2c50df232df8967a44b41ebd8b166b0782506672
[libav.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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 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 #define PMAX(a,b) \
26 "pshufw $0x0E," #a ", " #b " \n\t"\
27 PMAXW(b, a)\
28 "pshufw $0x01," #a ", " #b " \n\t"\
29 PMAXW(b, a)
30 #else
31 #define SPREADW(a) \
32 "punpcklwd " #a ", " #a " \n\t"\
33 "punpcklwd " #a ", " #a " \n\t"
34 #define PMAXW(a,b) \
35 "psubusw " #a ", " #b " \n\t"\
36 "paddw " #a ", " #b " \n\t"
37 #define PMAX(a,b) \
38 "movq " #a ", " #b " \n\t"\
39 "psrlq $32, " #a " \n\t"\
40 PMAXW(b, a)\
41 "movq " #a ", " #b " \n\t"\
42 "psrlq $16, " #a " \n\t"\
43 PMAXW(b, a)
44
45 #endif
46
47 static int RENAME(dct_quantize)(MpegEncContext *s,
48 DCTELEM *block, int n,
49 int qscale, int *overflow)
50 {
51 long last_non_zero_p1;
52 int level=0, q; //=0 is cuz gcc says uninitalized ...
53 const uint16_t *qmat, *bias;
54 DECLARE_ALIGNED_8(int16_t, temp_block[64]);
55
56 assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
57
58 //s->fdct (block);
59 RENAMEl(ff_fdct) (block); //cant be anything else ...
60
61 if(s->dct_error_sum)
62 s->denoise_dct(s, block);
63
64 if (s->mb_intra) {
65 int dummy;
66 if (n < 4)
67 q = s->y_dc_scale;
68 else
69 q = s->c_dc_scale;
70 /* note: block[0] is assumed to be positive */
71 if (!s->h263_aic) {
72 #if 1
73 asm volatile (
74 "mul %%ecx \n\t"
75 : "=d" (level), "=a"(dummy)
76 : "a" ((block[0]>>2) + q), "c" (inverse[q<<1])
77 );
78 #else
79 asm volatile (
80 "xorl %%edx, %%edx \n\t"
81 "divw %%cx \n\t"
82 "movzwl %%ax, %%eax \n\t"
83 : "=a" (level)
84 : "a" ((block[0]>>2) + q), "c" (q<<1)
85 : "%edx"
86 );
87 #endif
88 } else
89 /* For AIC we skip quant/dequant of INTRADC */
90 level = (block[0] + 4)>>3;
91
92 block[0]=0; //avoid fake overflow
93 // temp_block[0] = (block[0] + (q >> 1)) / q;
94 last_non_zero_p1 = 1;
95 bias = s->q_intra_matrix16[qscale][1];
96 qmat = s->q_intra_matrix16[qscale][0];
97 } else {
98 last_non_zero_p1 = 0;
99 bias = s->q_inter_matrix16[qscale][1];
100 qmat = s->q_inter_matrix16[qscale][0];
101 }
102
103 if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
104
105 asm volatile(
106 "movd %%"REG_a", %%mm3 \n\t" // last_non_zero_p1
107 SPREADW(%%mm3)
108 "pxor %%mm7, %%mm7 \n\t" // 0
109 "pxor %%mm4, %%mm4 \n\t" // 0
110 "movq (%2), %%mm5 \n\t" // qmat[0]
111 "pxor %%mm6, %%mm6 \n\t"
112 "psubw (%3), %%mm6 \n\t" // -bias[0]
113 "mov $-128, %%"REG_a" \n\t"
114 ".balign 16 \n\t"
115 "1: \n\t"
116 "pxor %%mm1, %%mm1 \n\t" // 0
117 "movq (%1, %%"REG_a"), %%mm0 \n\t" // block[i]
118 "pcmpgtw %%mm0, %%mm1 \n\t" // block[i] <= 0 ? 0xFF : 0x00
119 "pxor %%mm1, %%mm0 \n\t"
120 "psubw %%mm1, %%mm0 \n\t" // ABS(block[i])
121 "psubusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
122 "pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
123 "por %%mm0, %%mm4 \n\t"
124 "pxor %%mm1, %%mm0 \n\t"
125 "psubw %%mm1, %%mm0 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
126 "movq %%mm0, (%5, %%"REG_a") \n\t"
127 "pcmpeqw %%mm7, %%mm0 \n\t" // out==0 ? 0xFF : 0x00
128 "movq (%4, %%"REG_a"), %%mm1 \n\t"
129 "movq %%mm7, (%1, %%"REG_a") \n\t" // 0
130 "pandn %%mm1, %%mm0 \n\t"
131 PMAXW(%%mm0, %%mm3)
132 "add $8, %%"REG_a" \n\t"
133 " js 1b \n\t"
134 PMAX(%%mm3, %%mm0)
135 "movd %%mm3, %%"REG_a" \n\t"
136 "movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
137 : "+a" (last_non_zero_p1)
138 : "r" (block+64), "r" (qmat), "r" (bias),
139 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
140 );
141 // note the asm is split cuz gcc doesnt like that many operands ...
142 asm volatile(
143 "movd %1, %%mm1 \n\t" // max_qcoeff
144 SPREADW(%%mm1)
145 "psubusw %%mm1, %%mm4 \n\t"
146 "packuswb %%mm4, %%mm4 \n\t"
147 "movd %%mm4, %0 \n\t" // *overflow
148 : "=g" (*overflow)
149 : "g" (s->max_qcoeff)
150 );
151 }else{ // FMT_H263
152 asm volatile(
153 "movd %%"REG_a", %%mm3 \n\t" // last_non_zero_p1
154 SPREADW(%%mm3)
155 "pxor %%mm7, %%mm7 \n\t" // 0
156 "pxor %%mm4, %%mm4 \n\t" // 0
157 "mov $-128, %%"REG_a" \n\t"
158 ".balign 16 \n\t"
159 "1: \n\t"
160 "pxor %%mm1, %%mm1 \n\t" // 0
161 "movq (%1, %%"REG_a"), %%mm0 \n\t" // block[i]
162 "pcmpgtw %%mm0, %%mm1 \n\t" // block[i] <= 0 ? 0xFF : 0x00
163 "pxor %%mm1, %%mm0 \n\t"
164 "psubw %%mm1, %%mm0 \n\t" // ABS(block[i])
165 "movq (%3, %%"REG_a"), %%mm6 \n\t" // bias[0]
166 "paddusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
167 "movq (%2, %%"REG_a"), %%mm5 \n\t" // qmat[i]
168 "pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
169 "por %%mm0, %%mm4 \n\t"
170 "pxor %%mm1, %%mm0 \n\t"
171 "psubw %%mm1, %%mm0 \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
172 "movq %%mm0, (%5, %%"REG_a") \n\t"
173 "pcmpeqw %%mm7, %%mm0 \n\t" // out==0 ? 0xFF : 0x00
174 "movq (%4, %%"REG_a"), %%mm1 \n\t"
175 "movq %%mm7, (%1, %%"REG_a") \n\t" // 0
176 "pandn %%mm1, %%mm0 \n\t"
177 PMAXW(%%mm0, %%mm3)
178 "add $8, %%"REG_a" \n\t"
179 " js 1b \n\t"
180 PMAX(%%mm3, %%mm0)
181 "movd %%mm3, %%"REG_a" \n\t"
182 "movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
183 : "+a" (last_non_zero_p1)
184 : "r" (block+64), "r" (qmat+64), "r" (bias+64),
185 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
186 );
187 // note the asm is split cuz gcc doesnt like that many operands ...
188 asm volatile(
189 "movd %1, %%mm1 \n\t" // max_qcoeff
190 SPREADW(%%mm1)
191 "psubusw %%mm1, %%mm4 \n\t"
192 "packuswb %%mm4, %%mm4 \n\t"
193 "movd %%mm4, %0 \n\t" // *overflow
194 : "=g" (*overflow)
195 : "g" (s->max_qcoeff)
196 );
197 }
198
199 if(s->mb_intra) block[0]= level;
200 else block[0]= temp_block[0];
201
202 if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
203 if(last_non_zero_p1 <= 1) goto end;
204 block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
205 block[0x20] = temp_block[0x10];
206 if(last_non_zero_p1 <= 4) goto end;
207 block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
208 block[0x09] = temp_block[0x03];
209 if(last_non_zero_p1 <= 7) goto end;
210 block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
211 block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
212 if(last_non_zero_p1 <= 11) goto end;
213 block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
214 block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
215 block[0x0C] = temp_block[0x05];
216 if(last_non_zero_p1 <= 16) goto end;
217 block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
218 block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
219 block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
220 block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
221 if(last_non_zero_p1 <= 24) goto end;
222 block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
223 block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
224 block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
225 block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
226 if(last_non_zero_p1 <= 32) goto end;
227 block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
228 block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
229 block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
230 block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
231 if(last_non_zero_p1 <= 40) goto end;
232 block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
233 block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
234 block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
235 block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
236 if(last_non_zero_p1 <= 48) goto end;
237 block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
238 block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
239 block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
240 block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
241 if(last_non_zero_p1 <= 56) goto end;
242 block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
243 block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
244 block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
245 block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
246 }else if(s->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
247 if(last_non_zero_p1 <= 1) goto end;
248 block[0x04] = temp_block[0x01];
249 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
250 if(last_non_zero_p1 <= 4) goto end;
251 block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
252 block[0x05] = temp_block[0x03];
253 if(last_non_zero_p1 <= 7) goto end;
254 block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
255 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
256 if(last_non_zero_p1 <= 11) goto end;
257 block[0x1C] = temp_block[0x19];
258 block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
259 block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
260 if(last_non_zero_p1 <= 16) goto end;
261 block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
262 block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
263 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
264 block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
265 if(last_non_zero_p1 <= 24) goto end;
266 block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
267 block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
268 block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
269 block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
270 if(last_non_zero_p1 <= 32) goto end;
271 block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
272 block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
273 block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
274 block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
275 if(last_non_zero_p1 <= 40) goto end;
276 block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
277 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
278 block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
279 block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
280 if(last_non_zero_p1 <= 48) goto end;
281 block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
282 block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
283 block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
284 block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
285 if(last_non_zero_p1 <= 56) goto end;
286 block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
287 block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
288 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
289 block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
290 }else{
291 if(last_non_zero_p1 <= 1) goto end;
292 block[0x01] = temp_block[0x01];
293 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
294 if(last_non_zero_p1 <= 4) goto end;
295 block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
296 block[0x03] = temp_block[0x03];
297 if(last_non_zero_p1 <= 7) goto end;
298 block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
299 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
300 if(last_non_zero_p1 <= 11) goto end;
301 block[0x19] = temp_block[0x19];
302 block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
303 block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
304 if(last_non_zero_p1 <= 16) goto end;
305 block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
306 block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
307 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
308 block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
309 if(last_non_zero_p1 <= 24) goto end;
310 block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
311 block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
312 block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
313 block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
314 if(last_non_zero_p1 <= 32) goto end;
315 block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
316 block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
317 block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
318 block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
319 if(last_non_zero_p1 <= 40) goto end;
320 block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
321 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
322 block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
323 block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
324 if(last_non_zero_p1 <= 48) goto end;
325 block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
326 block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
327 block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
328 block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
329 if(last_non_zero_p1 <= 56) goto end;
330 block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
331 block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
332 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
333 block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
334 }
335 end:
336 /*
337 for(i=0; i<last_non_zero_p1; i++)
338 {
339 int j= zigzag_direct_noperm[i];
340 block[block_permute_op(j)]= temp_block[j];
341 }
342 */
343
344 return last_non_zero_p1 - 1;
345 }