2 * The simplest mpeg encoder (well, it was the simplest!)
3 * Copyright (c) 2000,2001 Gerard Lantau.
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 * Optimized for ia32 cpus by Nick Kurshev <nickols_k@mail.ru>
20 * h263 dequantizer by Michael Niedermayer <michaelni@gmx.at>
23 #include "../dsputil.h"
24 #include "../mpegvideo.h"
25 #include "../avcodec.h"
26 #include "../mangle.h"
28 extern UINT8 zigzag_end[64];
29 extern void (*draw_edges)(UINT8 *buf, int wrap, int width, int height, int w);
30 extern int (*dct_quantize)(MpegEncContext *s, DCTELEM *block, int n, int qscale);
32 extern UINT8 zigzag_direct_noperm[64];
33 extern UINT16 inv_zigzag_direct16[64];
34 extern UINT32 inverse[256];
38 /* XXX: GL: I don't understand why this function needs optimization
39 (it is called only once per frame!), so I disabled it */
41 void MPV_frame_start(MpegEncContext *s)
43 if (s->pict_type == B_TYPE) {
45 "movl (%1), %%eax\n\t"
46 "movl 4(%1), %%edx\n\t"
47 "movl 8(%1), %%ecx\n\t"
48 "movl %%eax, (%0)\n\t"
49 "movl %%edx, 4(%0)\n\t"
50 "movl %%ecx, 8(%0)\n\t"
52 :"r"(s->current_picture), "r"(s->aux_picture)
53 :"eax","edx","ecx","memory");
55 /* swap next and last */
57 "movl (%1), %%eax\n\t"
58 "movl 4(%1), %%edx\n\t"
59 "movl 8(%1), %%ecx\n\t"
60 "xchgl (%0), %%eax\n\t"
61 "xchgl 4(%0), %%edx\n\t"
62 "xchgl 8(%0), %%ecx\n\t"
63 "movl %%eax, (%1)\n\t"
64 "movl %%edx, 4(%1)\n\t"
65 "movl %%ecx, 8(%1)\n\t"
66 "movl %%eax, (%2)\n\t"
67 "movl %%edx, 4(%2)\n\t"
68 "movl %%ecx, 8(%2)\n\t"
70 :"r"(s->last_picture), "r"(s->next_picture), "r"(s->current_picture)
71 :"eax","edx","ecx","memory");
76 static const unsigned long long int mm_wabs __attribute__ ((aligned(8))) = 0xffffffffffffffffULL;
77 static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL;
80 static void dct_unquantize_h263_mmx(MpegEncContext *s,
81 DCTELEM *block, int n, int qscale)
83 int i, level, qmul, qadd, nCoeffs;
85 qmul = s->qscale << 1;
86 qadd = (s->qscale - 1) | 1;
90 block[0] = block[0] * s->y_dc_scale;
92 block[0] = block[0] * s->c_dc_scale;
98 level = level * qmul - qadd;
100 level = level * qmul + qadd;
108 nCoeffs= zigzag_end[ s->block_last_index[n] ];
110 //printf("%d %d ", qmul, qadd);
112 "movd %1, %%mm6 \n\t" //qmul
113 "packssdw %%mm6, %%mm6 \n\t"
114 "packssdw %%mm6, %%mm6 \n\t"
115 "movd %2, %%mm5 \n\t" //qadd
116 "pxor %%mm7, %%mm7 \n\t"
117 "packssdw %%mm5, %%mm5 \n\t"
118 "packssdw %%mm5, %%mm5 \n\t"
119 "psubw %%mm5, %%mm7 \n\t"
120 "pxor %%mm4, %%mm4 \n\t"
123 "movq (%0, %3), %%mm0 \n\t"
124 "movq 8(%0, %3), %%mm1 \n\t"
126 "pmullw %%mm6, %%mm0 \n\t"
127 "pmullw %%mm6, %%mm1 \n\t"
129 "movq (%0, %3), %%mm2 \n\t"
130 "movq 8(%0, %3), %%mm3 \n\t"
132 "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
133 "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
135 "pxor %%mm2, %%mm0 \n\t"
136 "pxor %%mm3, %%mm1 \n\t"
138 "paddw %%mm7, %%mm0 \n\t"
139 "paddw %%mm7, %%mm1 \n\t"
141 "pxor %%mm0, %%mm2 \n\t"
142 "pxor %%mm1, %%mm3 \n\t"
144 "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0
145 "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0
147 "pandn %%mm2, %%mm0 \n\t"
148 "pandn %%mm3, %%mm1 \n\t"
150 "movq %%mm0, (%0, %3) \n\t"
151 "movq %%mm1, 8(%0, %3) \n\t"
155 ::"r" (block+nCoeffs), "g"(qmul), "g" (qadd), "r" (2*(i-nCoeffs))
163 Note: looking at PARANOID:
164 "enable all paranoid tests for rounding, overflows, etc..."
167 if (level < -2048 || level > 2047)
168 fprintf(stderr, "unquant error %d %d\n", i, level);
170 We can suppose that result of two multiplications can't be greate of 0xFFFF
171 i.e. is 16-bit, so we use here only PMULLW instruction and can avoid
172 a complex multiplication.
173 =====================================================
174 Full formula for multiplication of 2 integer numbers
175 which are represent as high:low words:
176 input: value1 = high1:low1
178 output: value3 = value1*value2
179 value3=high3:low3 (on overflow: modulus 2^32 wrap-around)
180 this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4
181 but this algorithm will compute only 0x66cb0ce4
182 this limited by 16-bit size of operands
183 ---------------------------------
186 tlow1 = tlow1 + tlow2
187 high3:low3 = low1*low2
190 static void dct_unquantize_mpeg1_mmx(MpegEncContext *s,
191 DCTELEM *block, int n, int qscale)
193 int i, level, nCoeffs;
194 const UINT16 *quant_matrix;
196 if(s->alternate_scan) nCoeffs= 64;
197 else nCoeffs= nCoeffs= zigzag_end[ s->block_last_index[n] ];
201 block[0] = block[0] * s->y_dc_scale;
203 block[0] = block[0] * s->c_dc_scale;
204 /* isnt used anymore (we have a h263 unquantizer since some time)
205 if (s->out_format == FMT_H263) {
209 /* XXX: only mpeg1 */
210 quant_matrix = s->intra_matrix;
212 /* Align on 4 elements boundary */
217 if (level < 0) level = -level;
218 level = (int)(level * qscale * quant_matrix[i]) >> 3;
219 level = (level - 1) | 1;
220 if (block[i] < 0) level = -level;
226 "movd %0, %%mm6\n\t" /* mm6 = qscale | 0 */
227 "punpckldq %%mm6, %%mm6\n\t" /* mm6 = qscale | qscale */
229 "movq %%mm6, %%mm7\n\t"
231 "packssdw %%mm6, %%mm7\n\t" /* mm7 = qscale | qscale | qscale | qscale */
232 "pxor %%mm6, %%mm6\n\t"
233 ::"g"(qscale),"m"(mm_wone),"m"(mm_wabs):"memory");
234 for(;i<nCoeffs;i+=4) {
237 "movq %%mm7, %%mm1\n\t"
238 "movq %%mm0, %%mm2\n\t"
239 "movq %%mm0, %%mm3\n\t"
240 "pcmpgtw %%mm6, %%mm2\n\t"
241 "pmullw %2, %%mm1\n\t"
242 "pandn %%mm4, %%mm2\n\t"
243 "por %%mm5, %%mm2\n\t"
244 "pmullw %%mm2, %%mm0\n\t" /* mm0 = abs(block[i]). */
246 "pcmpeqw %%mm6, %%mm3\n\t"
247 "pmullw %%mm0, %%mm1\n\t"
248 "psraw $3, %%mm1\n\t"
249 "psubw %%mm5, %%mm1\n\t" /* block[i] --; */
250 "pandn %%mm4, %%mm3\n\t" /* fake of pcmpneqw : mm0 != 0 then mm1 = -1 */
251 "por %%mm5, %%mm1\n\t" /* block[i] |= 1 */
252 "pmullw %%mm2, %%mm1\n\t" /* change signs again */
254 "pand %%mm3, %%mm1\n\t" /* nullify if was zero */
257 :"m"(block[i]), "m"(quant_matrix[i])
263 quant_matrix = s->non_intra_matrix;
264 /* Align on 4 elements boundary */
269 if (level < 0) level = -level;
270 level = (((level << 1) + 1) * qscale *
271 ((int) quant_matrix[i])) >> 4;
272 level = (level - 1) | 1;
273 if(block[i] < 0) level = -level;
279 "pcmpeqw %%mm7, %%mm7 \n\t"
280 "psrlw $15, %%mm7 \n\t"
281 "movd %2, %%mm6 \n\t"
282 "packssdw %%mm6, %%mm6 \n\t"
283 "packssdw %%mm6, %%mm6 \n\t"
286 "movq (%0, %3), %%mm0 \n\t"
287 "movq 8(%0, %3), %%mm1 \n\t"
288 "movq (%1, %3), %%mm4 \n\t"
289 "movq 8(%1, %3), %%mm5 \n\t"
290 "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i]
291 "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i]
292 "pxor %%mm2, %%mm2 \n\t"
293 "pxor %%mm3, %%mm3 \n\t"
294 "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0
295 "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0
296 "pxor %%mm2, %%mm0 \n\t"
297 "pxor %%mm3, %%mm1 \n\t"
298 "psubw %%mm2, %%mm0 \n\t" // abs(block[i])
299 "psubw %%mm3, %%mm1 \n\t" // abs(block[i])
300 "paddw %%mm0, %%mm0 \n\t" // abs(block[i])*2
301 "paddw %%mm1, %%mm1 \n\t" // abs(block[i])*2
302 "paddw %%mm7, %%mm0 \n\t" // abs(block[i])*2 + 1
303 "paddw %%mm7, %%mm1 \n\t" // abs(block[i])*2 + 1
304 "pmullw %%mm4, %%mm0 \n\t" // (abs(block[i])*2 + 1)*q
305 "pmullw %%mm5, %%mm1 \n\t" // (abs(block[i])*2 + 1)*q
306 "pxor %%mm4, %%mm4 \n\t"
307 "pxor %%mm5, %%mm5 \n\t" // FIXME slow
308 "pcmpeqw (%0, %3), %%mm4 \n\t" // block[i] == 0 ? -1 : 0
309 "pcmpeqw 8(%0, %3), %%mm5 \n\t" // block[i] == 0 ? -1 : 0
310 "psraw $4, %%mm0 \n\t"
311 "psraw $4, %%mm1 \n\t"
312 "psubw %%mm7, %%mm0 \n\t"
313 "psubw %%mm7, %%mm1 \n\t"
314 "por %%mm7, %%mm0 \n\t"
315 "por %%mm7, %%mm1 \n\t"
316 "pxor %%mm2, %%mm0 \n\t"
317 "pxor %%mm3, %%mm1 \n\t"
318 "psubw %%mm2, %%mm0 \n\t"
319 "psubw %%mm3, %%mm1 \n\t"
320 "pandn %%mm0, %%mm4 \n\t"
321 "pandn %%mm1, %%mm5 \n\t"
322 "movq %%mm4, (%0, %3) \n\t"
323 "movq %%mm5, 8(%0, %3) \n\t"
327 ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "g" (qscale), "r" (2*(i-nCoeffs))
333 /* draw the edges of width 'w' of an image of size width, height
334 this mmx version can only handle w==8 || w==16 */
335 static void draw_edges_mmx(UINT8 *buf, int wrap, int width, int height, int w)
337 UINT8 *ptr, *last_line;
340 last_line = buf + (height - 1) * wrap;
347 "movd (%0), %%mm0 \n\t"
348 "punpcklbw %%mm0, %%mm0 \n\t"
349 "punpcklwd %%mm0, %%mm0 \n\t"
350 "punpckldq %%mm0, %%mm0 \n\t"
351 "movq %%mm0, -8(%0) \n\t"
352 "movq -8(%0, %2), %%mm1 \n\t"
353 "punpckhbw %%mm1, %%mm1 \n\t"
354 "punpckhwd %%mm1, %%mm1 \n\t"
355 "punpckhdq %%mm1, %%mm1 \n\t"
356 "movq %%mm1, (%0, %2) \n\t"
361 : "r" (wrap), "r" (width), "r" (ptr + wrap*height)
368 "movd (%0), %%mm0 \n\t"
369 "punpcklbw %%mm0, %%mm0 \n\t"
370 "punpcklwd %%mm0, %%mm0 \n\t"
371 "punpckldq %%mm0, %%mm0 \n\t"
372 "movq %%mm0, -8(%0) \n\t"
373 "movq %%mm0, -16(%0) \n\t"
374 "movq -8(%0, %2), %%mm1 \n\t"
375 "punpckhbw %%mm1, %%mm1 \n\t"
376 "punpckhwd %%mm1, %%mm1 \n\t"
377 "punpckhdq %%mm1, %%mm1 \n\t"
378 "movq %%mm1, (%0, %2) \n\t"
379 "movq %%mm1, 8(%0, %2) \n\t"
384 : "r" (wrap), "r" (width), "r" (ptr + wrap*height)
389 /* top and bottom (and hopefully also the corners) */
390 ptr= buf - (i + 1) * wrap - w;
393 "movq (%1, %0), %%mm0 \n\t"
394 "movq %%mm0, (%0) \n\t"
395 "movq %%mm0, (%0, %2) \n\t"
396 "movq %%mm0, (%0, %2, 2) \n\t"
397 "movq %%mm0, (%0, %3) \n\t"
402 : "r" ((int)buf - (int)ptr - w), "r" (-wrap), "r" (-wrap*3), "r" (ptr+width+2*w)
404 ptr= last_line + (i + 1) * wrap - w;
407 "movq (%1, %0), %%mm0 \n\t"
408 "movq %%mm0, (%0) \n\t"
409 "movq %%mm0, (%0, %2) \n\t"
410 "movq %%mm0, (%0, %2, 2) \n\t"
411 "movq %%mm0, (%0, %3) \n\t"
416 : "r" ((int)last_line - (int)ptr - w), "r" (wrap), "r" (wrap*3), "r" (ptr+width+2*w)
421 static volatile int esp_temp;
423 void unused_var_warning_killer(){
428 #define RENAME(a) a ## _MMX
429 #include "mpegvideo_mmx_template.c"
433 #define RENAME(a) a ## _MMX2
434 #include "mpegvideo_mmx_template.c"
436 void MPV_common_init_mmx(MpegEncContext *s)
438 if (mm_flags & MM_MMX) {
439 if (s->out_format == FMT_H263)
440 s->dct_unquantize = dct_unquantize_h263_mmx;
442 s->dct_unquantize = dct_unquantize_mpeg1_mmx;
444 draw_edges = draw_edges_mmx;
446 if(mm_flags & MM_MMXEXT){
447 dct_quantize= dct_quantize_MMX2;
449 dct_quantize= dct_quantize_MMX;