%mm4, "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_D") \n\t"
2095 "add $4, %%"REG_S" \n\t"
2096
2097 "movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_D") \n\t"
2098 "add $"DCTSIZE_S"*2*4, %%"REG_D" \n\t" //4 rows
2099 "dec %%"REG_c" \n\t"
2100 "jnz 6b \n\t"
2101
2102 : "+S"(pixels), "+D"(data), "+c"(cnt) //input regs
2103 : "a"(line_size)
2104 : "%"REG_d);
2105 }
2106
2107 #endif // HAVE_MMX