2 * MPEG video MMX templates
4 * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
6 * This file is part of FFmpeg.
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.
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.
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
25 #include "libavutil/internal.h"
26 #include "libavutil/mem_internal.h"
27 #include "libavutil/x86/asm.h"
28 #include "libavcodec/mpegutils.h"
29 #include "libavcodec/mpegvideo.h"
41 #if COMPILE_TEMPLATE_SSE2
42 #define MMREG_WIDTH "16"
46 "pshuflw $0, "a", "a" \n\t"\
47 "punpcklwd "a", "a" \n\t"
48 #define PMAXW(a,b) "pmaxsw "a", "b" \n\t"
50 "movhlps "a", "b" \n\t"\
52 "pshuflw $0x0E, "a", "b" \n\t"\
54 "pshuflw $0x01, "a", "b" \n\t"\
57 #define MMREG_WIDTH "8"
60 #if COMPILE_TEMPLATE_MMXEXT
61 #define SPREADW(a) "pshufw $0, "a", "a" \n\t"
62 #define PMAXW(a,b) "pmaxsw "a", "b" \n\t"
64 "pshufw $0x0E, "a", "b" \n\t"\
66 "pshufw $0x01, "a", "b" \n\t"\
70 "punpcklwd "a", "a" \n\t"\
71 "punpcklwd "a", "a" \n\t"
73 "psubusw "a", "b" \n\t"\
77 "psrlq $32, "a" \n\t"\
80 "psrlq $16, "a" \n\t"\
86 #if COMPILE_TEMPLATE_SSSE3
87 #define SAVE_SIGN(a,b) \
88 "movdqa "b", "a" \n\t"\
90 #define RESTORE_SIGN(a,b) \
91 "psignw "a", "b" \n\t"
93 #define SAVE_SIGN(a,b) \
95 "pcmpgtw "b", "a" \n\t" /* block[i] <= 0 ? 0xFF : 0x00 */\
97 "psubw "a", "b" \n\t" /* ABS(block[i]) */
98 #define RESTORE_SIGN(a,b) \
100 "psubw "a", "b" \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
103 static int RENAME(dct_quantize)(MpegEncContext *s,
104 int16_t *block, int n,
105 int qscale, int *overflow)
107 x86_reg last_non_zero_p1;
108 int level=0, q; //=0 is because gcc says uninitialized ...
109 const uint16_t *qmat, *bias;
110 LOCAL_ALIGNED_16(int16_t, temp_block, [64]);
112 av_assert2((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
115 RENAME_FDCT(ff_fdct)(block); // cannot be anything else ...
118 s->denoise_dct(s, block);
124 bias = s->q_intra_matrix16[qscale][1];
125 qmat = s->q_intra_matrix16[qscale][0];
128 bias = s->q_chroma_intra_matrix16[qscale][1];
129 qmat = s->q_chroma_intra_matrix16[qscale][0];
131 /* note: block[0] is assumed to be positive */
135 : "=d" (level), "=a"(dummy)
136 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
139 /* For AIC we skip quant/dequant of INTRADC */
140 level = (block[0] + 4)>>3;
142 block[0]=0; //avoid fake overflow
143 // temp_block[0] = (block[0] + (q >> 1)) / q;
144 last_non_zero_p1 = 1;
146 last_non_zero_p1 = 0;
147 bias = s->q_inter_matrix16[qscale][1];
148 qmat = s->q_inter_matrix16[qscale][0];
151 if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
154 "movd %%"FF_REG_a", "MM"3 \n\t" // last_non_zero_p1
156 "pxor "MM"7, "MM"7 \n\t" // 0
157 "pxor "MM"4, "MM"4 \n\t" // 0
158 MOVQ" (%2), "MM"5 \n\t" // qmat[0]
159 "pxor "MM"6, "MM"6 \n\t"
160 "psubw (%3), "MM"6 \n\t" // -bias[0]
161 "mov $-128, %%"FF_REG_a" \n\t"
164 MOVQ" (%1, %%"FF_REG_a"), "MM"0 \n\t" // block[i]
165 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
166 "psubusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
167 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
168 "por "MM"0, "MM"4 \n\t"
169 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
170 MOVQ" "MM"0, (%5, %%"FF_REG_a") \n\t"
171 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
172 MOVQ" (%4, %%"FF_REG_a"), "MM"1 \n\t"
173 MOVQ" "MM"7, (%1, %%"FF_REG_a") \n\t" // 0
174 "pandn "MM"1, "MM"0 \n\t"
176 "add $"MMREG_WIDTH", %%"FF_REG_a" \n\t"
179 "movd "MM"3, %%"FF_REG_a" \n\t"
180 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
181 : "+a" (last_non_zero_p1)
182 : "r" (block+64), "r" (qmat), "r" (bias),
183 "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
184 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
185 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
189 "movd %%"FF_REG_a", "MM"3 \n\t" // last_non_zero_p1
191 "pxor "MM"7, "MM"7 \n\t" // 0
192 "pxor "MM"4, "MM"4 \n\t" // 0
193 "mov $-128, %%"FF_REG_a" \n\t"
196 MOVQ" (%1, %%"FF_REG_a"), "MM"0 \n\t" // block[i]
197 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
198 MOVQ" (%3, %%"FF_REG_a"), "MM"6 \n\t" // bias[0]
199 "paddusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
200 MOVQ" (%2, %%"FF_REG_a"), "MM"5 \n\t" // qmat[i]
201 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
202 "por "MM"0, "MM"4 \n\t"
203 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
204 MOVQ" "MM"0, (%5, %%"FF_REG_a") \n\t"
205 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
206 MOVQ" (%4, %%"FF_REG_a"), "MM"1 \n\t"
207 MOVQ" "MM"7, (%1, %%"FF_REG_a") \n\t" // 0
208 "pandn "MM"1, "MM"0 \n\t"
210 "add $"MMREG_WIDTH", %%"FF_REG_a" \n\t"
213 "movd "MM"3, %%"FF_REG_a" \n\t"
214 "movzbl %%al, %%eax \n\t" // last_non_zero_p1
215 : "+a" (last_non_zero_p1)
216 : "r" (block+64), "r" (qmat+64), "r" (bias+64),
217 "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
218 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
219 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
223 "movd %1, "MM"1 \n\t" // max_qcoeff
225 "psubusw "MM"1, "MM"4 \n\t"
226 "packuswb "MM"4, "MM"4 \n\t"
227 #if COMPILE_TEMPLATE_SSE2
228 "packsswb "MM"4, "MM"4 \n\t"
230 "movd "MM"4, %0 \n\t" // *overflow
232 : "g" (s->max_qcoeff)
235 if(s->mb_intra) block[0]= level;
236 else block[0]= temp_block[0];
238 if (s->idsp.perm_type == FF_IDCT_PERM_SIMPLE) {
239 if(last_non_zero_p1 <= 1) goto end;
240 block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
241 block[0x20] = temp_block[0x10];
242 if(last_non_zero_p1 <= 4) goto end;
243 block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
244 block[0x09] = temp_block[0x03];
245 if(last_non_zero_p1 <= 7) goto end;
246 block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
247 block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
248 if(last_non_zero_p1 <= 11) goto end;
249 block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
250 block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
251 block[0x0C] = temp_block[0x05];
252 if(last_non_zero_p1 <= 16) goto end;
253 block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
254 block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
255 block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
256 block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
257 if(last_non_zero_p1 <= 24) goto end;
258 block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
259 block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
260 block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
261 block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
262 if(last_non_zero_p1 <= 32) goto end;
263 block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
264 block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
265 block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
266 block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
267 if(last_non_zero_p1 <= 40) goto end;
268 block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
269 block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
270 block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
271 block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
272 if(last_non_zero_p1 <= 48) goto end;
273 block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
274 block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
275 block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
276 block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
277 if(last_non_zero_p1 <= 56) goto end;
278 block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
279 block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
280 block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
281 block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
282 }else if(s->idsp.perm_type == FF_IDCT_PERM_LIBMPEG2){
283 if(last_non_zero_p1 <= 1) goto end;
284 block[0x04] = temp_block[0x01];
285 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
286 if(last_non_zero_p1 <= 4) goto end;
287 block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
288 block[0x05] = temp_block[0x03];
289 if(last_non_zero_p1 <= 7) goto end;
290 block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
291 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
292 if(last_non_zero_p1 <= 11) goto end;
293 block[0x1C] = temp_block[0x19];
294 block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
295 block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
296 if(last_non_zero_p1 <= 16) goto end;
297 block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
298 block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
299 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
300 block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
301 if(last_non_zero_p1 <= 24) goto end;
302 block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
303 block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
304 block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
305 block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
306 if(last_non_zero_p1 <= 32) goto end;
307 block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
308 block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
309 block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
310 block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
311 if(last_non_zero_p1 <= 40) goto end;
312 block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
313 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
314 block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
315 block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
316 if(last_non_zero_p1 <= 48) goto end;
317 block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
318 block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
319 block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
320 block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
321 if(last_non_zero_p1 <= 56) goto end;
322 block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
323 block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
324 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
325 block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
326 } else if (s->idsp.perm_type == FF_IDCT_PERM_NONE) {
327 if(last_non_zero_p1 <= 1) goto end;
328 block[0x01] = temp_block[0x01];
329 block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
330 if(last_non_zero_p1 <= 4) goto end;
331 block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
332 block[0x03] = temp_block[0x03];
333 if(last_non_zero_p1 <= 7) goto end;
334 block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
335 block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
336 if(last_non_zero_p1 <= 11) goto end;
337 block[0x19] = temp_block[0x19];
338 block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
339 block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
340 if(last_non_zero_p1 <= 16) goto end;
341 block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
342 block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
343 block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
344 block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
345 if(last_non_zero_p1 <= 24) goto end;
346 block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
347 block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
348 block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
349 block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
350 if(last_non_zero_p1 <= 32) goto end;
351 block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
352 block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
353 block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
354 block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
355 if(last_non_zero_p1 <= 40) goto end;
356 block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
357 block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
358 block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
359 block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
360 if(last_non_zero_p1 <= 48) goto end;
361 block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
362 block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
363 block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
364 block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
365 if(last_non_zero_p1 <= 56) goto end;
366 block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
367 block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
368 block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
369 block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
370 } else if (s->idsp.perm_type == FF_IDCT_PERM_TRANSPOSE) {
371 if(last_non_zero_p1 <= 1) goto end;
372 block[0x08] = temp_block[0x01];
373 block[0x01] = temp_block[0x08]; block[0x02] = temp_block[0x10];
374 if(last_non_zero_p1 <= 4) goto end;
375 block[0x09] = temp_block[0x09]; block[0x10] = temp_block[0x02];
376 block[0x18] = temp_block[0x03];
377 if(last_non_zero_p1 <= 7) goto end;
378 block[0x11] = temp_block[0x0A]; block[0x0A] = temp_block[0x11];
379 block[0x03] = temp_block[0x18]; block[0x04] = temp_block[0x20];
380 if(last_non_zero_p1 <= 11) goto end;
381 block[0x0B] = temp_block[0x19];
382 block[0x12] = temp_block[0x12]; block[0x19] = temp_block[0x0B];
383 block[0x20] = temp_block[0x04]; block[0x28] = temp_block[0x05];
384 if(last_non_zero_p1 <= 16) goto end;
385 block[0x21] = temp_block[0x0C]; block[0x1A] = temp_block[0x13];
386 block[0x13] = temp_block[0x1A]; block[0x0C] = temp_block[0x21];
387 block[0x05] = temp_block[0x28]; block[0x06] = temp_block[0x30];
388 block[0x0D] = temp_block[0x29]; block[0x14] = temp_block[0x22];
389 if(last_non_zero_p1 <= 24) goto end;
390 block[0x1B] = temp_block[0x1B]; block[0x22] = temp_block[0x14];
391 block[0x29] = temp_block[0x0D]; block[0x30] = temp_block[0x06];
392 block[0x38] = temp_block[0x07]; block[0x31] = temp_block[0x0E];
393 block[0x2A] = temp_block[0x15]; block[0x23] = temp_block[0x1C];
394 if(last_non_zero_p1 <= 32) goto end;
395 block[0x1C] = temp_block[0x23]; block[0x15] = temp_block[0x2A];
396 block[0x0E] = temp_block[0x31]; block[0x07] = temp_block[0x38];
397 block[0x0F] = temp_block[0x39]; block[0x16] = temp_block[0x32];
398 block[0x1D] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
399 if(last_non_zero_p1 <= 40) goto end;
400 block[0x2B] = temp_block[0x1D]; block[0x32] = temp_block[0x16];
401 block[0x39] = temp_block[0x0F]; block[0x3A] = temp_block[0x17];
402 block[0x33] = temp_block[0x1E]; block[0x2C] = temp_block[0x25];
403 block[0x25] = temp_block[0x2C]; block[0x1E] = temp_block[0x33];
404 if(last_non_zero_p1 <= 48) goto end;
405 block[0x17] = temp_block[0x3A]; block[0x1F] = temp_block[0x3B];
406 block[0x26] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
407 block[0x34] = temp_block[0x26]; block[0x3B] = temp_block[0x1F];
408 block[0x3C] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
409 if(last_non_zero_p1 <= 56) goto end;
410 block[0x2E] = temp_block[0x35]; block[0x27] = temp_block[0x3C];
411 block[0x2F] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
412 block[0x3D] = temp_block[0x2F]; block[0x3E] = temp_block[0x37];
413 block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
415 av_log(s, AV_LOG_DEBUG, "s->idsp.perm_type: %d\n",
416 (int)s->idsp.perm_type);
417 av_assert0(s->idsp.perm_type == FF_IDCT_PERM_NONE ||
418 s->idsp.perm_type == FF_IDCT_PERM_LIBMPEG2 ||
419 s->idsp.perm_type == FF_IDCT_PERM_SIMPLE ||
420 s->idsp.perm_type == FF_IDCT_PERM_TRANSPOSE);
423 return last_non_zero_p1 - 1;