1 /*
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include "./vpx_dsp_rtcd.h"
12
13 #include "vpx_config.h"
14 #include "vp8_rtcd.h"
15 #include "encodemb.h"
16 #include "vp8/common/reconinter.h"
17 #include "vp8/encoder/quantize.h"
18 #include "tokenize.h"
19 #include "vp8/common/invtrans.h"
20 #include "vpx_mem/vpx_mem.h"
21 #include "rdopt.h"
22
vp8_subtract_b(BLOCK * be,BLOCKD * bd,int pitch)23 void vp8_subtract_b(BLOCK *be, BLOCKD *bd, int pitch) {
24 unsigned char *src_ptr = (*(be->base_src) + be->src);
25 short *diff_ptr = be->src_diff;
26 unsigned char *pred_ptr = bd->predictor;
27 int src_stride = be->src_stride;
28
29 vpx_subtract_block(4, 4, diff_ptr, pitch, src_ptr, src_stride,
30 pred_ptr, pitch);
31 }
32
vp8_subtract_mbuv(short * diff,unsigned char * usrc,unsigned char * vsrc,int src_stride,unsigned char * upred,unsigned char * vpred,int pred_stride)33 void vp8_subtract_mbuv(short *diff, unsigned char *usrc, unsigned char *vsrc,
34 int src_stride, unsigned char *upred,
35 unsigned char *vpred, int pred_stride) {
36 short *udiff = diff + 256;
37 short *vdiff = diff + 320;
38
39 vpx_subtract_block(8, 8, udiff, 8, usrc, src_stride, upred, pred_stride);
40 vpx_subtract_block(8, 8, vdiff, 8, vsrc, src_stride, vpred, pred_stride);
41 }
42
vp8_subtract_mby(short * diff,unsigned char * src,int src_stride,unsigned char * pred,int pred_stride)43 void vp8_subtract_mby(short *diff, unsigned char *src, int src_stride,
44 unsigned char *pred, int pred_stride) {
45 vpx_subtract_block(16, 16, diff, 16, src, src_stride, pred, pred_stride);
46 }
47
vp8_subtract_mb(MACROBLOCK * x)48 static void vp8_subtract_mb(MACROBLOCK *x)
49 {
50 BLOCK *b = &x->block[0];
51
52 vp8_subtract_mby(x->src_diff, *(b->base_src),
53 b->src_stride, x->e_mbd.dst.y_buffer, x->e_mbd.dst.y_stride);
54 vp8_subtract_mbuv(x->src_diff, x->src.u_buffer,
55 x->src.v_buffer, x->src.uv_stride, x->e_mbd.dst.u_buffer,
56 x->e_mbd.dst.v_buffer, x->e_mbd.dst.uv_stride);
57 }
58
build_dcblock(MACROBLOCK * x)59 static void build_dcblock(MACROBLOCK *x)
60 {
61 short *src_diff_ptr = &x->src_diff[384];
62 int i;
63
64 for (i = 0; i < 16; i++)
65 {
66 src_diff_ptr[i] = x->coeff[i * 16];
67 }
68 }
69
vp8_transform_mbuv(MACROBLOCK * x)70 void vp8_transform_mbuv(MACROBLOCK *x)
71 {
72 int i;
73
74 for (i = 16; i < 24; i += 2)
75 {
76 x->short_fdct8x4(&x->block[i].src_diff[0],
77 &x->block[i].coeff[0], 16);
78 }
79 }
80
81
vp8_transform_intra_mby(MACROBLOCK * x)82 void vp8_transform_intra_mby(MACROBLOCK *x)
83 {
84 int i;
85
86 for (i = 0; i < 16; i += 2)
87 {
88 x->short_fdct8x4(&x->block[i].src_diff[0],
89 &x->block[i].coeff[0], 32);
90 }
91
92 /* build dc block from 16 y dc values */
93 build_dcblock(x);
94
95 /* do 2nd order transform on the dc block */
96 x->short_walsh4x4(&x->block[24].src_diff[0],
97 &x->block[24].coeff[0], 8);
98
99 }
100
101
transform_mb(MACROBLOCK * x)102 static void transform_mb(MACROBLOCK *x)
103 {
104 int i;
105
106 for (i = 0; i < 16; i += 2)
107 {
108 x->short_fdct8x4(&x->block[i].src_diff[0],
109 &x->block[i].coeff[0], 32);
110 }
111
112 /* build dc block from 16 y dc values */
113 if (x->e_mbd.mode_info_context->mbmi.mode != SPLITMV)
114 build_dcblock(x);
115
116 for (i = 16; i < 24; i += 2)
117 {
118 x->short_fdct8x4(&x->block[i].src_diff[0],
119 &x->block[i].coeff[0], 16);
120 }
121
122 /* do 2nd order transform on the dc block */
123 if (x->e_mbd.mode_info_context->mbmi.mode != SPLITMV)
124 x->short_walsh4x4(&x->block[24].src_diff[0],
125 &x->block[24].coeff[0], 8);
126
127 }
128
129
transform_mby(MACROBLOCK * x)130 static void transform_mby(MACROBLOCK *x)
131 {
132 int i;
133
134 for (i = 0; i < 16; i += 2)
135 {
136 x->short_fdct8x4(&x->block[i].src_diff[0],
137 &x->block[i].coeff[0], 32);
138 }
139
140 /* build dc block from 16 y dc values */
141 if (x->e_mbd.mode_info_context->mbmi.mode != SPLITMV)
142 {
143 build_dcblock(x);
144 x->short_walsh4x4(&x->block[24].src_diff[0],
145 &x->block[24].coeff[0], 8);
146 }
147 }
148
149
150
151 #define RDTRUNC(RM,DM,R,D) ( (128+(R)*(RM)) & 0xFF )
152
153 typedef struct vp8_token_state vp8_token_state;
154
155 struct vp8_token_state{
156 int rate;
157 int error;
158 signed char next;
159 signed char token;
160 short qc;
161 };
162
163 /* TODO: experiments to find optimal multiple numbers */
164 #define Y1_RD_MULT 4
165 #define UV_RD_MULT 2
166 #define Y2_RD_MULT 16
167
168 static const int plane_rd_mult[4]=
169 {
170 Y1_RD_MULT,
171 Y2_RD_MULT,
172 UV_RD_MULT,
173 Y1_RD_MULT
174 };
175
optimize_b(MACROBLOCK * mb,int ib,int type,ENTROPY_CONTEXT * a,ENTROPY_CONTEXT * l)176 static void optimize_b(MACROBLOCK *mb, int ib, int type,
177 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l)
178 {
179 BLOCK *b;
180 BLOCKD *d;
181 vp8_token_state tokens[17][2];
182 unsigned best_mask[2];
183 const short *dequant_ptr;
184 const short *coeff_ptr;
185 short *qcoeff_ptr;
186 short *dqcoeff_ptr;
187 int eob;
188 int i0;
189 int rc;
190 int x;
191 int sz = 0;
192 int next;
193 int rdmult;
194 int rddiv;
195 int final_eob;
196 int rd_cost0;
197 int rd_cost1;
198 int rate0;
199 int rate1;
200 int error0;
201 int error1;
202 int t0;
203 int t1;
204 int best;
205 int band;
206 int pt;
207 int i;
208 int err_mult = plane_rd_mult[type];
209
210 b = &mb->block[ib];
211 d = &mb->e_mbd.block[ib];
212
213 dequant_ptr = d->dequant;
214 coeff_ptr = b->coeff;
215 qcoeff_ptr = d->qcoeff;
216 dqcoeff_ptr = d->dqcoeff;
217 i0 = !type;
218 eob = *d->eob;
219
220 /* Now set up a Viterbi trellis to evaluate alternative roundings. */
221 rdmult = mb->rdmult * err_mult;
222 if(mb->e_mbd.mode_info_context->mbmi.ref_frame==INTRA_FRAME)
223 rdmult = (rdmult * 9)>>4;
224
225 rddiv = mb->rddiv;
226 best_mask[0] = best_mask[1] = 0;
227 /* Initialize the sentinel node of the trellis. */
228 tokens[eob][0].rate = 0;
229 tokens[eob][0].error = 0;
230 tokens[eob][0].next = 16;
231 tokens[eob][0].token = DCT_EOB_TOKEN;
232 tokens[eob][0].qc = 0;
233 *(tokens[eob] + 1) = *(tokens[eob] + 0);
234 next = eob;
235 for (i = eob; i-- > i0;)
236 {
237 int base_bits;
238 int d2;
239 int dx;
240
241 rc = vp8_default_zig_zag1d[i];
242 x = qcoeff_ptr[rc];
243 /* Only add a trellis state for non-zero coefficients. */
244 if (x)
245 {
246 int shortcut=0;
247 error0 = tokens[next][0].error;
248 error1 = tokens[next][1].error;
249 /* Evaluate the first possibility for this state. */
250 rate0 = tokens[next][0].rate;
251 rate1 = tokens[next][1].rate;
252 t0 = (vp8_dct_value_tokens_ptr + x)->Token;
253 /* Consider both possible successor states. */
254 if (next < 16)
255 {
256 band = vp8_coef_bands[i + 1];
257 pt = vp8_prev_token_class[t0];
258 rate0 +=
259 mb->token_costs[type][band][pt][tokens[next][0].token];
260 rate1 +=
261 mb->token_costs[type][band][pt][tokens[next][1].token];
262 }
263 rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);
264 rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);
265 if (rd_cost0 == rd_cost1)
266 {
267 rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);
268 rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);
269 }
270 /* And pick the best. */
271 best = rd_cost1 < rd_cost0;
272 base_bits = *(vp8_dct_value_cost_ptr + x);
273 dx = dqcoeff_ptr[rc] - coeff_ptr[rc];
274 d2 = dx*dx;
275 tokens[i][0].rate = base_bits + (best ? rate1 : rate0);
276 tokens[i][0].error = d2 + (best ? error1 : error0);
277 tokens[i][0].next = next;
278 tokens[i][0].token = t0;
279 tokens[i][0].qc = x;
280 best_mask[0] |= best << i;
281 /* Evaluate the second possibility for this state. */
282 rate0 = tokens[next][0].rate;
283 rate1 = tokens[next][1].rate;
284
285 if((abs(x)*dequant_ptr[rc]>abs(coeff_ptr[rc])) &&
286 (abs(x)*dequant_ptr[rc]<abs(coeff_ptr[rc])+dequant_ptr[rc]))
287 shortcut = 1;
288 else
289 shortcut = 0;
290
291 if(shortcut)
292 {
293 sz = -(x < 0);
294 x -= 2*sz + 1;
295 }
296
297 /* Consider both possible successor states. */
298 if (!x)
299 {
300 /* If we reduced this coefficient to zero, check to see if
301 * we need to move the EOB back here.
302 */
303 t0 = tokens[next][0].token == DCT_EOB_TOKEN ?
304 DCT_EOB_TOKEN : ZERO_TOKEN;
305 t1 = tokens[next][1].token == DCT_EOB_TOKEN ?
306 DCT_EOB_TOKEN : ZERO_TOKEN;
307 }
308 else
309 {
310 t0=t1 = (vp8_dct_value_tokens_ptr + x)->Token;
311 }
312 if (next < 16)
313 {
314 band = vp8_coef_bands[i + 1];
315 if(t0!=DCT_EOB_TOKEN)
316 {
317 pt = vp8_prev_token_class[t0];
318 rate0 += mb->token_costs[type][band][pt][
319 tokens[next][0].token];
320 }
321 if(t1!=DCT_EOB_TOKEN)
322 {
323 pt = vp8_prev_token_class[t1];
324 rate1 += mb->token_costs[type][band][pt][
325 tokens[next][1].token];
326 }
327 }
328
329 rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);
330 rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);
331 if (rd_cost0 == rd_cost1)
332 {
333 rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);
334 rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);
335 }
336 /* And pick the best. */
337 best = rd_cost1 < rd_cost0;
338 base_bits = *(vp8_dct_value_cost_ptr + x);
339
340 if(shortcut)
341 {
342 dx -= (dequant_ptr[rc] + sz) ^ sz;
343 d2 = dx*dx;
344 }
345 tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
346 tokens[i][1].error = d2 + (best ? error1 : error0);
347 tokens[i][1].next = next;
348 tokens[i][1].token =best?t1:t0;
349 tokens[i][1].qc = x;
350 best_mask[1] |= best << i;
351 /* Finally, make this the new head of the trellis. */
352 next = i;
353 }
354 /* There's no choice to make for a zero coefficient, so we don't
355 * add a new trellis node, but we do need to update the costs.
356 */
357 else
358 {
359 band = vp8_coef_bands[i + 1];
360 t0 = tokens[next][0].token;
361 t1 = tokens[next][1].token;
362 /* Update the cost of each path if we're past the EOB token. */
363 if (t0 != DCT_EOB_TOKEN)
364 {
365 tokens[next][0].rate += mb->token_costs[type][band][0][t0];
366 tokens[next][0].token = ZERO_TOKEN;
367 }
368 if (t1 != DCT_EOB_TOKEN)
369 {
370 tokens[next][1].rate += mb->token_costs[type][band][0][t1];
371 tokens[next][1].token = ZERO_TOKEN;
372 }
373 /* Don't update next, because we didn't add a new node. */
374 }
375 }
376
377 /* Now pick the best path through the whole trellis. */
378 band = vp8_coef_bands[i + 1];
379 VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
380 rate0 = tokens[next][0].rate;
381 rate1 = tokens[next][1].rate;
382 error0 = tokens[next][0].error;
383 error1 = tokens[next][1].error;
384 t0 = tokens[next][0].token;
385 t1 = tokens[next][1].token;
386 rate0 += mb->token_costs[type][band][pt][t0];
387 rate1 += mb->token_costs[type][band][pt][t1];
388 rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);
389 rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);
390 if (rd_cost0 == rd_cost1)
391 {
392 rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);
393 rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);
394 }
395 best = rd_cost1 < rd_cost0;
396 final_eob = i0 - 1;
397 for (i = next; i < eob; i = next)
398 {
399 x = tokens[i][best].qc;
400 if (x)
401 final_eob = i;
402 rc = vp8_default_zig_zag1d[i];
403 qcoeff_ptr[rc] = x;
404 dqcoeff_ptr[rc] = x * dequant_ptr[rc];
405 next = tokens[i][best].next;
406 best = (best_mask[best] >> i) & 1;
407 }
408 final_eob++;
409
410 *a = *l = (final_eob != !type);
411 *d->eob = (char)final_eob;
412 }
check_reset_2nd_coeffs(MACROBLOCKD * x,int type,ENTROPY_CONTEXT * a,ENTROPY_CONTEXT * l)413 static void check_reset_2nd_coeffs(MACROBLOCKD *x, int type,
414 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l)
415 {
416 int sum=0;
417 int i;
418 BLOCKD *bd = &x->block[24];
419
420 if(bd->dequant[0]>=35 && bd->dequant[1]>=35)
421 return;
422
423 for(i=0;i<(*bd->eob);i++)
424 {
425 int coef = bd->dqcoeff[vp8_default_zig_zag1d[i]];
426 sum+= (coef>=0)?coef:-coef;
427 if(sum>=35)
428 return;
429 }
430 /**************************************************************************
431 our inverse hadamard transform effectively is weighted sum of all 16 inputs
432 with weight either 1 or -1. It has a last stage scaling of (sum+3)>>3. And
433 dc only idct is (dc+4)>>3. So if all the sums are between -35 and 29, the
434 output after inverse wht and idct will be all zero. A sum of absolute value
435 smaller than 35 guarantees all 16 different (+1/-1) weighted sums in wht
436 fall between -35 and +35.
437 **************************************************************************/
438 if(sum < 35)
439 {
440 for(i=0;i<(*bd->eob);i++)
441 {
442 int rc = vp8_default_zig_zag1d[i];
443 bd->qcoeff[rc]=0;
444 bd->dqcoeff[rc]=0;
445 }
446 *bd->eob = 0;
447 *a = *l = (*bd->eob != !type);
448 }
449 }
450
optimize_mb(MACROBLOCK * x)451 static void optimize_mb(MACROBLOCK *x)
452 {
453 int b;
454 int type;
455 int has_2nd_order;
456
457 ENTROPY_CONTEXT_PLANES t_above, t_left;
458 ENTROPY_CONTEXT *ta;
459 ENTROPY_CONTEXT *tl;
460
461 memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
462 memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
463
464 ta = (ENTROPY_CONTEXT *)&t_above;
465 tl = (ENTROPY_CONTEXT *)&t_left;
466
467 has_2nd_order = (x->e_mbd.mode_info_context->mbmi.mode != B_PRED
468 && x->e_mbd.mode_info_context->mbmi.mode != SPLITMV);
469 type = has_2nd_order ? PLANE_TYPE_Y_NO_DC : PLANE_TYPE_Y_WITH_DC;
470
471 for (b = 0; b < 16; b++)
472 {
473 optimize_b(x, b, type,
474 ta + vp8_block2above[b], tl + vp8_block2left[b]);
475 }
476
477 for (b = 16; b < 24; b++)
478 {
479 optimize_b(x, b, PLANE_TYPE_UV,
480 ta + vp8_block2above[b], tl + vp8_block2left[b]);
481 }
482
483 if (has_2nd_order)
484 {
485 b=24;
486 optimize_b(x, b, PLANE_TYPE_Y2,
487 ta + vp8_block2above[b], tl + vp8_block2left[b]);
488 check_reset_2nd_coeffs(&x->e_mbd, PLANE_TYPE_Y2,
489 ta + vp8_block2above[b], tl + vp8_block2left[b]);
490 }
491 }
492
493
vp8_optimize_mby(MACROBLOCK * x)494 void vp8_optimize_mby(MACROBLOCK *x)
495 {
496 int b;
497 int type;
498 int has_2nd_order;
499
500 ENTROPY_CONTEXT_PLANES t_above, t_left;
501 ENTROPY_CONTEXT *ta;
502 ENTROPY_CONTEXT *tl;
503
504 if (!x->e_mbd.above_context)
505 return;
506
507 if (!x->e_mbd.left_context)
508 return;
509
510 memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
511 memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
512
513 ta = (ENTROPY_CONTEXT *)&t_above;
514 tl = (ENTROPY_CONTEXT *)&t_left;
515
516 has_2nd_order = (x->e_mbd.mode_info_context->mbmi.mode != B_PRED
517 && x->e_mbd.mode_info_context->mbmi.mode != SPLITMV);
518 type = has_2nd_order ? PLANE_TYPE_Y_NO_DC : PLANE_TYPE_Y_WITH_DC;
519
520 for (b = 0; b < 16; b++)
521 {
522 optimize_b(x, b, type,
523 ta + vp8_block2above[b], tl + vp8_block2left[b]);
524 }
525
526
527 if (has_2nd_order)
528 {
529 b=24;
530 optimize_b(x, b, PLANE_TYPE_Y2,
531 ta + vp8_block2above[b], tl + vp8_block2left[b]);
532 check_reset_2nd_coeffs(&x->e_mbd, PLANE_TYPE_Y2,
533 ta + vp8_block2above[b], tl + vp8_block2left[b]);
534 }
535 }
536
vp8_optimize_mbuv(MACROBLOCK * x)537 void vp8_optimize_mbuv(MACROBLOCK *x)
538 {
539 int b;
540 ENTROPY_CONTEXT_PLANES t_above, t_left;
541 ENTROPY_CONTEXT *ta;
542 ENTROPY_CONTEXT *tl;
543
544 if (!x->e_mbd.above_context)
545 return;
546
547 if (!x->e_mbd.left_context)
548 return;
549
550 memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
551 memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
552
553 ta = (ENTROPY_CONTEXT *)&t_above;
554 tl = (ENTROPY_CONTEXT *)&t_left;
555
556 for (b = 16; b < 24; b++)
557 {
558 optimize_b(x, b, PLANE_TYPE_UV,
559 ta + vp8_block2above[b], tl + vp8_block2left[b]);
560 }
561 }
562
vp8_encode_inter16x16(MACROBLOCK * x)563 void vp8_encode_inter16x16(MACROBLOCK *x)
564 {
565 vp8_build_inter_predictors_mb(&x->e_mbd);
566
567 vp8_subtract_mb(x);
568
569 transform_mb(x);
570
571 vp8_quantize_mb(x);
572
573 if (x->optimize)
574 optimize_mb(x);
575 }
576
577 /* this funciton is used by first pass only */
vp8_encode_inter16x16y(MACROBLOCK * x)578 void vp8_encode_inter16x16y(MACROBLOCK *x)
579 {
580 BLOCK *b = &x->block[0];
581
582 vp8_build_inter16x16_predictors_mby(&x->e_mbd, x->e_mbd.dst.y_buffer,
583 x->e_mbd.dst.y_stride);
584
585 vp8_subtract_mby(x->src_diff, *(b->base_src),
586 b->src_stride, x->e_mbd.dst.y_buffer, x->e_mbd.dst.y_stride);
587
588 transform_mby(x);
589
590 vp8_quantize_mby(x);
591
592 vp8_inverse_transform_mby(&x->e_mbd);
593 }
594