2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
12 #include "vpx_config.h"
15 #include "vp8/common/reconinter.h"
18 #include "vp8/common/invtrans.h"
19 #include "vpx_mem/vpx_mem.h"
22 void vp8_subtract_b_c(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;
31 for (r = 0; r < 4; r++)
33 for (c = 0; c < 4; c++)
35 diff_ptr[c] = src_ptr[c] - pred_ptr[c];
40 src_ptr += src_stride;
44 void vp8_subtract_mbuv_c(short *diff, unsigned char *usrc, unsigned char *vsrc,
45 int src_stride, unsigned char *upred,
46 unsigned char *vpred, int pred_stride)
48 short *udiff = diff + 256;
49 short *vdiff = diff + 320;
53 for (r = 0; r < 8; r++)
55 for (c = 0; c < 8; c++)
57 udiff[c] = usrc[c] - upred[c];
65 for (r = 0; r < 8; r++)
67 for (c = 0; c < 8; c++)
69 vdiff[c] = vsrc[c] - vpred[c];
78 void vp8_subtract_mby_c(short *diff, unsigned char *src, int src_stride,
79 unsigned char *pred, int pred_stride)
83 for (r = 0; r < 16; r++)
85 for (c = 0; c < 16; c++)
87 diff[c] = src[c] - pred[c];
96 static void vp8_subtract_mb(MACROBLOCK *x)
98 BLOCK *b = &x->block[0];
100 vp8_subtract_mby(x->src_diff, *(b->base_src),
101 b->src_stride, x->e_mbd.dst.y_buffer, x->e_mbd.dst.y_stride);
102 vp8_subtract_mbuv(x->src_diff, x->src.u_buffer,
103 x->src.v_buffer, x->src.uv_stride, x->e_mbd.dst.u_buffer,
104 x->e_mbd.dst.v_buffer, x->e_mbd.dst.uv_stride);
107 static void build_dcblock(MACROBLOCK *x)
109 short *src_diff_ptr = &x->src_diff[384];
112 for (i = 0; i < 16; i++)
114 src_diff_ptr[i] = x->coeff[i * 16];
118 void vp8_transform_mbuv(MACROBLOCK *x)
122 for (i = 16; i < 24; i += 2)
124 x->short_fdct8x4(&x->block[i].src_diff[0],
125 &x->block[i].coeff[0], 16);
130 void vp8_transform_intra_mby(MACROBLOCK *x)
134 for (i = 0; i < 16; i += 2)
136 x->short_fdct8x4(&x->block[i].src_diff[0],
137 &x->block[i].coeff[0], 32);
140 // build dc block from 16 y dc values
143 // do 2nd order transform on the dc block
144 x->short_walsh4x4(&x->block[24].src_diff[0],
145 &x->block[24].coeff[0], 8);
150 static void transform_mb(MACROBLOCK *x)
154 for (i = 0; i < 16; i += 2)
156 x->short_fdct8x4(&x->block[i].src_diff[0],
157 &x->block[i].coeff[0], 32);
160 // build dc block from 16 y dc values
161 if (x->e_mbd.mode_info_context->mbmi.mode != SPLITMV)
164 for (i = 16; i < 24; i += 2)
166 x->short_fdct8x4(&x->block[i].src_diff[0],
167 &x->block[i].coeff[0], 16);
170 // do 2nd order transform on the dc block
171 if (x->e_mbd.mode_info_context->mbmi.mode != SPLITMV)
172 x->short_walsh4x4(&x->block[24].src_diff[0],
173 &x->block[24].coeff[0], 8);
178 static void transform_mby(MACROBLOCK *x)
182 for (i = 0; i < 16; i += 2)
184 x->short_fdct8x4(&x->block[i].src_diff[0],
185 &x->block[i].coeff[0], 32);
188 // build dc block from 16 y dc values
189 if (x->e_mbd.mode_info_context->mbmi.mode != SPLITMV)
192 x->short_walsh4x4(&x->block[24].src_diff[0],
193 &x->block[24].coeff[0], 8);
199 #define RDTRUNC(RM,DM,R,D) ( (128+(R)*(RM)) & 0xFF )
201 typedef struct vp8_token_state vp8_token_state;
203 struct vp8_token_state{
211 // TODO: experiments to find optimal multiple numbers
214 #define Y2_RD_MULT 16
216 static const int plane_rd_mult[4]=
224 static void optimize_b(MACROBLOCK *mb, int ib, int type,
225 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l)
229 vp8_token_state tokens[17][2];
230 unsigned best_mask[2];
231 const short *dequant_ptr;
232 const short *coeff_ptr;
256 int err_mult = plane_rd_mult[type];
259 d = &mb->e_mbd.block[ib];
261 /* Enable this to test the effect of RDO as a replacement for the dynamic
262 * zero bin instead of an augmentation of it.
265 vp8_strict_quantize_b(b, d);
268 dequant_ptr = d->dequant;
269 coeff_ptr = b->coeff;
270 qcoeff_ptr = d->qcoeff;
271 dqcoeff_ptr = d->dqcoeff;
275 /* Now set up a Viterbi trellis to evaluate alternative roundings. */
276 rdmult = mb->rdmult * err_mult;
277 if(mb->e_mbd.mode_info_context->mbmi.ref_frame==INTRA_FRAME)
278 rdmult = (rdmult * 9)>>4;
281 best_mask[0] = best_mask[1] = 0;
282 /* Initialize the sentinel node of the trellis. */
283 tokens[eob][0].rate = 0;
284 tokens[eob][0].error = 0;
285 tokens[eob][0].next = 16;
286 tokens[eob][0].token = DCT_EOB_TOKEN;
287 tokens[eob][0].qc = 0;
288 *(tokens[eob] + 1) = *(tokens[eob] + 0);
290 for (i = eob; i-- > i0;)
296 rc = vp8_default_zig_zag1d[i];
298 /* Only add a trellis state for non-zero coefficients. */
302 error0 = tokens[next][0].error;
303 error1 = tokens[next][1].error;
304 /* Evaluate the first possibility for this state. */
305 rate0 = tokens[next][0].rate;
306 rate1 = tokens[next][1].rate;
307 t0 = (vp8_dct_value_tokens_ptr + x)->Token;
308 /* Consider both possible successor states. */
311 band = vp8_coef_bands[i + 1];
312 pt = vp8_prev_token_class[t0];
314 mb->token_costs[type][band][pt][tokens[next][0].token];
316 mb->token_costs[type][band][pt][tokens[next][1].token];
318 rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);
319 rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);
320 if (rd_cost0 == rd_cost1)
322 rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);
323 rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);
325 /* And pick the best. */
326 best = rd_cost1 < rd_cost0;
327 base_bits = *(vp8_dct_value_cost_ptr + x);
328 dx = dqcoeff_ptr[rc] - coeff_ptr[rc];
330 tokens[i][0].rate = base_bits + (best ? rate1 : rate0);
331 tokens[i][0].error = d2 + (best ? error1 : error0);
332 tokens[i][0].next = next;
333 tokens[i][0].token = t0;
335 best_mask[0] |= best << i;
336 /* Evaluate the second possibility for this state. */
337 rate0 = tokens[next][0].rate;
338 rate1 = tokens[next][1].rate;
340 if((abs(x)*dequant_ptr[rc]>abs(coeff_ptr[rc])) &&
341 (abs(x)*dequant_ptr[rc]<abs(coeff_ptr[rc])+dequant_ptr[rc]))
352 /* Consider both possible successor states. */
355 /* If we reduced this coefficient to zero, check to see if
356 * we need to move the EOB back here.
358 t0 = tokens[next][0].token == DCT_EOB_TOKEN ?
359 DCT_EOB_TOKEN : ZERO_TOKEN;
360 t1 = tokens[next][1].token == DCT_EOB_TOKEN ?
361 DCT_EOB_TOKEN : ZERO_TOKEN;
365 t0=t1 = (vp8_dct_value_tokens_ptr + x)->Token;
369 band = vp8_coef_bands[i + 1];
370 if(t0!=DCT_EOB_TOKEN)
372 pt = vp8_prev_token_class[t0];
373 rate0 += mb->token_costs[type][band][pt][
374 tokens[next][0].token];
376 if(t1!=DCT_EOB_TOKEN)
378 pt = vp8_prev_token_class[t1];
379 rate1 += mb->token_costs[type][band][pt][
380 tokens[next][1].token];
384 rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);
385 rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);
386 if (rd_cost0 == rd_cost1)
388 rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);
389 rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);
391 /* And pick the best. */
392 best = rd_cost1 < rd_cost0;
393 base_bits = *(vp8_dct_value_cost_ptr + x);
397 dx -= (dequant_ptr[rc] + sz) ^ sz;
400 tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
401 tokens[i][1].error = d2 + (best ? error1 : error0);
402 tokens[i][1].next = next;
403 tokens[i][1].token =best?t1:t0;
405 best_mask[1] |= best << i;
406 /* Finally, make this the new head of the trellis. */
409 /* There's no choice to make for a zero coefficient, so we don't
410 * add a new trellis node, but we do need to update the costs.
414 band = vp8_coef_bands[i + 1];
415 t0 = tokens[next][0].token;
416 t1 = tokens[next][1].token;
417 /* Update the cost of each path if we're past the EOB token. */
418 if (t0 != DCT_EOB_TOKEN)
420 tokens[next][0].rate += mb->token_costs[type][band][0][t0];
421 tokens[next][0].token = ZERO_TOKEN;
423 if (t1 != DCT_EOB_TOKEN)
425 tokens[next][1].rate += mb->token_costs[type][band][0][t1];
426 tokens[next][1].token = ZERO_TOKEN;
428 /* Don't update next, because we didn't add a new node. */
432 /* Now pick the best path through the whole trellis. */
433 band = vp8_coef_bands[i + 1];
434 VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
435 rate0 = tokens[next][0].rate;
436 rate1 = tokens[next][1].rate;
437 error0 = tokens[next][0].error;
438 error1 = tokens[next][1].error;
439 t0 = tokens[next][0].token;
440 t1 = tokens[next][1].token;
441 rate0 += mb->token_costs[type][band][pt][t0];
442 rate1 += mb->token_costs[type][band][pt][t1];
443 rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);
444 rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);
445 if (rd_cost0 == rd_cost1)
447 rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);
448 rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);
450 best = rd_cost1 < rd_cost0;
452 for (i = next; i < eob; i = next)
454 x = tokens[i][best].qc;
457 rc = vp8_default_zig_zag1d[i];
459 dqcoeff_ptr[rc] = x * dequant_ptr[rc];
460 next = tokens[i][best].next;
461 best = (best_mask[best] >> i) & 1;
465 *a = *l = (final_eob != !type);
466 *d->eob = (char)final_eob;
468 static void check_reset_2nd_coeffs(MACROBLOCKD *x, int type,
469 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l)
473 BLOCKD *bd = &x->block[24];
475 if(bd->dequant[0]>=35 && bd->dequant[1]>=35)
478 for(i=0;i<(*bd->eob);i++)
480 int coef = bd->dqcoeff[vp8_default_zig_zag1d[i]];
481 sum+= (coef>=0)?coef:-coef;
485 /**************************************************************************
486 our inverse hadamard transform effectively is weighted sum of all 16 inputs
487 with weight either 1 or -1. It has a last stage scaling of (sum+3)>>3. And
488 dc only idct is (dc+4)>>3. So if all the sums are between -35 and 29, the
489 output after inverse wht and idct will be all zero. A sum of absolute value
490 smaller than 35 guarantees all 16 different (+1/-1) weighted sums in wht
491 fall between -35 and +35.
492 **************************************************************************/
495 for(i=0;i<(*bd->eob);i++)
497 int rc = vp8_default_zig_zag1d[i];
502 *a = *l = (*bd->eob != !type);
506 static void optimize_mb(MACROBLOCK *x)
512 ENTROPY_CONTEXT_PLANES t_above, t_left;
516 vpx_memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
517 vpx_memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
519 ta = (ENTROPY_CONTEXT *)&t_above;
520 tl = (ENTROPY_CONTEXT *)&t_left;
522 has_2nd_order = (x->e_mbd.mode_info_context->mbmi.mode != B_PRED
523 && x->e_mbd.mode_info_context->mbmi.mode != SPLITMV);
524 type = has_2nd_order ? PLANE_TYPE_Y_NO_DC : PLANE_TYPE_Y_WITH_DC;
526 for (b = 0; b < 16; b++)
528 optimize_b(x, b, type,
529 ta + vp8_block2above[b], tl + vp8_block2left[b]);
532 for (b = 16; b < 24; b++)
534 optimize_b(x, b, PLANE_TYPE_UV,
535 ta + vp8_block2above[b], tl + vp8_block2left[b]);
541 optimize_b(x, b, PLANE_TYPE_Y2,
542 ta + vp8_block2above[b], tl + vp8_block2left[b]);
543 check_reset_2nd_coeffs(&x->e_mbd, PLANE_TYPE_Y2,
544 ta + vp8_block2above[b], tl + vp8_block2left[b]);
549 void vp8_optimize_mby(MACROBLOCK *x)
555 ENTROPY_CONTEXT_PLANES t_above, t_left;
559 if (!x->e_mbd.above_context)
562 if (!x->e_mbd.left_context)
565 vpx_memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
566 vpx_memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
568 ta = (ENTROPY_CONTEXT *)&t_above;
569 tl = (ENTROPY_CONTEXT *)&t_left;
571 has_2nd_order = (x->e_mbd.mode_info_context->mbmi.mode != B_PRED
572 && x->e_mbd.mode_info_context->mbmi.mode != SPLITMV);
573 type = has_2nd_order ? PLANE_TYPE_Y_NO_DC : PLANE_TYPE_Y_WITH_DC;
575 for (b = 0; b < 16; b++)
577 optimize_b(x, b, type,
578 ta + vp8_block2above[b], tl + vp8_block2left[b]);
585 optimize_b(x, b, PLANE_TYPE_Y2,
586 ta + vp8_block2above[b], tl + vp8_block2left[b]);
587 check_reset_2nd_coeffs(&x->e_mbd, PLANE_TYPE_Y2,
588 ta + vp8_block2above[b], tl + vp8_block2left[b]);
592 void vp8_optimize_mbuv(MACROBLOCK *x)
595 ENTROPY_CONTEXT_PLANES t_above, t_left;
599 if (!x->e_mbd.above_context)
602 if (!x->e_mbd.left_context)
605 vpx_memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
606 vpx_memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
608 ta = (ENTROPY_CONTEXT *)&t_above;
609 tl = (ENTROPY_CONTEXT *)&t_left;
611 for (b = 16; b < 24; b++)
613 optimize_b(x, b, PLANE_TYPE_UV,
614 ta + vp8_block2above[b], tl + vp8_block2left[b]);
618 void vp8_encode_inter16x16(MACROBLOCK *x)
620 vp8_build_inter_predictors_mb(&x->e_mbd);
632 /* this funciton is used by first pass only */
633 void vp8_encode_inter16x16y(MACROBLOCK *x)
635 BLOCK *b = &x->block[0];
637 vp8_build_inter16x16_predictors_mby(&x->e_mbd, x->e_mbd.dst.y_buffer,
638 x->e_mbd.dst.y_stride);
640 vp8_subtract_mby(x->src_diff, *(b->base_src),
641 b->src_stride, x->e_mbd.dst.y_buffer, x->e_mbd.dst.y_stride);
647 vp8_inverse_transform_mby(&x->e_mbd);