FFmpeg
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
mpegvideoenc_template.c
Go to the documentation of this file.
1 /*
2  * MPEG video MMX templates
3  *
4  * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * FFmpeg is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with FFmpeg; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22 
23 #undef MMREG_WIDTH
24 #undef MM
25 #undef MOVQ
26 #undef SPREADW
27 #undef PMAXW
28 #undef PMAX
29 #undef SAVE_SIGN
30 #undef RESTORE_SIGN
31 
32 #if COMPILE_TEMPLATE_SSE2
33 #define MMREG_WIDTH "16"
34 #define MM "%%xmm"
35 #define MOVQ "movdqa"
36 #define SPREADW(a) \
37  "pshuflw $0, "a", "a" \n\t"\
38  "punpcklwd "a", "a" \n\t"
39 #define PMAXW(a,b) "pmaxsw "a", "b" \n\t"
40 #define PMAX(a,b) \
41  "movhlps "a", "b" \n\t"\
42  PMAXW(b, a)\
43  "pshuflw $0x0E, "a", "b" \n\t"\
44  PMAXW(b, a)\
45  "pshuflw $0x01, "a", "b" \n\t"\
46  PMAXW(b, a)
47 #else
48 #define MMREG_WIDTH "8"
49 #define MM "%%mm"
50 #define MOVQ "movq"
51 #if COMPILE_TEMPLATE_MMXEXT
52 #define SPREADW(a) "pshufw $0, "a", "a" \n\t"
53 #define PMAXW(a,b) "pmaxsw "a", "b" \n\t"
54 #define PMAX(a,b) \
55  "pshufw $0x0E, "a", "b" \n\t"\
56  PMAXW(b, a)\
57  "pshufw $0x01, "a", "b" \n\t"\
58  PMAXW(b, a)
59 #else
60 #define SPREADW(a) \
61  "punpcklwd "a", "a" \n\t"\
62  "punpcklwd "a", "a" \n\t"
63 #define PMAXW(a,b) \
64  "psubusw "a", "b" \n\t"\
65  "paddw "a", "b" \n\t"
66 #define PMAX(a,b) \
67  "movq "a", "b" \n\t"\
68  "psrlq $32, "a" \n\t"\
69  PMAXW(b, a)\
70  "movq "a", "b" \n\t"\
71  "psrlq $16, "a" \n\t"\
72  PMAXW(b, a)
73 
74 #endif
75 #endif
76 
77 #if COMPILE_TEMPLATE_SSSE3
78 #define SAVE_SIGN(a,b) \
79  "movdqa "b", "a" \n\t"\
80  "pabsw "b", "b" \n\t"
81 #define RESTORE_SIGN(a,b) \
82  "psignw "a", "b" \n\t"
83 #else
84 #define SAVE_SIGN(a,b) \
85  "pxor "a", "a" \n\t"\
86  "pcmpgtw "b", "a" \n\t" /* block[i] <= 0 ? 0xFF : 0x00 */\
87  "pxor "a", "b" \n\t"\
88  "psubw "a", "b" \n\t" /* ABS(block[i]) */
89 #define RESTORE_SIGN(a,b) \
90  "pxor "a", "b" \n\t"\
91  "psubw "a", "b" \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
92 #endif
93 
94 static int RENAME(dct_quantize)(MpegEncContext *s,
95  int16_t *block, int n,
96  int qscale, int *overflow)
97 {
98  x86_reg last_non_zero_p1;
99  int level=0, q; //=0 is because gcc says uninitialized ...
100  const uint16_t *qmat, *bias;
101  LOCAL_ALIGNED_16(int16_t, temp_block, [64]);
102 
103  av_assert2((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
104 
105  //s->fdct (block);
106  RENAMEl(ff_fdct) (block); //cannot be anything else ...
107 
108  if(s->dct_error_sum)
109  s->denoise_dct(s, block);
110 
111  if (s->mb_intra) {
112  int dummy;
113  if (n < 4){
114  q = s->y_dc_scale;
115  bias = s->q_intra_matrix16[qscale][1];
116  qmat = s->q_intra_matrix16[qscale][0];
117  }else{
118  q = s->c_dc_scale;
119  bias = s->q_chroma_intra_matrix16[qscale][1];
120  qmat = s->q_chroma_intra_matrix16[qscale][0];
121  }
122  /* note: block[0] is assumed to be positive */
123  if (!s->h263_aic) {
124  __asm__ volatile (
125  "mul %%ecx \n\t"
126  : "=d" (level), "=a"(dummy)
127  : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
128  );
129  } else
130  /* For AIC we skip quant/dequant of INTRADC */
131  level = (block[0] + 4)>>3;
132 
133  block[0]=0; //avoid fake overflow
134 // temp_block[0] = (block[0] + (q >> 1)) / q;
135  last_non_zero_p1 = 1;
136  } else {
137  last_non_zero_p1 = 0;
138  bias = s->q_inter_matrix16[qscale][1];
139  qmat = s->q_inter_matrix16[qscale][0];
140  }
141 
142  if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
143 
144  __asm__ volatile(
145  "movd %%"REG_a", "MM"3 \n\t" // last_non_zero_p1
146  SPREADW(MM"3")
147  "pxor "MM"7, "MM"7 \n\t" // 0
148  "pxor "MM"4, "MM"4 \n\t" // 0
149  MOVQ" (%2), "MM"5 \n\t" // qmat[0]
150  "pxor "MM"6, "MM"6 \n\t"
151  "psubw (%3), "MM"6 \n\t" // -bias[0]
152  "mov $-128, %%"REG_a" \n\t"
153  ".p2align 4 \n\t"
154  "1: \n\t"
155  MOVQ" (%1, %%"REG_a"), "MM"0 \n\t" // block[i]
156  SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
157  "psubusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
158  "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
159  "por "MM"0, "MM"4 \n\t"
160  RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
161  MOVQ" "MM"0, (%5, %%"REG_a") \n\t"
162  "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
163  MOVQ" (%4, %%"REG_a"), "MM"1 \n\t"
164  MOVQ" "MM"7, (%1, %%"REG_a") \n\t" // 0
165  "pandn "MM"1, "MM"0 \n\t"
166  PMAXW(MM"0", MM"3")
167  "add $"MMREG_WIDTH", %%"REG_a" \n\t"
168  " js 1b \n\t"
169  PMAX(MM"3", MM"0")
170  "movd "MM"3, %%"REG_a" \n\t"
171  "movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
172  : "+a" (last_non_zero_p1)
173  : "r" (block+64), "r" (qmat), "r" (bias),
174  "r" (ff_inv_zigzag_direct16+64), "r" (temp_block+64)
175  XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
176  "%xmm4", "%xmm5", "%xmm6", "%xmm7")
177  );
178  }else{ // FMT_H263
179  __asm__ volatile(
180  "movd %%"REG_a", "MM"3 \n\t" // last_non_zero_p1
181  SPREADW(MM"3")
182  "pxor "MM"7, "MM"7 \n\t" // 0
183  "pxor "MM"4, "MM"4 \n\t" // 0
184  "mov $-128, %%"REG_a" \n\t"
185  ".p2align 4 \n\t"
186  "1: \n\t"
187  MOVQ" (%1, %%"REG_a"), "MM"0 \n\t" // block[i]
188  SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
189  MOVQ" (%3, %%"REG_a"), "MM"6 \n\t" // bias[0]
190  "paddusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
191  MOVQ" (%2, %%"REG_a"), "MM"5 \n\t" // qmat[i]
192  "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
193  "por "MM"0, "MM"4 \n\t"
194  RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
195  MOVQ" "MM"0, (%5, %%"REG_a") \n\t"
196  "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
197  MOVQ" (%4, %%"REG_a"), "MM"1 \n\t"
198  MOVQ" "MM"7, (%1, %%"REG_a") \n\t" // 0
199  "pandn "MM"1, "MM"0 \n\t"
200  PMAXW(MM"0", MM"3")
201  "add $"MMREG_WIDTH", %%"REG_a" \n\t"
202  " js 1b \n\t"
203  PMAX(MM"3", MM"0")
204  "movd "MM"3, %%"REG_a" \n\t"
205  "movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
206  : "+a" (last_non_zero_p1)
207  : "r" (block+64), "r" (qmat+64), "r" (bias+64),
208  "r" (ff_inv_zigzag_direct16+64), "r" (temp_block+64)
209  XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
210  "%xmm4", "%xmm5", "%xmm6", "%xmm7")
211  );
212  }
213  __asm__ volatile(
214  "movd %1, "MM"1 \n\t" // max_qcoeff
215  SPREADW(MM"1")
216  "psubusw "MM"1, "MM"4 \n\t"
217  "packuswb "MM"4, "MM"4 \n\t"
218 #if COMPILE_TEMPLATE_SSE2
219  "packuswb "MM"4, "MM"4 \n\t"
220 #endif
221  "movd "MM"4, %0 \n\t" // *overflow
222  : "=g" (*overflow)
223  : "g" (s->max_qcoeff)
224  );
225 
226  if(s->mb_intra) block[0]= level;
227  else block[0]= temp_block[0];
228 
229  if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
230  if(last_non_zero_p1 <= 1) goto end;
231  block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
232  block[0x20] = temp_block[0x10];
233  if(last_non_zero_p1 <= 4) goto end;
234  block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
235  block[0x09] = temp_block[0x03];
236  if(last_non_zero_p1 <= 7) goto end;
237  block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
238  block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
239  if(last_non_zero_p1 <= 11) goto end;
240  block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
241  block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
242  block[0x0C] = temp_block[0x05];
243  if(last_non_zero_p1 <= 16) goto end;
244  block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
245  block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
246  block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
247  block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
248  if(last_non_zero_p1 <= 24) goto end;
249  block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
250  block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
251  block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
252  block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
253  if(last_non_zero_p1 <= 32) goto end;
254  block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
255  block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
256  block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
257  block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
258  if(last_non_zero_p1 <= 40) goto end;
259  block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
260  block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
261  block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
262  block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
263  if(last_non_zero_p1 <= 48) goto end;
264  block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
265  block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
266  block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
267  block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
268  if(last_non_zero_p1 <= 56) goto end;
269  block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
270  block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
271  block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
272  block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
273  }else if(s->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
274  if(last_non_zero_p1 <= 1) goto end;
275  block[0x04] = temp_block[0x01];
276  block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
277  if(last_non_zero_p1 <= 4) goto end;
278  block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
279  block[0x05] = temp_block[0x03];
280  if(last_non_zero_p1 <= 7) goto end;
281  block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
282  block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
283  if(last_non_zero_p1 <= 11) goto end;
284  block[0x1C] = temp_block[0x19];
285  block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
286  block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
287  if(last_non_zero_p1 <= 16) goto end;
288  block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
289  block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
290  block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
291  block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
292  if(last_non_zero_p1 <= 24) goto end;
293  block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
294  block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
295  block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
296  block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
297  if(last_non_zero_p1 <= 32) goto end;
298  block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
299  block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
300  block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
301  block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
302  if(last_non_zero_p1 <= 40) goto end;
303  block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
304  block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
305  block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
306  block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
307  if(last_non_zero_p1 <= 48) goto end;
308  block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
309  block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
310  block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
311  block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
312  if(last_non_zero_p1 <= 56) goto end;
313  block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
314  block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
315  block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
316  block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
317  }else{
318  if(last_non_zero_p1 <= 1) goto end;
319  block[0x01] = temp_block[0x01];
320  block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
321  if(last_non_zero_p1 <= 4) goto end;
322  block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
323  block[0x03] = temp_block[0x03];
324  if(last_non_zero_p1 <= 7) goto end;
325  block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
326  block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
327  if(last_non_zero_p1 <= 11) goto end;
328  block[0x19] = temp_block[0x19];
329  block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
330  block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
331  if(last_non_zero_p1 <= 16) goto end;
332  block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
333  block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
334  block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
335  block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
336  if(last_non_zero_p1 <= 24) goto end;
337  block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
338  block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
339  block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
340  block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
341  if(last_non_zero_p1 <= 32) goto end;
342  block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
343  block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
344  block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
345  block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
346  if(last_non_zero_p1 <= 40) goto end;
347  block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
348  block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
349  block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
350  block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
351  if(last_non_zero_p1 <= 48) goto end;
352  block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
353  block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
354  block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
355  block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
356  if(last_non_zero_p1 <= 56) goto end;
357  block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
358  block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
359  block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
360  block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
361  }
362  end:
363  return last_non_zero_p1 - 1;
364 }