H.261 encoder by (Maarten Daniels <maarten dot daniels at luc dot ac dot be>)
[ffmpeg.git] / libavcodec / i386 / mpegvideo_mmx_template.c
1 /*
2  * MPEG video MMX templates
3  *
4  * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
5  *
6  * This library is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2 of the License, or (at your option) any later version.
10  *
11  * This library is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with this library; if not, write to the Free Software
18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
19  */
20 #undef SPREADW
21 #undef PMAXW
22 #ifdef HAVE_MMX2
23 #define SPREADW(a) "pshufw $0, " #a ", " #a " \n\t"
24 #define PMAXW(a,b) "pmaxsw " #a ", " #b " \n\t"
25
26 #else
27 #define SPREADW(a) \
28         "punpcklwd " #a ", " #a " \n\t"\
29         "punpcklwd " #a ", " #a " \n\t"
30 #define PMAXW(a,b) \
31         "psubusw " #a ", " #b " \n\t"\
32         "paddw " #a ", " #b " \n\t"
33 #endif
34
35 static int RENAME(dct_quantize)(MpegEncContext *s,
36                             DCTELEM *block, int n,
37                             int qscale, int *overflow)
38 {
39     long last_non_zero_p1;
40     int level=0, q; //=0 is cuz gcc says uninitalized ...
41     const uint16_t *qmat, *bias;
42     __align8 int16_t temp_block[64];
43     
44     assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
45
46     //s->fdct (block);
47     RENAMEl(ff_fdct) (block); //cant be anything else ...
48
49     if(s->dct_error_sum)
50         s->denoise_dct(s, block);
51
52     if (s->mb_intra) {
53         int dummy;
54         if (n < 4)
55             q = s->y_dc_scale;
56         else
57             q = s->c_dc_scale;
58         /* note: block[0] is assumed to be positive */
59         if (!s->h263_aic) {
60 #if 1
61         asm volatile (
62                 "mul %%ecx              \n\t"
63                 : "=d" (level), "=a"(dummy)
64                 : "a" ((block[0]>>2) + q), "c" (inverse[q<<1])
65         );
66 #else
67         asm volatile (
68                 "xorl %%edx, %%edx      \n\t"
69                 "divw %%cx              \n\t"
70                 "movzwl %%ax, %%eax     \n\t"
71                 : "=a" (level)
72                 : "a" ((block[0]>>2) + q), "c" (q<<1)
73                 : "%edx"
74         );
75 #endif
76         } else
77             /* For AIC we skip quant/dequant of INTRADC */
78             level = (block[0] + 4)>>3;
79             
80         block[0]=0; //avoid fake overflow
81 //        temp_block[0] = (block[0] + (q >> 1)) / q;
82         last_non_zero_p1 = 1;
83         bias = s->q_intra_matrix16[qscale][1];
84         qmat = s->q_intra_matrix16[qscale][0];
85     } else {
86         last_non_zero_p1 = 0;
87         bias = s->q_inter_matrix16[qscale][1];
88         qmat = s->q_inter_matrix16[qscale][0];
89     }
90
91     if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
92     
93         asm volatile(
94             "movd %%"REG_a", %%mm3              \n\t" // last_non_zero_p1
95             SPREADW(%%mm3)
96             "pxor %%mm7, %%mm7                  \n\t" // 0
97             "pxor %%mm4, %%mm4                  \n\t" // 0
98             "movq (%2), %%mm5                   \n\t" // qmat[0]
99             "pxor %%mm6, %%mm6                  \n\t"
100             "psubw (%3), %%mm6                  \n\t" // -bias[0]
101             "mov $-128, %%"REG_a"               \n\t"
102             ".balign 16                         \n\t"
103             "1:                                 \n\t"
104             "pxor %%mm1, %%mm1                  \n\t" // 0
105             "movq (%1, %%"REG_a"), %%mm0        \n\t" // block[i]
106             "pcmpgtw %%mm0, %%mm1               \n\t" // block[i] <= 0 ? 0xFF : 0x00
107             "pxor %%mm1, %%mm0                  \n\t" 
108             "psubw %%mm1, %%mm0                 \n\t" // ABS(block[i])
109             "psubusw %%mm6, %%mm0               \n\t" // ABS(block[i]) + bias[0]
110             "pmulhw %%mm5, %%mm0                \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
111             "por %%mm0, %%mm4                   \n\t" 
112             "pxor %%mm1, %%mm0                  \n\t" 
113             "psubw %%mm1, %%mm0                 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
114             "movq %%mm0, (%5, %%"REG_a")        \n\t"
115             "pcmpeqw %%mm7, %%mm0               \n\t" // out==0 ? 0xFF : 0x00
116             "movq (%4, %%"REG_a"), %%mm1        \n\t" 
117             "movq %%mm7, (%1, %%"REG_a")        \n\t" // 0
118             "pandn %%mm1, %%mm0                 \n\t"
119             PMAXW(%%mm0, %%mm3)
120             "add $8, %%"REG_a"                  \n\t"
121             " js 1b                             \n\t"
122             "movq %%mm3, %%mm0                  \n\t"
123             "psrlq $32, %%mm3                   \n\t"
124             PMAXW(%%mm0, %%mm3)
125             "movq %%mm3, %%mm0                  \n\t"
126             "psrlq $16, %%mm3                   \n\t"
127             PMAXW(%%mm0, %%mm3)
128             "movd %%mm3, %%"REG_a"              \n\t"
129             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
130             : "+a" (last_non_zero_p1)
131             : "r" (block+64), "r" (qmat), "r" (bias),
132               "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
133         );
134         // note the asm is split cuz gcc doesnt like that many operands ...
135         asm volatile(
136             "movd %1, %%mm1                     \n\t" // max_qcoeff
137             SPREADW(%%mm1)
138             "psubusw %%mm1, %%mm4               \n\t" 
139             "packuswb %%mm4, %%mm4              \n\t"
140             "movd %%mm4, %0                     \n\t" // *overflow
141         : "=g" (*overflow)
142         : "g" (s->max_qcoeff)
143         );
144     }else{ // FMT_H263
145         asm volatile(
146             "movd %%"REG_a", %%mm3              \n\t" // last_non_zero_p1
147             SPREADW(%%mm3)
148             "pxor %%mm7, %%mm7                  \n\t" // 0
149             "pxor %%mm4, %%mm4                  \n\t" // 0
150             "mov $-128, %%"REG_a"               \n\t"
151             ".balign 16                         \n\t"
152             "1:                                 \n\t"
153             "pxor %%mm1, %%mm1                  \n\t" // 0
154             "movq (%1, %%"REG_a"), %%mm0        \n\t" // block[i]
155             "pcmpgtw %%mm0, %%mm1               \n\t" // block[i] <= 0 ? 0xFF : 0x00
156             "pxor %%mm1, %%mm0                  \n\t" 
157             "psubw %%mm1, %%mm0                 \n\t" // ABS(block[i])
158             "movq (%3, %%"REG_a"), %%mm6        \n\t" // bias[0]
159             "paddusw %%mm6, %%mm0               \n\t" // ABS(block[i]) + bias[0]
160             "movq (%2, %%"REG_a"), %%mm5                \n\t" // qmat[i]
161             "pmulhw %%mm5, %%mm0                \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
162             "por %%mm0, %%mm4                   \n\t" 
163             "pxor %%mm1, %%mm0                  \n\t" 
164             "psubw %%mm1, %%mm0                 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
165             "movq %%mm0, (%5, %%"REG_a")        \n\t"
166             "pcmpeqw %%mm7, %%mm0               \n\t" // out==0 ? 0xFF : 0x00
167             "movq (%4, %%"REG_a"), %%mm1                \n\t" 
168             "movq %%mm7, (%1, %%"REG_a")                \n\t" // 0
169             "pandn %%mm1, %%mm0                 \n\t"
170             PMAXW(%%mm0, %%mm3)
171             "add $8, %%"REG_a"                  \n\t"
172             " js 1b                             \n\t"
173             "movq %%mm3, %%mm0                  \n\t"
174             "psrlq $32, %%mm3                   \n\t"
175             PMAXW(%%mm0, %%mm3)
176             "movq %%mm3, %%mm0                  \n\t"
177             "psrlq $16, %%mm3                   \n\t"
178             PMAXW(%%mm0, %%mm3)
179             "movd %%mm3, %%"REG_a"              \n\t"
180             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
181             : "+a" (last_non_zero_p1)
182             : "r" (block+64), "r" (qmat+64), "r" (bias+64),
183               "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
184         );
185         // note the asm is split cuz gcc doesnt like that many operands ...
186         asm volatile(
187             "movd %1, %%mm1                     \n\t" // max_qcoeff
188             SPREADW(%%mm1)
189             "psubusw %%mm1, %%mm4               \n\t" 
190             "packuswb %%mm4, %%mm4              \n\t"
191             "movd %%mm4, %0                     \n\t" // *overflow
192         : "=g" (*overflow)
193         : "g" (s->max_qcoeff)
194         );
195     }
196
197     if(s->mb_intra) block[0]= level;
198     else            block[0]= temp_block[0];
199
200     if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
201         if(last_non_zero_p1 <= 1) goto end;
202         block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08]; 
203         block[0x20] = temp_block[0x10]; 
204         if(last_non_zero_p1 <= 4) goto end;
205         block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02]; 
206         block[0x09] = temp_block[0x03]; 
207         if(last_non_zero_p1 <= 7) goto end;
208         block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11]; 
209         block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20]; 
210         if(last_non_zero_p1 <= 11) goto end;
211         block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12]; 
212         block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04]; 
213         block[0x0C] = temp_block[0x05]; 
214         if(last_non_zero_p1 <= 16) goto end;
215         block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13]; 
216         block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21]; 
217         block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30]; 
218         block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22]; 
219         if(last_non_zero_p1 <= 24) goto end;
220         block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14]; 
221         block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06]; 
222         block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E]; 
223         block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C]; 
224         if(last_non_zero_p1 <= 32) goto end;
225         block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A]; 
226         block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38]; 
227         block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32]; 
228         block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24]; 
229         if(last_non_zero_p1 <= 40) goto end;
230         block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16]; 
231         block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17]; 
232         block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25]; 
233         block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33]; 
234         if(last_non_zero_p1 <= 48) goto end;
235         block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B]; 
236         block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D]; 
237         block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F]; 
238         block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E]; 
239         if(last_non_zero_p1 <= 56) goto end;
240         block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C]; 
241         block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36]; 
242         block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37]; 
243         block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
244     }else if(s->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
245         if(last_non_zero_p1 <= 1) goto end;
246         block[0x04] = temp_block[0x01]; 
247         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10]; 
248         if(last_non_zero_p1 <= 4) goto end;
249         block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02]; 
250         block[0x05] = temp_block[0x03]; 
251         if(last_non_zero_p1 <= 7) goto end;
252         block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11]; 
253         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20]; 
254         if(last_non_zero_p1 <= 11) goto end;
255         block[0x1C] = temp_block[0x19]; 
256         block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B]; 
257         block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05]; 
258         if(last_non_zero_p1 <= 16) goto end;
259         block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13]; 
260         block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21]; 
261         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30]; 
262         block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22]; 
263         if(last_non_zero_p1 <= 24) goto end;
264         block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14]; 
265         block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06]; 
266         block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E]; 
267         block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C]; 
268         if(last_non_zero_p1 <= 32) goto end;
269         block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A]; 
270         block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38]; 
271         block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32]; 
272         block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24]; 
273         if(last_non_zero_p1 <= 40) goto end;
274         block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16]; 
275         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17]; 
276         block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25]; 
277         block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33]; 
278         if(last_non_zero_p1 <= 48) goto end;
279         block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B]; 
280         block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D]; 
281             block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F]; 
282         block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E]; 
283         if(last_non_zero_p1 <= 56) goto end;
284         block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C]; 
285         block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36]; 
286         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37]; 
287         block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
288     }else{
289         if(last_non_zero_p1 <= 1) goto end;
290         block[0x01] = temp_block[0x01]; 
291         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10]; 
292         if(last_non_zero_p1 <= 4) goto end;
293         block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02]; 
294         block[0x03] = temp_block[0x03]; 
295         if(last_non_zero_p1 <= 7) goto end;
296         block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11]; 
297         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20]; 
298         if(last_non_zero_p1 <= 11) goto end;
299         block[0x19] = temp_block[0x19]; 
300         block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B]; 
301         block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05]; 
302         if(last_non_zero_p1 <= 16) goto end;
303         block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13]; 
304         block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21]; 
305         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30]; 
306         block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22]; 
307         if(last_non_zero_p1 <= 24) goto end;
308         block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14]; 
309         block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06]; 
310         block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E]; 
311         block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C]; 
312         if(last_non_zero_p1 <= 32) goto end;
313         block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A]; 
314         block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38]; 
315         block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32]; 
316         block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24]; 
317         if(last_non_zero_p1 <= 40) goto end;
318         block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16]; 
319         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17]; 
320         block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25]; 
321         block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33]; 
322         if(last_non_zero_p1 <= 48) goto end;
323         block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B]; 
324         block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D]; 
325         block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F]; 
326         block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E]; 
327         if(last_non_zero_p1 <= 56) goto end;
328         block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C]; 
329         block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36]; 
330         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37]; 
331         block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
332     }
333     end:
334 /*
335     for(i=0; i<last_non_zero_p1; i++)
336     {
337        int j= zigzag_direct_noperm[i];
338        block[block_permute_op(j)]= temp_block[j];
339     }
340 */
341
342     return last_non_zero_p1 - 1;
343 }