Mercurial > libavcodec.hg

comparison x86/mpegvideo_mmx_template.c @ 8430:7768bdfd4f7b libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Rename libavcodec/i386/ --> libavcodec/x86/. It contains optimizations that are not specific to i386 and libavutil uses this naming scheme already.
author diego
date Mon, 22 Dec 2008 09:12:42 +0000
parents
children 7a463923ecd1
comparison
equal deleted inserted replaced
8429:b3ecaba81501 8430:7768bdfd4f7b
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 defined(HAVE_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 defined(HAVE_MMX2)
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 #ifdef HAVE_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 DCTELEM *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 DECLARE_ALIGNED_16(int16_t, temp_block[64]);
102
103 assert((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 else
116 q = s->c_dc_scale;
117 /* note: block[0] is assumed to be positive */
118 if (!s->h263_aic) {
119 #if 1
120 __asm__ volatile (
121 "mul %%ecx \n\t"
122 : "=d" (level), "=a"(dummy)
123 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
124 );
125 #else
126 __asm__ volatile (
127 "xorl %%edx, %%edx \n\t"
128 "divw %%cx \n\t"
129 "movzwl %%ax, %%eax \n\t"
130 : "=a" (level)
131 : "a" ((block[0]>>2) + q), "c" (q<<1)
132 : "%edx"
133 );
134 #endif
135 } else
136 /* For AIC we skip quant/dequant of INTRADC */
137 level = (block[0] + 4)>>3;
138
139 block[0]=0; //avoid fake overflow
140 // temp_block[0] = (block[0] + (q >> 1)) / q;
141 last_non_zero_p1 = 1;
142 bias = s->q_intra_matrix16[qscale][1];
143 qmat = s->q_intra_matrix16[qscale][0];
144 } else {
145 last_non_zero_p1 = 0;
146 bias = s->q_inter_matrix16[qscale][1];
147 qmat = s->q_inter_matrix16[qscale][0];
148 }
149
150 if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
151
152 __asm__ volatile(
153 "movd %%"REG_a", "MM"3 \n\t" // last_non_zero_p1
154 SPREADW(MM"3")
155 "pxor "MM"7, "MM"7 \n\t" // 0
156 "pxor "MM"4, "MM"4 \n\t" // 0
157 MOVQ" (%2), "MM"5 \n\t" // qmat[0]
158 "pxor "MM"6, "MM"6 \n\t"
159 "psubw (%3), "MM"6 \n\t" // -bias[0]
160 "mov $-128, %%"REG_a" \n\t"
161 ASMALIGN(4)
162 "1: \n\t"
163 MOVQ" (%1, %%"REG_a"), "MM"0 \n\t" // block[i]
164 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
165 "psubusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
166 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
167 "por "MM"0, "MM"4 \n\t"
168 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
169 MOVQ" "MM"0, (%5, %%"REG_a") \n\t"
170 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
171 MOVQ" (%4, %%"REG_a"), "MM"1 \n\t"
172 MOVQ" "MM"7, (%1, %%"REG_a") \n\t" // 0
173 "pandn "MM"1, "MM"0 \n\t"
174 PMAXW(MM"0", MM"3")
175 "add $"MMREG_WIDTH", %%"REG_a" \n\t"
176 " js 1b \n\t"
177 PMAX(MM"3", MM"0")
178 "movd "MM"3, %%"REG_a" \n\t"
179 "movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
180 : "+a" (last_non_zero_p1)
181 : "r" (block+64), "r" (qmat), "r" (bias),
182 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
183 );
184 }else{ // FMT_H263
185 __asm__ volatile(
186 "movd %%"REG_a", "MM"3 \n\t" // last_non_zero_p1
187 SPREADW(MM"3")
188 "pxor "MM"7, "MM"7 \n\t" // 0
189 "pxor "MM"4, "MM"4 \n\t" // 0
190 "mov $-128, %%"REG_a" \n\t"
191 ASMALIGN(4)
192 "1: \n\t"
193 MOVQ" (%1, %%"REG_a"), "MM"0 \n\t" // block[i]
194 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
195 MOVQ" (%3, %%"REG_a"), "MM"6 \n\t" // bias[0]
196 "paddusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
197 MOVQ" (%2, %%"REG_a"), "MM"5 \n\t" // qmat[i]
198 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
199 "por "MM"0, "MM"4 \n\t"
200 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
201 MOVQ" "MM"0, (%5, %%"REG_a") \n\t"
202 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
203 MOVQ" (%4, %%"REG_a"), "MM"1 \n\t"
204 MOVQ" "MM"7, (%1, %%"REG_a") \n\t" // 0
205 "pandn "MM"1, "MM"0 \n\t"
206 PMAXW(MM"0", MM"3")
207 "add $"MMREG_WIDTH", %%"REG_a" \n\t"
208 " js 1b \n\t"
209 PMAX(MM"3", MM"0")
210 "movd "MM"3, %%"REG_a" \n\t"
211 "movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
212 : "+a" (last_non_zero_p1)
213 : "r" (block+64), "r" (qmat+64), "r" (bias+64),
214 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
215 );
216 }
217 __asm__ volatile(
218 "movd %1, "MM"1 \n\t" // max_qcoeff
219 SPREADW(MM"1")
220 "psubusw "MM"1, "MM"4 \n\t"
221 "packuswb "MM"4, "MM"4 \n\t"
222 #ifdef HAVE_SSE2
223 "packuswb "MM"4, "MM"4 \n\t"
224 #endif
225 "movd "MM"4, %0 \n\t" // *overflow
226 : "=g" (*overflow)
227 : "g" (s->max_qcoeff)
228 );
229
230 if(s->mb_intra) block[0]= level;
231 else block[0]= temp_block[0];
232
233 if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
234 if(last_non_zero_p1 <= 1) goto end;
235 block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
236 block[0x20] = temp_block[0x10];
237 if(last_non_zero_p1 <= 4) goto end;
238 block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
239 block[0x09] = temp_block[0x03];
240 if(last_non_zero_p1 <= 7) goto end;
241 block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
242 block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
243 if(last_non_zero_p1 <= 11) goto end;
244 block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
245 block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
246 block[0x0C] = temp_block[0x05];
247 if(last_non_zero_p1 <= 16) goto end;
248 block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
249 block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
250 block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
251 block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
252 if(last_non_zero_p1 <= 24) goto end;
253 block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
254 block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
255 block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
256 block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
257 if(last_non_zero_p1 <= 32) goto end;
258 block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
259 block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
260 block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
261 block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
262 if(last_non_zero_p1 <= 40) goto end;
263 block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
264 block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
265 block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
266 block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
267 if(last_non_zero_p1 <= 48) goto end;
268 block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
269 block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
270 block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
271 block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
272 if(last_non_zero_p1 <= 56) goto end;
273 block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
274 block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
275 block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
276 block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
277 }else if(s->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
278 if(last_non_zero_p1 <= 1) goto end;
279 block[0x04] = temp_block[0x01];
280 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
281 if(last_non_zero_p1 <= 4) goto end;
282 block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
283 block[0x05] = temp_block[0x03];
284 if(last_non_zero_p1 <= 7) goto end;
285 block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
286 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
287 if(last_non_zero_p1 <= 11) goto end;
288 block[0x1C] = temp_block[0x19];
289 block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
290 block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
291 if(last_non_zero_p1 <= 16) goto end;
292 block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
293 block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
294 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
295 block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
296 if(last_non_zero_p1 <= 24) goto end;
297 block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
298 block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
299 block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
300 block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
301 if(last_non_zero_p1 <= 32) goto end;
302 block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
303 block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
304 block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
305 block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
306 if(last_non_zero_p1 <= 40) goto end;
307 block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
308 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
309 block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
310 block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
311 if(last_non_zero_p1 <= 48) goto end;
312 block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
313 block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
314 block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
315 block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
316 if(last_non_zero_p1 <= 56) goto end;
317 block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
318 block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
319 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
320 block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
321 }else{
322 if(last_non_zero_p1 <= 1) goto end;
323 block[0x01] = temp_block[0x01];
324 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
325 if(last_non_zero_p1 <= 4) goto end;
326 block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
327 block[0x03] = temp_block[0x03];
328 if(last_non_zero_p1 <= 7) goto end;
329 block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
330 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
331 if(last_non_zero_p1 <= 11) goto end;
332 block[0x19] = temp_block[0x19];
333 block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
334 block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
335 if(last_non_zero_p1 <= 16) goto end;
336 block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
337 block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
338 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
339 block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
340 if(last_non_zero_p1 <= 24) goto end;
341 block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
342 block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
343 block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
344 block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
345 if(last_non_zero_p1 <= 32) goto end;
346 block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
347 block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
348 block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
349 block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
350 if(last_non_zero_p1 <= 40) goto end;
351 block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
352 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
353 block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
354 block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
355 if(last_non_zero_p1 <= 48) goto end;
356 block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
357 block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
358 block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
359 block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
360 if(last_non_zero_p1 <= 56) goto end;
361 block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
362 block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
363 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
364 block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
365 }
366 end:
367 /*
368 for(i=0; i<last_non_zero_p1; i++)
369 {
370 int j= zigzag_direct_noperm[i];
371 block[block_permute_op(j)]= temp_block[j];
372 }
373 */
374
375 return last_non_zero_p1 - 1;
376 }