2 * Copyright (C) 2003 Michael Niedermayer <michaelni@gmx.at>
3 * Copyright (C) 2005 Nikolaj Poroshin <porosh3@psu.ru>
5 * This file is part of MPlayer.
7 * MPlayer is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
12 * MPlayer is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License along
18 * with MPlayer; if not, write to the Free Software Foundation, Inc.,
19 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
23 * This implementation is based on an algorithm described in
24 * "Aria Nosratinia Embedded Post-Processing for
25 * Enhancement of Compressed Images (1999)"
26 * (http://citeseer.nj.nec.com/nosratinia99embedded.html)
27 * Futher, with splitting (i)dct into hor/ver passes, one of them can be
28 * performed once per block, not pixel. This allows for much better speed.
32 Heavily optimized version of SPP filter by Nikolaj
44 #include "cpudetect.h"
50 #include "img_format.h"
53 #include "libvo/fastmemcpy.h"
55 #include "libavutil/internal.h"
56 #include "libavutil/intreadwrite.h"
57 #include "libavutil/mem.h"
58 #include "libavcodec/avcodec.h"
59 #include "libavcodec/dsputil.h"
64 //===========================================================================//
67 static const short custom_threshold
[64]=
68 // values (296) can't be too high
69 // -it causes too big quant dependence
70 // or maybe overflow(check), which results in some flashing
71 { 71, 296, 295, 237, 71, 40, 38, 19,
72 245, 193, 185, 121, 102, 73, 53, 27,
73 158, 129, 141, 107, 97, 73, 50, 26,
74 102, 116, 109, 98, 82, 66, 45, 23,
75 71, 94, 95, 81, 70, 56, 38, 20,
76 56, 77, 74, 66, 56, 44, 30, 15,
77 38, 53, 50, 45, 38, 30, 21, 11,
78 20, 27, 26, 23, 20, 15, 11, 5
81 static const uint8_t __attribute__((aligned(32))) dither
[8][8]={
82 { 0, 48, 12, 60, 3, 51, 15, 63, },
83 { 32, 16, 44, 28, 35, 19, 47, 31, },
84 { 8, 56, 4, 52, 11, 59, 7, 55, },
85 { 40, 24, 36, 20, 43, 27, 39, 23, },
86 { 2, 50, 14, 62, 1, 49, 13, 61, },
87 { 34, 18, 46, 30, 33, 17, 45, 29, },
88 { 10, 58, 6, 54, 9, 57, 5, 53, },
89 { 42, 26, 38, 22, 41, 25, 37, 21, },
92 struct vf_priv_s
{ //align 16 !
93 uint64_t threshold_mtx_noq
[8*2];
94 uint64_t threshold_mtx
[8*2];//used in both C & MMX (& later SSE2) versions
110 //This func reads from 1 slice, 1 and clears 0 & 1
111 static void store_slice_c(uint8_t *dst
, int16_t *src
, int dst_stride
, int src_stride
, int width
, int height
, int log2_scale
)
114 temp= (src[x + pos] + (d[pos]>>log2_scale))>>(6-log2_scale); \
115 src[x + pos]=src[x + pos - 8*src_stride]=0; \
116 if(temp & 0x100) temp= ~(temp>>31); \
119 for(y
=0; y
<height
; y
++){
120 const uint8_t *d
= dither
[y
];
121 for(x
=0; x
<width
; x
+=8){
137 //This func reads from 2 slices, 0 & 2 and clears 2-nd
138 static void store_slice2_c(uint8_t *dst
, int16_t *src
, int dst_stride
, int src_stride
, int width
, int height
, int log2_scale
)
140 #define STORE2(pos) \
141 temp= (src[x + pos] + src[x + pos + 16*src_stride] + (d[pos]>>log2_scale))>>(6-log2_scale); \
142 src[x + pos + 16*src_stride]=0; \
143 if(temp & 0x100) temp= ~(temp>>31); \
146 for(y
=0; y
<height
; y
++){
147 const uint8_t *d
= dither
[y
];
148 for(x
=0; x
<width
; x
+=8){
164 static void mul_thrmat_c(struct vf_priv_s
*p
,int q
)
168 ((short*)p
->threshold_mtx
)[a
]=q
* ((short*)p
->threshold_mtx_noq
)[a
];//ints faster in C
171 static void column_fidct_c(int16_t* thr_adr
, DCTELEM
*data
, DCTELEM
*output
, int cnt
);
172 static void row_idct_c(DCTELEM
* workspace
,
173 int16_t* output_adr
, int output_stride
, int cnt
);
174 static void row_fdct_c(DCTELEM
*data
, const uint8_t *pixels
, int line_size
, int cnt
);
176 //this is rather ugly, but there is no need for function pointers
177 #define store_slice_s store_slice_c
178 #define store_slice2_s store_slice2_c
179 #define mul_thrmat_s mul_thrmat_c
180 #define column_fidct_s column_fidct_c
181 #define row_idct_s row_idct_c
182 #define row_fdct_s row_fdct_c
186 //This func reads from 1 slice, 1 and clears 0 & 1
187 static void store_slice_mmx(uint8_t *dst
, int16_t *src
, long dst_stride
, long src_stride
, long width
, long height
, long log2_scale
)
189 const uint8_t *od
=&dither
[0][0];
190 const uint8_t *end
=&dither
[height
][0];
191 width
= (width
+7)&~7;
193 //src_stride=(src_stride-width)*2;
195 "mov %5, %%"REG_d
" \n\t"
196 "mov %6, %%"REG_S
" \n\t"
197 "mov %7, %%"REG_D
" \n\t"
198 "mov %1, %%"REG_a
" \n\t"
199 "movd %%"REG_d
", %%mm5 \n\t"
200 "xor $-1, %%"REG_d
" \n\t"
201 "mov %%"REG_a
", %%"REG_c
" \n\t"
202 "add $7, %%"REG_d
" \n\t"
204 "sub %0, %%"REG_c
" \n\t"
205 "add %%"REG_c
", %%"REG_c
" \n\t"
206 "movd %%"REG_d
", %%mm2 \n\t"
207 "mov %%"REG_c
", %1 \n\t"
208 "mov %2, %%"REG_d
" \n\t"
209 "shl $4, %%"REG_a
" \n\t"
212 "movq (%%"REG_d
"), %%mm3 \n\t"
213 "movq %%mm3, %%mm4 \n\t"
214 "pxor %%mm7, %%mm7 \n\t"
215 "punpcklbw %%mm7, %%mm3 \n\t"
216 "punpckhbw %%mm7, %%mm4 \n\t"
217 "mov %0, %%"REG_c
" \n\t"
218 "psraw %%mm5, %%mm3 \n\t"
219 "psraw %%mm5, %%mm4 \n\t"
221 "movq %%mm7, (%%"REG_S
",%%"REG_a
",) \n\t"
222 "movq (%%"REG_S
"), %%mm0 \n\t"
223 "movq 8(%%"REG_S
"), %%mm1 \n\t"
225 "movq %%mm7, 8(%%"REG_S
",%%"REG_a
",) \n\t"
226 "paddw %%mm3, %%mm0 \n\t"
227 "paddw %%mm4, %%mm1 \n\t"
229 "movq %%mm7, (%%"REG_S
") \n\t"
230 "psraw %%mm2, %%mm0 \n\t"
231 "psraw %%mm2, %%mm1 \n\t"
233 "movq %%mm7, 8(%%"REG_S
") \n\t"
234 "packuswb %%mm1, %%mm0 \n\t"
235 "add $16, %%"REG_S
" \n\t"
237 "movq %%mm0, (%%"REG_D
") \n\t"
238 "add $8, %%"REG_D
" \n\t"
239 "sub $8, %%"REG_c
" \n\t"
241 "add %1, %%"REG_S
" \n\t"
242 "add $8, %%"REG_d
" \n\t"
243 "add %3, %%"REG_D
" \n\t"
244 "cmp %4, %%"REG_d
" \n\t"
248 : "m" (width
), "m" (src_stride
), "g" (od
), "m" (dst_stride
), "g" (end
),
249 "m" (log2_scale
), "m" (src
), "m" (dst
) //input
250 : "%"REG_a
, "%"REG_c
, "%"REG_d
, "%"REG_S
, "%"REG_D
254 //This func reads from 2 slices, 0 & 2 and clears 2-nd
255 static void store_slice2_mmx(uint8_t *dst
, int16_t *src
, long dst_stride
, long src_stride
, long width
, long height
, long log2_scale
)
257 const uint8_t *od
=&dither
[0][0];
258 const uint8_t *end
=&dither
[height
][0];
259 width
= (width
+7)&~7;
261 //src_stride=(src_stride-width)*2;
263 "mov %5, %%"REG_d
" \n\t"
264 "mov %6, %%"REG_S
" \n\t"
265 "mov %7, %%"REG_D
" \n\t"
266 "mov %1, %%"REG_a
" \n\t"
267 "movd %%"REG_d
", %%mm5 \n\t"
268 "xor $-1, %%"REG_d
" \n\t"
269 "mov %%"REG_a
", %%"REG_c
" \n\t"
270 "add $7, %%"REG_d
" \n\t"
271 "sub %0, %%"REG_c
" \n\t"
272 "add %%"REG_c
", %%"REG_c
" \n\t"
273 "movd %%"REG_d
", %%mm2 \n\t"
274 "mov %%"REG_c
", %1 \n\t"
275 "mov %2, %%"REG_d
" \n\t"
276 "shl $5, %%"REG_a
" \n\t"
279 "movq (%%"REG_d
"), %%mm3 \n\t"
280 "movq %%mm3, %%mm4 \n\t"
281 "pxor %%mm7, %%mm7 \n\t"
282 "punpcklbw %%mm7, %%mm3 \n\t"
283 "punpckhbw %%mm7, %%mm4 \n\t"
284 "mov %0, %%"REG_c
" \n\t"
285 "psraw %%mm5, %%mm3 \n\t"
286 "psraw %%mm5, %%mm4 \n\t"
288 "movq (%%"REG_S
"), %%mm0 \n\t"
289 "movq 8(%%"REG_S
"), %%mm1 \n\t"
290 "paddw %%mm3, %%mm0 \n\t"
292 "paddw (%%"REG_S
",%%"REG_a
",), %%mm0 \n\t"
293 "paddw %%mm4, %%mm1 \n\t"
294 "movq 8(%%"REG_S
",%%"REG_a
",), %%mm6 \n\t"
296 "movq %%mm7, (%%"REG_S
",%%"REG_a
",) \n\t"
297 "psraw %%mm2, %%mm0 \n\t"
298 "paddw %%mm6, %%mm1 \n\t"
300 "movq %%mm7, 8(%%"REG_S
",%%"REG_a
",) \n\t"
301 "psraw %%mm2, %%mm1 \n\t"
302 "packuswb %%mm1, %%mm0 \n\t"
304 "movq %%mm0, (%%"REG_D
") \n\t"
305 "add $16, %%"REG_S
" \n\t"
306 "add $8, %%"REG_D
" \n\t"
307 "sub $8, %%"REG_c
" \n\t"
309 "add %1, %%"REG_S
" \n\t"
310 "add $8, %%"REG_d
" \n\t"
311 "add %3, %%"REG_D
" \n\t"
312 "cmp %4, %%"REG_d
" \n\t"
316 : "m" (width
), "m" (src_stride
), "g" (od
), "m" (dst_stride
), "g" (end
),
317 "m" (log2_scale
), "m" (src
), "m" (dst
) //input
318 : "%"REG_a
, "%"REG_c
, "%"REG_d
, "%"REG_D
, "%"REG_S
322 static void mul_thrmat_mmx(struct vf_priv_s
*p
, int q
)
324 uint64_t *adr
=&p
->threshold_mtx_noq
[0];
326 "movd %0, %%mm7 \n\t"
327 "add $8*8*2, %%"REG_D
" \n\t"
328 "movq 0*8(%%"REG_S
"), %%mm0 \n\t"
329 "punpcklwd %%mm7, %%mm7 \n\t"
330 "movq 1*8(%%"REG_S
"), %%mm1 \n\t"
331 "punpckldq %%mm7, %%mm7 \n\t"
332 "pmullw %%mm7, %%mm0 \n\t"
334 "movq 2*8(%%"REG_S
"), %%mm2 \n\t"
335 "pmullw %%mm7, %%mm1 \n\t"
337 "movq 3*8(%%"REG_S
"), %%mm3 \n\t"
338 "pmullw %%mm7, %%mm2 \n\t"
340 "movq %%mm0, 0*8(%%"REG_D
") \n\t"
341 "movq 4*8(%%"REG_S
"), %%mm4 \n\t"
342 "pmullw %%mm7, %%mm3 \n\t"
344 "movq %%mm1, 1*8(%%"REG_D
") \n\t"
345 "movq 5*8(%%"REG_S
"), %%mm5 \n\t"
346 "pmullw %%mm7, %%mm4 \n\t"
348 "movq %%mm2, 2*8(%%"REG_D
") \n\t"
349 "movq 6*8(%%"REG_S
"), %%mm6 \n\t"
350 "pmullw %%mm7, %%mm5 \n\t"
352 "movq %%mm3, 3*8(%%"REG_D
") \n\t"
353 "movq 7*8+0*8(%%"REG_S
"), %%mm0 \n\t"
354 "pmullw %%mm7, %%mm6 \n\t"
356 "movq %%mm4, 4*8(%%"REG_D
") \n\t"
357 "movq 7*8+1*8(%%"REG_S
"), %%mm1 \n\t"
358 "pmullw %%mm7, %%mm0 \n\t"
360 "movq %%mm5, 5*8(%%"REG_D
") \n\t"
361 "movq 7*8+2*8(%%"REG_S
"), %%mm2 \n\t"
362 "pmullw %%mm7, %%mm1 \n\t"
364 "movq %%mm6, 6*8(%%"REG_D
") \n\t"
365 "movq 7*8+3*8(%%"REG_S
"), %%mm3 \n\t"
366 "pmullw %%mm7, %%mm2 \n\t"
368 "movq %%mm0, 7*8+0*8(%%"REG_D
") \n\t"
369 "movq 7*8+4*8(%%"REG_S
"), %%mm4 \n\t"
370 "pmullw %%mm7, %%mm3 \n\t"
372 "movq %%mm1, 7*8+1*8(%%"REG_D
") \n\t"
373 "movq 7*8+5*8(%%"REG_S
"), %%mm5 \n\t"
374 "pmullw %%mm7, %%mm4 \n\t"
376 "movq %%mm2, 7*8+2*8(%%"REG_D
") \n\t"
377 "movq 7*8+6*8(%%"REG_S
"), %%mm6 \n\t"
378 "pmullw %%mm7, %%mm5 \n\t"
380 "movq %%mm3, 7*8+3*8(%%"REG_D
") \n\t"
381 "movq 14*8+0*8(%%"REG_S
"), %%mm0 \n\t"
382 "pmullw %%mm7, %%mm6 \n\t"
384 "movq %%mm4, 7*8+4*8(%%"REG_D
") \n\t"
385 "movq 14*8+1*8(%%"REG_S
"), %%mm1 \n\t"
386 "pmullw %%mm7, %%mm0 \n\t"
388 "movq %%mm5, 7*8+5*8(%%"REG_D
") \n\t"
389 "pmullw %%mm7, %%mm1 \n\t"
391 "movq %%mm6, 7*8+6*8(%%"REG_D
") \n\t"
392 "movq %%mm0, 14*8+0*8(%%"REG_D
") \n\t"
393 "movq %%mm1, 14*8+1*8(%%"REG_D
") \n\t"
395 : "+g" (q
), "+S" (adr
), "+D" (adr
)
400 static void column_fidct_mmx(int16_t* thr_adr
, DCTELEM
*data
, DCTELEM
*output
, int cnt
);
401 static void row_idct_mmx(DCTELEM
* workspace
,
402 int16_t* output_adr
, int output_stride
, int cnt
);
403 static void row_fdct_mmx(DCTELEM
*data
, const uint8_t *pixels
, int line_size
, int cnt
);
405 #define store_slice_s store_slice_mmx
406 #define store_slice2_s store_slice2_mmx
407 #define mul_thrmat_s mul_thrmat_mmx
408 #define column_fidct_s column_fidct_mmx
409 #define row_idct_s row_idct_mmx
410 #define row_fdct_s row_fdct_mmx
413 static void filter(struct vf_priv_s
*p
, uint8_t *dst
, uint8_t *src
,
414 int dst_stride
, int src_stride
,
415 int width
, int height
,
416 uint8_t *qp_store
, int qp_stride
, int is_luma
)
418 int x
, x0
, y
, es
, qy
, t
;
419 const int stride
= is_luma
? p
->temp_stride
: (width
+16);//((width+16+15)&(~15))
420 const int step
=6-p
->log2_count
;
421 const int qps
= 3 + is_luma
;
422 int32_t __attribute__((aligned(32))) block_align
[4*8*BLOCKSZ
+ 4*8*BLOCKSZ
];
423 DCTELEM
*block
= (DCTELEM
*)block_align
;
424 DCTELEM
*block3
=(DCTELEM
*)(block_align
+4*8*BLOCKSZ
);
426 memset(block3
, 0, 4*8*BLOCKSZ
);
428 //p->src=src-src_stride*8-8;//!
429 if (!src
|| !dst
) return; // HACK avoid crash for Y8 colourspace
430 for(y
=0; y
<height
; y
++){
431 int index
= 8 + 8*stride
+ y
*stride
;
432 fast_memcpy(p
->src
+ index
, src
+ y
*src_stride
, width
);//this line can be avoided by using DR & user fr.buffers
434 p
->src
[index
- x
- 1]= p
->src
[index
+ x
];
435 p
->src
[index
+ width
+ x
]= p
->src
[index
+ width
- x
- 1];
439 fast_memcpy(p
->src
+ ( 7-y
)*stride
, p
->src
+ ( y
+8)*stride
, stride
);
440 fast_memcpy(p
->src
+ (height
+8+y
)*stride
, p
->src
+ (height
-y
+7)*stride
, stride
);
442 //FIXME (try edge emu)
445 memset(p
->temp
+ 8 +y
*stride
, 0,width
*sizeof(int16_t));
447 for(y
=step
; y
<height
+8; y
+=step
){ //step= 1,2
449 if (qy
>height
-1) qy
=height
-1;
451 qy
=(qy
>>qps
)*qp_stride
;
452 row_fdct_s(block
, p
->src
+ y
*stride
+2-(y
&1), stride
, 2);
453 for(x0
=0; x0
<width
+8-8*(BLOCKSZ
-1); x0
+=8*(BLOCKSZ
-1)){
454 row_fdct_s(block
+8*8, p
->src
+ y
*stride
+8+x0
+2-(y
&1), stride
, 2*(BLOCKSZ
-1));
456 column_fidct_s((int16_t*)(&p
->threshold_mtx
[0]), block
+0*8, block3
+0*8, 8*(BLOCKSZ
-1)); //yes, this is a HOTSPOT
458 for (x
=0; x
<8*(BLOCKSZ
-1); x
+=8) {
459 t
=x
+x0
-2; //correct t=x+x0-2-(y&1), but its the same
460 if (t
<0) t
=0;//t always < width-2
461 t
=qp_store
[qy
+(t
>>qps
)];
462 if(p
->mpeg2
) t
>>=1; //copy p->mpeg2,prev_q to locals?
463 if (t
!=p
->prev_q
) p
->prev_q
=t
, mul_thrmat_s(p
, t
);
464 column_fidct_s((int16_t*)(&p
->threshold_mtx
[0]), block
+x
*8, block3
+x
*8, 8); //yes, this is a HOTSPOT
466 row_idct_s(block3
+0*8, p
->temp
+ (y
&15)*stride
+x0
+2-(y
&1), stride
, 2*(BLOCKSZ
-1));
467 memmove(block
, block
+(BLOCKSZ
-1)*64, 8*8*sizeof(DCTELEM
)); //cycling
468 memmove(block3
, block3
+(BLOCKSZ
-1)*64, 6*8*sizeof(DCTELEM
));
471 es
=width
+8-x0
; // 8, ...
473 row_fdct_s(block
+8*8, p
->src
+ y
*stride
+8+x0
+2-(y
&1), stride
, (es
-4)>>2);
474 column_fidct_s((int16_t*)(&p
->threshold_mtx
[0]), block
, block3
, es
&(~1));
475 row_idct_s(block3
+0*8, p
->temp
+ (y
&15)*stride
+x0
+2-(y
&1), stride
, es
>>2);
476 {const int y1
=y
-8+step
;//l5-7 l4-6
478 if (y1
&8) store_slice_s(dst
+ (y1
-8)*dst_stride
, p
->temp
+ 8 +8*stride
,
479 dst_stride
, stride
, width
, 8, 5-p
->log2_count
);
480 else store_slice2_s(dst
+ (y1
-8)*dst_stride
, p
->temp
+ 8 +0*stride
,
481 dst_stride
, stride
, width
, 8, 5-p
->log2_count
);
485 if (y
&7) { // == height & 7
486 if (y
&8) store_slice_s(dst
+ ((y
-8)&~7)*dst_stride
, p
->temp
+ 8 +8*stride
,
487 dst_stride
, stride
, width
, y
&7, 5-p
->log2_count
);
488 else store_slice2_s(dst
+ ((y
-8)&~7)*dst_stride
, p
->temp
+ 8 +0*stride
,
489 dst_stride
, stride
, width
, y
&7, 5-p
->log2_count
);
493 static int config(struct vf_instance_s
* vf
,
494 int width
, int height
, int d_width
, int d_height
,
495 unsigned int flags
, unsigned int outfmt
)
497 int h
= (height
+16+15)&(~15);
499 vf
->priv
->temp_stride
= (width
+16+15)&(~15);
500 vf
->priv
->temp
= (int16_t*)av_mallocz(vf
->priv
->temp_stride
*3*8*sizeof(int16_t));
501 //this can also be avoided, see above
502 vf
->priv
->src
= (uint8_t*)av_malloc(vf
->priv
->temp_stride
*h
*sizeof(uint8_t));
504 return vf_next_config(vf
,width
,height
,d_width
,d_height
,flags
,outfmt
);
507 static void get_image(struct vf_instance_s
* vf
, mp_image_t
*mpi
)
509 if(mpi
->flags
&MP_IMGFLAG_PRESERVE
) return; // don't change
510 // ok, we can do pp in-place (or pp disabled):
511 vf
->dmpi
=vf_get_image(vf
->next
,mpi
->imgfmt
,
512 mpi
->type
, mpi
->flags
, mpi
->width
, mpi
->height
);
513 mpi
->planes
[0]=vf
->dmpi
->planes
[0];
514 mpi
->stride
[0]=vf
->dmpi
->stride
[0];
515 mpi
->width
=vf
->dmpi
->width
;
516 if(mpi
->flags
&MP_IMGFLAG_PLANAR
){
517 mpi
->planes
[1]=vf
->dmpi
->planes
[1];
518 mpi
->planes
[2]=vf
->dmpi
->planes
[2];
519 mpi
->stride
[1]=vf
->dmpi
->stride
[1];
520 mpi
->stride
[2]=vf
->dmpi
->stride
[2];
522 mpi
->flags
|=MP_IMGFLAG_DIRECT
;
525 static int put_image(struct vf_instance_s
* vf
, mp_image_t
*mpi
, double pts
)
528 if(!(mpi
->flags
&MP_IMGFLAG_DIRECT
)){
529 // no DR, so get a new image! hope we'll get DR buffer:
530 dmpi
=vf_get_image(vf
->next
,mpi
->imgfmt
,
532 MP_IMGFLAG_ACCEPT_STRIDE
|MP_IMGFLAG_PREFER_ALIGNED_STRIDE
,
533 mpi
->width
,mpi
->height
);
534 vf_clone_mpi_attributes(dmpi
, mpi
);
539 vf
->priv
->mpeg2
= mpi
->qscale_type
;
540 if(mpi
->pict_type
!= 3 && mpi
->qscale
&& !vf
->priv
->qp
){
541 if(!vf
->priv
->non_b_qp
)
542 vf
->priv
->non_b_qp
= malloc(mpi
->qstride
* ((mpi
->h
+ 15) >> 4));
543 fast_memcpy(vf
->priv
->non_b_qp
, mpi
->qscale
, mpi
->qstride
* ((mpi
->h
+ 15) >> 4));
545 if(vf
->priv
->log2_count
|| !(mpi
->flags
&MP_IMGFLAG_DIRECT
)){
546 char *qp_tab
= vf
->priv
->non_b_qp
;
547 if(vf
->priv
->bframes
|| !qp_tab
)
550 if(qp_tab
|| vf
->priv
->qp
){
551 filter(vf
->priv
, dmpi
->planes
[0], mpi
->planes
[0], dmpi
->stride
[0], mpi
->stride
[0],
552 mpi
->w
, mpi
->h
, qp_tab
, mpi
->qstride
, 1);
553 filter(vf
->priv
, dmpi
->planes
[1], mpi
->planes
[1], dmpi
->stride
[1], mpi
->stride
[1],
554 mpi
->w
>>mpi
->chroma_x_shift
, mpi
->h
>>mpi
->chroma_y_shift
, qp_tab
, mpi
->qstride
, 0);
555 filter(vf
->priv
, dmpi
->planes
[2], mpi
->planes
[2], dmpi
->stride
[2], mpi
->stride
[2],
556 mpi
->w
>>mpi
->chroma_x_shift
, mpi
->h
>>mpi
->chroma_y_shift
, qp_tab
, mpi
->qstride
, 0);
558 memcpy_pic(dmpi
->planes
[0], mpi
->planes
[0], mpi
->w
, mpi
->h
, dmpi
->stride
[0], mpi
->stride
[0]);
559 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]);
560 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]);
565 if(gCpuCaps
.hasMMX
) __asm__
volatile ("emms\n\t");
568 if(gCpuCaps
.hasMMX2
) __asm__
volatile ("sfence\n\t");
570 return vf_next_put_image(vf
,dmpi
, pts
);
573 static void uninit(struct vf_instance_s
* vf
)
575 if(!vf
->priv
) return;
577 if(vf
->priv
->temp
) av_free(vf
->priv
->temp
);
578 vf
->priv
->temp
= NULL
;
579 if(vf
->priv
->src
) av_free(vf
->priv
->src
);
581 //if(vf->priv->avctx) free(vf->priv->avctx);
582 //vf->priv->avctx= NULL;
583 if(vf
->priv
->non_b_qp
) free(vf
->priv
->non_b_qp
);
584 vf
->priv
->non_b_qp
= NULL
;
590 //===========================================================================//
592 static int query_format(struct vf_instance_s
* vf
, unsigned int fmt
)
606 return vf_next_query_format(vf
,fmt
);
611 static int control(struct vf_instance_s
* vf
, int request
, void* data
)
614 case VFCTRL_QUERY_MAX_PP_LEVEL
:
616 case VFCTRL_SET_PP_LEVEL
:
617 vf
->priv
->log2_count
= *((unsigned int*)data
);
618 if (vf
->priv
->log2_count
< 4) vf
->priv
->log2_count
=4;
621 return vf_next_control(vf
,request
,data
);
624 static int open(vf_instance_t
*vf
, char* args
)
627 int custom_threshold_m
[64];
631 vf
->put_image
=put_image
;
632 vf
->get_image
=get_image
;
633 vf
->query_format
=query_format
;
635 vf
->control
= control
;
636 vf
->priv
=av_mallocz(sizeof(struct vf_priv_s
));//assumes align 16 !
640 //vf->priv->avctx= avcodec_alloc_context();
641 //dsputil_init(&vf->priv->dsp, vf->priv->avctx);
643 vf
->priv
->log2_count
= 4;
644 vf
->priv
->bframes
= 0;
646 if (args
) sscanf(args
, "%d:%d:%d:%d", &log2c
, &vf
->priv
->qp
, &i
, &vf
->priv
->bframes
);
648 if( log2c
>=4 && log2c
<=5 )
649 vf
->priv
->log2_count
= log2c
;
650 else if( log2c
>= 6 )
651 vf
->priv
->log2_count
= 5;
656 if (i
< -15) i
= -15;
659 bias
= (1<<4)+i
; //regulable
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);
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);
675 if (vf
->priv
->qp
) vf
->priv
->prev_q
=vf
->priv
->qp
, mul_thrmat_s(vf
->priv
, vf
->priv
->qp
);
680 const vf_info_t vf_info_fspp
= {
681 "fast simple postprocess",
683 "Michael Niedermayer, Nikolaj Poroshin",
689 //====================================================================
690 //Specific spp's dct, idct and threshold functions
691 //I'd prefer to have them in the separate file.
693 //#define MANGLE(a) #a
695 //typedef int16_t DCTELEM; //! only int16_t
698 #define DCTSIZE_S "8"
700 #define FIX(x,s) ((int) ((x) * (1<<s) + 0.5)&0xffff)
701 #define C64(x) ((uint64_t)((x)|(x)<<16))<<32 | (uint64_t)(x) | (uint64_t)(x)<<16
702 #define FIX64(x,s) C64(FIX(x,s))
704 #define MULTIPLY16H(x,k) (((x)*(k))>>16)
705 #define THRESHOLD(r,x,t) if(((unsigned)((x)+t))>t*2) r=(x);else r=0;
706 #define DESCALE(x,n) (((x) + (1 << ((n)-1))) >> n)
710 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_382683433
)=FIX64(0.382683433, 14);
711 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_541196100
)=FIX64(0.541196100, 14);
712 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_707106781
)=FIX64(0.707106781, 14);
713 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_306562965
)=FIX64(1.306562965, 14);
715 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562_A
)=FIX64(1.414213562, 14);
717 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_847759065
)=FIX64(1.847759065, 13);
718 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_2_613125930
)=FIX64(-2.613125930, 13); //-
719 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562
)=FIX64(1.414213562, 13);
720 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_082392200
)=FIX64(1.082392200, 13);
721 //for t3,t5,t7 == 0 shortcut
722 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_847759065
)=FIX64(0.847759065, 14);
723 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_566454497
)=FIX64(0.566454497, 14);
724 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_198912367
)=FIX64(0.198912367, 14);
726 DECLARE_ASM_CONST(8, uint64_t, MM_DESCALE_RND
)=C64(4);
727 DECLARE_ASM_CONST(8, uint64_t, MM_2
)=C64(2);
729 #else /* !HAVE_MMX */
731 typedef int32_t int_simd16_t
;
732 static const int16_t FIX_0_382683433
=FIX(0.382683433, 14);
733 static const int16_t FIX_0_541196100
=FIX(0.541196100, 14);
734 static const int16_t FIX_0_707106781
=FIX(0.707106781, 14);
735 static const int16_t FIX_1_306562965
=FIX(1.306562965, 14);
736 static const int16_t FIX_1_414213562_A
=FIX(1.414213562, 14);
737 static const int16_t FIX_1_847759065
=FIX(1.847759065, 13);
738 static const int16_t FIX_2_613125930
=FIX(-2.613125930, 13); //-
739 static const int16_t FIX_1_414213562
=FIX(1.414213562, 13);
740 static const int16_t FIX_1_082392200
=FIX(1.082392200, 13);
746 static void column_fidct_c(int16_t* thr_adr
, DCTELEM
*data
, DCTELEM
*output
, int cnt
)
748 int_simd16_t tmp0
, tmp1
, tmp2
, tmp3
, tmp4
, tmp5
, tmp6
, tmp7
;
749 int_simd16_t tmp10
, tmp11
, tmp12
, tmp13
;
750 int_simd16_t z1
,z2
,z3
,z4
,z5
, z10
, z11
, z12
, z13
;
751 int_simd16_t d0
, d1
, d2
, d3
, d4
, d5
, d6
, d7
;
761 for (; cnt
> 0; cnt
-=2) { //start positions
762 threshold
=(int16_t*)thr_adr
;//threshold_mtx
763 for (ctr
= DCTSIZE
; ctr
> 0; ctr
--) {
764 // Process columns from input, add to output.
765 tmp0
= dataptr
[DCTSIZE
*0] + dataptr
[DCTSIZE
*7];
766 tmp7
= dataptr
[DCTSIZE
*0] - dataptr
[DCTSIZE
*7];
768 tmp1
= dataptr
[DCTSIZE
*1] + dataptr
[DCTSIZE
*6];
769 tmp6
= dataptr
[DCTSIZE
*1] - dataptr
[DCTSIZE
*6];
771 tmp2
= dataptr
[DCTSIZE
*2] + dataptr
[DCTSIZE
*5];
772 tmp5
= dataptr
[DCTSIZE
*2] - dataptr
[DCTSIZE
*5];
774 tmp3
= dataptr
[DCTSIZE
*3] + dataptr
[DCTSIZE
*4];
775 tmp4
= dataptr
[DCTSIZE
*3] - dataptr
[DCTSIZE
*4];
787 z1
= MULTIPLY16H((tmp12
+ tmp13
) <<2, FIX_0_707106781
);
793 THRESHOLD(tmp0
, d0
, threshold
[0*8]);
794 THRESHOLD(tmp1
, d2
, threshold
[2*8]);
795 THRESHOLD(tmp2
, d4
, threshold
[4*8]);
796 THRESHOLD(tmp3
, d6
, threshold
[6*8]);
798 tmp10
= (tmp0
+ tmp2
)>>2;
799 tmp11
= (tmp0
- tmp2
)>>2;
801 tmp13
= (tmp1
+ tmp3
)>>2; //+2 ! (psnr decides)
802 tmp12
= MULTIPLY16H((tmp1
- tmp3
), FIX_1_414213562_A
) - tmp13
; //<<2
804 tmp0
= tmp10
+ tmp13
; //->temps
805 tmp3
= tmp10
- tmp13
; //->temps
806 tmp1
= tmp11
+ tmp12
; //->temps
807 tmp2
= tmp11
- tmp12
; //->temps
815 z5
= MULTIPLY16H((tmp10
- tmp12
)<<2, FIX_0_382683433
);
816 z2
= MULTIPLY16H(tmp10
<<2, FIX_0_541196100
) + z5
;
817 z4
= MULTIPLY16H(tmp12
<<2, FIX_1_306562965
) + z5
;
818 z3
= MULTIPLY16H(tmp11
<<2, FIX_0_707106781
);
830 THRESHOLD(tmp4
, d1
, threshold
[1*8]);
831 THRESHOLD(tmp5
, d3
, threshold
[3*8]);
832 THRESHOLD(tmp6
, d5
, threshold
[5*8]);
833 THRESHOLD(tmp7
, d7
, threshold
[7*8]);
835 //Simd version uses here a shortcut for the tmp5,tmp6,tmp7 == 0
837 z10
= (tmp6
- tmp5
)<<1;
839 z12
= (tmp4
- tmp7
)<<1;
841 tmp7
= (z11
+ z13
)>>2; //+2 !
842 tmp11
= MULTIPLY16H((z11
- z13
)<<1, FIX_1_414213562
);
843 z5
= MULTIPLY16H(z10
+ z12
, FIX_1_847759065
);
844 tmp10
= MULTIPLY16H(z12
, FIX_1_082392200
) - z5
;
845 tmp12
= MULTIPLY16H(z10
, FIX_2_613125930
) + z5
; // - !!
851 wsptr
[DCTSIZE
*0]+= (tmp0
+ tmp7
);
852 wsptr
[DCTSIZE
*1]+= (tmp1
+ tmp6
);
853 wsptr
[DCTSIZE
*2]+= (tmp2
+ tmp5
);
854 wsptr
[DCTSIZE
*3]+= (tmp3
- tmp4
);
855 wsptr
[DCTSIZE
*4]+= (tmp3
+ tmp4
);
856 wsptr
[DCTSIZE
*5]+= (tmp2
- tmp5
);
857 wsptr
[DCTSIZE
*6]= (tmp1
- tmp6
);
858 wsptr
[DCTSIZE
*7]= (tmp0
- tmp7
);
860 dataptr
++; //next column
864 dataptr
+=8; //skip each second start pos
871 static void column_fidct_mmx(int16_t* thr_adr
, DCTELEM
*data
, DCTELEM
*output
, int cnt
)
873 uint64_t __attribute__((aligned(8))) temps
[4];
877 "movq "DCTSIZE_S
"*0*2(%%"REG_S
"), %%mm1 \n\t"
879 "movq "DCTSIZE_S
"*3*2(%%"REG_S
"), %%mm7 \n\t"
880 "movq %%mm1, %%mm0 \n\t"
882 "paddw "DCTSIZE_S
"*7*2(%%"REG_S
"), %%mm1 \n\t" //t0
883 "movq %%mm7, %%mm3 \n\t"
885 "paddw "DCTSIZE_S
"*4*2(%%"REG_S
"), %%mm7 \n\t" //t3
886 "movq %%mm1, %%mm5 \n\t"
888 "movq "DCTSIZE_S
"*1*2(%%"REG_S
"), %%mm6 \n\t"
889 "psubw %%mm7, %%mm1 \n\t" //t13
891 "movq "DCTSIZE_S
"*2*2(%%"REG_S
"), %%mm2 \n\t"
892 "movq %%mm6, %%mm4 \n\t"
894 "paddw "DCTSIZE_S
"*6*2(%%"REG_S
"), %%mm6 \n\t" //t1
895 "paddw %%mm7, %%mm5 \n\t" //t10
897 "paddw "DCTSIZE_S
"*5*2(%%"REG_S
"), %%mm2 \n\t" //t2
898 "movq %%mm6, %%mm7 \n\t"
900 "paddw %%mm2, %%mm6 \n\t" //t11
901 "psubw %%mm2, %%mm7 \n\t" //t12
903 "movq %%mm5, %%mm2 \n\t"
904 "paddw %%mm6, %%mm5 \n\t" //d0
905 // i0 t13 t12 i3 i1 d0 - d4
906 "psubw %%mm6, %%mm2 \n\t" //d4
907 "paddw %%mm1, %%mm7 \n\t"
909 "movq 4*16(%%"REG_d
"), %%mm6 \n\t"
910 "psllw $2, %%mm7 \n\t"
912 "psubw 0*16(%%"REG_d
"), %%mm5 \n\t"
913 "psubw %%mm6, %%mm2 \n\t"
915 "paddusw 0*16(%%"REG_d
"), %%mm5 \n\t"
916 "paddusw %%mm6, %%mm2 \n\t"
918 "pmulhw "MANGLE(MM_FIX_0_707106781
)", %%mm7 \n\t"
920 "paddw 0*16(%%"REG_d
"), %%mm5 \n\t"
921 "paddw %%mm6, %%mm2 \n\t"
923 "psubusw 0*16(%%"REG_d
"), %%mm5 \n\t"
924 "psubusw %%mm6, %%mm2 \n\t"
926 //This func is totally compute-bound, operates at huge speed. So, DC shortcut
927 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
928 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
929 "paddw "MANGLE(MM_2
)", %%mm5 \n\t"
930 "movq %%mm2, %%mm6 \n\t"
932 "paddw %%mm5, %%mm2 \n\t"
933 "psubw %%mm6, %%mm5 \n\t"
935 "movq %%mm1, %%mm6 \n\t"
936 "paddw %%mm7, %%mm1 \n\t" //d2
938 "psubw 2*16(%%"REG_d
"), %%mm1 \n\t"
939 "psubw %%mm7, %%mm6 \n\t" //d6
941 "movq 6*16(%%"REG_d
"), %%mm7 \n\t"
942 "psraw $2, %%mm5 \n\t"
944 "paddusw 2*16(%%"REG_d
"), %%mm1 \n\t"
945 "psubw %%mm7, %%mm6 \n\t"
946 // t7 d2 /t11 t4 t6 - d6 /t10
948 "paddw 2*16(%%"REG_d
"), %%mm1 \n\t"
949 "paddusw %%mm7, %%mm6 \n\t"
951 "psubusw 2*16(%%"REG_d
"), %%mm1 \n\t"
952 "paddw %%mm7, %%mm6 \n\t"
954 "psubw "DCTSIZE_S
"*4*2(%%"REG_S
"), %%mm3 \n\t"
955 "psubusw %%mm7, %%mm6 \n\t"
957 //movq [edi+"DCTSIZE_S"*2*2], mm1
958 //movq [edi+"DCTSIZE_S"*6*2], mm6
959 "movq %%mm1, %%mm7 \n\t"
960 "psraw $2, %%mm2 \n\t"
962 "psubw "DCTSIZE_S
"*6*2(%%"REG_S
"), %%mm4 \n\t"
963 "psubw %%mm6, %%mm1 \n\t"
965 "psubw "DCTSIZE_S
"*7*2(%%"REG_S
"), %%mm0 \n\t"
966 "paddw %%mm7, %%mm6 \n\t" //'t13
968 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
969 "movq %%mm2, %%mm7 \n\t"
971 "pmulhw "MANGLE(MM_FIX_1_414213562_A
)", %%mm1 \n\t"
972 "paddw %%mm6, %%mm2 \n\t" //'t0
974 "movq %%mm2, 0*8+%3 \n\t" //!
975 "psubw %%mm6, %%mm7 \n\t" //'t3
977 "movq "DCTSIZE_S
"*2*2(%%"REG_S
"), %%mm2 \n\t"
978 "psubw %%mm6, %%mm1 \n\t" //'t12
980 "psubw "DCTSIZE_S
"*5*2(%%"REG_S
"), %%mm2 \n\t" //t5
981 "movq %%mm5, %%mm6 \n\t"
983 "movq %%mm7, 3*8+%3 \n\t"
984 "paddw %%mm2, %%mm3 \n\t" //t10
986 "paddw %%mm4, %%mm2 \n\t" //t11
987 "paddw %%mm0, %%mm4 \n\t" //t12
989 "movq %%mm3, %%mm7 \n\t"
990 "psubw %%mm4, %%mm3 \n\t"
992 "psllw $2, %%mm3 \n\t"
993 "psllw $2, %%mm7 \n\t" //opt for P6
995 "pmulhw "MANGLE(MM_FIX_0_382683433
)", %%mm3 \n\t"
996 "psllw $2, %%mm4 \n\t"
998 "pmulhw "MANGLE(MM_FIX_0_541196100
)", %%mm7 \n\t"
999 "psllw $2, %%mm2 \n\t"
1001 "pmulhw "MANGLE(MM_FIX_1_306562965
)", %%mm4 \n\t"
1002 "paddw %%mm1, %%mm5 \n\t" //'t1
1004 "pmulhw "MANGLE(MM_FIX_0_707106781
)", %%mm2 \n\t"
1005 "psubw %%mm1, %%mm6 \n\t" //'t2
1006 // t7 't12 't11 t4 t6 - 't13 't10 ---
1008 "paddw %%mm3, %%mm7 \n\t" //z2
1010 "movq %%mm5, 1*8+%3 \n\t"
1011 "paddw %%mm3, %%mm4 \n\t" //z4
1013 "movq 3*16(%%"REG_d
"), %%mm3 \n\t"
1014 "movq %%mm0, %%mm1 \n\t"
1016 "movq %%mm6, 2*8+%3 \n\t"
1017 "psubw %%mm2, %%mm1 \n\t" //z13
1020 "paddw %%mm2, %%mm0 \n\t" //z11
1021 "movq %%mm1, %%mm5 \n\t"
1023 "movq 5*16(%%"REG_d
"), %%mm2 \n\t"
1024 "psubw %%mm7, %%mm1 \n\t" //d3
1026 "paddw %%mm7, %%mm5 \n\t" //d5
1027 "psubw %%mm3, %%mm1 \n\t"
1029 "movq 1*16(%%"REG_d
"), %%mm7 \n\t"
1030 "psubw %%mm2, %%mm5 \n\t"
1032 "movq %%mm0, %%mm6 \n\t"
1033 "paddw %%mm4, %%mm0 \n\t" //d1
1035 "paddusw %%mm3, %%mm1 \n\t"
1036 "psubw %%mm4, %%mm6 \n\t" //d7
1038 // d1 d3 - - - d5 d7 -
1039 "movq 7*16(%%"REG_d
"), %%mm4 \n\t"
1040 "psubw %%mm7, %%mm0 \n\t"
1042 "psubw %%mm4, %%mm6 \n\t"
1043 "paddusw %%mm2, %%mm5 \n\t"
1045 "paddusw %%mm4, %%mm6 \n\t"
1046 "paddw %%mm3, %%mm1 \n\t"
1048 "paddw %%mm2, %%mm5 \n\t"
1049 "paddw %%mm4, %%mm6 \n\t"
1051 "psubusw %%mm3, %%mm1 \n\t"
1052 "psubusw %%mm2, %%mm5 \n\t"
1054 "psubusw %%mm4, %%mm6 \n\t"
1055 "movq %%mm1, %%mm4 \n\t"
1057 "por %%mm5, %%mm4 \n\t"
1058 "paddusw %%mm7, %%mm0 \n\t"
1060 "por %%mm6, %%mm4 \n\t"
1061 "paddw %%mm7, %%mm0 \n\t"
1063 "packssdw %%mm4, %%mm4 \n\t"
1064 "psubusw %%mm7, %%mm0 \n\t"
1066 "movd %%mm4, %%"REG_a
" \n\t"
1067 "or %%"REG_a
", %%"REG_a
" \n\t"
1069 //movq [edi+"DCTSIZE_S"*3*2], mm1
1070 //movq [edi+"DCTSIZE_S"*5*2], mm5
1071 //movq [edi+"DCTSIZE_S"*1*2], mm0
1072 //movq [edi+"DCTSIZE_S"*7*2], mm6
1073 // t4 t5 - - - t6 t7 -
1074 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
1075 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
1076 "movq 0*8+%3, %%mm4 \n\t"
1077 "movq %%mm0, %%mm1 \n\t"
1079 "pmulhw "MANGLE(MM_FIX_0_847759065
)", %%mm0 \n\t" //tmp6
1080 "movq %%mm1, %%mm2 \n\t"
1082 "movq "DCTSIZE_S
"*0*2(%%"REG_D
"), %%mm5 \n\t"
1083 "movq %%mm2, %%mm3 \n\t"
1085 "pmulhw "MANGLE(MM_FIX_0_566454497
)", %%mm1 \n\t" //tmp5
1086 "paddw %%mm4, %%mm5 \n\t"
1088 "movq 1*8+%3, %%mm6 \n\t"
1090 "psraw $2, %%mm3 \n\t" //tmp7
1092 "pmulhw "MANGLE(MM_FIX_0_198912367
)", %%mm2 \n\t" //-tmp4
1093 "psubw %%mm3, %%mm4 \n\t"
1095 "movq "DCTSIZE_S
"*1*2(%%"REG_D
"), %%mm7 \n\t"
1096 "paddw %%mm3, %%mm5 \n\t"
1098 "movq %%mm4, "DCTSIZE_S
"*7*2(%%"REG_D
") \n\t"
1099 "paddw %%mm6, %%mm7 \n\t"
1101 "movq 2*8+%3, %%mm3 \n\t"
1102 "psubw %%mm0, %%mm6 \n\t"
1104 "movq "DCTSIZE_S
"*2*2(%%"REG_D
"), %%mm4 \n\t"
1105 "paddw %%mm0, %%mm7 \n\t"
1107 "movq %%mm5, "DCTSIZE_S
"*0*2(%%"REG_D
") \n\t"
1108 "paddw %%mm3, %%mm4 \n\t"
1110 "movq %%mm6, "DCTSIZE_S
"*6*2(%%"REG_D
") \n\t"
1111 "psubw %%mm1, %%mm3 \n\t"
1113 "movq "DCTSIZE_S
"*5*2(%%"REG_D
"), %%mm5 \n\t"
1114 "paddw %%mm1, %%mm4 \n\t"
1116 "movq "DCTSIZE_S
"*3*2(%%"REG_D
"), %%mm6 \n\t"
1117 "paddw %%mm3, %%mm5 \n\t"
1119 "movq 3*8+%3, %%mm0 \n\t"
1120 "add $8, %%"REG_S
" \n\t"
1122 "movq %%mm7, "DCTSIZE_S
"*1*2(%%"REG_D
") \n\t"
1123 "paddw %%mm0, %%mm6 \n\t"
1125 "movq %%mm4, "DCTSIZE_S
"*2*2(%%"REG_D
") \n\t"
1126 "psubw %%mm2, %%mm0 \n\t"
1128 "movq "DCTSIZE_S
"*4*2(%%"REG_D
"), %%mm7 \n\t"
1129 "paddw %%mm2, %%mm6 \n\t"
1131 "movq %%mm5, "DCTSIZE_S
"*5*2(%%"REG_D
") \n\t"
1132 "paddw %%mm0, %%mm7 \n\t"
1134 "movq %%mm6, "DCTSIZE_S
"*3*2(%%"REG_D
") \n\t"
1136 "movq %%mm7, "DCTSIZE_S
"*4*2(%%"REG_D
") \n\t"
1137 "add $8, %%"REG_D
" \n\t"
1142 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
1146 "movq %%mm5, %%mm3 \n\t"
1147 "psubw %%mm1, %%mm5 \n\t"
1149 "psllw $1, %%mm5 \n\t" //'z10
1150 "paddw %%mm1, %%mm3 \n\t" //'z13
1152 "movq %%mm0, %%mm2 \n\t"
1153 "psubw %%mm6, %%mm0 \n\t"
1155 "movq %%mm5, %%mm1 \n\t"
1156 "psllw $1, %%mm0 \n\t" //'z12
1158 "pmulhw "MANGLE(MM_FIX_2_613125930
)", %%mm1 \n\t" //-
1159 "paddw %%mm0, %%mm5 \n\t"
1161 "pmulhw "MANGLE(MM_FIX_1_847759065
)", %%mm5 \n\t" //'z5
1162 "paddw %%mm6, %%mm2 \n\t" //'z11
1164 "pmulhw "MANGLE(MM_FIX_1_082392200
)", %%mm0 \n\t"
1165 "movq %%mm2, %%mm7 \n\t"
1168 "movq 0*8+%3, %%mm4 \n\t"
1169 "psubw %%mm3, %%mm2 \n\t"
1171 "psllw $1, %%mm2 \n\t"
1172 "paddw %%mm3, %%mm7 \n\t" //'t7
1174 "pmulhw "MANGLE(MM_FIX_1_414213562
)", %%mm2 \n\t" //'t11
1175 "movq %%mm4, %%mm6 \n\t"
1177 "psraw $2, %%mm7 \n\t"
1179 "paddw "DCTSIZE_S
"*0*2(%%"REG_D
"), %%mm4 \n\t"
1180 "psubw %%mm7, %%mm6 \n\t"
1182 "movq 1*8+%3, %%mm3 \n\t"
1183 "paddw %%mm7, %%mm4 \n\t"
1185 "movq %%mm6, "DCTSIZE_S
"*7*2(%%"REG_D
") \n\t"
1186 "paddw %%mm5, %%mm1 \n\t" //'t12
1188 "movq %%mm4, "DCTSIZE_S
"*0*2(%%"REG_D
") \n\t"
1189 "psubw %%mm7, %%mm1 \n\t" //'t6
1191 "movq 2*8+%3, %%mm7 \n\t"
1192 "psubw %%mm5, %%mm0 \n\t" //'t10
1194 "movq 3*8+%3, %%mm6 \n\t"
1195 "movq %%mm3, %%mm5 \n\t"
1197 "paddw "DCTSIZE_S
"*1*2(%%"REG_D
"), %%mm3 \n\t"
1198 "psubw %%mm1, %%mm5 \n\t"
1200 "psubw %%mm1, %%mm2 \n\t" //'t5
1201 "paddw %%mm1, %%mm3 \n\t"
1203 "movq %%mm5, "DCTSIZE_S
"*6*2(%%"REG_D
") \n\t"
1204 "movq %%mm7, %%mm4 \n\t"
1206 "paddw "DCTSIZE_S
"*2*2(%%"REG_D
"), %%mm7 \n\t"
1207 "psubw %%mm2, %%mm4 \n\t"
1209 "paddw "DCTSIZE_S
"*5*2(%%"REG_D
"), %%mm4 \n\t"
1210 "paddw %%mm2, %%mm7 \n\t"
1212 "movq %%mm3, "DCTSIZE_S
"*1*2(%%"REG_D
") \n\t"
1213 "paddw %%mm2, %%mm0 \n\t" //'t4
1215 // 't4 't6 't5 - - - - 't7
1216 "movq %%mm7, "DCTSIZE_S
"*2*2(%%"REG_D
") \n\t"
1217 "movq %%mm6, %%mm1 \n\t"
1219 "paddw "DCTSIZE_S
"*4*2(%%"REG_D
"), %%mm6 \n\t"
1220 "psubw %%mm0, %%mm1 \n\t"
1222 "paddw "DCTSIZE_S
"*3*2(%%"REG_D
"), %%mm1 \n\t"
1223 "paddw %%mm0, %%mm6 \n\t"
1225 "movq %%mm4, "DCTSIZE_S
"*5*2(%%"REG_D
") \n\t"
1226 "add $8, %%"REG_S
" \n\t"
1228 "movq %%mm6, "DCTSIZE_S
"*4*2(%%"REG_D
") \n\t"
1230 "movq %%mm1, "DCTSIZE_S
"*3*2(%%"REG_D
") \n\t"
1231 "add $8, %%"REG_D
" \n\t"
1234 //=part 2 (the same)===========================================================
1235 "movq "DCTSIZE_S
"*0*2(%%"REG_S
"), %%mm1 \n\t"
1237 "movq "DCTSIZE_S
"*3*2(%%"REG_S
"), %%mm7 \n\t"
1238 "movq %%mm1, %%mm0 \n\t"
1240 "paddw "DCTSIZE_S
"*7*2(%%"REG_S
"), %%mm1 \n\t" //t0
1241 "movq %%mm7, %%mm3 \n\t"
1243 "paddw "DCTSIZE_S
"*4*2(%%"REG_S
"), %%mm7 \n\t" //t3
1244 "movq %%mm1, %%mm5 \n\t"
1246 "movq "DCTSIZE_S
"*1*2(%%"REG_S
"), %%mm6 \n\t"
1247 "psubw %%mm7, %%mm1 \n\t" //t13
1249 "movq "DCTSIZE_S
"*2*2(%%"REG_S
"), %%mm2 \n\t"
1250 "movq %%mm6, %%mm4 \n\t"
1252 "paddw "DCTSIZE_S
"*6*2(%%"REG_S
"), %%mm6 \n\t" //t1
1253 "paddw %%mm7, %%mm5 \n\t" //t10
1255 "paddw "DCTSIZE_S
"*5*2(%%"REG_S
"), %%mm2 \n\t" //t2
1256 "movq %%mm6, %%mm7 \n\t"
1258 "paddw %%mm2, %%mm6 \n\t" //t11
1259 "psubw %%mm2, %%mm7 \n\t" //t12
1261 "movq %%mm5, %%mm2 \n\t"
1262 "paddw %%mm6, %%mm5 \n\t" //d0
1263 // i0 t13 t12 i3 i1 d0 - d4
1264 "psubw %%mm6, %%mm2 \n\t" //d4
1265 "paddw %%mm1, %%mm7 \n\t"
1267 "movq 1*8+4*16(%%"REG_d
"), %%mm6 \n\t"
1268 "psllw $2, %%mm7 \n\t"
1270 "psubw 1*8+0*16(%%"REG_d
"), %%mm5 \n\t"
1271 "psubw %%mm6, %%mm2 \n\t"
1273 "paddusw 1*8+0*16(%%"REG_d
"), %%mm5 \n\t"
1274 "paddusw %%mm6, %%mm2 \n\t"
1276 "pmulhw "MANGLE(MM_FIX_0_707106781
)", %%mm7 \n\t"
1278 "paddw 1*8+0*16(%%"REG_d
"), %%mm5 \n\t"
1279 "paddw %%mm6, %%mm2 \n\t"
1281 "psubusw 1*8+0*16(%%"REG_d
"), %%mm5 \n\t"
1282 "psubusw %%mm6, %%mm2 \n\t"
1284 //This func is totally compute-bound, operates at huge speed. So, DC shortcut
1285 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
1286 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
1287 "paddw "MANGLE(MM_2
)", %%mm5 \n\t"
1288 "movq %%mm2, %%mm6 \n\t"
1290 "paddw %%mm5, %%mm2 \n\t"
1291 "psubw %%mm6, %%mm5 \n\t"
1293 "movq %%mm1, %%mm6 \n\t"
1294 "paddw %%mm7, %%mm1 \n\t" //d2
1296 "psubw 1*8+2*16(%%"REG_d
"), %%mm1 \n\t"
1297 "psubw %%mm7, %%mm6 \n\t" //d6
1299 "movq 1*8+6*16(%%"REG_d
"), %%mm7 \n\t"
1300 "psraw $2, %%mm5 \n\t"
1302 "paddusw 1*8+2*16(%%"REG_d
"), %%mm1 \n\t"
1303 "psubw %%mm7, %%mm6 \n\t"
1304 // t7 d2 /t11 t4 t6 - d6 /t10
1306 "paddw 1*8+2*16(%%"REG_d
"), %%mm1 \n\t"
1307 "paddusw %%mm7, %%mm6 \n\t"
1309 "psubusw 1*8+2*16(%%"REG_d
"), %%mm1 \n\t"
1310 "paddw %%mm7, %%mm6 \n\t"
1312 "psubw "DCTSIZE_S
"*4*2(%%"REG_S
"), %%mm3 \n\t"
1313 "psubusw %%mm7, %%mm6 \n\t"
1315 //movq [edi+"DCTSIZE_S"*2*2], mm1
1316 //movq [edi+"DCTSIZE_S"*6*2], mm6
1317 "movq %%mm1, %%mm7 \n\t"
1318 "psraw $2, %%mm2 \n\t"
1320 "psubw "DCTSIZE_S
"*6*2(%%"REG_S
"), %%mm4 \n\t"
1321 "psubw %%mm6, %%mm1 \n\t"
1323 "psubw "DCTSIZE_S
"*7*2(%%"REG_S
"), %%mm0 \n\t"
1324 "paddw %%mm7, %%mm6 \n\t" //'t13
1326 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
1327 "movq %%mm2, %%mm7 \n\t"
1329 "pmulhw "MANGLE(MM_FIX_1_414213562_A
)", %%mm1 \n\t"
1330 "paddw %%mm6, %%mm2 \n\t" //'t0
1332 "movq %%mm2, 0*8+%3 \n\t" //!
1333 "psubw %%mm6, %%mm7 \n\t" //'t3
1335 "movq "DCTSIZE_S
"*2*2(%%"REG_S
"), %%mm2 \n\t"
1336 "psubw %%mm6, %%mm1 \n\t" //'t12
1338 "psubw "DCTSIZE_S
"*5*2(%%"REG_S
"), %%mm2 \n\t" //t5
1339 "movq %%mm5, %%mm6 \n\t"
1341 "movq %%mm7, 3*8+%3 \n\t"
1342 "paddw %%mm2, %%mm3 \n\t" //t10
1344 "paddw %%mm4, %%mm2 \n\t" //t11
1345 "paddw %%mm0, %%mm4 \n\t" //t12
1347 "movq %%mm3, %%mm7 \n\t"
1348 "psubw %%mm4, %%mm3 \n\t"
1350 "psllw $2, %%mm3 \n\t"
1351 "psllw $2, %%mm7 \n\t" //opt for P6
1353 "pmulhw "MANGLE(MM_FIX_0_382683433
)", %%mm3 \n\t"
1354 "psllw $2, %%mm4 \n\t"
1356 "pmulhw "MANGLE(MM_FIX_0_541196100
)", %%mm7 \n\t"
1357 "psllw $2, %%mm2 \n\t"
1359 "pmulhw "MANGLE(MM_FIX_1_306562965
)", %%mm4 \n\t"
1360 "paddw %%mm1, %%mm5 \n\t" //'t1
1362 "pmulhw "MANGLE(MM_FIX_0_707106781
)", %%mm2 \n\t"
1363 "psubw %%mm1, %%mm6 \n\t" //'t2
1364 // t7 't12 't11 t4 t6 - 't13 't10 ---
1366 "paddw %%mm3, %%mm7 \n\t" //z2
1368 "movq %%mm5, 1*8+%3 \n\t"
1369 "paddw %%mm3, %%mm4 \n\t" //z4
1371 "movq 1*8+3*16(%%"REG_d
"), %%mm3 \n\t"
1372 "movq %%mm0, %%mm1 \n\t"
1374 "movq %%mm6, 2*8+%3 \n\t"
1375 "psubw %%mm2, %%mm1 \n\t" //z13
1378 "paddw %%mm2, %%mm0 \n\t" //z11
1379 "movq %%mm1, %%mm5 \n\t"
1381 "movq 1*8+5*16(%%"REG_d
"), %%mm2 \n\t"
1382 "psubw %%mm7, %%mm1 \n\t" //d3
1384 "paddw %%mm7, %%mm5 \n\t" //d5
1385 "psubw %%mm3, %%mm1 \n\t"
1387 "movq 1*8+1*16(%%"REG_d
"), %%mm7 \n\t"
1388 "psubw %%mm2, %%mm5 \n\t"
1390 "movq %%mm0, %%mm6 \n\t"
1391 "paddw %%mm4, %%mm0 \n\t" //d1
1393 "paddusw %%mm3, %%mm1 \n\t"
1394 "psubw %%mm4, %%mm6 \n\t" //d7
1396 // d1 d3 - - - d5 d7 -
1397 "movq 1*8+7*16(%%"REG_d
"), %%mm4 \n\t"
1398 "psubw %%mm7, %%mm0 \n\t"
1400 "psubw %%mm4, %%mm6 \n\t"
1401 "paddusw %%mm2, %%mm5 \n\t"
1403 "paddusw %%mm4, %%mm6 \n\t"
1404 "paddw %%mm3, %%mm1 \n\t"
1406 "paddw %%mm2, %%mm5 \n\t"
1407 "paddw %%mm4, %%mm6 \n\t"
1409 "psubusw %%mm3, %%mm1 \n\t"
1410 "psubusw %%mm2, %%mm5 \n\t"
1412 "psubusw %%mm4, %%mm6 \n\t"
1413 "movq %%mm1, %%mm4 \n\t"
1415 "por %%mm5, %%mm4 \n\t"
1416 "paddusw %%mm7, %%mm0 \n\t"
1418 "por %%mm6, %%mm4 \n\t"
1419 "paddw %%mm7, %%mm0 \n\t"
1421 "packssdw %%mm4, %%mm4 \n\t"
1422 "psubusw %%mm7, %%mm0 \n\t"
1424 "movd %%mm4, %%"REG_a
" \n\t"
1425 "or %%"REG_a
", %%"REG_a
" \n\t"
1427 //movq [edi+"DCTSIZE_S"*3*2], mm1
1428 //movq [edi+"DCTSIZE_S"*5*2], mm5
1429 //movq [edi+"DCTSIZE_S"*1*2], mm0
1430 //movq [edi+"DCTSIZE_S"*7*2], mm6
1431 // t4 t5 - - - t6 t7 -
1432 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
1433 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
1434 "movq 0*8+%3, %%mm4 \n\t"
1435 "movq %%mm0, %%mm1 \n\t"
1437 "pmulhw "MANGLE(MM_FIX_0_847759065
)", %%mm0 \n\t" //tmp6
1438 "movq %%mm1, %%mm2 \n\t"
1440 "movq "DCTSIZE_S
"*0*2(%%"REG_D
"), %%mm5 \n\t"
1441 "movq %%mm2, %%mm3 \n\t"
1443 "pmulhw "MANGLE(MM_FIX_0_566454497
)", %%mm1 \n\t" //tmp5
1444 "paddw %%mm4, %%mm5 \n\t"
1446 "movq 1*8+%3, %%mm6 \n\t"
1448 "psraw $2, %%mm3 \n\t" //tmp7
1450 "pmulhw "MANGLE(MM_FIX_0_198912367
)", %%mm2 \n\t" //-tmp4
1451 "psubw %%mm3, %%mm4 \n\t"
1453 "movq "DCTSIZE_S
"*1*2(%%"REG_D
"), %%mm7 \n\t"
1454 "paddw %%mm3, %%mm5 \n\t"
1456 "movq %%mm4, "DCTSIZE_S
"*7*2(%%"REG_D
") \n\t"
1457 "paddw %%mm6, %%mm7 \n\t"
1459 "movq 2*8+%3, %%mm3 \n\t"
1460 "psubw %%mm0, %%mm6 \n\t"
1462 "movq "DCTSIZE_S
"*2*2(%%"REG_D
"), %%mm4 \n\t"
1463 "paddw %%mm0, %%mm7 \n\t"
1465 "movq %%mm5, "DCTSIZE_S
"*0*2(%%"REG_D
") \n\t"
1466 "paddw %%mm3, %%mm4 \n\t"
1468 "movq %%mm6, "DCTSIZE_S
"*6*2(%%"REG_D
") \n\t"
1469 "psubw %%mm1, %%mm3 \n\t"
1471 "movq "DCTSIZE_S
"*5*2(%%"REG_D
"), %%mm5 \n\t"
1472 "paddw %%mm1, %%mm4 \n\t"
1474 "movq "DCTSIZE_S
"*3*2(%%"REG_D
"), %%mm6 \n\t"
1475 "paddw %%mm3, %%mm5 \n\t"
1477 "movq 3*8+%3, %%mm0 \n\t"
1478 "add $24, %%"REG_S
" \n\t"
1480 "movq %%mm7, "DCTSIZE_S
"*1*2(%%"REG_D
") \n\t"
1481 "paddw %%mm0, %%mm6 \n\t"
1483 "movq %%mm4, "DCTSIZE_S
"*2*2(%%"REG_D
") \n\t"
1484 "psubw %%mm2, %%mm0 \n\t"
1486 "movq "DCTSIZE_S
"*4*2(%%"REG_D
"), %%mm7 \n\t"
1487 "paddw %%mm2, %%mm6 \n\t"
1489 "movq %%mm5, "DCTSIZE_S
"*5*2(%%"REG_D
") \n\t"
1490 "paddw %%mm0, %%mm7 \n\t"
1492 "movq %%mm6, "DCTSIZE_S
"*3*2(%%"REG_D
") \n\t"
1494 "movq %%mm7, "DCTSIZE_S
"*4*2(%%"REG_D
") \n\t"
1495 "add $24, %%"REG_D
" \n\t"
1496 "sub $2, %%"REG_c
" \n\t"
1502 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
1506 "movq %%mm5, %%mm3 \n\t"
1507 "psubw %%mm1, %%mm5 \n\t"
1509 "psllw $1, %%mm5 \n\t" //'z10
1510 "paddw %%mm1, %%mm3 \n\t" //'z13
1512 "movq %%mm0, %%mm2 \n\t"
1513 "psubw %%mm6, %%mm0 \n\t"
1515 "movq %%mm5, %%mm1 \n\t"
1516 "psllw $1, %%mm0 \n\t" //'z12
1518 "pmulhw "MANGLE(MM_FIX_2_613125930
)", %%mm1 \n\t" //-
1519 "paddw %%mm0, %%mm5 \n\t"
1521 "pmulhw "MANGLE(MM_FIX_1_847759065
)", %%mm5 \n\t" //'z5
1522 "paddw %%mm6, %%mm2 \n\t" //'z11
1524 "pmulhw "MANGLE(MM_FIX_1_082392200
)", %%mm0 \n\t"
1525 "movq %%mm2, %%mm7 \n\t"
1528 "movq 0*8+%3, %%mm4 \n\t"
1529 "psubw %%mm3, %%mm2 \n\t"
1531 "psllw $1, %%mm2 \n\t"
1532 "paddw %%mm3, %%mm7 \n\t" //'t7
1534 "pmulhw "MANGLE(MM_FIX_1_414213562
)", %%mm2 \n\t" //'t11
1535 "movq %%mm4, %%mm6 \n\t"
1537 "psraw $2, %%mm7 \n\t"
1539 "paddw "DCTSIZE_S
"*0*2(%%"REG_D
"), %%mm4 \n\t"
1540 "psubw %%mm7, %%mm6 \n\t"
1542 "movq 1*8+%3, %%mm3 \n\t"
1543 "paddw %%mm7, %%mm4 \n\t"
1545 "movq %%mm6, "DCTSIZE_S
"*7*2(%%"REG_D
") \n\t"
1546 "paddw %%mm5, %%mm1 \n\t" //'t12
1548 "movq %%mm4, "DCTSIZE_S
"*0*2(%%"REG_D
") \n\t"
1549 "psubw %%mm7, %%mm1 \n\t" //'t6
1551 "movq 2*8+%3, %%mm7 \n\t"
1552 "psubw %%mm5, %%mm0 \n\t" //'t10
1554 "movq 3*8+%3, %%mm6 \n\t"
1555 "movq %%mm3, %%mm5 \n\t"
1557 "paddw "DCTSIZE_S
"*1*2(%%"REG_D
"), %%mm3 \n\t"
1558 "psubw %%mm1, %%mm5 \n\t"
1560 "psubw %%mm1, %%mm2 \n\t" //'t5
1561 "paddw %%mm1, %%mm3 \n\t"
1563 "movq %%mm5, "DCTSIZE_S
"*6*2(%%"REG_D
") \n\t"
1564 "movq %%mm7, %%mm4 \n\t"
1566 "paddw "DCTSIZE_S
"*2*2(%%"REG_D
"), %%mm7 \n\t"
1567 "psubw %%mm2, %%mm4 \n\t"
1569 "paddw "DCTSIZE_S
"*5*2(%%"REG_D
"), %%mm4 \n\t"
1570 "paddw %%mm2, %%mm7 \n\t"
1572 "movq %%mm3, "DCTSIZE_S
"*1*2(%%"REG_D
") \n\t"
1573 "paddw %%mm2, %%mm0 \n\t" //'t4
1575 // 't4 't6 't5 - - - - 't7
1576 "movq %%mm7, "DCTSIZE_S
"*2*2(%%"REG_D
") \n\t"
1577 "movq %%mm6, %%mm1 \n\t"
1579 "paddw "DCTSIZE_S
"*4*2(%%"REG_D
"), %%mm6 \n\t"
1580 "psubw %%mm0, %%mm1 \n\t"
1582 "paddw "DCTSIZE_S
"*3*2(%%"REG_D
"), %%mm1 \n\t"
1583 "paddw %%mm0, %%mm6 \n\t"
1585 "movq %%mm4, "DCTSIZE_S
"*5*2(%%"REG_D
") \n\t"
1586 "add $24, %%"REG_S
" \n\t"
1588 "movq %%mm6, "DCTSIZE_S
"*4*2(%%"REG_D
") \n\t"
1590 "movq %%mm1, "DCTSIZE_S
"*3*2(%%"REG_D
") \n\t"
1591 "add $24, %%"REG_D
" \n\t"
1592 "sub $2, %%"REG_c
" \n\t"
1596 : "+S"(data
), "+D"(output
), "+c"(cnt
), "=o"(temps
)
1606 static void row_idct_c(DCTELEM
* workspace
,
1607 int16_t* output_adr
, int output_stride
, int cnt
)
1609 int_simd16_t tmp0
, tmp1
, tmp2
, tmp3
, tmp4
, tmp5
, tmp6
, tmp7
;
1610 int_simd16_t tmp10
, tmp11
, tmp12
, tmp13
;
1611 int_simd16_t z5
, z10
, z11
, z12
, z13
;
1617 outptr
= output_adr
;
1618 for (; cnt
> 0; cnt
--) {
1620 //Simd version reads 4x4 block and transposes it
1621 tmp10
= ( wsptr
[2] + wsptr
[3]);
1622 tmp11
= ( wsptr
[2] - wsptr
[3]);
1624 tmp13
= ( wsptr
[0] + wsptr
[1]);
1625 tmp12
= (MULTIPLY16H( wsptr
[0] - wsptr
[1], FIX_1_414213562_A
)<<2) - tmp13
;//this shift order to avoid overflow
1627 tmp0
= tmp10
+ tmp13
; //->temps
1628 tmp3
= tmp10
- tmp13
; //->temps
1629 tmp1
= tmp11
+ tmp12
;
1630 tmp2
= tmp11
- tmp12
;
1633 //Also transpose, with previous:
1635 // ---- ---- idct ||||
1636 // ---- ---- ---> ||||
1638 z13
= wsptr
[4] + wsptr
[5];
1639 z10
= wsptr
[4] - wsptr
[5];
1640 z11
= wsptr
[6] + wsptr
[7];
1641 z12
= wsptr
[6] - wsptr
[7];
1644 tmp11
= MULTIPLY16H(z11
- z13
, FIX_1_414213562
);
1646 z5
= MULTIPLY16H(z10
+ z12
, FIX_1_847759065
);
1647 tmp10
= MULTIPLY16H(z12
, FIX_1_082392200
) - z5
;
1648 tmp12
= MULTIPLY16H(z10
, FIX_2_613125930
) + z5
; // - FIX_
1650 tmp6
= (tmp12
<<3) - tmp7
;
1651 tmp5
= (tmp11
<<3) - tmp6
;
1652 tmp4
= (tmp10
<<3) + tmp5
;
1654 // Final output stage: descale and write column
1655 outptr
[0*output_stride
]+= DESCALE(tmp0
+ tmp7
, 3);
1656 outptr
[1*output_stride
]+= DESCALE(tmp1
+ tmp6
, 3);
1657 outptr
[2*output_stride
]+= DESCALE(tmp2
+ tmp5
, 3);
1658 outptr
[3*output_stride
]+= DESCALE(tmp3
- tmp4
, 3);
1659 outptr
[4*output_stride
]+= DESCALE(tmp3
+ tmp4
, 3);
1660 outptr
[5*output_stride
]+= DESCALE(tmp2
- tmp5
, 3);
1661 outptr
[6*output_stride
]+= DESCALE(tmp1
- tmp6
, 3); //no += ?
1662 outptr
[7*output_stride
]+= DESCALE(tmp0
- tmp7
, 3); //no += ?
1665 wsptr
+= DCTSIZE
; // advance pointer to next row
1669 #else /* HAVE_MMX */
1671 static void row_idct_mmx (DCTELEM
* workspace
,
1672 int16_t* output_adr
, int output_stride
, int cnt
)
1674 uint64_t __attribute__((aligned(8))) temps
[4];
1676 "lea (%%"REG_a
",%%"REG_a
",2), %%"REG_d
" \n\t"
1678 "movq "DCTSIZE_S
"*0*2(%%"REG_S
"), %%mm0 \n\t"
1681 "movq "DCTSIZE_S
"*1*2(%%"REG_S
"), %%mm1 \n\t"
1682 "movq %%mm0, %%mm4 \n\t"
1684 "movq "DCTSIZE_S
"*2*2(%%"REG_S
"), %%mm2 \n\t"
1685 "punpcklwd %%mm1, %%mm0 \n\t"
1687 "movq "DCTSIZE_S
"*3*2(%%"REG_S
"), %%mm3 \n\t"
1688 "punpckhwd %%mm1, %%mm4 \n\t"
1691 "movq %%mm2, %%mm7 \n\t"
1692 "punpcklwd %%mm3, %%mm2 \n\t"
1694 "movq %%mm0, %%mm6 \n\t"
1695 "punpckldq %%mm2, %%mm0 \n\t" //0
1697 "punpckhdq %%mm2, %%mm6 \n\t" //1
1698 "movq %%mm0, %%mm5 \n\t"
1700 "punpckhwd %%mm3, %%mm7 \n\t"
1701 "psubw %%mm6, %%mm0 \n\t"
1703 "pmulhw "MANGLE(MM_FIX_1_414213562_A
)", %%mm0 \n\t"
1704 "movq %%mm4, %%mm2 \n\t"
1706 "punpckldq %%mm7, %%mm4 \n\t" //2
1707 "paddw %%mm6, %%mm5 \n\t"
1709 "punpckhdq %%mm7, %%mm2 \n\t" //3
1710 "movq %%mm4, %%mm1 \n\t"
1712 "psllw $2, %%mm0 \n\t"
1713 "paddw %%mm2, %%mm4 \n\t" //t10
1715 "movq "DCTSIZE_S
"*0*2+"DCTSIZE_S
"(%%"REG_S
"), %%mm3 \n\t"
1716 "psubw %%mm2, %%mm1 \n\t" //t11
1718 "movq "DCTSIZE_S
"*1*2+"DCTSIZE_S
"(%%"REG_S
"), %%mm2 \n\t"
1719 "psubw %%mm5, %%mm0 \n\t"
1721 "movq %%mm4, %%mm6 \n\t"
1722 "paddw %%mm5, %%mm4 \n\t" //t0
1724 "psubw %%mm5, %%mm6 \n\t" //t3
1725 "movq %%mm1, %%mm7 \n\t"
1727 "movq "DCTSIZE_S
"*2*2+"DCTSIZE_S
"(%%"REG_S
"), %%mm5 \n\t"
1728 "paddw %%mm0, %%mm1 \n\t" //t1
1730 "movq %%mm4, 0*8+%3 \n\t" //t0
1731 "movq %%mm3, %%mm4 \n\t"
1733 "movq %%mm6, 1*8+%3 \n\t" //t3
1734 "punpcklwd %%mm2, %%mm3 \n\t"
1737 "movq "DCTSIZE_S
"*3*2+"DCTSIZE_S
"(%%"REG_S
"), %%mm6 \n\t"
1738 "punpckhwd %%mm2, %%mm4 \n\t"
1740 "movq %%mm5, %%mm2 \n\t"
1741 "punpcklwd %%mm6, %%mm5 \n\t"
1743 "psubw %%mm0, %%mm7 \n\t" //t2
1744 "punpckhwd %%mm6, %%mm2 \n\t"
1746 "movq %%mm3, %%mm0 \n\t"
1747 "punpckldq %%mm5, %%mm3 \n\t" //4
1749 "punpckhdq %%mm5, %%mm0 \n\t" //5
1750 "movq %%mm4, %%mm5 \n\t"
1753 "movq %%mm3, %%mm6 \n\t"
1754 "punpckldq %%mm2, %%mm4 \n\t" //6
1756 "psubw %%mm0, %%mm3 \n\t" //z10
1757 "punpckhdq %%mm2, %%mm5 \n\t" //7
1759 "paddw %%mm0, %%mm6 \n\t" //z13
1760 "movq %%mm4, %%mm2 \n\t"
1762 "movq %%mm3, %%mm0 \n\t"
1763 "psubw %%mm5, %%mm4 \n\t" //z12
1765 "pmulhw "MANGLE(MM_FIX_2_613125930
)", %%mm0 \n\t" //-
1766 "paddw %%mm4, %%mm3 \n\t"
1768 "pmulhw "MANGLE(MM_FIX_1_847759065
)", %%mm3 \n\t" //z5
1769 "paddw %%mm5, %%mm2 \n\t" //z11 >
1771 "pmulhw "MANGLE(MM_FIX_1_082392200
)", %%mm4 \n\t"
1772 "movq %%mm2, %%mm5 \n\t"
1774 "psubw %%mm6, %%mm2 \n\t"
1775 "paddw %%mm6, %%mm5 \n\t" //t7
1777 "pmulhw "MANGLE(MM_FIX_1_414213562
)", %%mm2 \n\t" //t11
1778 "paddw %%mm3, %%mm0 \n\t" //t12
1780 "psllw $3, %%mm0 \n\t"
1781 "psubw %%mm3, %%mm4 \n\t" //t10
1783 "movq 0*8+%3, %%mm6 \n\t"
1784 "movq %%mm1, %%mm3 \n\t"
1786 "psllw $3, %%mm4 \n\t"
1787 "psubw %%mm5, %%mm0 \n\t" //t6
1789 "psllw $3, %%mm2 \n\t"
1790 "paddw %%mm0, %%mm1 \n\t" //d1
1792 "psubw %%mm0, %%mm2 \n\t" //t5
1793 "psubw %%mm0, %%mm3 \n\t" //d6
1795 "paddw %%mm2, %%mm4 \n\t" //t4
1796 "movq %%mm7, %%mm0 \n\t"
1798 "paddw %%mm2, %%mm7 \n\t" //d2
1799 "psubw %%mm2, %%mm0 \n\t" //d5
1801 "movq "MANGLE(MM_DESCALE_RND
)", %%mm2 \n\t" //4
1802 "psubw %%mm5, %%mm6 \n\t" //d7
1804 "paddw 0*8+%3, %%mm5 \n\t" //d0
1805 "paddw %%mm2, %%mm1 \n\t"
1807 "paddw %%mm2, %%mm5 \n\t"
1808 "psraw $3, %%mm1 \n\t"
1810 "paddw %%mm2, %%mm7 \n\t"
1811 "psraw $3, %%mm5 \n\t"
1813 "paddw (%%"REG_D
"), %%mm5 \n\t"
1814 "psraw $3, %%mm7 \n\t"
1816 "paddw (%%"REG_D
",%%"REG_a
",), %%mm1 \n\t"
1817 "paddw %%mm2, %%mm0 \n\t"
1819 "paddw (%%"REG_D
",%%"REG_a
",2), %%mm7 \n\t"
1820 "paddw %%mm2, %%mm3 \n\t"
1822 "movq %%mm5, (%%"REG_D
") \n\t"
1823 "paddw %%mm2, %%mm6 \n\t"
1825 "movq %%mm1, (%%"REG_D
",%%"REG_a
",) \n\t"
1826 "psraw $3, %%mm0 \n\t"
1828 "movq %%mm7, (%%"REG_D
",%%"REG_a
",2) \n\t"
1829 "add %%"REG_d
", %%"REG_D
" \n\t" //3*ls
1831 "movq 1*8+%3, %%mm5 \n\t" //t3
1832 "psraw $3, %%mm3 \n\t"
1834 "paddw (%%"REG_D
",%%"REG_a
",2), %%mm0 \n\t"
1835 "psubw %%mm4, %%mm5 \n\t" //d3
1837 "paddw (%%"REG_D
",%%"REG_d
",), %%mm3 \n\t"
1838 "psraw $3, %%mm6 \n\t"
1840 "paddw 1*8+%3, %%mm4 \n\t" //d4
1841 "paddw %%mm2, %%mm5 \n\t"
1843 "paddw (%%"REG_D
",%%"REG_a
",4), %%mm6 \n\t"
1844 "paddw %%mm2, %%mm4 \n\t"
1846 "movq %%mm0, (%%"REG_D
",%%"REG_a
",2) \n\t"
1847 "psraw $3, %%mm5 \n\t"
1849 "paddw (%%"REG_D
"), %%mm5 \n\t"
1850 "psraw $3, %%mm4 \n\t"
1852 "paddw (%%"REG_D
",%%"REG_a
",), %%mm4 \n\t"
1853 "add $"DCTSIZE_S
"*2*4, %%"REG_S
" \n\t" //4 rows
1855 "movq %%mm3, (%%"REG_D
",%%"REG_d
",) \n\t"
1856 "movq %%mm6, (%%"REG_D
",%%"REG_a
",4) \n\t"
1857 "movq %%mm5, (%%"REG_D
") \n\t"
1858 "movq %%mm4, (%%"REG_D
",%%"REG_a
",) \n\t"
1860 "sub %%"REG_d
", %%"REG_D
" \n\t"
1861 "add $8, %%"REG_D
" \n\t"
1862 "dec %%"REG_c
" \n\t"
1865 : "+S"(workspace
), "+D"(output_adr
), "+c"(cnt
), "=o"(temps
)
1866 : "a"(output_stride
*sizeof(short))
1875 static void row_fdct_c(DCTELEM
*data
, const uint8_t *pixels
, int line_size
, int cnt
)
1877 int_simd16_t tmp0
, tmp1
, tmp2
, tmp3
, tmp4
, tmp5
, tmp6
, tmp7
;
1878 int_simd16_t tmp10
, tmp11
, tmp12
, tmp13
;
1879 int_simd16_t z1
, z2
, z3
, z4
, z5
, z11
, z13
;
1883 // Pass 1: process rows.
1886 for (; cnt
> 0; cnt
--) {
1887 tmp0
= pixels
[line_size
*0] + pixels
[line_size
*7];
1888 tmp7
= pixels
[line_size
*0] - pixels
[line_size
*7];
1889 tmp1
= pixels
[line_size
*1] + pixels
[line_size
*6];
1890 tmp6
= pixels
[line_size
*1] - pixels
[line_size
*6];
1891 tmp2
= pixels
[line_size
*2] + pixels
[line_size
*5];
1892 tmp5
= pixels
[line_size
*2] - pixels
[line_size
*5];
1893 tmp3
= pixels
[line_size
*3] + pixels
[line_size
*4];
1894 tmp4
= pixels
[line_size
*3] - pixels
[line_size
*4];
1898 tmp10
= tmp0
+ tmp3
;
1899 tmp13
= tmp0
- tmp3
;
1900 tmp11
= tmp1
+ tmp2
;
1901 tmp12
= tmp1
- tmp2
;
1902 //Even columns are written first, this leads to different order of columns
1903 //in column_fidct(), but they are processed independently, so all ok.
1904 //Later in the row_idct() columns readed at the same order.
1905 dataptr
[2] = tmp10
+ tmp11
;
1906 dataptr
[3] = tmp10
- tmp11
;
1908 z1
= MULTIPLY16H((tmp12
+ tmp13
)<<2, FIX_0_707106781
);
1909 dataptr
[0] = tmp13
+ z1
;
1910 dataptr
[1] = tmp13
- z1
;
1914 tmp10
= (tmp4
+ tmp5
) <<2;
1915 tmp11
= (tmp5
+ tmp6
) <<2;
1916 tmp12
= (tmp6
+ tmp7
) <<2;
1918 z5
= MULTIPLY16H(tmp10
- tmp12
, FIX_0_382683433
);
1919 z2
= MULTIPLY16H(tmp10
, FIX_0_541196100
) + z5
;
1920 z4
= MULTIPLY16H(tmp12
, FIX_1_306562965
) + z5
;
1921 z3
= MULTIPLY16H(tmp11
, FIX_0_707106781
);
1926 dataptr
[4] = z13
+ z2
;
1927 dataptr
[5] = z13
- z2
;
1928 dataptr
[6] = z11
+ z4
;
1929 dataptr
[7] = z11
- z4
;
1931 pixels
++; // advance pointer to next column
1936 #else /* HAVE_MMX */
1938 static void row_fdct_mmx(DCTELEM
*data
, const uint8_t *pixels
, int line_size
, int cnt
)
1940 uint64_t __attribute__((aligned(8))) temps
[4];
1942 "lea (%%"REG_a
",%%"REG_a
",2), %%"REG_d
" \n\t"
1944 "movd (%%"REG_S
"), %%mm0 \n\t"
1945 "pxor %%mm7, %%mm7 \n\t"
1947 "movd (%%"REG_S
",%%"REG_a
",), %%mm1 \n\t"
1948 "punpcklbw %%mm7, %%mm0 \n\t"
1950 "movd (%%"REG_S
",%%"REG_a
",2), %%mm2 \n\t"
1951 "punpcklbw %%mm7, %%mm1 \n\t"
1953 "punpcklbw %%mm7, %%mm2 \n\t"
1954 "add %%"REG_d
", %%"REG_S
" \n\t"
1956 "movq %%mm0, %%mm5 \n\t"
1959 "movd (%%"REG_S
",%%"REG_a
",4), %%mm3 \n\t" //7 ;prefetch!
1960 "movq %%mm1, %%mm6 \n\t"
1962 "movd (%%"REG_S
",%%"REG_d
",), %%mm4 \n\t" //6
1963 "punpcklbw %%mm7, %%mm3 \n\t"
1965 "psubw %%mm3, %%mm5 \n\t"
1966 "punpcklbw %%mm7, %%mm4 \n\t"
1968 "paddw %%mm3, %%mm0 \n\t"
1969 "psubw %%mm4, %%mm6 \n\t"
1971 "movd (%%"REG_S
",%%"REG_a
",2), %%mm3 \n\t" //5
1972 "paddw %%mm4, %%mm1 \n\t"
1974 "movq %%mm5, 0*8+%3 \n\t" //t7
1975 "punpcklbw %%mm7, %%mm3 \n\t"
1977 "movq %%mm6, 1*8+%3 \n\t" //t6
1978 "movq %%mm2, %%mm4 \n\t"
1980 "movd (%%"REG_S
"), %%mm5 \n\t" //3
1981 "paddw %%mm3, %%mm2 \n\t"
1983 "movd (%%"REG_S
",%%"REG_a
",), %%mm6 \n\t" //4
1984 "punpcklbw %%mm7, %%mm5 \n\t"
1986 "psubw %%mm3, %%mm4 \n\t"
1987 "punpcklbw %%mm7, %%mm6 \n\t"
1989 "movq %%mm5, %%mm3 \n\t"
1990 "paddw %%mm6, %%mm5 \n\t" //t3
1992 "psubw %%mm6, %%mm3 \n\t" //t4 ; t0 t1 t2 t4 t5 t3 - -
1993 "movq %%mm0, %%mm6 \n\t"
1995 "movq %%mm1, %%mm7 \n\t"
1996 "psubw %%mm5, %%mm0 \n\t" //t13
1998 "psubw %%mm2, %%mm1 \n\t"
1999 "paddw %%mm2, %%mm7 \n\t" //t11
2001 "paddw %%mm0, %%mm1 \n\t"
2002 "movq %%mm7, %%mm2 \n\t"
2004 "psllw $2, %%mm1 \n\t"
2005 "paddw %%mm5, %%mm6 \n\t" //t10
2007 "pmulhw "MANGLE(MM_FIX_0_707106781
)", %%mm1 \n\t"
2008 "paddw %%mm6, %%mm7 \n\t" //d2
2010 "psubw %%mm2, %%mm6 \n\t" //d3
2011 "movq %%mm0, %%mm5 \n\t"
2014 "movq %%mm7, %%mm2 \n\t"
2015 "punpcklwd %%mm6, %%mm7 \n\t"
2017 "paddw %%mm1, %%mm0 \n\t" //d0
2018 "punpckhwd %%mm6, %%mm2 \n\t"
2020 "psubw %%mm1, %%mm5 \n\t" //d1
2021 "movq %%mm0, %%mm6 \n\t"
2023 "movq 1*8+%3, %%mm1 \n\t"
2024 "punpcklwd %%mm5, %%mm0 \n\t"
2026 "punpckhwd %%mm5, %%mm6 \n\t"
2027 "movq %%mm0, %%mm5 \n\t"
2029 "punpckldq %%mm7, %%mm0 \n\t" //0
2030 "paddw %%mm4, %%mm3 \n\t"
2032 "punpckhdq %%mm7, %%mm5 \n\t" //1
2033 "movq %%mm6, %%mm7 \n\t"
2035 "movq %%mm0, "DCTSIZE_S
"*0*2(%%"REG_D
") \n\t"
2036 "punpckldq %%mm2, %%mm6 \n\t" //2
2038 "movq %%mm5, "DCTSIZE_S
"*1*2(%%"REG_D
") \n\t"
2039 "punpckhdq %%mm2, %%mm7 \n\t" //3
2041 "movq %%mm6, "DCTSIZE_S
"*2*2(%%"REG_D
") \n\t"
2042 "paddw %%mm1, %%mm4 \n\t"
2044 "movq %%mm7, "DCTSIZE_S
"*3*2(%%"REG_D
") \n\t"
2045 "psllw $2, %%mm3 \n\t" //t10
2047 "movq 0*8+%3, %%mm2 \n\t"
2048 "psllw $2, %%mm4 \n\t" //t11
2050 "pmulhw "MANGLE(MM_FIX_0_707106781
)", %%mm4 \n\t" //z3
2051 "paddw %%mm2, %%mm1 \n\t"
2053 "psllw $2, %%mm1 \n\t" //t12
2054 "movq %%mm3, %%mm0 \n\t"
2056 "pmulhw "MANGLE(MM_FIX_0_541196100
)", %%mm0 \n\t"
2057 "psubw %%mm1, %%mm3 \n\t"
2059 "pmulhw "MANGLE(MM_FIX_0_382683433
)", %%mm3 \n\t" //z5
2060 "movq %%mm2, %%mm5 \n\t"
2062 "pmulhw "MANGLE(MM_FIX_1_306562965
)", %%mm1 \n\t"
2063 "psubw %%mm4, %%mm2 \n\t" //z13
2065 "paddw %%mm4, %%mm5 \n\t" //z11
2066 "movq %%mm2, %%mm6 \n\t"
2068 "paddw %%mm3, %%mm0 \n\t" //z2
2069 "movq %%mm5, %%mm7 \n\t"
2071 "paddw %%mm0, %%mm2 \n\t" //d4
2072 "psubw %%mm0, %%mm6 \n\t" //d5
2074 "movq %%mm2, %%mm4 \n\t"
2075 "paddw %%mm3, %%mm1 \n\t" //z4
2078 "punpcklwd %%mm6, %%mm2 \n\t"
2079 "paddw %%mm1, %%mm5 \n\t" //d6
2081 "punpckhwd %%mm6, %%mm4 \n\t"
2082 "psubw %%mm1, %%mm7 \n\t" //d7
2084 "movq %%mm5, %%mm6 \n\t"
2085 "punpcklwd %%mm7, %%mm5 \n\t"
2087 "punpckhwd %%mm7, %%mm6 \n\t"
2088 "movq %%mm2, %%mm7 \n\t"
2090 "punpckldq %%mm5, %%mm2 \n\t" //4
2091 "sub %%"REG_d
", %%"REG_S
" \n\t"
2093 "punpckhdq %%mm5, %%mm7 \n\t" //5
2094 "movq %%mm4, %%mm5 \n\t"
2096 "movq %%mm2, "DCTSIZE_S
"*0*2+"DCTSIZE_S
"(%%"REG_D
") \n\t"
2097 "punpckldq %%mm6, %%mm4 \n\t" //6
2099 "movq %%mm7, "DCTSIZE_S
"*1*2+"DCTSIZE_S
"(%%"REG_D
") \n\t"
2100 "punpckhdq %%mm6, %%mm5 \n\t" //7
2102 "movq %%mm4, "DCTSIZE_S
"*2*2+"DCTSIZE_S
"(%%"REG_D
") \n\t"
2103 "add $4, %%"REG_S
" \n\t"
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"
2110 : "+S"(pixels
), "+D"(data
), "+c"(cnt
), "=o"(temps
)