2 * MPEG video MMX templates
4 * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
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.
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.
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
23 #define SPREADW(a) "pshufw $0, " #a ", " #a " \n\t"
24 #define PMAXW(a,b) "pmaxsw " #a ", " #b " \n\t"
28 "punpcklwd " #a ", " #a " \n\t"\
29 "punpcklwd " #a ", " #a " \n\t"
31 "psubusw " #a ", " #b " \n\t"\
32 "paddw " #a ", " #b " \n\t"
35 static int RENAME(dct_quantize)(MpegEncContext *s,
36 DCTELEM *block, int n,
37 int qscale, int *overflow)
39 int level=0, last_non_zero_p1, q; //=0 is cuz gcc says uninitalized ...
40 const uint16_t *qmat, *bias;
41 __align8 int16_t temp_block[64];
43 assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
46 ff_fdct_mmx (block); //cant be anything else ...
54 /* note: block[0] is assumed to be positive */
59 : "=d" (level), "=a"(dummy)
60 : "a" ((block[0]>>2) + q), "c" (inverse[q<<1])
64 "xorl %%edx, %%edx \n\t"
66 "movzwl %%ax, %%eax \n\t"
68 : "a" ((block[0]>>2) + q), "c" (q<<1)
73 /* For AIC we skip quant/dequant of INTRADC */
74 level = (block[0] + 4)>>3;
76 block[0]=0; //avoid fake overflow
77 // temp_block[0] = (block[0] + (q >> 1)) / q;
79 bias = s->q_intra_matrix16_bias[qscale];
80 qmat = s->q_intra_matrix16[qscale];
83 bias = s->q_inter_matrix16_bias[qscale];
84 qmat = s->q_inter_matrix16[qscale];
87 if(s->out_format == FMT_H263 && s->mpeg_quant==0){
90 "movd %%eax, %%mm3 \n\t" // last_non_zero_p1
92 "pxor %%mm7, %%mm7 \n\t" // 0
93 "pxor %%mm4, %%mm4 \n\t" // 0
94 "movq (%2), %%mm5 \n\t" // qmat[0]
95 "pxor %%mm6, %%mm6 \n\t"
96 "psubw (%3), %%mm6 \n\t" // -bias[0]
97 "movl $-128, %%eax \n\t"
100 "pxor %%mm1, %%mm1 \n\t" // 0
101 "movq (%1, %%eax), %%mm0 \n\t" // block[i]
102 "pcmpgtw %%mm0, %%mm1 \n\t" // block[i] <= 0 ? 0xFF : 0x00
103 "pxor %%mm1, %%mm0 \n\t"
104 "psubw %%mm1, %%mm0 \n\t" // ABS(block[i])
105 "psubusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
106 "pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
107 "por %%mm0, %%mm4 \n\t"
108 "pxor %%mm1, %%mm0 \n\t"
109 "psubw %%mm1, %%mm0 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
110 "movq %%mm0, (%5, %%eax) \n\t"
111 "pcmpeqw %%mm7, %%mm0 \n\t" // out==0 ? 0xFF : 0x00
112 "movq (%4, %%eax), %%mm1 \n\t"
113 "movq %%mm7, (%1, %%eax) \n\t" // 0
114 "pandn %%mm1, %%mm0 \n\t"
116 "addl $8, %%eax \n\t"
118 "movq %%mm3, %%mm0 \n\t"
119 "psrlq $32, %%mm3 \n\t"
121 "movq %%mm3, %%mm0 \n\t"
122 "psrlq $16, %%mm3 \n\t"
124 "movd %%mm3, %%eax \n\t"
125 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
126 : "+a" (last_non_zero_p1)
127 : "r" (block+64), "r" (qmat), "r" (bias),
128 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
130 // note the asm is split cuz gcc doesnt like that many operands ...
132 "movd %1, %%mm1 \n\t" // max_qcoeff
134 "psubusw %%mm1, %%mm4 \n\t"
135 "packuswb %%mm4, %%mm4 \n\t"
136 "movd %%mm4, %0 \n\t" // *overflow
138 : "g" (s->max_qcoeff)
142 "movd %%eax, %%mm3 \n\t" // last_non_zero_p1
144 "pxor %%mm7, %%mm7 \n\t" // 0
145 "pxor %%mm4, %%mm4 \n\t" // 0
146 "movl $-128, %%eax \n\t"
149 "pxor %%mm1, %%mm1 \n\t" // 0
150 "movq (%1, %%eax), %%mm0 \n\t" // block[i]
151 "pcmpgtw %%mm0, %%mm1 \n\t" // block[i] <= 0 ? 0xFF : 0x00
152 "pxor %%mm1, %%mm0 \n\t"
153 "psubw %%mm1, %%mm0 \n\t" // ABS(block[i])
154 "movq (%3, %%eax), %%mm6 \n\t" // bias[0]
155 "paddusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
156 "movq (%2, %%eax), %%mm5 \n\t" // qmat[i]
157 "pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
158 "por %%mm0, %%mm4 \n\t"
159 "pxor %%mm1, %%mm0 \n\t"
160 "psubw %%mm1, %%mm0 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
161 "movq %%mm0, (%5, %%eax) \n\t"
162 "pcmpeqw %%mm7, %%mm0 \n\t" // out==0 ? 0xFF : 0x00
163 "movq (%4, %%eax), %%mm1 \n\t"
164 "movq %%mm7, (%1, %%eax) \n\t" // 0
165 "pandn %%mm1, %%mm0 \n\t"
167 "addl $8, %%eax \n\t"
169 "movq %%mm3, %%mm0 \n\t"
170 "psrlq $32, %%mm3 \n\t"
172 "movq %%mm3, %%mm0 \n\t"
173 "psrlq $16, %%mm3 \n\t"
175 "movd %%mm3, %%eax \n\t"
176 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
177 : "+a" (last_non_zero_p1)
178 : "r" (block+64), "r" (qmat+64), "r" (bias+64),
179 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
181 // note the asm is split cuz gcc doesnt like that many operands ...
183 "movd %1, %%mm1 \n\t" // max_qcoeff
185 "psubusw %%mm1, %%mm4 \n\t"
186 "packuswb %%mm4, %%mm4 \n\t"
187 "movd %%mm4, %0 \n\t" // *overflow
189 : "g" (s->max_qcoeff)
193 if(s->mb_intra) block[0]= level;
194 else block[0]= temp_block[0];
196 if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
197 if(last_non_zero_p1 <= 1) goto end;
198 block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
199 block[0x20] = temp_block[0x10];
200 if(last_non_zero_p1 <= 4) goto end;
201 block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
202 block[0x09] = temp_block[0x03];
203 if(last_non_zero_p1 <= 7) goto end;
204 block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
205 block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
206 if(last_non_zero_p1 <= 11) goto end;
207 block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
208 block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
209 block[0x0C] = temp_block[0x05];
210 if(last_non_zero_p1 <= 16) goto end;
211 block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
212 block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
213 block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
214 block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
215 if(last_non_zero_p1 <= 24) goto end;
216 block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
217 block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
218 block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
219 block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
220 if(last_non_zero_p1 <= 32) goto end;
221 block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
222 block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
223 block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
224 block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
225 if(last_non_zero_p1 <= 40) goto end;
226 block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
227 block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
228 block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
229 block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
230 if(last_non_zero_p1 <= 48) goto end;
231 block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
232 block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
233 block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
234 block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
235 if(last_non_zero_p1 <= 56) goto end;
236 block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
237 block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
238 block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
239 block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
240 }else if(s->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
241 if(last_non_zero_p1 <= 1) goto end;
242 block[0x04] = temp_block[0x01];
243 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
244 if(last_non_zero_p1 <= 4) goto end;
245 block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
246 block[0x05] = temp_block[0x03];
247 if(last_non_zero_p1 <= 7) goto end;
248 block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
249 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
250 if(last_non_zero_p1 <= 11) goto end;
251 block[0x1C] = temp_block[0x19];
252 block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
253 block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
254 if(last_non_zero_p1 <= 16) goto end;
255 block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
256 block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
257 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
258 block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
259 if(last_non_zero_p1 <= 24) goto end;
260 block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
261 block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
262 block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
263 block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
264 if(last_non_zero_p1 <= 32) goto end;
265 block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
266 block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
267 block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
268 block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
269 if(last_non_zero_p1 <= 40) goto end;
270 block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
271 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
272 block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
273 block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
274 if(last_non_zero_p1 <= 48) goto end;
275 block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
276 block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
277 block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
278 block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
279 if(last_non_zero_p1 <= 56) goto end;
280 block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
281 block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
282 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
283 block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
285 if(last_non_zero_p1 <= 1) goto end;
286 block[0x01] = temp_block[0x01];
287 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
288 if(last_non_zero_p1 <= 4) goto end;
289 block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
290 block[0x03] = temp_block[0x03];
291 if(last_non_zero_p1 <= 7) goto end;
292 block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
293 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
294 if(last_non_zero_p1 <= 11) goto end;
295 block[0x19] = temp_block[0x19];
296 block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
297 block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
298 if(last_non_zero_p1 <= 16) goto end;
299 block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
300 block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
301 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
302 block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
303 if(last_non_zero_p1 <= 24) goto end;
304 block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
305 block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
306 block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
307 block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
308 if(last_non_zero_p1 <= 32) goto end;
309 block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
310 block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
311 block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
312 block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
313 if(last_non_zero_p1 <= 40) goto end;
314 block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
315 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
316 block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
317 block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
318 if(last_non_zero_p1 <= 48) goto end;
319 block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
320 block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
321 block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
322 block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
323 if(last_non_zero_p1 <= 56) goto end;
324 block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
325 block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
326 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
327 block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
331 for(i=0; i<last_non_zero_p1; i++)
333 int j= zigzag_direct_noperm[i];
334 block[block_permute_op(j)]= temp_block[j];
338 return last_non_zero_p1 - 1;