Statistics
| Branch: | Revision:

ffmpeg / libavcodec / i386 / mpegvideo_mmx.c @ 2f349de2

History | View | Annotate | Download (12.6 KB)

1
/*
2
 * The simplest mpeg encoder (well, it was the simplest!)
3
 * Copyright (c) 2000,2001 Gerard Lantau.
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., 675 Mass Ave, Cambridge, MA 02139, USA.
18
 *
19
 * Optimized for ia32 cpus by Nick Kurshev <nickols_k@mail.ru>
20
 * h263 dequantizer by Michael Niedermayer <michaelni@gmx.at>
21
 */
22

    
23
#include "../dsputil.h"
24
#include "../mpegvideo.h"
25
#include "../avcodec.h"
26
#include "../mangle.h"
27

    
28
extern UINT8 zigzag_end[64];
29
extern void (*draw_edges)(UINT8 *buf, int wrap, int width, int height, int w);
30
extern int (*dct_quantize)(MpegEncContext *s, DCTELEM *block, int n, int qscale);
31

    
32
extern UINT8 zigzag_direct_noperm[64];
33
extern UINT16 inv_zigzag_direct16[64];
34
extern UINT32 inverse[256];
35

    
36
#if 0
37

38
/* XXX: GL: I don't understand why this function needs optimization
39
   (it is called only once per frame!), so I disabled it */
40

41
void MPV_frame_start(MpegEncContext *s)
42
{
43
    if (s->pict_type == B_TYPE) {
44
        __asm __volatile(
45
            "movl        (%1), %%eax\n\t"
46
            "movl        4(%1), %%edx\n\t"
47
            "movl        8(%1), %%ecx\n\t"
48
            "movl        %%eax, (%0)\n\t"
49
            "movl        %%edx, 4(%0)\n\t"
50
            "movl        %%ecx, 8(%0)\n\t"
51
            :
52
            :"r"(s->current_picture), "r"(s->aux_picture)
53
            :"eax","edx","ecx","memory");
54
    } else {
55
            /* swap next and last */
56
        __asm __volatile(
57
            "movl        (%1), %%eax\n\t"
58
            "movl        4(%1), %%edx\n\t"
59
            "movl        8(%1), %%ecx\n\t"
60
            "xchgl        (%0), %%eax\n\t"
61
            "xchgl        4(%0), %%edx\n\t"
62
            "xchgl        8(%0), %%ecx\n\t"
63
            "movl        %%eax, (%1)\n\t"
64
            "movl        %%edx, 4(%1)\n\t"
65
            "movl        %%ecx, 8(%1)\n\t"
66
            "movl        %%eax, (%2)\n\t"
67
            "movl        %%edx, 4(%2)\n\t"
68
            "movl        %%ecx, 8(%2)\n\t"
69
            :
70
            :"r"(s->last_picture), "r"(s->next_picture), "r"(s->current_picture)
71
            :"eax","edx","ecx","memory");
72
    }
73
}
74
#endif
75

    
76
static const unsigned long long int mm_wabs __attribute__ ((aligned(8))) = 0xffffffffffffffffULL;
77
static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL;
78

    
79

    
80
static void dct_unquantize_h263_mmx(MpegEncContext *s,
81
                                  DCTELEM *block, int n, int qscale)
82
{
83
    int i, level, qmul, qadd, nCoeffs;
84
    
85
    qmul = s->qscale << 1;
86
    qadd = (s->qscale - 1) | 1;
87

    
88
    if (s->mb_intra) {
89
        if (n < 4)
90
            block[0] = block[0] * s->y_dc_scale;
91
        else
92
            block[0] = block[0] * s->c_dc_scale;
93

    
94
        for(i=1; i<8; i++) {
95
                level = block[i];
96
                if (level) {
97
                        if (level < 0) {
98
                                level = level * qmul - qadd;
99
                        } else {
100
                                level = level * qmul + qadd;
101
                        }
102
                        block[i] = level;
103
                }
104
        }
105
        nCoeffs=64;
106
    } else {
107
        i = 0;
108
        nCoeffs= zigzag_end[ s->block_last_index[n] ];
109
    }
110
//printf("%d %d  ", qmul, qadd);
111
asm volatile(
112
                "movd %1, %%mm6                        \n\t" //qmul
113
                "packssdw %%mm6, %%mm6                \n\t"
114
                "packssdw %%mm6, %%mm6                \n\t"
115
                "movd %2, %%mm5                        \n\t" //qadd
116
                "pxor %%mm7, %%mm7                \n\t"
117
                "packssdw %%mm5, %%mm5                \n\t"
118
                "packssdw %%mm5, %%mm5                \n\t"
119
                "psubw %%mm5, %%mm7                \n\t"
120
                "pxor %%mm4, %%mm4                \n\t"
121
                ".balign 16\n\t"
122
                "1:                                \n\t"
123
                "movq (%0, %3), %%mm0                \n\t"
124
                "movq 8(%0, %3), %%mm1                \n\t"
125

    
126
                "pmullw %%mm6, %%mm0                \n\t"
127
                "pmullw %%mm6, %%mm1                \n\t"
128

    
129
                "movq (%0, %3), %%mm2                \n\t"
130
                "movq 8(%0, %3), %%mm3                \n\t"
131

    
132
                "pcmpgtw %%mm4, %%mm2                \n\t" // block[i] < 0 ? -1 : 0
133
                "pcmpgtw %%mm4, %%mm3                \n\t" // block[i] < 0 ? -1 : 0
134

    
135
                "pxor %%mm2, %%mm0                \n\t"
136
                "pxor %%mm3, %%mm1                \n\t"
137

    
138
                "paddw %%mm7, %%mm0                \n\t"
139
                "paddw %%mm7, %%mm1                \n\t"
140

    
141
                "pxor %%mm0, %%mm2                \n\t"
142
                "pxor %%mm1, %%mm3                \n\t"
143

    
144
                "pcmpeqw %%mm7, %%mm0                \n\t" // block[i] == 0 ? -1 : 0
145
                "pcmpeqw %%mm7, %%mm1                \n\t" // block[i] == 0 ? -1 : 0
146

    
147
                "pandn %%mm2, %%mm0                \n\t"
148
                "pandn %%mm3, %%mm1                \n\t"
149

    
150
                "movq %%mm0, (%0, %3)                \n\t"
151
                "movq %%mm1, 8(%0, %3)                \n\t"
152

    
153
                "addl $16, %3                        \n\t"
154
                "js 1b                                \n\t"
155
                ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(i-nCoeffs))
156
                : "memory"
157
        );
158
}
159

    
160

    
161
/*
162
  NK:
163
  Note: looking at PARANOID:
164
  "enable all paranoid tests for rounding, overflows, etc..."
165

166
#ifdef PARANOID
167
                if (level < -2048 || level > 2047)
168
                    fprintf(stderr, "unquant error %d %d\n", i, level);
169
#endif
170
  We can suppose that result of two multiplications can't be greate of 0xFFFF
171
  i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
172
  a complex multiplication.
173
=====================================================
174
 Full formula for multiplication of 2 integer numbers
175
 which are represent as high:low words:
176
 input: value1 = high1:low1
177
        value2 = high2:low2
178
 output: value3 = value1*value2
179
 value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
180
 this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
181
 but this algorithm will compute only 0x66cb0ce4
182
 this limited by 16-bit size of operands
183
 ---------------------------------
184
 tlow1 = high1*low2
185
 tlow2 = high2*low1
186
 tlow1 = tlow1 + tlow2
187
 high3:low3 = low1*low2
188
 high3 += tlow1
189
*/
190
static void dct_unquantize_mpeg1_mmx(MpegEncContext *s,
191
                                     DCTELEM *block, int n, int qscale)
192
{
193
    int i, level, nCoeffs;
194
    const UINT16 *quant_matrix;
195
    
196
    if(s->alternate_scan) nCoeffs= 64;
197
    else nCoeffs= nCoeffs= zigzag_end[ s->block_last_index[n] ];
198

    
199
    if (s->mb_intra) {
200
        if (n < 4) 
201
            block[0] = block[0] * s->y_dc_scale;
202
        else
203
            block[0] = block[0] * s->c_dc_scale;
204
        /* isnt used anymore (we have a h263 unquantizer since some time)
205
        if (s->out_format == FMT_H263) {
206
            i = 1;
207
            goto unquant_even;
208
        }*/
209
        /* XXX: only mpeg1 */
210
        quant_matrix = s->intra_matrix;
211
        i=1;
212
        /* Align on 4 elements boundary */
213
        while(i&3)
214
        {
215
            level = block[i];
216
            if (level) {
217
                if (level < 0) level = -level;
218
                    level = (int)(level * qscale * quant_matrix[i]) >> 3;
219
                    level = (level - 1) | 1;
220
                if (block[i] < 0) level = -level;
221
                block[i] = level;
222
            }
223
            i++;
224
        }
225
        __asm __volatile(
226
        "movd        %0, %%mm6\n\t"       /* mm6 = qscale | 0  */
227
        "punpckldq %%mm6, %%mm6\n\t" /* mm6 = qscale | qscale */
228
        "movq        %2, %%mm4\n\t"
229
        "movq        %%mm6, %%mm7\n\t"
230
        "movq        %1, %%mm5\n\t"
231
        "packssdw %%mm6, %%mm7\n\t" /* mm7 = qscale | qscale | qscale | qscale */
232
        "pxor        %%mm6, %%mm6\n\t"
233
        ::"g"(qscale),"m"(mm_wone),"m"(mm_wabs):"memory");
234
        for(;i<nCoeffs;i+=4) {
235
                __asm __volatile(
236
                        "movq        %1, %%mm0\n\t"
237
                        "movq        %%mm7, %%mm1\n\t"
238
                        "movq        %%mm0, %%mm2\n\t"
239
                        "movq        %%mm0, %%mm3\n\t"
240
                        "pcmpgtw %%mm6, %%mm2\n\t"
241
                        "pmullw        %2, %%mm1\n\t"
242
                        "pandn        %%mm4, %%mm2\n\t"
243
                        "por        %%mm5, %%mm2\n\t"
244
                        "pmullw        %%mm2, %%mm0\n\t" /* mm0 = abs(block[i]). */
245

    
246
                        "pcmpeqw %%mm6, %%mm3\n\t"
247
                        "pmullw        %%mm0, %%mm1\n\t"
248
                        "psraw        $3, %%mm1\n\t"
249
                        "psubw        %%mm5, %%mm1\n\t"   /* block[i] --; */
250
                        "pandn        %%mm4, %%mm3\n\t"  /* fake of pcmpneqw : mm0 != 0 then mm1 = -1 */
251
                        "por        %%mm5, %%mm1\n\t"   /* block[i] |= 1 */
252
                        "pmullw %%mm2, %%mm1\n\t"   /* change signs again */
253

    
254
                        "pand        %%mm3, %%mm1\n\t" /* nullify if was zero */
255
                        "movq        %%mm1, %0"
256
                        :"=m"(block[i])
257
                        :"m"(block[i]), "m"(quant_matrix[i])
258
                        :"memory");
259
        }
260
    } else {
261
        i = 0;
262
//    unquant_even:
263
        quant_matrix = s->non_intra_matrix;
264
        /* Align on 4 elements boundary */
265
        while(i&7)
266
        {
267
            level = block[i];
268
            if (level) {
269
                if (level < 0) level = -level;
270
                    level = (((level << 1) + 1) * qscale *
271
                             ((int) quant_matrix[i])) >> 4;
272
                    level = (level - 1) | 1;
273
                if(block[i] < 0) level = -level;
274
                block[i] = level;
275
            }
276
            i++;
277
        }
278
asm volatile(
279
                "pcmpeqw %%mm7, %%mm7                \n\t"
280
                "psrlw $15, %%mm7                \n\t"
281
                "movd %2, %%mm6                        \n\t"
282
                "packssdw %%mm6, %%mm6                \n\t"
283
                "packssdw %%mm6, %%mm6                \n\t"
284
                ".balign 16\n\t"
285
                "1:                                \n\t"
286
                "movq (%0, %3), %%mm0                \n\t"
287
                "movq 8(%0, %3), %%mm1                \n\t"
288
                "movq (%1, %3), %%mm4                \n\t"
289
                "movq 8(%1, %3), %%mm5                \n\t"
290
                "pmullw %%mm6, %%mm4                \n\t" // q=qscale*quant_matrix[i]
291
                "pmullw %%mm6, %%mm5                \n\t" // q=qscale*quant_matrix[i]
292
                "pxor %%mm2, %%mm2                \n\t"
293
                "pxor %%mm3, %%mm3                \n\t"
294
                "pcmpgtw %%mm0, %%mm2                \n\t" // block[i] < 0 ? -1 : 0
295
                "pcmpgtw %%mm1, %%mm3                \n\t" // block[i] < 0 ? -1 : 0
296
                "pxor %%mm2, %%mm0                \n\t"
297
                "pxor %%mm3, %%mm1                \n\t"
298
                "psubw %%mm2, %%mm0                \n\t" // abs(block[i])
299
                "psubw %%mm3, %%mm1                \n\t" // abs(block[i])
300
                "paddw %%mm0, %%mm0                \n\t" // abs(block[i])*2
301
                "paddw %%mm1, %%mm1                \n\t" // abs(block[i])*2
302
                "paddw %%mm7, %%mm0                \n\t" // abs(block[i])*2 + 1
303
                "paddw %%mm7, %%mm1                \n\t" // abs(block[i])*2 + 1
304
                "pmullw %%mm4, %%mm0                \n\t" // (abs(block[i])*2 + 1)*q
305
                "pmullw %%mm5, %%mm1                \n\t" // (abs(block[i])*2 + 1)*q
306
                "pxor %%mm4, %%mm4                \n\t"
307
                "pxor %%mm5, %%mm5                \n\t" // FIXME slow
308
                "pcmpeqw (%0, %3), %%mm4        \n\t" // block[i] == 0 ? -1 : 0
309
                "pcmpeqw 8(%0, %3), %%mm5        \n\t" // block[i] == 0 ? -1 : 0
310
                "psraw $4, %%mm0                \n\t"
311
                "psraw $4, %%mm1                \n\t"
312
                "psubw %%mm7, %%mm0                \n\t"
313
                "psubw %%mm7, %%mm1                \n\t"
314
                "por %%mm7, %%mm0                \n\t"
315
                "por %%mm7, %%mm1                \n\t"
316
                "pxor %%mm2, %%mm0                \n\t"
317
                "pxor %%mm3, %%mm1                \n\t"
318
                "psubw %%mm2, %%mm0                \n\t"
319
                "psubw %%mm3, %%mm1                \n\t"
320
                "pandn %%mm0, %%mm4                \n\t"
321
                "pandn %%mm1, %%mm5                \n\t"
322
                "movq %%mm4, (%0, %3)                \n\t"
323
                "movq %%mm5, 8(%0, %3)                \n\t"
324

    
325
                "addl $16, %3                        \n\t"
326
                "js 1b                                \n\t"
327
                ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "r" (2*(i-nCoeffs))
328
                : "memory"
329
        );
330
    }
331
}
332

    
333
/* draw the edges of width 'w' of an image of size width, height 
334
   this mmx version can only handle w==8 || w==16 */
335
static void draw_edges_mmx(UINT8 *buf, int wrap, int width, int height, int w)
336
{
337
    UINT8 *ptr, *last_line;
338
    int i;
339

    
340
    last_line = buf + (height - 1) * wrap;
341
    /* left and right */
342
    ptr = buf;
343
    if(w==8)
344
    {
345
        asm volatile(
346
                "1:                                \n\t"
347
                "movd (%0), %%mm0                \n\t"
348
                "punpcklbw %%mm0, %%mm0                \n\t" 
349
                "punpcklwd %%mm0, %%mm0                \n\t"
350
                "punpckldq %%mm0, %%mm0                \n\t"
351
                "movq %%mm0, -8(%0)                \n\t"
352
                "movq -8(%0, %2), %%mm1                \n\t"
353
                "punpckhbw %%mm1, %%mm1                \n\t"
354
                "punpckhwd %%mm1, %%mm1                \n\t"
355
                "punpckhdq %%mm1, %%mm1                \n\t"
356
                "movq %%mm1, (%0, %2)                \n\t"
357
                "addl %1, %0                        \n\t"
358
                "cmpl %3, %0                        \n\t"
359
                " jb 1b                                \n\t"
360
                : "+r" (ptr)
361
                : "r" (wrap), "r" (width), "r" (ptr + wrap*height)
362
        );
363
    }
364
    else
365
    {
366
        asm volatile(
367
                "1:                                \n\t"
368
                "movd (%0), %%mm0                \n\t"
369
                "punpcklbw %%mm0, %%mm0                \n\t" 
370
                "punpcklwd %%mm0, %%mm0                \n\t"
371
                "punpckldq %%mm0, %%mm0                \n\t"
372
                "movq %%mm0, -8(%0)                \n\t"
373
                "movq %%mm0, -16(%0)                \n\t"
374
                "movq -8(%0, %2), %%mm1                \n\t"
375
                "punpckhbw %%mm1, %%mm1                \n\t"
376
                "punpckhwd %%mm1, %%mm1                \n\t"
377
                "punpckhdq %%mm1, %%mm1                \n\t"
378
                "movq %%mm1, (%0, %2)                \n\t"
379
                "movq %%mm1, 8(%0, %2)                \n\t"
380
                "addl %1, %0                        \n\t"
381
                "cmpl %3, %0                        \n\t"
382
                " jb 1b                                \n\t"                
383
                : "+r" (ptr)
384
                : "r" (wrap), "r" (width), "r" (ptr + wrap*height)
385
        );
386
    }
387
    
388
    for(i=0;i<w;i+=4) {
389
        /* top and bottom (and hopefully also the corners) */
390
        ptr= buf - (i + 1) * wrap - w;
391
        asm volatile(
392
                "1:                                \n\t"
393
                "movq (%1, %0), %%mm0                \n\t"
394
                "movq %%mm0, (%0)                \n\t"
395
                "movq %%mm0, (%0, %2)                \n\t"
396
                "movq %%mm0, (%0, %2, 2)        \n\t"
397
                "movq %%mm0, (%0, %3)                \n\t"
398
                "addl $8, %0                        \n\t"
399
                "cmpl %4, %0                        \n\t"
400
                " jb 1b                                \n\t"
401
                : "+r" (ptr)
402
                : "r" ((int)buf - (int)ptr - w), "r" (-wrap), "r" (-wrap*3), "r" (ptr+width+2*w)
403
        );
404
        ptr= last_line + (i + 1) * wrap - w;
405
        asm volatile(
406
                "1:                                \n\t"
407
                "movq (%1, %0), %%mm0                \n\t"
408
                "movq %%mm0, (%0)                \n\t"
409
                "movq %%mm0, (%0, %2)                \n\t"
410
                "movq %%mm0, (%0, %2, 2)        \n\t"
411
                "movq %%mm0, (%0, %3)                \n\t"
412
                "addl $8, %0                        \n\t"
413
                "cmpl %4, %0                        \n\t"
414
                " jb 1b                                \n\t"
415
                : "+r" (ptr)
416
                : "r" ((int)last_line - (int)ptr - w), "r" (wrap), "r" (wrap*3), "r" (ptr+width+2*w)
417
        );
418
    }
419
}
420

    
421
static volatile int esp_temp;
422

    
423
void unused_var_warning_killer(){
424
        esp_temp++;
425
}
426

    
427
#undef HAVE_MMX2
428
#define RENAME(a) a ## _MMX
429
#include "mpegvideo_mmx_template.c"
430

    
431
#define HAVE_MMX2
432
#undef RENAME
433
#define RENAME(a) a ## _MMX2
434
#include "mpegvideo_mmx_template.c"
435

    
436
void MPV_common_init_mmx(MpegEncContext *s)
437
{
438
    if (mm_flags & MM_MMX) {
439
        if (s->out_format == FMT_H263)
440
                s->dct_unquantize = dct_unquantize_h263_mmx;
441
        else
442
                s->dct_unquantize = dct_unquantize_mpeg1_mmx;
443
        
444
        draw_edges = draw_edges_mmx;
445

    
446
        if(mm_flags & MM_MMXEXT){
447
                dct_quantize= dct_quantize_MMX2;
448
        }else{
449
                dct_quantize= dct_quantize_MMX;
450
        }
451
    }
452
}