?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
fix
[mplayer/glamo.git] / libmpcodecs / vf_fspp.c
blobf6ba41547c256a55fe4f02de0474589bb37ae598
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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 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 #ifdef USE_LIBAVCODEC
42
43 #include "mp_msg.h"
44 #include "cpudetect.h"
45
46 #ifdef USE_LIBAVCODEC_SO
47 #include <ffmpeg/avcodec.h>
48 #include <ffmpeg/dsputil.h>
49 #else
50 #include "libavcodec/avcodec.h"
51 #include "libavcodec/dsputil.h"
52 #endif
53
54 #ifdef HAVE_MALLOC_H
55 #include <malloc.h>
56 #endif
57
58 #include "img_format.h"
59 #include "mp_image.h"
60 #include "vf.h"
61 #include "libvo/fastmemcpy.h"
62
63 //===========================================================================//
64 #define BLOCKSZ 12
65
66 static const short custom_threshold[64]=
67 // values (296) can't be too high
68 // -it causes too big quant dependence
69 // or maybe overflow(check), which results in some flashing
70 { 71, 296, 295, 237, 71, 40, 38, 19,
71 245, 193, 185, 121, 102, 73, 53, 27,
72 158, 129, 141, 107, 97, 73, 50, 26,
73 102, 116, 109, 98, 82, 66, 45, 23,
74 71, 94, 95, 81, 70, 56, 38, 20,
75 56, 77, 74, 66, 56, 44, 30, 15,
76 38, 53, 50, 45, 38, 30, 21, 11,
77 20, 27, 26, 23, 20, 15, 11, 5
78 };
79
80 static const uint8_t __attribute__((aligned(32))) dither[8][8]={
81 { 0, 48, 12, 60, 3, 51, 15, 63, },
82 { 32, 16, 44, 28, 35, 19, 47, 31, },
83 { 8, 56, 4, 52, 11, 59, 7, 55, },
84 { 40, 24, 36, 20, 43, 27, 39, 23, },
85 { 2, 50, 14, 62, 1, 49, 13, 61, },
86 { 34, 18, 46, 30, 33, 17, 45, 29, },
87 { 10, 58, 6, 54, 9, 57, 5, 53, },
88 { 42, 26, 38, 22, 41, 25, 37, 21, },
89 };
90
91 struct vf_priv_s { //align 16 !
92 uint64_t threshold_mtx_noq[8*2];
93 uint64_t threshold_mtx[8*2];//used in both C & MMX (& later SSE2) versions
94
95 int log2_count;
96 int temp_stride;
97 int qp;
98 int mpeg2;
99 int prev_q;
100 uint8_t *src;
101 int16_t *temp;
102 //int mode;
103 };
104
105
106 #ifndef HAVE_MMX
107
108 //This func reads from 1 slice, 1 and clears 0 & 1
109 static void store_slice_c(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale)
110 {int y, x;
111 #define STORE(pos) \
112 temp= (src[x + pos] + (d[pos]>>log2_scale))>>(6-log2_scale); \
113 src[x + pos]=src[x + pos - 8*src_stride]=0; \
114 if(temp & 0x100) temp= ~(temp>>31); \
115 dst[x + pos]= temp;
116
117 for(y=0; y<height; y++){
118 const uint8_t *d= dither[y];
119 for(x=0; x<width; x+=8){
120 int temp;
121 STORE(0);
122 STORE(1);
123 STORE(2);
124 STORE(3);
125 STORE(4);
126 STORE(5);
127 STORE(6);
128 STORE(7);
129 }
130 src+=src_stride;
131 dst+=dst_stride;
132 }
133 }
134
135 //This func reads from 2 slices, 0 & 2 and clears 2-nd
136 static void store_slice2_c(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale)
137 {int y, x;
138 #define STORE2(pos) \
139 temp= (src[x + pos] + src[x + pos + 16*src_stride] + (d[pos]>>log2_scale))>>(6-log2_scale); \
140 src[x + pos + 16*src_stride]=0; \
141 if(temp & 0x100) temp= ~(temp>>31); \
142 dst[x + pos]= temp;
143
144 for(y=0; y<height; y++){
145 const uint8_t *d= dither[y];
146 for(x=0; x<width; x+=8){
147 int temp;
148 STORE2(0);
149 STORE2(1);
150 STORE2(2);
151 STORE2(3);
152 STORE2(4);
153 STORE2(5);
154 STORE2(6);
155 STORE2(7);
156 }
157 src+=src_stride;
158 dst+=dst_stride;
159 }
160 }
161
162 static void mul_thrmat_c(struct vf_priv_s *p,int q)
163 {
164 int a;
165 for(a=0;a<64;a++)
166 ((short*)p->threshold_mtx)[a]=q * ((short*)p->threshold_mtx_noq)[a];//ints faster in C
167 }
168
169 static void column_fidct_c(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt);
170 static void row_idct_c(DCTELEM* workspace,
171 int16_t* output_adr, int output_stride, int cnt);
172 static void row_fdct_c(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt);
173
174 //this is rather ugly, but there is no need for function pointers
175 #define store_slice_s store_slice_c
176 #define store_slice2_s store_slice2_c
177 #define mul_thrmat_s mul_thrmat_c
178 #define column_fidct_s column_fidct_c
179 #define row_idct_s row_idct_c
180 #define row_fdct_s row_fdct_c
181
182 #else /* HAVE_MMX */
183
184 //This func reads from 1 slice, 1 and clears 0 & 1
185 static void store_slice_mmx(uint8_t *dst, int16_t *src, long dst_stride, long src_stride, long width, long height, long log2_scale)
186 {
187 const uint8_t *od=&dither[0][0];
188 const uint8_t *end=&dither[height][0];
189 width = (width+7)&~7;
190 dst_stride-=width;
191 //src_stride=(src_stride-width)*2;
192 asm volatile(
193 "mov %5, %%"REG_d" \n\t"
194 "mov %6, %%"REG_S" \n\t"
195 "mov %7, %%"REG_D" \n\t"
196 "mov %1, %%"REG_a" \n\t"
197 "movd %%"REG_d", %%mm5 \n\t"
198 "xor $-1, %%"REG_d" \n\t"
199 "mov %%"REG_a", %%"REG_c" \n\t"
200 "add $7, %%"REG_d" \n\t"
201 "neg %%"REG_a" \n\t"
202 "sub %0, %%"REG_c" \n\t"
203 "add %%"REG_c", %%"REG_c" \n\t"
204 "movd %%"REG_d", %%mm2 \n\t"
205 "mov %%"REG_c", %1 \n\t"
206 "mov %2, %%"REG_d" \n\t"
207 "shl $4, %%"REG_a" \n\t"
208
209 "2: \n\t"
210 "movq (%%"REG_d"), %%mm3 \n\t"
211 "movq %%mm3, %%mm4 \n\t"
212 "pxor %%mm7, %%mm7 \n\t"
213 "punpcklbw %%mm7, %%mm3 \n\t"
214 "punpckhbw %%mm7, %%mm4 \n\t"
215 "mov %0, %%"REG_c" \n\t"
216 "psraw %%mm5, %%mm3 \n\t"
217 "psraw %%mm5, %%mm4 \n\t"
218 "1: \n\t"
219 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t"
220 "movq (%%"REG_S"), %%mm0 \n\t"
221 "movq 8(%%"REG_S"), %%mm1 \n\t"
222
223 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t"
224 "paddw %%mm3, %%mm0 \n\t"
225 "paddw %%mm4, %%mm1 \n\t"
226
227 "movq %%mm7, (%%"REG_S") \n\t"
228 "psraw %%mm2, %%mm0 \n\t"
229 "psraw %%mm2, %%mm1 \n\t"
230
231 "movq %%mm7, 8(%%"REG_S") \n\t"
232 "packuswb %%mm1, %%mm0 \n\t"
233 "add $16, %%"REG_S" \n\t"
234
235 "movq %%mm0, (%%"REG_D") \n\t"
236 "add $8, %%"REG_D" \n\t"
237 "sub $8, %%"REG_c" \n\t"
238 "jg 1b \n\t"
239 "add %1, %%"REG_S" \n\t"
240 "add $8, %%"REG_d" \n\t"
241 "add %3, %%"REG_D" \n\t"
242 "cmp %4, %%"REG_d" \n\t"
243 "jl 2b \n\t"
244
245 :
246 : "m" (width), "m" (src_stride), "g" (od), "m" (dst_stride), "g" (end),
247 "m" (log2_scale), "m" (src), "m" (dst) //input
248 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
249 );
250 }
251
252 //This func reads from 2 slices, 0 & 2 and clears 2-nd
253 static void store_slice2_mmx(uint8_t *dst, int16_t *src, long dst_stride, long src_stride, long width, long height, long log2_scale)
254 {
255 const uint8_t *od=&dither[0][0];
256 const uint8_t *end=&dither[height][0];
257 width = (width+7)&~7;
258 dst_stride-=width;
259 //src_stride=(src_stride-width)*2;
260 asm volatile(
261 "mov %5, %%"REG_d" \n\t"
262 "mov %6, %%"REG_S" \n\t"
263 "mov %7, %%"REG_D" \n\t"
264 "mov %1, %%"REG_a" \n\t"
265 "movd %%"REG_d", %%mm5 \n\t"
266 "xor $-1, %%"REG_d" \n\t"
267 "mov %%"REG_a", %%"REG_c" \n\t"
268 "add $7, %%"REG_d" \n\t"
269 "sub %0, %%"REG_c" \n\t"
270 "add %%"REG_c", %%"REG_c" \n\t"
271 "movd %%"REG_d", %%mm2 \n\t"
272 "mov %%"REG_c", %1 \n\t"
273 "mov %2, %%"REG_d" \n\t"
274 "shl $5, %%"REG_a" \n\t"
275
276 "2: \n\t"
277 "movq (%%"REG_d"), %%mm3 \n\t"
278 "movq %%mm3, %%mm4 \n\t"
279 "pxor %%mm7, %%mm7 \n\t"
280 "punpcklbw %%mm7, %%mm3 \n\t"
281 "punpckhbw %%mm7, %%mm4 \n\t"
282 "mov %0, %%"REG_c" \n\t"
283 "psraw %%mm5, %%mm3 \n\t"
284 "psraw %%mm5, %%mm4 \n\t"
285 "1: \n\t"
286 "movq (%%"REG_S"), %%mm0 \n\t"
287 "movq 8(%%"REG_S"), %%mm1 \n\t"
288 "paddw %%mm3, %%mm0 \n\t"
289
290 "paddw (%%"REG_S",%%"REG_a",), %%mm0 \n\t"
291 "paddw %%mm4, %%mm1 \n\t"
292 "movq 8(%%"REG_S",%%"REG_a",), %%mm6 \n\t"
293
294 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t"
295 "psraw %%mm2, %%mm0 \n\t"
296 "paddw %%mm6, %%mm1 \n\t"
297
298 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t"
299 "psraw %%mm2, %%mm1 \n\t"
300 "packuswb %%mm1, %%mm0 \n\t"
301
302 "movq %%mm0, (%%"REG_D") \n\t"
303 "add $16, %%"REG_S" \n\t"
304 "add $8, %%"REG_D" \n\t"
305 "sub $8, %%"REG_c" \n\t"
306 "jg 1b \n\t"
307 "add %1, %%"REG_S" \n\t"
308 "add $8, %%"REG_d" \n\t"
309 "add %3, %%"REG_D" \n\t"
310 "cmp %4, %%"REG_d" \n\t"
311 "jl 2b \n\t"
312
313 :
314 : "m" (width), "m" (src_stride), "g" (od), "m" (dst_stride), "g" (end),
315 "m" (log2_scale), "m" (src), "m" (dst) //input
316 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_D, "%"REG_S
317 );
318 }
319
320 static void mul_thrmat_mmx(struct vf_priv_s *p, int q)
321 {
322 uint64_t *adr=&p->threshold_mtx_noq[0];
323 asm volatile(
324 "movd %0, %%mm7 \n\t"
325 "add $8*8*2, %%"REG_D" \n\t"
326 "movq 0*8(%%"REG_S"), %%mm0 \n\t"
327 "punpcklwd %%mm7, %%mm7 \n\t"
328 "movq 1*8(%%"REG_S"), %%mm1 \n\t"
329 "punpckldq %%mm7, %%mm7 \n\t"
330 "pmullw %%mm7, %%mm0 \n\t"
331
332 "movq 2*8(%%"REG_S"), %%mm2 \n\t"
333 "pmullw %%mm7, %%mm1 \n\t"
334
335 "movq 3*8(%%"REG_S"), %%mm3 \n\t"
336 "pmullw %%mm7, %%mm2 \n\t"
337
338 "movq %%mm0, 0*8(%%"REG_D") \n\t"
339 "movq 4*8(%%"REG_S"), %%mm4 \n\t"
340 "pmullw %%mm7, %%mm3 \n\t"
341
342 "movq %%mm1, 1*8(%%"REG_D") \n\t"
343 "movq 5*8(%%"REG_S"), %%mm5 \n\t"
344 "pmullw %%mm7, %%mm4 \n\t"
345
346 "movq %%mm2, 2*8(%%"REG_D") \n\t"
347 "movq 6*8(%%"REG_S"), %%mm6 \n\t"
348 "pmullw %%mm7, %%mm5 \n\t"
349
350 "movq %%mm3, 3*8(%%"REG_D") \n\t"
351 "movq 7*8+0*8(%%"REG_S"), %%mm0 \n\t"
352 "pmullw %%mm7, %%mm6 \n\t"
353
354 "movq %%mm4, 4*8(%%"REG_D") \n\t"
355 "movq 7*8+1*8(%%"REG_S"), %%mm1 \n\t"
356 "pmullw %%mm7, %%mm0 \n\t"
357
358 "movq %%mm5, 5*8(%%"REG_D") \n\t"
359 "movq 7*8+2*8(%%"REG_S"), %%mm2 \n\t"
360 "pmullw %%mm7, %%mm1 \n\t"
361
362 "movq %%mm6, 6*8(%%"REG_D") \n\t"
363 "movq 7*8+3*8(%%"REG_S"), %%mm3 \n\t"
364 "pmullw %%mm7, %%mm2 \n\t"
365
366 "movq %%mm0, 7*8+0*8(%%"REG_D") \n\t"
367 "movq 7*8+4*8(%%"REG_S"), %%mm4 \n\t"
368 "pmullw %%mm7, %%mm3 \n\t"
369
370 "movq %%mm1, 7*8+1*8(%%"REG_D") \n\t"
371 "movq 7*8+5*8(%%"REG_S"), %%mm5 \n\t"
372 "pmullw %%mm7, %%mm4 \n\t"
373
374 "movq %%mm2, 7*8+2*8(%%"REG_D") \n\t"
375 "movq 7*8+6*8(%%"REG_S"), %%mm6 \n\t"
376 "pmullw %%mm7, %%mm5 \n\t"
377
378 "movq %%mm3, 7*8+3*8(%%"REG_D") \n\t"
379 "movq 14*8+0*8(%%"REG_S"), %%mm0 \n\t"
380 "pmullw %%mm7, %%mm6 \n\t"
381
382 "movq %%mm4, 7*8+4*8(%%"REG_D") \n\t"
383 "movq 14*8+1*8(%%"REG_S"), %%mm1 \n\t"
384 "pmullw %%mm7, %%mm0 \n\t"
385
386 "movq %%mm5, 7*8+5*8(%%"REG_D") \n\t"
387 "pmullw %%mm7, %%mm1 \n\t"
388
389 "movq %%mm6, 7*8+6*8(%%"REG_D") \n\t"
390 "movq %%mm0, 14*8+0*8(%%"REG_D") \n\t"
391 "movq %%mm1, 14*8+1*8(%%"REG_D") \n\t"
392
393 : "+g" (q), "+S" (adr), "+D" (adr)
394 :
395 );
396 }
397
398 static void column_fidct_mmx(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt);
399 static void row_idct_mmx(DCTELEM* workspace,
400 int16_t* output_adr, int output_stride, int cnt);
401 static void row_fdct_mmx(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt);
402
403 #define store_slice_s store_slice_mmx
404 #define store_slice2_s store_slice2_mmx
405 #define mul_thrmat_s mul_thrmat_mmx
406 #define column_fidct_s column_fidct_mmx
407 #define row_idct_s row_idct_mmx
408 #define row_fdct_s row_fdct_mmx
409 #endif // HAVE_MMX
410
411 static void filter(struct vf_priv_s *p, uint8_t *dst, uint8_t *src,
412 int dst_stride, int src_stride,
413 int width, int height,
414 uint8_t *qp_store, int qp_stride, int is_luma)
415 {
416 int x, x0, y, es, qy, t;
417 const int stride= is_luma ? p->temp_stride : (width+16);//((width+16+15)&(~15))
418 const int step=6-p->log2_count;
419 const int qps= 3 + is_luma;
420 int32_t __attribute__((aligned(32))) block_align[4*8*BLOCKSZ+ 4*8*BLOCKSZ];
421 DCTELEM *block= (DCTELEM *)block_align;
422 DCTELEM *block3=(DCTELEM *)(block_align+4*8*BLOCKSZ);
423
424 memset(block3, 0, 4*8*BLOCKSZ);
425
426 //p->src=src-src_stride*8-8;//!
427 if (!src || !dst) return; // HACK avoid crash for Y8 colourspace
428 for(y=0; y<height; y++){
429 int index= 8 + 8*stride + y*stride;
430 memcpy(p->src + index, src + y*src_stride, width);//this line can be avoided by using DR & user fr.buffers
431 for(x=0; x<8; x++){
432 p->src[index - x - 1]= p->src[index + x ];
433 p->src[index + width + x ]= p->src[index + width - x - 1];
434 }
435 }
436 for(y=0; y<8; y++){
437 memcpy(p->src + ( 7-y)*stride, p->src + ( y+8)*stride, stride);
438 memcpy(p->src + (height+8+y)*stride, p->src + (height-y+7)*stride, stride);
439 }
440 //FIXME (try edge emu)
441
442 for(y=8; y<24; y++)
443 memset(p->temp+ 8 +y*stride, 0,width*sizeof(int16_t));
444
445 for(y=step; y<height+8; y+=step){ //step= 1,2
446 qy=y-4;
447 if (qy>height-1) qy=height-1;
448 if (qy<0) qy=0;
449 qy=(qy>>qps)*qp_stride;
450 row_fdct_s(block, p->src + y*stride +2-(y&1), stride, 2);
451 for(x0=0; x0<width+8-8*(BLOCKSZ-1); x0+=8*(BLOCKSZ-1)){
452 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, 2*(BLOCKSZ-1));
453 if(p->qp)
454 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+0*8, block3+0*8, 8*(BLOCKSZ-1)); //yes, this is a HOTSPOT
455 else
456 for (x=0; x<8*(BLOCKSZ-1); x+=8) {
457 t=x+x0-2; //correct t=x+x0-2-(y&1), but its the same
458 if (t<0) t=0;//t always < width-2
459 t=qp_store[qy+(t>>qps)];
460 if(p->mpeg2) t>>=1; //copy p->mpeg2,prev_q to locals?
461 if (t!=p->prev_q) p->prev_q=t, mul_thrmat_s(p, t);
462 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+x*8, block3+x*8, 8); //yes, this is a HOTSPOT
463 }
464 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, 2*(BLOCKSZ-1));
465 memcpy(block, block+(BLOCKSZ-1)*64, 8*8*sizeof(DCTELEM)); //cycling
466 memcpy(block3, block3+(BLOCKSZ-1)*64, 6*8*sizeof(DCTELEM));
467 }
468 //
469 es=width+8-x0; // 8, ...
470 if (es>8)
471 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, (es-4)>>2);
472 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block, block3, es-0);
473 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, es>>2);
474 {const int y1=y-8+step;//l5-7 l4-6
475 if (!(y1&7) && y1) {
476 if (y1&8) store_slice_s(dst + (y1-8)*dst_stride, p->temp+ 8 +8*stride,
477 dst_stride, stride, width, 8, 5-p->log2_count);
478 else store_slice2_s(dst + (y1-8)*dst_stride, p->temp+ 8 +0*stride,
479 dst_stride, stride, width, 8, 5-p->log2_count);
480 } }
481 }
482
483 if (y&7) { // == height & 7
484 if (y&8) store_slice_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +8*stride,
485 dst_stride, stride, width, y&7, 5-p->log2_count);
486 else store_slice2_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +0*stride,
487 dst_stride, stride, width, y&7, 5-p->log2_count);
488 }
489 }
490
491 static int config(struct vf_instance_s* vf,
492 int width, int height, int d_width, int d_height,
493 unsigned int flags, unsigned int outfmt)
494 {
495 int h= (height+16+15)&(~15);
496
497 vf->priv->temp_stride= (width+16+15)&(~15);
498 vf->priv->temp= (int16_t*)av_mallocz(vf->priv->temp_stride*3*8*sizeof(int16_t));
499 //this can also be avoided, see above
500 vf->priv->src = (uint8_t*)av_malloc(vf->priv->temp_stride*h*sizeof(uint8_t));
501
502 return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt);
503 }
504
505 static void get_image(struct vf_instance_s* vf, mp_image_t *mpi)
506 {
507 if(mpi->flags&MP_IMGFLAG_PRESERVE) return; // don't change
508 // ok, we can do pp in-place (or pp disabled):
509 vf->dmpi=vf_get_image(vf->next,mpi->imgfmt,
510 mpi->type, mpi->flags, mpi->width, mpi->height);
511 mpi->planes[0]=vf->dmpi->planes[0];
512 mpi->stride[0]=vf->dmpi->stride[0];
513 mpi->width=vf->dmpi->width;
514 if(mpi->flags&MP_IMGFLAG_PLANAR){
515 mpi->planes[1]=vf->dmpi->planes[1];
516 mpi->planes[2]=vf->dmpi->planes[2];
517 mpi->stride[1]=vf->dmpi->stride[1];
518 mpi->stride[2]=vf->dmpi->stride[2];
519 }
520 mpi->flags|=MP_IMGFLAG_DIRECT;
521 }
522
523 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi)
524 {
525 mp_image_t *dmpi;
526 if(!(mpi->flags&MP_IMGFLAG_DIRECT)){
527 // no DR, so get a new image! hope we'll get DR buffer:
528 dmpi=vf_get_image(vf->next,mpi->imgfmt,
529 MP_IMGTYPE_TEMP,
530 MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE,
531 mpi->width,mpi->height);
532 vf_clone_mpi_attributes(dmpi, mpi);
533 }else{
534 dmpi=vf->dmpi;
535 }
536
537 vf->priv->mpeg2= mpi->qscale_type;
538 if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){
539 if(mpi->qscale || vf->priv->qp){
540 filter(vf->priv, dmpi->planes[0], mpi->planes[0], dmpi->stride[0], mpi->stride[0], mpi->w, mpi->h, mpi->qscale, mpi->qstride, 1);
541 filter(vf->priv, dmpi->planes[1], mpi->planes[1], dmpi->stride[1], mpi->stride[1], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, mpi->qscale, mpi->qstride, 0);
542 filter(vf->priv, dmpi->planes[2], mpi->planes[2], dmpi->stride[2], mpi->stride[2], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, mpi->qscale, mpi->qstride, 0);
543 }else{
544 memcpy_pic(dmpi->planes[0], mpi->planes[0], mpi->w, mpi->h, dmpi->stride[0], mpi->stride[0]);
545 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]);
546 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]);
547 }
548 }
549
550 #ifdef HAVE_MMX
551 if(gCpuCaps.hasMMX) asm volatile ("emms\n\t");
552 #endif
553 #ifdef HAVE_MMX2
554 if(gCpuCaps.hasMMX2) asm volatile ("sfence\n\t");
555 #endif
556 return vf_next_put_image(vf,dmpi);
557 }
558
559 static void uninit(struct vf_instance_s* vf)
560 {
561 if(!vf->priv) return;
562
563 if(vf->priv->temp) av_free(vf->priv->temp);
564 vf->priv->temp= NULL;
565 if(vf->priv->src) av_free(vf->priv->src);
566 vf->priv->src= NULL;
567 //if(vf->priv->avctx) free(vf->priv->avctx);
568 //vf->priv->avctx= NULL;
569
570 av_free(vf->priv);
571 vf->priv=NULL;
572 }
573
574 //===========================================================================//
575
576 static int query_format(struct vf_instance_s* vf, unsigned int fmt)
577 {
578 switch(fmt){
579 case IMGFMT_YVU9:
580 case IMGFMT_IF09:
581 case IMGFMT_YV12:
582 case IMGFMT_I420:
583 case IMGFMT_IYUV:
584 case IMGFMT_CLPL:
585 case IMGFMT_Y800:
586 case IMGFMT_Y8:
587 case IMGFMT_444P:
588 case IMGFMT_422P:
589 case IMGFMT_411P:
590 return vf_next_query_format(vf,fmt);
591 }
592 return 0;
593 }
594
595 /*
596 static unsigned int fmt_list[]={
597 IMGFMT_YVU9,
598 IMGFMT_IF09,
599 IMGFMT_YV12,
600 IMGFMT_I420,
601 IMGFMT_IYUV,
602 IMGFMT_CLPL,
603 IMGFMT_Y800,
604 IMGFMT_Y8,
605 IMGFMT_444P,
606 IMGFMT_422P,
607 IMGFMT_411P,
608 0
609 };
610 */
611
612 static int control(struct vf_instance_s* vf, int request, void* data)
613 {
614 switch(request){
615 case VFCTRL_QUERY_MAX_PP_LEVEL:
616 return 5;
617 case VFCTRL_SET_PP_LEVEL:
618 vf->priv->log2_count= *((unsigned int*)data);
619 if (vf->priv->log2_count < 4) vf->priv->log2_count=4;
620 return CONTROL_TRUE;
621 }
622 return vf_next_control(vf,request,data);
623 }
624
625 static int open(vf_instance_t *vf, char* args)
626 {
627 int i=0, bias;
628 int custom_threshold_m[64];
629 int log2c=-1;
630
631 vf->config=config;
632 vf->put_image=put_image;
633 vf->get_image=get_image;
634 vf->query_format=query_format;
635 vf->uninit=uninit;
636 vf->control= control;
637 vf->priv=av_mallocz(sizeof(struct vf_priv_s));//assumes align 16 !
638
639 avcodec_init();
640
641 //vf->priv->avctx= avcodec_alloc_context();
642 //dsputil_init(&vf->priv->dsp, vf->priv->avctx);
643
644 vf->priv->log2_count= 4;
645
646 if (args) sscanf(args, "%d:%d:%d", &log2c, &vf->priv->qp, &i);
647
648 if( log2c >=4 && log2c <=5 )
649 vf->priv->log2_count = log2c;
650 else if( log2c >= 6 )
651 vf->priv->log2_count = 5;
652
653 if(vf->priv->qp < 0)
654 vf->priv->qp = 0;
655
656 if (i < -15) i = -15;
657 if (i > 32) i = 32;
658
659 bias= (1<<4)+i; //regulable
660 vf->priv->prev_q=0;
661 //
662 for(i=0;i<64;i++) //FIXME: tune custom_threshold[] and remove this !
663 custom_threshold_m[i]=(int)(custom_threshold[i]*(bias/71.)+ 0.5);
664 for(i=0;i<8;i++){
665 vf->priv->threshold_mtx_noq[2*i]=(uint64_t)custom_threshold_m[i*8+2]
666 |(((uint64_t)custom_threshold_m[i*8+6])<<16)
667 |(((uint64_t)custom_threshold_m[i*8+0])<<32)
668 |(((uint64_t)custom_threshold_m[i*8+4])<<48);
669 vf->priv->threshold_mtx_noq[2*i+1]=(uint64_t)custom_threshold_m[i*8+5]
670 |(((uint64_t)custom_threshold_m[i*8+3])<<16)
671 |(((uint64_t)custom_threshold_m[i*8+1])<<32)
672 |(((uint64_t)custom_threshold_m[i*8+7])<<48);
673 }
674
675 if (vf->priv->qp) vf->priv->prev_q=vf->priv->qp, mul_thrmat_s(vf->priv, vf->priv->qp);
676
677 return 1;
678 }
679
680 vf_info_t vf_info_fspp = {
681 "fast simple postprocess",
682 "fspp",
683 "Michael Niedermayer, Nikolaj Poroshin",
684 "",
685 open,
686 NULL
687 };
688
689 //====================================================================
690 //Specific spp's dct, idct and threshold functions
691 //I'd prefer to have them in the separate file.
692
693 #include "mangle.h"
694 //#define MANGLE(a) #a
695
696 //typedef int16_t DCTELEM; //! only int16_t
697
698 #define DCTSIZE 8
699 #define DCTSIZE_S "8"
700
701 #define FIX(x,s) ((int) ((x) * (1<<s) + 0.5)&0xffff)
702 #define C64(x) ((uint64_t)((x)|(x)<<16))<<32 | (uint64_t)(x) | (uint64_t)(x)<<16
703 #define FIX64(x,s) C64(FIX(x,s))
704
705 #define MULTIPLY16H(x,k) (((x)*(k))>>16)
706 #define THRESHOLD(r,x,t) if(((unsigned)((x)+t))>t*2) r=(x);else r=0;
707 #define DESCALE(x,n) (((x) + (1 << ((n)-1))) >> n)
708
709 #ifdef HAVE_MMX
710
711 static uint64_t attribute_used __attribute__((aligned(8))) temps[4];//!!
712
713 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_382683433=FIX64(0.382683433, 14);
714 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_541196100=FIX64(0.541196100, 14);
715 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_707106781=FIX64(0.707106781, 14);
716 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_306562965=FIX64(1.306562965, 14);
717
718 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_414213562_A=FIX64(1.414213562, 14);
719
720 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_847759065=FIX64(1.847759065, 13);
721 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_2_613125930=FIX64(-2.613125930, 13); //-
722 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_414213562=FIX64(1.414213562, 13);
723 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_082392200=FIX64(1.082392200, 13);
724 //for t3,t5,t7 == 0 shortcut
725 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_847759065=FIX64(0.847759065, 14);
726 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_566454497=FIX64(0.566454497, 14);
727 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_198912367=FIX64(0.198912367, 14);
728
729 static uint64_t attribute_used __attribute__((aligned(8))) MM_DESCALE_RND=C64(4);
730 static uint64_t attribute_used __attribute__((aligned(8))) MM_2=C64(2);
731
732 #else /* !HAVE_MMX */
733
734 typedef int32_t int_simd16_t;
735 static int16_t FIX_0_382683433=FIX(0.382683433, 14);
736 static int16_t FIX_0_541196100=FIX(0.541196100, 14);
737 static int16_t FIX_0_707106781=FIX(0.707106781, 14);
738 static int16_t FIX_1_306562965=FIX(1.306562965, 14);
739 static int16_t FIX_1_414213562_A=FIX(1.414213562, 14);
740 static int16_t FIX_1_847759065=FIX(1.847759065, 13);
741 static int16_t FIX_2_613125930=FIX(-2.613125930, 13); //-
742 static int16_t FIX_1_414213562=FIX(1.414213562, 13);
743 static int16_t FIX_1_082392200=FIX(1.082392200, 13);
744
745 #endif
746
747 #ifndef HAVE_MMX
748
749 static void column_fidct_c(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt)
750 {
751 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
752 int_simd16_t tmp10, tmp11, tmp12, tmp13;
753 int_simd16_t z1,z2,z3,z4,z5, z10, z11, z12, z13;
754 int_simd16_t d0, d1, d2, d3, d4, d5, d6, d7;
755
756 DCTELEM* dataptr;
757 DCTELEM* wsptr;
758 int16_t *threshold;
759 int ctr;
760
761 dataptr = data;
762 wsptr = output;
763
764 for (; cnt > 0; cnt-=2) { //start positions
765 threshold=(int16_t*)thr_adr;//threshold_mtx
766 for (ctr = DCTSIZE; ctr > 0; ctr--) {
767 // Process columns from input, add to output.
768 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
769 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
770
771 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
772 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
773
774 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
775 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
776
777 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
778 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
779
780 // Even part of FDCT
781
782 tmp10 = tmp0 + tmp3;
783 tmp13 = tmp0 - tmp3;
784 tmp11 = tmp1 + tmp2;
785 tmp12 = tmp1 - tmp2;
786
787 d0 = tmp10 + tmp11;
788 d4 = tmp10 - tmp11;
789
790 z1 = MULTIPLY16H((tmp12 + tmp13) <<2, FIX_0_707106781);
791 d2 = tmp13 + z1;
792 d6 = tmp13 - z1;
793
794 // Even part of IDCT
795
796 THRESHOLD(tmp0, d0, threshold[0*8]);
797 THRESHOLD(tmp1, d2, threshold[2*8]);
798 THRESHOLD(tmp2, d4, threshold[4*8]);
799 THRESHOLD(tmp3, d6, threshold[6*8]);
800 tmp0+=2;
801 tmp10 = (tmp0 + tmp2)>>2;
802 tmp11 = (tmp0 - tmp2)>>2;
803
804 tmp13 = (tmp1 + tmp3)>>2; //+2 ! (psnr decides)
805 tmp12 = MULTIPLY16H((tmp1 - tmp3), FIX_1_414213562_A) - tmp13; //<<2
806
807 tmp0 = tmp10 + tmp13; //->temps
808 tmp3 = tmp10 - tmp13; //->temps
809 tmp1 = tmp11 + tmp12; //->temps
810 tmp2 = tmp11 - tmp12; //->temps
811
812 // Odd part of FDCT
813
814 tmp10 = tmp4 + tmp5;
815 tmp11 = tmp5 + tmp6;
816 tmp12 = tmp6 + tmp7;
817
818 z5 = MULTIPLY16H((tmp10 - tmp12)<<2, FIX_0_382683433);
819 z2 = MULTIPLY16H(tmp10 <<2, FIX_0_541196100) + z5;
820 z4 = MULTIPLY16H(tmp12 <<2, FIX_1_306562965) + z5;
821 z3 = MULTIPLY16H(tmp11 <<2, FIX_0_707106781);
822
823 z11 = tmp7 + z3;
824 z13 = tmp7 - z3;
825
826 d5 = z13 + z2;
827 d3 = z13 - z2;
828 d1 = z11 + z4;
829 d7 = z11 - z4;
830
831 // Odd part of IDCT
832
833 THRESHOLD(tmp4, d1, threshold[1*8]);
834 THRESHOLD(tmp5, d3, threshold[3*8]);
835 THRESHOLD(tmp6, d5, threshold[5*8]);
836 THRESHOLD(tmp7, d7, threshold[7*8]);
837
838 //Simd version uses here a shortcut for the tmp5,tmp6,tmp7 == 0
839 z13 = tmp6 + tmp5;
840 z10 = (tmp6 - tmp5)<<1;
841 z11 = tmp4 + tmp7;
842 z12 = (tmp4 - tmp7)<<1;
843
844 tmp7 = (z11 + z13)>>2; //+2 !
845 tmp11 = MULTIPLY16H((z11 - z13)<<1, FIX_1_414213562);
846 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
847 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
848 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - !!
849
850 tmp6 = tmp12 - tmp7;
851 tmp5 = tmp11 - tmp6;
852 tmp4 = tmp10 + tmp5;
853
854 wsptr[DCTSIZE*0]+= (tmp0 + tmp7);
855 wsptr[DCTSIZE*1]+= (tmp1 + tmp6);
856 wsptr[DCTSIZE*2]+= (tmp2 + tmp5);
857 wsptr[DCTSIZE*3]+= (tmp3 - tmp4);
858 wsptr[DCTSIZE*4]+= (tmp3 + tmp4);
859 wsptr[DCTSIZE*5]+= (tmp2 - tmp5);
860 wsptr[DCTSIZE*6]= (tmp1 - tmp6);
861 wsptr[DCTSIZE*7]= (tmp0 - tmp7);
862 //
863 dataptr++; //next column
864 wsptr++;
865 threshold++;
866 }
867 dataptr+=8; //skip each second start pos
868 wsptr +=8;
869 }
870 }
871
872 #else /* HAVE_MMX */
873
874 static void column_fidct_mmx(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt)
875 {
876 asm volatile(
877 ".align 16 \n\t"
878 "1: \n\t"
879 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
880 //
881 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
882 "movq %%mm1, %%mm0 \n\t"
883
884 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0
885 "movq %%mm7, %%mm3 \n\t"
886
887 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
888 "movq %%mm1, %%mm5 \n\t"
889
890 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
891 "psubw %%mm7, %%mm1 \n\t" //t13
892
893 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
894 "movq %%mm6, %%mm4 \n\t"
895
896 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
897 "paddw %%mm7, %%mm5 \n\t" //t10
898
899 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
900 "movq %%mm6, %%mm7 \n\t"
901
902 "paddw %%mm2, %%mm6 \n\t" //t11
903 "psubw %%mm2, %%mm7 \n\t" //t12
904
905 "movq %%mm5, %%mm2 \n\t"
906 "paddw %%mm6, %%mm5 \n\t" //d0
907 // i0 t13 t12 i3 i1 d0 - d4
908 "psubw %%mm6, %%mm2 \n\t" //d4
909 "paddw %%mm1, %%mm7 \n\t"
910
911 "movq 4*16(%%"REG_d"), %%mm6 \n\t"
912 "psllw $2, %%mm7 \n\t"
913
914 "psubw 0*16(%%"REG_d"), %%mm5 \n\t"
915 "psubw %%mm6, %%mm2 \n\t"
916
917 "paddusw 0*16(%%"REG_d"), %%mm5 \n\t"
918 "paddusw %%mm6, %%mm2 \n\t"
919
920 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t"
921 //
922 "paddw 0*16(%%"REG_d"), %%mm5 \n\t"
923 "paddw %%mm6, %%mm2 \n\t"
924
925 "psubusw 0*16(%%"REG_d"), %%mm5 \n\t"
926 "psubusw %%mm6, %%mm2 \n\t"
927
928 //This func is totally compute-bound, operates at huge speed. So, DC shortcut
929 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
930 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
931 "paddw "MANGLE(MM_2)", %%mm5 \n\t"
932 "movq %%mm2, %%mm6 \n\t"
933
934 "paddw %%mm5, %%mm2 \n\t"
935 "psubw %%mm6, %%mm5 \n\t"
936
937 "movq %%mm1, %%mm6 \n\t"
938 "paddw %%mm7, %%mm1 \n\t" //d2
939
940 "psubw 2*16(%%"REG_d"), %%mm1 \n\t"
941 "psubw %%mm7, %%mm6 \n\t" //d6
942
943 "movq 6*16(%%"REG_d"), %%mm7 \n\t"
944 "psraw $2, %%mm5 \n\t"
945
946 "paddusw 2*16(%%"REG_d"), %%mm1 \n\t"
947 "psubw %%mm7, %%mm6 \n\t"
948 // t7 d2 /t11 t4 t6 - d6 /t10
949
950 "paddw 2*16(%%"REG_d"), %%mm1 \n\t"
951 "paddusw %%mm7, %%mm6 \n\t"
952
953 "psubusw 2*16(%%"REG_d"), %%mm1 \n\t"
954 "paddw %%mm7, %%mm6 \n\t"
955
956 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t"
957 "psubusw %%mm7, %%mm6 \n\t"
958
959 //movq [edi+"DCTSIZE_S"*2*2], mm1
960 //movq [edi+"DCTSIZE_S"*6*2], mm6
961 "movq %%mm1, %%mm7 \n\t"
962 "psraw $2, %%mm2 \n\t"
963
964 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t"
965 "psubw %%mm6, %%mm1 \n\t"
966
967 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t"
968 "paddw %%mm7, %%mm6 \n\t" //'t13
969
970 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
971 "movq %%mm2, %%mm7 \n\t"
972
973 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
974 "paddw %%mm6, %%mm2 \n\t" //'t0
975
976 "movq %%mm2, "MANGLE(temps)"+0*8 \n\t" //!
977 "psubw %%mm6, %%mm7 \n\t" //'t3
978
979 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
980 "psubw %%mm6, %%mm1 \n\t" //'t12
981
982 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5
983 "movq %%mm5, %%mm6 \n\t"
984
985 "movq %%mm7, "MANGLE(temps)"+3*8 \n\t"
986 "paddw %%mm2, %%mm3 \n\t" //t10
987
988 "paddw %%mm4, %%mm2 \n\t" //t11
989 "paddw %%mm0, %%mm4 \n\t" //t12
990
991 "movq %%mm3, %%mm7 \n\t"
992 "psubw %%mm4, %%mm3 \n\t"
993
994 "psllw $2, %%mm3 \n\t"
995 "psllw $2, %%mm7 \n\t" //opt for P6
996
997 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
998 "psllw $2, %%mm4 \n\t"
999
1000 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t"
1001 "psllw $2, %%mm2 \n\t"
1002
1003 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
1004 "paddw %%mm1, %%mm5 \n\t" //'t1
1005
1006 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t"
1007 "psubw %%mm1, %%mm6 \n\t" //'t2
1008 // t7 't12 't11 t4 t6 - 't13 't10 ---
1009
1010 "paddw %%mm3, %%mm7 \n\t" //z2
1011
1012 "movq %%mm5, "MANGLE(temps)"+1*8 \n\t"
1013 "paddw %%mm3, %%mm4 \n\t" //z4
1014
1015 "movq 3*16(%%"REG_d"), %%mm3 \n\t"
1016 "movq %%mm0, %%mm1 \n\t"
1017
1018 "movq %%mm6, "MANGLE(temps)"+2*8 \n\t"
1019 "psubw %%mm2, %%mm1 \n\t" //z13
1020
1021 //===
1022 "paddw %%mm2, %%mm0 \n\t" //z11
1023 "movq %%mm1, %%mm5 \n\t"
1024
1025 "movq 5*16(%%"REG_d"), %%mm2 \n\t"
1026 "psubw %%mm7, %%mm1 \n\t" //d3
1027
1028 "paddw %%mm7, %%mm5 \n\t" //d5
1029 "psubw %%mm3, %%mm1 \n\t"
1030
1031 "movq 1*16(%%"REG_d"), %%mm7 \n\t"
1032 "psubw %%mm2, %%mm5 \n\t"
1033
1034 "movq %%mm0, %%mm6 \n\t"
1035 "paddw %%mm4, %%mm0 \n\t" //d1
1036
1037 "paddusw %%mm3, %%mm1 \n\t"
1038 "psubw %%mm4, %%mm6 \n\t" //d7
1039
1040 // d1 d3 - - - d5 d7 -
1041 "movq 7*16(%%"REG_d"), %%mm4 \n\t"
1042 "psubw %%mm7, %%mm0 \n\t"
1043
1044 "psubw %%mm4, %%mm6 \n\t"
1045 "paddusw %%mm2, %%mm5 \n\t"
1046
1047 "paddusw %%mm4, %%mm6 \n\t"
1048 "paddw %%mm3, %%mm1 \n\t"
1049
1050 "paddw %%mm2, %%mm5 \n\t"
1051 "paddw %%mm4, %%mm6 \n\t"
1052
1053 "psubusw %%mm3, %%mm1 \n\t"
1054 "psubusw %%mm2, %%mm5 \n\t"
1055
1056 "psubusw %%mm4, %%mm6 \n\t"
1057 "movq %%mm1, %%mm4 \n\t"
1058
1059 "por %%mm5, %%mm4 \n\t"
1060 "paddusw %%mm7, %%mm0 \n\t"
1061
1062 "por %%mm6, %%mm4 \n\t"
1063 "paddw %%mm7, %%mm0 \n\t"
1064
1065 "packssdw %%mm4, %%mm4 \n\t"
1066 "psubusw %%mm7, %%mm0 \n\t"
1067
1068 "movd %%mm4, %%"REG_a" \n\t"
1069 "or %%"REG_a", %%"REG_a" \n\t"
1070 "jnz 2f \n\t"
1071 //movq [edi+"DCTSIZE_S"*3*2], mm1
1072 //movq [edi+"DCTSIZE_S"*5*2], mm5
1073 //movq [edi+"DCTSIZE_S"*1*2], mm0
1074 //movq [edi+"DCTSIZE_S"*7*2], mm6
1075 // t4 t5 - - - t6 t7 -
1076 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
1077 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
1078 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t"
1079 "movq %%mm0, %%mm1 \n\t"
1080
1081 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
1082 "movq %%mm1, %%mm2 \n\t"
1083
1084 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
1085 "movq %%mm2, %%mm3 \n\t"
1086
1087 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
1088 "paddw %%mm4, %%mm5 \n\t"
1089
1090 "movq "MANGLE(temps)"+1*8, %%mm6 \n\t"
1091 //paddw mm3, MM_2
1092 "psraw $2, %%mm3 \n\t" //tmp7
1093
1094 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
1095 "psubw %%mm3, %%mm4 \n\t"
1096
1097 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
1098 "paddw %%mm3, %%mm5 \n\t"
1099
1100 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
1101 "paddw %%mm6, %%mm7 \n\t"
1102
1103 "movq "MANGLE(temps)"+2*8, %%mm3 \n\t"
1104 "psubw %%mm0, %%mm6 \n\t"
1105
1106 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
1107 "paddw %%mm0, %%mm7 \n\t"
1108
1109 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
1110 "paddw %%mm3, %%mm4 \n\t"
1111
1112 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
1113 "psubw %%mm1, %%mm3 \n\t"
1114
1115 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
1116 "paddw %%mm1, %%mm4 \n\t"
1117
1118 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
1119 "paddw %%mm3, %%mm5 \n\t"
1120
1121 "movq "MANGLE(temps)"+3*8, %%mm0 \n\t"
1122 "add $8, %%"REG_S" \n\t"
1123
1124 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
1125 "paddw %%mm0, %%mm6 \n\t"
1126
1127 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
1128 "psubw %%mm2, %%mm0 \n\t"
1129
1130 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
1131 "paddw %%mm2, %%mm6 \n\t"
1132
1133 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
1134 "paddw %%mm0, %%mm7 \n\t"
1135
1136 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
1137
1138 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
1139 "add $8, %%"REG_D" \n\t"
1140 "jmp 4f \n\t"
1141
1142 "2: \n\t"
1143 //--- non DC2
1144 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
1145 //psraw mm5, 2
1146 //psraw mm0, 2
1147 //psraw mm6, 2
1148 "movq %%mm5, %%mm3 \n\t"
1149 "psubw %%mm1, %%mm5 \n\t"
1150
1151 "psllw $1, %%mm5 \n\t" //'z10
1152 "paddw %%mm1, %%mm3 \n\t" //'z13
1153
1154 "movq %%mm0, %%mm2 \n\t"
1155 "psubw %%mm6, %%mm0 \n\t"
1156
1157 "movq %%mm5, %%mm1 \n\t"
1158 "psllw $1, %%mm0 \n\t" //'z12
1159
1160 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
1161 "paddw %%mm0, %%mm5 \n\t"
1162
1163 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
1164 "paddw %%mm6, %%mm2 \n\t" //'z11
1165
1166 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
1167 "movq %%mm2, %%mm7 \n\t"
1168
1169 //---
1170 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t"
1171 "psubw %%mm3, %%mm2 \n\t"
1172
1173 "psllw $1, %%mm2 \n\t"
1174 "paddw %%mm3, %%mm7 \n\t" //'t7
1175
1176 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
1177 "movq %%mm4, %%mm6 \n\t"
1178 //paddw mm7, MM_2
1179 "psraw $2, %%mm7 \n\t"
1180
1181 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
1182 "psubw %%mm7, %%mm6 \n\t"
1183
1184 "movq "MANGLE(temps)"+1*8, %%mm3 \n\t"
1185 "paddw %%mm7, %%mm4 \n\t"
1186
1187 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
1188 "paddw %%mm5, %%mm1 \n\t" //'t12
1189
1190 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
1191 "psubw %%mm7, %%mm1 \n\t" //'t6
1192
1193 "movq "MANGLE(temps)"+2*8, %%mm7 \n\t"
1194 "psubw %%mm5, %%mm0 \n\t" //'t10
1195
1196 "movq "MANGLE(temps)"+3*8, %%mm6 \n\t"
1197 "movq %%mm3, %%mm5 \n\t"
1198
1199 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
1200 "psubw %%mm1, %%mm5 \n\t"
1201
1202 "psubw %%mm1, %%mm2 \n\t" //'t5
1203 "paddw %%mm1, %%mm3 \n\t"
1204
1205 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
1206 "movq %%mm7, %%mm4 \n\t"
1207
1208 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
1209 "psubw %%mm2, %%mm4 \n\t"
1210
1211 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
1212 "paddw %%mm2, %%mm7 \n\t"
1213
1214 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
1215 "paddw %%mm2, %%mm0 \n\t" //'t4
1216
1217 // 't4 't6 't5 - - - - 't7
1218 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
1219 "movq %%mm6, %%mm1 \n\t"
1220
1221 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
1222 "psubw %%mm0, %%mm1 \n\t"
1223
1224 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
1225 "paddw %%mm0, %%mm6 \n\t"
1226
1227 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
1228 "add $8, %%"REG_S" \n\t"
1229
1230 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
1231
1232 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
1233 "add $8, %%"REG_D" \n\t"
1234
1235 "4: \n\t"
1236 //=part 2 (the same)===========================================================
1237 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
1238 //
1239 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
1240 "movq %%mm1, %%mm0 \n\t"
1241
1242 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0
1243 "movq %%mm7, %%mm3 \n\t"
1244
1245 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
1246 "movq %%mm1, %%mm5 \n\t"
1247
1248 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
1249 "psubw %%mm7, %%mm1 \n\t" //t13
1250
1251 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
1252 "movq %%mm6, %%mm4 \n\t"
1253
1254 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
1255 "paddw %%mm7, %%mm5 \n\t" //t10
1256
1257 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
1258 "movq %%mm6, %%mm7 \n\t"
1259
1260 "paddw %%mm2, %%mm6 \n\t" //t11
1261 "psubw %%mm2, %%mm7 \n\t" //t12
1262
1263 "movq %%mm5, %%mm2 \n\t"
1264 "paddw %%mm6, %%mm5 \n\t" //d0
1265 // i0 t13 t12 i3 i1 d0 - d4
1266 "psubw %%mm6, %%mm2 \n\t" //d4
1267 "paddw %%mm1, %%mm7 \n\t"
1268
1269 "movq 1*8+4*16(%%"REG_d"), %%mm6 \n\t"
1270 "psllw $2, %%mm7 \n\t"
1271
1272 "psubw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
1273 "psubw %%mm6, %%mm2 \n\t"
1274
1275 "paddusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
1276 "paddusw %%mm6, %%mm2 \n\t"
1277
1278 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t"
1279 //
1280 "paddw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
1281 "paddw %%mm6, %%mm2 \n\t"
1282
1283 "psubusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
1284 "psubusw %%mm6, %%mm2 \n\t"
1285
1286 //This func is totally compute-bound, operates at huge speed. So, DC shortcut
1287 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
1288 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
1289 "paddw "MANGLE(MM_2)", %%mm5 \n\t"
1290 "movq %%mm2, %%mm6 \n\t"
1291
1292 "paddw %%mm5, %%mm2 \n\t"
1293 "psubw %%mm6, %%mm5 \n\t"
1294
1295 "movq %%mm1, %%mm6 \n\t"
1296 "paddw %%mm7, %%mm1 \n\t" //d2
1297
1298 "psubw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
1299 "psubw %%mm7, %%mm6 \n\t" //d6
1300
1301 "movq 1*8+6*16(%%"REG_d"), %%mm7 \n\t"
1302 "psraw $2, %%mm5 \n\t"
1303
1304 "paddusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
1305 "psubw %%mm7, %%mm6 \n\t"
1306 // t7 d2 /t11 t4 t6 - d6 /t10
1307
1308 "paddw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
1309 "paddusw %%mm7, %%mm6 \n\t"
1310
1311 "psubusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
1312 "paddw %%mm7, %%mm6 \n\t"
1313
1314 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t"
1315 "psubusw %%mm7, %%mm6 \n\t"
1316
1317 //movq [edi+"DCTSIZE_S"*2*2], mm1
1318 //movq [edi+"DCTSIZE_S"*6*2], mm6
1319 "movq %%mm1, %%mm7 \n\t"
1320 "psraw $2, %%mm2 \n\t"
1321
1322 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t"
1323 "psubw %%mm6, %%mm1 \n\t"
1324
1325 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t"
1326 "paddw %%mm7, %%mm6 \n\t" //'t13
1327
1328 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
1329 "movq %%mm2, %%mm7 \n\t"
1330
1331 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
1332 "paddw %%mm6, %%mm2 \n\t" //'t0
1333
1334 "movq %%mm2, "MANGLE(temps)"+0*8 \n\t" //!
1335 "psubw %%mm6, %%mm7 \n\t" //'t3
1336
1337 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
1338 "psubw %%mm6, %%mm1 \n\t" //'t12
1339
1340 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5
1341 "movq %%mm5, %%mm6 \n\t"
1342
1343 "movq %%mm7, "MANGLE(temps)"+3*8 \n\t"
1344 "paddw %%mm2, %%mm3 \n\t" //t10
1345
1346 "paddw %%mm4, %%mm2 \n\t" //t11
1347 "paddw %%mm0, %%mm4 \n\t" //t12
1348
1349 "movq %%mm3, %%mm7 \n\t"
1350 "psubw %%mm4, %%mm3 \n\t"
1351
1352 "psllw $2, %%mm3 \n\t"
1353 "psllw $2, %%mm7 \n\t" //opt for P6
1354
1355 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
1356 "psllw $2, %%mm4 \n\t"
1357
1358 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t"
1359 "psllw $2, %%mm2 \n\t"
1360
1361 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
1362 "paddw %%mm1, %%mm5 \n\t" //'t1
1363
1364 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t"
1365 "psubw %%mm1, %%mm6 \n\t" //'t2
1366 // t7 't12 't11 t4 t6 - 't13 't10 ---
1367
1368 "paddw %%mm3, %%mm7 \n\t" //z2
1369
1370 "movq %%mm5, "MANGLE(temps)"+1*8 \n\t"
1371 "paddw %%mm3, %%mm4 \n\t" //z4
1372
1373 "movq 1*8+3*16(%%"REG_d"), %%mm3 \n\t"
1374 "movq %%mm0, %%mm1 \n\t"
1375
1376 "movq %%mm6, "MANGLE(temps)"+2*8 \n\t"
1377 "psubw %%mm2, %%mm1 \n\t" //z13
1378
1379 //===
1380 "paddw %%mm2, %%mm0 \n\t" //z11
1381 "movq %%mm1, %%mm5 \n\t"
1382
1383 "movq 1*8+5*16(%%"REG_d"), %%mm2 \n\t"
1384 "psubw %%mm7, %%mm1 \n\t" //d3
1385
1386 "paddw %%mm7, %%mm5 \n\t" //d5
1387 "psubw %%mm3, %%mm1 \n\t"
1388
1389 "movq 1*8+1*16(%%"REG_d"), %%mm7 \n\t"
1390 "psubw %%mm2, %%mm5 \n\t"
1391
1392 "movq %%mm0, %%mm6 \n\t"
1393 "paddw %%mm4, %%mm0 \n\t" //d1
1394
1395 "paddusw %%mm3, %%mm1 \n\t"
1396 "psubw %%mm4, %%mm6 \n\t" //d7
1397
1398 // d1 d3 - - - d5 d7 -
1399 "movq 1*8+7*16(%%"REG_d"), %%mm4 \n\t"
1400 "psubw %%mm7, %%mm0 \n\t"
1401
1402 "psubw %%mm4, %%mm6 \n\t"
1403 "paddusw %%mm2, %%mm5 \n\t"
1404
1405 "paddusw %%mm4, %%mm6 \n\t"
1406 "paddw %%mm3, %%mm1 \n\t"
1407
1408 "paddw %%mm2, %%mm5 \n\t"
1409 "paddw %%mm4, %%mm6 \n\t"
1410
1411 "psubusw %%mm3, %%mm1 \n\t"
1412 "psubusw %%mm2, %%mm5 \n\t"
1413
1414 "psubusw %%mm4, %%mm6 \n\t"
1415 "movq %%mm1, %%mm4 \n\t"
1416
1417 "por %%mm5, %%mm4 \n\t"
1418 "paddusw %%mm7, %%mm0 \n\t"
1419
1420 "por %%mm6, %%mm4 \n\t"
1421 "paddw %%mm7, %%mm0 \n\t"
1422
1423 "packssdw %%mm4, %%mm4 \n\t"
1424 "psubusw %%mm7, %%mm0 \n\t"
1425
1426 "movd %%mm4, %%"REG_a" \n\t"
1427 "or %%"REG_a", %%"REG_a" \n\t"
1428 "jnz 3f \n\t"
1429 //movq [edi+"DCTSIZE_S"*3*2], mm1
1430 //movq [edi+"DCTSIZE_S"*5*2], mm5
1431 //movq [edi+"DCTSIZE_S"*1*2], mm0
1432 //movq [edi+"DCTSIZE_S"*7*2], mm6
1433 // t4 t5 - - - t6 t7 -
1434 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
1435 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
1436 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t"
1437 "movq %%mm0, %%mm1 \n\t"
1438
1439 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
1440 "movq %%mm1, %%mm2 \n\t"
1441
1442 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
1443 "movq %%mm2, %%mm3 \n\t"
1444
1445 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
1446 "paddw %%mm4, %%mm5 \n\t"
1447
1448 "movq "MANGLE(temps)"+1*8, %%mm6 \n\t"
1449 //paddw mm3, MM_2
1450 "psraw $2, %%mm3 \n\t" //tmp7
1451
1452 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
1453 "psubw %%mm3, %%mm4 \n\t"
1454
1455 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
1456 "paddw %%mm3, %%mm5 \n\t"
1457
1458 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
1459 "paddw %%mm6, %%mm7 \n\t"
1460
1461 "movq "MANGLE(temps)"+2*8, %%mm3 \n\t"
1462 "psubw %%mm0, %%mm6 \n\t"
1463
1464 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
1465 "paddw %%mm0, %%mm7 \n\t"
1466
1467 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
1468 "paddw %%mm3, %%mm4 \n\t"
1469
1470 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
1471 "psubw %%mm1, %%mm3 \n\t"
1472
1473 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
1474 "paddw %%mm1, %%mm4 \n\t"
1475
1476 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
1477 "paddw %%mm3, %%mm5 \n\t"
1478
1479 "movq "MANGLE(temps)"+3*8, %%mm0 \n\t"
1480 "add $24, %%"REG_S" \n\t"
1481
1482 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
1483 "paddw %%mm0, %%mm6 \n\t"
1484
1485 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
1486 "psubw %%mm2, %%mm0 \n\t"
1487
1488 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
1489 "paddw %%mm2, %%mm6 \n\t"
1490
1491 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
1492 "paddw %%mm0, %%mm7 \n\t"
1493
1494 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
1495
1496 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
1497 "add $24, %%"REG_D" \n\t"
1498 "sub $2, %%"REG_c" \n\t"
1499 "jnz 1b \n\t"
1500 "jmp 5f \n\t"
1501
1502 "3: \n\t"
1503 //--- non DC2
1504 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
1505 //psraw mm5, 2
1506 //psraw mm0, 2
1507 //psraw mm6, 2
1508 "movq %%mm5, %%mm3 \n\t"
1509 "psubw %%mm1, %%mm5 \n\t"
1510
1511 "psllw $1, %%mm5 \n\t" //'z10
1512 "paddw %%mm1, %%mm3 \n\t" //'z13
1513
1514 "movq %%mm0, %%mm2 \n\t"
1515 "psubw %%mm6, %%mm0 \n\t"
1516
1517 "movq %%mm5, %%mm1 \n\t"
1518 "psllw $1, %%mm0 \n\t" //'z12
1519
1520 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
1521 "paddw %%mm0, %%mm5 \n\t"
1522
1523 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
1524 "paddw %%mm6, %%mm2 \n\t" //'z11
1525
1526 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
1527 "movq %%mm2, %%mm7 \n\t"
1528
1529 //---
1530 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t"
1531 "psubw %%mm3, %%mm2 \n\t"
1532
1533 "psllw $1, %%mm2 \n\t"
1534 "paddw %%mm3, %%mm7 \n\t" //'t7
1535
1536 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
1537 "movq %%mm4, %%mm6 \n\t"
1538 //paddw mm7, MM_2
1539 "psraw $2, %%mm7 \n\t"
1540
1541 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
1542 "psubw %%mm7, %%mm6 \n\t"
1543
1544 "movq "MANGLE(temps)"+1*8, %%mm3 \n\t"
1545 "paddw %%mm7, %%mm4 \n\t"
1546
1547 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
1548 "paddw %%mm5, %%mm1 \n\t" //'t12
1549
1550 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
1551 "psubw %%mm7, %%mm1 \n\t" //'t6
1552
1553 "movq "MANGLE(temps)"+2*8, %%mm7 \n\t"
1554 "psubw %%mm5, %%mm0 \n\t" //'t10
1555
1556 "movq "MANGLE(temps)"+3*8, %%mm6 \n\t"
1557 "movq %%mm3, %%mm5 \n\t"
1558
1559 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
1560 "psubw %%mm1, %%mm5 \n\t"
1561
1562 "psubw %%mm1, %%mm2 \n\t" //'t5
1563 "paddw %%mm1, %%mm3 \n\t"
1564
1565 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
1566 "movq %%mm7, %%mm4 \n\t"
1567
1568 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
1569 "psubw %%mm2, %%mm4 \n\t"
1570
1571 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
1572 "paddw %%mm2, %%mm7 \n\t"
1573
1574 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
1575 "paddw %%mm2, %%mm0 \n\t" //'t4
1576
1577 // 't4 't6 't5 - - - - 't7
1578 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
1579 "movq %%mm6, %%mm1 \n\t"
1580
1581 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
1582 "psubw %%mm0, %%mm1 \n\t"
1583
1584 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
1585 "paddw %%mm0, %%mm6 \n\t"
1586
1587 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
1588 "add $24, %%"REG_S" \n\t"
1589
1590 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
1591
1592 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
1593 "add $24, %%"REG_D" \n\t"
1594 "sub $2, %%"REG_c" \n\t"
1595 "jnz 1b \n\t"
1596 "5: \n\t"
1597
1598 : "+S"(data), "+D"(output), "+c"(cnt)// input regs
1599 : "d"(thr_adr)
1600 : "%"REG_a
1601 );
1602 }
1603
1604 #endif // HAVE_MMX
1605
1606 #ifndef HAVE_MMX
1607
1608 static void row_idct_c(DCTELEM* workspace,
1609 int16_t* output_adr, int output_stride, int cnt)
1610 {
1611 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
1612 int_simd16_t tmp10, tmp11, tmp12, tmp13;
1613 int_simd16_t z5, z10, z11, z12, z13;
1614 int16_t* outptr;
1615 DCTELEM* wsptr;
1616
1617 cnt*=4;
1618 wsptr = workspace;
1619 outptr = output_adr;
1620 for (; cnt > 0; cnt--) {
1621 // Even part
1622 //Simd version reads 4x4 block and transposes it
1623 tmp10 = ( wsptr[2] + wsptr[3]);
1624 tmp11 = ( wsptr[2] - wsptr[3]);
1625
1626 tmp13 = ( wsptr[0] + wsptr[1]);
1627 tmp12 = (MULTIPLY16H( wsptr[0] - wsptr[1], FIX_1_414213562_A)<<2) - tmp13;//this shift order to avoid overflow
1628
1629 tmp0 = tmp10 + tmp13; //->temps
1630 tmp3 = tmp10 - tmp13; //->temps
1631 tmp1 = tmp11 + tmp12;
1632 tmp2 = tmp11 - tmp12;
1633
1634 // Odd part
1635 //Also transpose, with previous:
1636 // ---- ---- ||||
1637 // ---- ---- idct ||||
1638 // ---- ---- ---> ||||
1639 // ---- ---- ||||
1640 z13 = wsptr[4] + wsptr[5];
1641 z10 = wsptr[4] - wsptr[5];
1642 z11 = wsptr[6] + wsptr[7];
1643 z12 = wsptr[6] - wsptr[7];
1644
1645 tmp7 = z11 + z13;
1646 tmp11 = MULTIPLY16H(z11 - z13, FIX_1_414213562);
1647
1648 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
1649 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
1650 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - FIX_
1651
1652 tmp6 = (tmp12<<3) - tmp7;
1653 tmp5 = (tmp11<<3) - tmp6;
1654 tmp4 = (tmp10<<3) + tmp5;
1655
1656 // Final output stage: descale and write column
1657 outptr[0*output_stride]+= DESCALE(tmp0 + tmp7, 3);
1658 outptr[1*output_stride]+= DESCALE(tmp1 + tmp6, 3);
1659 outptr[2*output_stride]+= DESCALE(tmp2 + tmp5, 3);
1660 outptr[3*output_stride]+= DESCALE(tmp3 - tmp4, 3);
1661 outptr[4*output_stride]+= DESCALE(tmp3 + tmp4, 3);
1662 outptr[5*output_stride]+= DESCALE(tmp2 - tmp5, 3);
1663 outptr[6*output_stride]+= DESCALE(tmp1 - tmp6, 3); //no += ?
1664 outptr[7*output_stride]+= DESCALE(tmp0 - tmp7, 3); //no += ?
1665 outptr++;
1666
1667 wsptr += DCTSIZE; // advance pointer to next row
1668 }
1669 }
1670
1671 #else /* HAVE_MMX */
1672
1673 static void row_idct_mmx (DCTELEM* workspace,
1674 int16_t* output_adr, int output_stride, int cnt)
1675 {
1676 asm volatile(
1677 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t"
1678 "1: \n\t"
1679 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm0 \n\t"
1680 //
1681
1682 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm1 \n\t"
1683 "movq %%mm0, %%mm4 \n\t"
1684
1685 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
1686 "punpcklwd %%mm1, %%mm0 \n\t"
1687
1688 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm3 \n\t"
1689 "punpckhwd %%mm1, %%mm4 \n\t"
1690
1691 //transpose 4x4
1692 "movq %%mm2, %%mm7 \n\t"
1693 "punpcklwd %%mm3, %%mm2 \n\t"
1694
1695 "movq %%mm0, %%mm6 \n\t"
1696 "punpckldq %%mm2, %%mm0 \n\t" //0
1697
1698 "punpckhdq %%mm2, %%mm6 \n\t" //1
1699 "movq %%mm0, %%mm5 \n\t"
1700
1701 "punpckhwd %%mm3, %%mm7 \n\t"
1702 "psubw %%mm6, %%mm0 \n\t"
1703
1704 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm0 \n\t"
1705 "movq %%mm4, %%mm2 \n\t"
1706
1707 "punpckldq %%mm7, %%mm4 \n\t" //2
1708 "paddw %%mm6, %%mm5 \n\t"
1709
1710 "punpckhdq %%mm7, %%mm2 \n\t" //3
1711 "movq %%mm4, %%mm1 \n\t"
1712
1713 "psllw $2, %%mm0 \n\t"
1714 "paddw %%mm2, %%mm4 \n\t" //t10
1715
1716 "movq "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_S"), %%mm3 \n\t"
1717 "psubw %%mm2, %%mm1 \n\t" //t11
1718
1719 "movq "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_S"), %%mm2 \n\t"
1720 "psubw %%mm5, %%mm0 \n\t"
1721
1722 "movq %%mm4, %%mm6 \n\t"
1723 "paddw %%mm5, %%mm4 \n\t" //t0
1724
1725 "psubw %%mm5, %%mm6 \n\t" //t3
1726 "movq %%mm1, %%mm7 \n\t"
1727
1728 "movq "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_S"), %%mm5 \n\t"
1729 "paddw %%mm0, %%mm1 \n\t" //t1
1730
1731 "movq %%mm4, "MANGLE(temps)"+0*8 \n\t" //t0
1732 "movq %%mm3, %%mm4 \n\t"
1733
1734 "movq %%mm6, "MANGLE(temps)"+1*8 \n\t" //t3
1735 "punpcklwd %%mm2, %%mm3 \n\t"
1736
1737 //transpose 4x4
1738 "movq "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_S"), %%mm6 \n\t"
1739 "punpckhwd %%mm2, %%mm4 \n\t"
1740
1741 "movq %%mm5, %%mm2 \n\t"
1742 "punpcklwd %%mm6, %%mm5 \n\t"
1743
1744 "psubw %%mm0, %%mm7 \n\t" //t2
1745 "punpckhwd %%mm6, %%mm2 \n\t"
1746
1747 "movq %%mm3, %%mm0 \n\t"
1748 "punpckldq %%mm5, %%mm3 \n\t" //4
1749
1750 "punpckhdq %%mm5, %%mm0 \n\t" //5
1751 "movq %%mm4, %%mm5 \n\t"
1752
1753 //
1754 "movq %%mm3, %%mm6 \n\t"
1755 "punpckldq %%mm2, %%mm4 \n\t" //6
1756
1757 "psubw %%mm0, %%mm3 \n\t" //z10
1758 "punpckhdq %%mm2, %%mm5 \n\t" //7
1759
1760 "paddw %%mm0, %%mm6 \n\t" //z13
1761 "movq %%mm4, %%mm2 \n\t"
1762
1763 "movq %%mm3, %%mm0 \n\t"
1764 "psubw %%mm5, %%mm4 \n\t" //z12
1765
1766 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm0 \n\t" //-
1767 "paddw %%mm4, %%mm3 \n\t"
1768
1769 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm3 \n\t" //z5
1770 "paddw %%mm5, %%mm2 \n\t" //z11 >
1771
1772 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm4 \n\t"
1773 "movq %%mm2, %%mm5 \n\t"
1774
1775 "psubw %%mm6, %%mm2 \n\t"
1776 "paddw %%mm6, %%mm5 \n\t" //t7
1777
1778 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //t11
1779 "paddw %%mm3, %%mm0 \n\t" //t12
1780
1781 "psllw $3, %%mm0 \n\t"
1782 "psubw %%mm3, %%mm4 \n\t" //t10
1783
1784 "movq "MANGLE(temps)"+0*8, %%mm6 \n\t"
1785 "movq %%mm1, %%mm3 \n\t"
1786
1787 "psllw $3, %%mm4 \n\t"
1788 "psubw %%mm5, %%mm0 \n\t" //t6
1789
1790 "psllw $3, %%mm2 \n\t"
1791 "paddw %%mm0, %%mm1 \n\t" //d1
1792
1793 "psubw %%mm0, %%mm2 \n\t" //t5
1794 "psubw %%mm0, %%mm3 \n\t" //d6
1795
1796 "paddw %%mm2, %%mm4 \n\t" //t4
1797 "movq %%mm7, %%mm0 \n\t"
1798
1799 "paddw %%mm2, %%mm7 \n\t" //d2
1800 "psubw %%mm2, %%mm0 \n\t" //d5
1801
1802 "movq "MANGLE(MM_DESCALE_RND)", %%mm2 \n\t" //4
1803 "psubw %%mm5, %%mm6 \n\t" //d7
1804
1805 "paddw "MANGLE(temps)"+0*8, %%mm5 \n\t" //d0
1806 "paddw %%mm2, %%mm1 \n\t"
1807
1808 "paddw %%mm2, %%mm5 \n\t"
1809 "psraw $3, %%mm1 \n\t"
1810
1811 "paddw %%mm2, %%mm7 \n\t"
1812 "psraw $3, %%mm5 \n\t"
1813
1814 "paddw (%%"REG_D"), %%mm5 \n\t"
1815 "psraw $3, %%mm7 \n\t"
1816
1817 "paddw (%%"REG_D",%%"REG_a",), %%mm1 \n\t"
1818 "paddw %%mm2, %%mm0 \n\t"
1819
1820 "paddw (%%"REG_D",%%"REG_a",2), %%mm7 \n\t"
1821 "paddw %%mm2, %%mm3 \n\t"
1822
1823 "movq %%mm5, (%%"REG_D") \n\t"
1824 "paddw %%mm2, %%mm6 \n\t"
1825
1826 "movq %%mm1, (%%"REG_D",%%"REG_a",) \n\t"
1827 "psraw $3, %%mm0 \n\t"
1828
1829 "movq %%mm7, (%%"REG_D",%%"REG_a",2) \n\t"
1830 "add %%"REG_d", %%"REG_D" \n\t" //3*ls
1831
1832 "movq "MANGLE(temps)"+1*8, %%mm5 \n\t" //t3
1833 "psraw $3, %%mm3 \n\t"
1834
1835 "paddw (%%"REG_D",%%"REG_a",2), %%mm0 \n\t"
1836 "psubw %%mm4, %%mm5 \n\t" //d3
1837
1838 "paddw (%%"REG_D",%%"REG_d",), %%mm3 \n\t"
1839 "psraw $3, %%mm6 \n\t"
1840
1841 "paddw "MANGLE(temps)"+1*8, %%mm4 \n\t" //d4
1842 "paddw %%mm2, %%mm5 \n\t"
1843
1844 "paddw (%%"REG_D",%%"REG_a",4), %%mm6 \n\t"
1845 "paddw %%mm2, %%mm4 \n\t"
1846
1847 "movq %%mm0, (%%"REG_D",%%"REG_a",2) \n\t"
1848 "psraw $3, %%mm5 \n\t"
1849
1850 "paddw (%%"REG_D"), %%mm5 \n\t"
1851 "psraw $3, %%mm4 \n\t"
1852
1853 "paddw (%%"REG_D",%%"REG_a",), %%mm4 \n\t"
1854 "add $"DCTSIZE_S"*2*4, %%"REG_S" \n\t" //4 rows
1855
1856 "movq %%mm3, (%%"REG_D",%%"REG_d",) \n\t"
1857 "movq %%mm6, (%%"REG_D",%%"REG_a",4) \n\t"
1858 "movq %%mm5, (%%"REG_D") \n\t"
1859 "movq %%mm4, (%%"REG_D",%%"REG_a",) \n\t"
1860
1861 "sub %%"REG_d", %%"REG_D" \n\t"
1862 "add $8, %%"REG_D" \n\t"
1863 "dec %%"REG_c" \n\t"
1864 "jnz 1b \n\t"
1865
1866 : "+S"(workspace), "+D"(output_adr), "+c"(cnt) //input regs
1867 : "a"(output_stride*sizeof(short))
1868 : "%"REG_d
1869 );
1870 }
1871
1872 #endif // HAVE_MMX
1873
1874 #ifndef HAVE_MMX
1875
1876 static void row_fdct_c(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt)
1877 {
1878 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
1879 int_simd16_t tmp10, tmp11, tmp12, tmp13;
1880 int_simd16_t z1, z2, z3, z4, z5, z11, z13;
1881 DCTELEM *dataptr;
1882
1883 cnt*=4;
1884 // Pass 1: process rows.
1885
1886 dataptr = data;
1887 for (; cnt > 0; cnt--) {
1888 tmp0 = pixels[line_size*0] + pixels[line_size*7];
1889 tmp7 = pixels[line_size*0] - pixels[line_size*7];
1890 tmp1 = pixels[line_size*1] + pixels[line_size*6];
1891 tmp6 = pixels[line_size*1] - pixels[line_size*6];
1892 tmp2 = pixels[line_size*2] + pixels[line_size*5];
1893 tmp5 = pixels[line_size*2] - pixels[line_size*5];
1894 tmp3 = pixels[line_size*3] + pixels[line_size*4];
1895 tmp4 = pixels[line_size*3] - pixels[line_size*4];
1896
1897 // Even part
1898
1899 tmp10 = tmp0 + tmp3;
1900 tmp13 = tmp0 - tmp3;
1901 tmp11 = tmp1 + tmp2;
1902 tmp12 = tmp1 - tmp2;
1903 //Even columns are written first, this leads to different order of columns
1904 //in column_fidct(), but they are processed independently, so all ok.
1905 //Later in the row_idct() columns readed at the same order.
1906 dataptr[2] = tmp10 + tmp11;
1907 dataptr[3] = tmp10 - tmp11;
1908
1909 z1 = MULTIPLY16H((tmp12 + tmp13)<<2, FIX_0_707106781);
1910 dataptr[0] = tmp13 + z1;
1911 dataptr[1] = tmp13 - z1;
1912
1913 // Odd part
1914
1915 tmp10 = (tmp4 + tmp5) <<2;
1916 tmp11 = (tmp5 + tmp6) <<2;
1917 tmp12 = (tmp6 + tmp7) <<2;
1918
1919 z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433);
1920 z2 = MULTIPLY16H(tmp10, FIX_0_541196100) + z5;
1921 z4 = MULTIPLY16H(tmp12, FIX_1_306562965) + z5;
1922 z3 = MULTIPLY16H(tmp11, FIX_0_707106781);
1923
1924 z11 = tmp7 + z3;
1925 z13 = tmp7 - z3;
1926
1927 dataptr[4] = z13 + z2;
1928 dataptr[5] = z13 - z2;
1929 dataptr[6] = z11 + z4;
1930 dataptr[7] = z11 - z4;
1931
1932 pixels++; // advance pointer to next column
1933 dataptr += DCTSIZE;
1934 }
1935 }
1936
1937 #else /* HAVE_MMX */
1938
1939 static void row_fdct_mmx(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt)
1940 {
1941 asm volatile(
1942 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t"
1943 "6: \n\t"
1944 "movd (%%"REG_S"), %%mm0 \n\t"
1945 "pxor %%mm7, %%mm7 \n\t"
1946
1947 "movd (%%"REG_S",%%"REG_a",), %%mm1 \n\t"
1948 "punpcklbw %%mm7, %%mm0 \n\t"
1949
1950 "movd (%%"REG_S",%%"REG_a",2), %%mm2 \n\t"
1951 "punpcklbw %%mm7, %%mm1 \n\t"
1952
1953 "punpcklbw %%mm7, %%mm2 \n\t"
1954 "add %%"REG_d", %%"REG_S" \n\t"
1955
1956 "movq %%mm0, %%mm5 \n\t"
1957 //
1958
1959 "movd (%%"REG_S",%%"REG_a",4), %%mm3 \n\t" //7 ;prefetch!
1960 "movq %%mm1, %%mm6 \n\t"
1961
1962 "movd (%%"REG_S",%%"REG_d",), %%mm4 \n\t" //6
1963 "punpcklbw %%mm7, %%mm3 \n\t"
1964
1965 "psubw %%mm3, %%mm5 \n\t"
1966 "punpcklbw %%mm7, %%mm4 \n\t"
1967
1968 "paddw %%mm3, %%mm0 \n\t"
1969 "psubw %%mm4, %%mm6 \n\t"
1970
1971 "movd (%%"REG_S",%%"REG_a",2), %%mm3 \n\t" //5
1972 "paddw %%mm4, %%mm1 \n\t"
1973
1974 "movq %%mm5, "MANGLE(temps)"+0*8 \n\t" //t7
1975 "punpcklbw %%mm7, %%mm3 \n\t"
1976
1977 "movq %%mm6, "MANGLE(temps)"+1*8 \n\t" //t6
1978 "movq %%mm2, %%mm4 \n\t"
1979
1980 "movd (%%"REG_S"), %%mm5 \n\t" //3
1981 "paddw %%mm3, %%mm2 \n\t"
1982
1983 "movd (%%"REG_S",%%"REG_a",), %%mm6 \n\t" //4
1984 "punpcklbw %%mm7, %%mm5 \n\t"
1985
1986 "psubw %%mm3, %%mm4 \n\t"
1987 "punpcklbw %%mm7, %%mm6 \n\t"
1988
1989 "movq %%mm5, %%mm3 \n\t"
1990 "paddw %%mm6, %%mm5 \n\t" //t3
1991
1992 "psubw %%mm6, %%mm3 \n\t" //t4 ; t0 t1 t2 t4 t5 t3 - -
1993 "movq %%mm0, %%mm6 \n\t"
1994
1995 "movq %%mm1, %%mm7 \n\t"
1996 "psubw %%mm5, %%mm0 \n\t" //t13
1997
1998 "psubw %%mm2, %%mm1 \n\t"
1999 "paddw %%mm2, %%mm7 \n\t" //t11
2000
2001 "paddw %%mm0, %%mm1 \n\t"
2002 "movq %%mm7, %%mm2 \n\t"
2003
2004 "psllw $2, %%mm1 \n\t"
2005 "paddw %%mm5, %%mm6 \n\t" //t10
2006
2007 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm1 \n\t"
2008 "paddw %%mm6, %%mm7 \n\t" //d2
2009
2010 "psubw %%mm2, %%mm6 \n\t" //d3
2011 "movq %%mm0, %%mm5 \n\t"
2012
2013 //transpose 4x4
2014 "movq %%mm7, %%mm2 \n\t"
2015 "punpcklwd %%mm6, %%mm7 \n\t"
2016
2017 "paddw %%mm1, %%mm0 \n\t" //d0
2018 "punpckhwd %%mm6, %%mm2 \n\t"
2019
2020 "psubw %%mm1, %%mm5 \n\t" //d1
2021 "movq %%mm0, %%mm6 \n\t"
2022
2023 "movq "MANGLE(temps)"+1*8, %%mm1 \n\t"
2024 "punpcklwd %%mm5, %%mm0 \n\t"
2025
2026 "punpckhwd %%mm5, %%mm6 \n\t"
2027 "movq %%mm0, %%mm5 \n\t"
2028
2029 "punpckldq %%mm7, %%mm0 \n\t" //0
2030 "paddw %%mm4, %%mm3 \n\t"
2031
2032 "punpckhdq %%mm7, %%mm5 \n\t" //1
2033 "movq %%mm6, %%mm7 \n\t"
2034
2035 "movq %%mm0, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
2036 "punpckldq %%mm2, %%mm6 \n\t" //2
2037
2038 "movq %%mm5, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
2039 "punpckhdq %%mm2, %%mm7 \n\t" //3
2040
2041 "movq %%mm6, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
2042 "paddw %%mm1, %%mm4 \n\t"
2043
2044 "movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
2045 "psllw $2, %%mm3 \n\t" //t10
2046
2047 "movq "MANGLE(temps)"+0*8, %%mm2 \n\t"
2048 "psllw $2, %%mm4 \n\t" //t11
2049
2050 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm4 \n\t" //z3
2051 "paddw %%mm2, %%mm1 \n\t"
2052
2053 "psllw $2, %%mm1 \n\t" //t12
2054 "movq %%mm3, %%mm0 \n\t"
2055
2056 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm0 \n\t"
2057 "psubw %%mm1, %%mm3 \n\t"
2058
2059 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" //z5
2060 "movq %%mm2, %%mm5 \n\t"
2061
2062 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm1 \n\t"
2063 "psubw %%mm4, %%mm2 \n\t" //z13
2064
2065 "paddw %%mm4, %%mm5 \n\t" //z11
2066 "movq %%mm2, %%mm6 \n\t"
2067
2068 "paddw %%mm3, %%mm0 \n\t" //z2
2069 "movq %%mm5, %%mm7 \n\t"
2070
2071 "paddw %%mm0, %%mm2 \n\t" //d4
2072 "psubw %%mm0, %%mm6 \n\t" //d5
2073
2074 "movq %%mm2, %%mm4 \n\t"
2075 "paddw %%mm3, %%mm1 \n\t" //z4
2076
2077 //transpose 4x4
2078 "punpcklwd %%mm6, %%mm2 \n\t"
2079 "paddw %%mm1, %%mm5 \n\t" //d6
2080
2081 "punpckhwd %%mm6, %%mm4 \n\t"
2082 "psubw %%mm1, %%mm7 \n\t" //d7
2083
2084 "movq %%mm5, %%mm6 \n\t"
2085 "punpcklwd %%mm7, %%mm5 \n\t"
2086
2087 "punpckhwd %%mm7, %%mm6 \n\t"
2088 "movq %%mm2, %%mm7 \n\t"
2089
2090 "punpckldq %%mm5, %%mm2 \n\t" //4
2091 "sub %%"REG_d", %%"REG_S" \n\t"
2092
2093 "punpckhdq %%mm5, %%mm7 \n\t" //5
2094 "movq %%mm4, %%mm5 \n\t"
2095
2096 "movq %%mm2, "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_D") \n\t"
2097 "punpckldq %%mm6, %%mm4 \n\t" //6
2098
2099 "movq %%mm7, "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_D") \n\t"
2100 "punpckhdq %%mm6, %%mm5 \n\t" //7
2101
2102 "movq %%mm4, "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_D") \n\t"
2103 "add $4, %%"REG_S" \n\t"
2104
2105 "movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_D") \n\t"
2106 "add $"DCTSIZE_S"*2*4, %%"REG_D" \n\t" //4 rows
2107 "dec %%"REG_c" \n\t"
2108 "jnz 6b \n\t"
2109
2110 : "+S"(pixels), "+D"(data), "+c"(cnt) //input regs
2111 : "a"(line_size)
2112 : "%"REG_d);
2113 }
2114
2115 #endif // HAVE_MMX
2116
2117 #endif //USE_LIBAVCODEC
mplayer with SMedia Glamo drivers