42 |
#include "../quant/quant.h" |
#include "../quant/quant.h" |
43 |
#include "../encoder.h" |
#include "../encoder.h" |
44 |
|
|
45 |
#include "../image/reduced.h" |
#include "../quant/quant_matrix.h" |
46 |
|
|
47 |
MBFIELDTEST_PTR MBFieldTest; |
MBFIELDTEST_PTR MBFieldTest; |
48 |
|
|
137 |
|
|
138 |
/* Quantize the block */ |
/* Quantize the block */ |
139 |
start_timer(); |
start_timer(); |
140 |
quant[mpeg](&data[0 * 64], &qcoeff[0 * 64], pMB->quant, scaler_lum); |
quant[mpeg](&data[0 * 64], &qcoeff[0 * 64], pMB->quant, scaler_lum, pParam->mpeg_quant_matrices); |
141 |
quant[mpeg](&data[1 * 64], &qcoeff[1 * 64], pMB->quant, scaler_lum); |
quant[mpeg](&data[1 * 64], &qcoeff[1 * 64], pMB->quant, scaler_lum, pParam->mpeg_quant_matrices); |
142 |
quant[mpeg](&data[2 * 64], &qcoeff[2 * 64], pMB->quant, scaler_lum); |
quant[mpeg](&data[2 * 64], &qcoeff[2 * 64], pMB->quant, scaler_lum, pParam->mpeg_quant_matrices); |
143 |
quant[mpeg](&data[3 * 64], &qcoeff[3 * 64], pMB->quant, scaler_lum); |
quant[mpeg](&data[3 * 64], &qcoeff[3 * 64], pMB->quant, scaler_lum, pParam->mpeg_quant_matrices); |
144 |
quant[mpeg](&data[4 * 64], &qcoeff[4 * 64], pMB->quant, scaler_chr); |
quant[mpeg](&data[4 * 64], &qcoeff[4 * 64], pMB->quant, scaler_chr, pParam->mpeg_quant_matrices); |
145 |
quant[mpeg](&data[5 * 64], &qcoeff[5 * 64], pMB->quant, scaler_chr); |
quant[mpeg](&data[5 * 64], &qcoeff[5 * 64], pMB->quant, scaler_chr, pParam->mpeg_quant_matrices); |
146 |
stop_quant_timer(); |
stop_quant_timer(); |
147 |
} |
} |
148 |
|
|
167 |
scaler_chr = get_dc_scaler(iQuant, 0); |
scaler_chr = get_dc_scaler(iQuant, 0); |
168 |
|
|
169 |
start_timer(); |
start_timer(); |
170 |
dequant[mpeg](&qcoeff[0 * 64], &data[0 * 64], iQuant, scaler_lum); |
dequant[mpeg](&qcoeff[0 * 64], &data[0 * 64], iQuant, scaler_lum, pParam->mpeg_quant_matrices); |
171 |
dequant[mpeg](&qcoeff[1 * 64], &data[1 * 64], iQuant, scaler_lum); |
dequant[mpeg](&qcoeff[1 * 64], &data[1 * 64], iQuant, scaler_lum, pParam->mpeg_quant_matrices); |
172 |
dequant[mpeg](&qcoeff[2 * 64], &data[2 * 64], iQuant, scaler_lum); |
dequant[mpeg](&qcoeff[2 * 64], &data[2 * 64], iQuant, scaler_lum, pParam->mpeg_quant_matrices); |
173 |
dequant[mpeg](&qcoeff[3 * 64], &data[3 * 64], iQuant, scaler_lum); |
dequant[mpeg](&qcoeff[3 * 64], &data[3 * 64], iQuant, scaler_lum, pParam->mpeg_quant_matrices); |
174 |
dequant[mpeg](&qcoeff[4 * 64], &data[4 * 64], iQuant, scaler_chr); |
dequant[mpeg](&qcoeff[4 * 64], &data[4 * 64], iQuant, scaler_chr, pParam->mpeg_quant_matrices); |
175 |
dequant[mpeg](&qcoeff[5 * 64], &data[5 * 64], iQuant, scaler_chr); |
dequant[mpeg](&qcoeff[5 * 64], &data[5 * 64], iQuant, scaler_chr, pParam->mpeg_quant_matrices); |
176 |
stop_iquant_timer(); |
stop_iquant_timer(); |
177 |
} |
} |
178 |
|
|
|
|
|
|
typedef int (*trellis_func_ptr_t)(int16_t *const Out, |
|
|
const int16_t *const In, |
|
|
int Q, |
|
|
const uint16_t * const Zigzag, |
|
|
int Non_Zero); |
|
|
|
|
179 |
static int |
static int |
180 |
dct_quantize_trellis_h263_c(int16_t *const Out, |
dct_quantize_trellis_c(int16_t *const Out, |
181 |
const int16_t *const In, |
const int16_t *const In, |
182 |
int Q, |
int Q, |
183 |
const uint16_t * const Zigzag, |
const uint16_t * const Zigzag, |
184 |
int Non_Zero); |
const uint16_t * const QuantMatrix, |
185 |
|
int Non_Zero, |
186 |
static int |
int Sum); |
|
dct_quantize_trellis_mpeg_c(int16_t *const Out, |
|
|
const int16_t *const In, |
|
|
int Q, |
|
|
const uint16_t * const Zigzag, |
|
|
int Non_Zero); |
|
187 |
|
|
188 |
/* Quantize all blocks -- Inter mode */ |
/* Quantize all blocks -- Inter mode */ |
189 |
static __inline uint8_t |
static __inline uint8_t |
207 |
quant_mpeg_inter |
quant_mpeg_inter |
208 |
}; |
}; |
209 |
|
|
|
trellis_func_ptr_t const trellis[2] = |
|
|
{ |
|
|
dct_quantize_trellis_h263_c, |
|
|
dct_quantize_trellis_mpeg_c |
|
|
}; |
|
|
|
|
210 |
mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
211 |
|
|
212 |
for (i = 0; i < 6; i++) { |
for (i = 0; i < 6; i++) { |
214 |
/* Quantize the block */ |
/* Quantize the block */ |
215 |
start_timer(); |
start_timer(); |
216 |
|
|
217 |
sum = quant[mpeg](&qcoeff[i*64], &data[i*64], pMB->quant); |
sum = quant[mpeg](&qcoeff[i*64], &data[i*64], pMB->quant, pParam->mpeg_quant_matrices); |
218 |
|
|
219 |
if(sum && (frame->vop_flags & XVID_VOP_TRELLISQUANT)) { |
if(sum && (pMB->quant > 2) && (frame->vop_flags & XVID_VOP_TRELLISQUANT)) { |
220 |
sum = trellis[mpeg](&qcoeff[i*64], &data[i*64], pMB->quant, &scan_tables[0][0], 63); |
const uint16_t *matrix; |
221 |
|
const static uint16_t h263matrix[] = |
222 |
|
{ |
223 |
|
16, 16, 16, 16, 16, 16, 16, 16, |
224 |
|
16, 16, 16, 16, 16, 16, 16, 16, |
225 |
|
16, 16, 16, 16, 16, 16, 16, 16, |
226 |
|
16, 16, 16, 16, 16, 16, 16, 16, |
227 |
|
16, 16, 16, 16, 16, 16, 16, 16, |
228 |
|
16, 16, 16, 16, 16, 16, 16, 16, |
229 |
|
16, 16, 16, 16, 16, 16, 16, 16, |
230 |
|
16, 16, 16, 16, 16, 16, 16, 16 |
231 |
|
}; |
232 |
|
|
233 |
|
matrix = (mpeg)?get_inter_matrix(pParam->mpeg_quant_matrices):h263matrix; |
234 |
|
sum = dct_quantize_trellis_c(&qcoeff[i*64], &data[i*64], |
235 |
|
pMB->quant, &scan_tables[0][0], |
236 |
|
matrix, |
237 |
|
63, |
238 |
|
sum); |
239 |
} |
} |
240 |
stop_quant_timer(); |
stop_quant_timer(); |
241 |
|
|
285 |
mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
286 |
|
|
287 |
start_timer(); |
start_timer(); |
288 |
if(cbp & (1 << (5 - 0))) dequant[mpeg](&data[0 * 64], &qcoeff[0 * 64], iQuant); |
if(cbp & (1 << (5 - 0))) dequant[mpeg](&data[0 * 64], &qcoeff[0 * 64], iQuant, pParam->mpeg_quant_matrices); |
289 |
if(cbp & (1 << (5 - 1))) dequant[mpeg](&data[1 * 64], &qcoeff[1 * 64], iQuant); |
if(cbp & (1 << (5 - 1))) dequant[mpeg](&data[1 * 64], &qcoeff[1 * 64], iQuant, pParam->mpeg_quant_matrices); |
290 |
if(cbp & (1 << (5 - 2))) dequant[mpeg](&data[2 * 64], &qcoeff[2 * 64], iQuant); |
if(cbp & (1 << (5 - 2))) dequant[mpeg](&data[2 * 64], &qcoeff[2 * 64], iQuant, pParam->mpeg_quant_matrices); |
291 |
if(cbp & (1 << (5 - 3))) dequant[mpeg](&data[3 * 64], &qcoeff[3 * 64], iQuant); |
if(cbp & (1 << (5 - 3))) dequant[mpeg](&data[3 * 64], &qcoeff[3 * 64], iQuant, pParam->mpeg_quant_matrices); |
292 |
if(cbp & (1 << (5 - 4))) dequant[mpeg](&data[4 * 64], &qcoeff[4 * 64], iQuant); |
if(cbp & (1 << (5 - 4))) dequant[mpeg](&data[4 * 64], &qcoeff[4 * 64], iQuant, pParam->mpeg_quant_matrices); |
293 |
if(cbp & (1 << (5 - 5))) dequant[mpeg](&data[5 * 64], &qcoeff[5 * 64], iQuant); |
if(cbp & (1 << (5 - 5))) dequant[mpeg](&data[5 * 64], &qcoeff[5 * 64], iQuant, pParam->mpeg_quant_matrices); |
294 |
stop_iquant_timer(); |
stop_iquant_timer(); |
295 |
} |
} |
296 |
|
|
309 |
uint32_t stride = pParam->edged_width; |
uint32_t stride = pParam->edged_width; |
310 |
uint32_t stride2 = stride / 2; |
uint32_t stride2 = stride / 2; |
311 |
uint32_t next_block = stride * 8; |
uint32_t next_block = stride * 8; |
|
int32_t cst; |
|
|
int vop_reduced; |
|
312 |
uint8_t *pY_Cur, *pU_Cur, *pV_Cur; |
uint8_t *pY_Cur, *pU_Cur, *pV_Cur; |
313 |
const IMAGE * const pCurrent = &frame->image; |
const IMAGE * const pCurrent = &frame->image; |
|
transfer_operation_8to16_t * const functions[2] = |
|
|
{ |
|
|
(transfer_operation_8to16_t *)transfer_8to16copy, |
|
|
(transfer_operation_8to16_t *)filter_18x18_to_8x8 |
|
|
}; |
|
|
transfer_operation_8to16_t *transfer_op = NULL; |
|
|
|
|
|
vop_reduced = !!(frame->vop_flags & XVID_VOP_REDUCED); |
|
314 |
|
|
315 |
/* Image pointers */ |
/* Image pointers */ |
316 |
pY_Cur = pCurrent->y + (y_pos << (4+vop_reduced)) * stride + (x_pos << (4+vop_reduced)); |
pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); |
317 |
pU_Cur = pCurrent->u + (y_pos << (3+vop_reduced)) * stride2 + (x_pos << (3+vop_reduced)); |
pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); |
318 |
pV_Cur = pCurrent->v + (y_pos << (3+vop_reduced)) * stride2 + (x_pos << (3+vop_reduced)); |
pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); |
|
|
|
|
/* Block size */ |
|
|
cst = 8<<vop_reduced; |
|
|
|
|
|
/* Operation function */ |
|
|
transfer_op = functions[vop_reduced]; |
|
319 |
|
|
320 |
/* Do the transfer */ |
/* Do the transfer */ |
321 |
start_timer(); |
start_timer(); |
322 |
transfer_op(&data[0 * 64], pY_Cur, stride); |
transfer_8to16copy(&data[0 * 64], pY_Cur, stride); |
323 |
transfer_op(&data[1 * 64], pY_Cur + cst, stride); |
transfer_8to16copy(&data[1 * 64], pY_Cur + 8, stride); |
324 |
transfer_op(&data[2 * 64], pY_Cur + next_block, stride); |
transfer_8to16copy(&data[2 * 64], pY_Cur + next_block, stride); |
325 |
transfer_op(&data[3 * 64], pY_Cur + next_block + cst, stride); |
transfer_8to16copy(&data[3 * 64], pY_Cur + next_block + 8, stride); |
326 |
transfer_op(&data[4 * 64], pU_Cur, stride2); |
transfer_8to16copy(&data[4 * 64], pU_Cur, stride2); |
327 |
transfer_op(&data[5 * 64], pV_Cur, stride2); |
transfer_8to16copy(&data[5 * 64], pV_Cur, stride2); |
328 |
stop_transfer_timer(); |
stop_transfer_timer(); |
329 |
} |
} |
330 |
|
|
342 |
uint32_t stride = pParam->edged_width; |
uint32_t stride = pParam->edged_width; |
343 |
uint32_t stride2 = stride / 2; |
uint32_t stride2 = stride / 2; |
344 |
uint32_t next_block = stride * 8; |
uint32_t next_block = stride * 8; |
|
uint32_t cst; |
|
|
int vop_reduced; |
|
345 |
const IMAGE * const pCurrent = &frame->image; |
const IMAGE * const pCurrent = &frame->image; |
346 |
|
|
347 |
/* Array of function pointers, indexed by [vop_reduced<<1+add] */ |
/* Array of function pointers, indexed by [add] */ |
348 |
transfer_operation_16to8_t * const functions[4] = |
transfer_operation_16to8_t * const functions[2] = |
349 |
{ |
{ |
350 |
(transfer_operation_16to8_t*)transfer_16to8copy, |
(transfer_operation_16to8_t*)transfer_16to8copy, |
351 |
(transfer_operation_16to8_t*)transfer_16to8add, |
(transfer_operation_16to8_t*)transfer_16to8add, |
|
(transfer_operation_16to8_t*)copy_upsampled_8x8_16to8, |
|
|
(transfer_operation_16to8_t*)add_upsampled_8x8_16to8 |
|
352 |
}; |
}; |
353 |
|
|
354 |
transfer_operation_16to8_t *transfer_op = NULL; |
transfer_operation_16to8_t *transfer_op = NULL; |
355 |
|
|
356 |
|
/* Image pointers */ |
357 |
|
pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); |
358 |
|
pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); |
359 |
|
pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); |
360 |
|
|
361 |
if (pMB->field_dct) { |
if (pMB->field_dct) { |
362 |
next_block = stride; |
next_block = stride; |
363 |
stride *= 2; |
stride *= 2; |
364 |
} |
} |
365 |
|
|
|
/* Makes this vars booleans */ |
|
|
vop_reduced = !!(frame->vop_flags & XVID_VOP_REDUCED); |
|
|
|
|
|
/* Image pointers */ |
|
|
pY_Cur = pCurrent->y + (y_pos << (4+vop_reduced)) * stride + (x_pos << (4+vop_reduced)); |
|
|
pU_Cur = pCurrent->u + (y_pos << (3+vop_reduced)) * stride2 + (x_pos << (3+vop_reduced)); |
|
|
pV_Cur = pCurrent->v + (y_pos << (3+vop_reduced)) * stride2 + (x_pos << (3+vop_reduced)); |
|
|
|
|
|
/* Block size */ |
|
|
cst = 8<<vop_reduced; |
|
|
|
|
366 |
/* Operation function */ |
/* Operation function */ |
367 |
transfer_op = functions[(vop_reduced<<1) + add]; |
transfer_op = functions[add]; |
368 |
|
|
369 |
/* Do the operation */ |
/* Do the operation */ |
370 |
start_timer(); |
start_timer(); |
371 |
if (cbp&32) transfer_op(pY_Cur, &data[0 * 64], stride); |
if (cbp&32) transfer_op(pY_Cur, &data[0 * 64], stride); |
372 |
if (cbp&16) transfer_op(pY_Cur + cst, &data[1 * 64], stride); |
if (cbp&16) transfer_op(pY_Cur + 8, &data[1 * 64], stride); |
373 |
if (cbp& 8) transfer_op(pY_Cur + next_block, &data[2 * 64], stride); |
if (cbp& 8) transfer_op(pY_Cur + next_block, &data[2 * 64], stride); |
374 |
if (cbp& 4) transfer_op(pY_Cur + next_block + cst, &data[3 * 64], stride); |
if (cbp& 4) transfer_op(pY_Cur + next_block + 8, &data[3 * 64], stride); |
375 |
if (cbp& 2) transfer_op(pU_Cur, &data[4 * 64], stride2); |
if (cbp& 2) transfer_op(pU_Cur, &data[4 * 64], stride2); |
376 |
if (cbp& 1) transfer_op(pV_Cur, &data[5 * 64], stride2); |
if (cbp& 1) transfer_op(pV_Cur, &data[5 * 64], stride2); |
377 |
stop_transfer_timer(); |
stop_transfer_timer(); |
625 |
* IEEE Transactions on Image Processing, Vol.9, No.8, Aug. 2000. |
* IEEE Transactions on Image Processing, Vol.9, No.8, Aug. 2000. |
626 |
* |
* |
627 |
* we are at stake with a simplified Bellmand-Ford / Dijkstra Single |
* we are at stake with a simplified Bellmand-Ford / Dijkstra Single |
628 |
* Source Shorted Path algo. But due to the underlying graph structure |
* Source Shortest Path algo. But due to the underlying graph structure |
629 |
* ("Trellis"), it can be turned into a dynamic programming algo, |
* ("Trellis"), it can be turned into a dynamic programming algo, |
630 |
* partially saving the explicit graph's nodes representation. And |
* partially saving the explicit graph's nodes representation. And |
631 |
* without using a heap, since the open frontier of the DAG is always |
* without using a heap, since the open frontier of the DAG is always |
632 |
* known, and of fixed sized. |
* known, and of fixed size. |
633 |
*--------------------------------------------------------------------------*/ |
*--------------------------------------------------------------------------*/ |
634 |
|
|
635 |
|
|
729 |
Code_Len24,Code_Len23,Code_Len22,Code_Len21, Code_Len3, Code_Len1, |
Code_Len24,Code_Len23,Code_Len22,Code_Len21, Code_Len3, Code_Len1, |
730 |
}; |
}; |
731 |
|
|
732 |
#define TL(q) 0xfe00/(q*q) |
/* TL_SHIFT controls the precision of the RD optimizations in trellis |
733 |
|
* valid range is [10..16]. The bigger, the more trellis is vulnerable |
734 |
|
* to overflows in cost formulas. |
735 |
|
* - 10 allows ac values up to 2^11 == 2048 |
736 |
|
* - 16 allows ac values up to 2^8 == 256 |
737 |
|
*/ |
738 |
|
#define TL_SHIFT 11 |
739 |
|
#define TL(q) ((0xfe00>>(16-TL_SHIFT))/(q*q)) |
740 |
|
|
741 |
static const int Trellis_Lambda_Tabs[31] = { |
static const int Trellis_Lambda_Tabs[31] = { |
742 |
TL( 1),TL( 2),TL( 3),TL( 4),TL( 5),TL( 6), TL( 7), |
TL( 1),TL( 2),TL( 3),TL( 4),TL( 5),TL( 6), TL( 7), |
756 |
return -1; |
return -1; |
757 |
} |
} |
758 |
|
|
|
static int __inline |
|
|
Compute_Sum(const int16_t *C, int last) |
|
|
{ |
|
|
int sum = 0; |
|
|
|
|
|
while(last--) |
|
|
sum += abs(C[last]); |
|
|
|
|
|
return(sum); |
|
|
} |
|
759 |
/* this routine has been strippen of all debug code */ |
/* this routine has been strippen of all debug code */ |
|
|
|
760 |
static int |
static int |
761 |
dct_quantize_trellis_h263_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero) |
dct_quantize_trellis_c(int16_t *const Out, |
762 |
|
const int16_t *const In, |
763 |
|
int Q, |
764 |
|
const uint16_t * const Zigzag, |
765 |
|
const uint16_t * const QuantMatrix, |
766 |
|
int Non_Zero, |
767 |
|
int Sum) |
768 |
{ |
{ |
769 |
|
|
770 |
/* |
/* Note: We should search last non-zero coeffs on *real* DCT input coeffs |
771 |
* Note: We should search last non-zero coeffs on *real* DCT input coeffs (In[]), |
* (In[]), not quantized one (Out[]). However, it only improves the result |
772 |
* not quantized one (Out[]). However, it only improves the result *very* |
* *very* slightly (~0.01dB), whereas speed drops to crawling level :) |
773 |
* slightly (~0.01dB), whereas speed drops to crawling level :) |
* Well, actually, taking 1 more coeff past Non_Zero into account sometimes |
774 |
* Well, actually, taking 1 more coeff past Non_Zero into account sometimes helps. |
* helps. */ |
|
*/ |
|
775 |
typedef struct { int16_t Run, Level; } NODE; |
typedef struct { int16_t Run, Level; } NODE; |
776 |
|
|
777 |
NODE Nodes[65], Last; |
NODE Nodes[65], Last; |
778 |
uint32_t Run_Costs0[64+1]; |
uint32_t Run_Costs0[64+1]; |
779 |
uint32_t * const Run_Costs = Run_Costs0 + 1; |
uint32_t * const Run_Costs = Run_Costs0 + 1; |
780 |
const int Mult = 2*Q; |
|
781 |
const int Bias = (Q-1) | 1; |
/* it's 1/lambda, actually */ |
782 |
const int Lev0 = Mult + Bias; |
const int Lambda = Trellis_Lambda_Tabs[Q-1]; |
|
const int Lambda = Trellis_Lambda_Tabs[Q-1]; /* it's 1/lambda, actually */ |
|
783 |
|
|
784 |
int Run_Start = -1; |
int Run_Start = -1; |
785 |
uint32_t Min_Cost = 2<<16; |
uint32_t Min_Cost = 2<<TL_SHIFT; |
786 |
|
|
787 |
int Last_Node = -1; |
int Last_Node = -1; |
788 |
uint32_t Last_Cost = 0; |
uint32_t Last_Cost = 0; |
789 |
|
|
790 |
int i, j, sum; |
int i, j; |
791 |
Run_Costs[-1] = 2<<16; /* source (w/ CBP penalty) */ |
|
792 |
|
/* source (w/ CBP penalty) */ |
793 |
|
Run_Costs[-1] = 2<<TL_SHIFT; |
794 |
|
|
795 |
Non_Zero = Find_Last(Out, Zigzag, Non_Zero); |
Non_Zero = Find_Last(Out, Zigzag, Non_Zero); |
796 |
if (Non_Zero<0) |
if (Non_Zero<0) |
797 |
return 0; /* Sum is zero if there are only zero coeffs */ |
return 0; /* Sum is zero if there are only zero coeffs */ |
798 |
|
|
799 |
for(i=0; i<=Non_Zero; i++) { |
for(i=0; i<=Non_Zero; i++) { |
800 |
|
const int q = ((Q*QuantMatrix[Zigzag[i]])>>4); |
801 |
|
const int Mult = 2*q; |
802 |
|
const int Bias = (q-1) | 1; |
803 |
|
const int Lev0 = Mult + Bias; |
804 |
|
|
805 |
const int AC = In[Zigzag[i]]; |
const int AC = In[Zigzag[i]]; |
806 |
const int Level1 = Out[Zigzag[i]]; |
const int Level1 = Out[Zigzag[i]]; |
807 |
const int Dist0 = Lambda* AC*AC; |
const unsigned int Dist0 = Lambda* AC*AC; |
808 |
uint32_t Best_Cost = 0xf0000000; |
uint32_t Best_Cost = 0xf0000000; |
809 |
Last_Cost += Dist0; |
Last_Cost += Dist0; |
810 |
|
|
824 |
Cost0 = Lambda*dQ*dQ; |
Cost0 = Lambda*dQ*dQ; |
825 |
|
|
826 |
Nodes[i].Run = 1; |
Nodes[i].Run = 1; |
827 |
Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0; |
Best_Cost = (Code_Len20[0]<<TL_SHIFT) + Run_Costs[i-1]+Cost0; |
828 |
for(Run=i-Run_Start; Run>0; --Run) { |
for(Run=i-Run_Start; Run>0; --Run) { |
829 |
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
830 |
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16); |
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<TL_SHIFT); |
831 |
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16); |
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<TL_SHIFT); |
832 |
|
|
833 |
/* |
/* TODO: what about tie-breaks? Should we favor short runs or |
|
* TODO: what about tie-breaks? Should we favor short runs or |
|
834 |
* long runs? Although the error is the same, it would not be |
* long runs? Although the error is the same, it would not be |
835 |
* spread the same way along high and low frequencies... |
* spread the same way along high and low frequencies... */ |
|
*/ |
|
836 |
|
|
837 |
/* (I'd say: favour short runs => hifreq errors (HVS) -- gruel ) */ |
/* Gruel: I'd say, favour short runs => hifreq errors (HVS) */ |
838 |
|
|
839 |
if (Cost<Best_Cost) { |
if (Cost<Best_Cost) { |
840 |
Best_Cost = Cost; |
Best_Cost = Cost; |
849 |
} |
} |
850 |
if (Last_Node==i) |
if (Last_Node==i) |
851 |
Last.Level = Nodes[i].Level; |
Last.Level = Nodes[i].Level; |
852 |
} else { /* "big" levels */ |
} else if (51U>(uint32_t)(Level1+25)) { |
853 |
|
/* "big" levels (not less than ESC3, though) */ |
854 |
const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last; |
const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last; |
855 |
int Level2; |
int Level2; |
856 |
int dQ1, dQ2; |
int dQ1, dQ2; |
886 |
uint32_t Cost1, Cost2; |
uint32_t Cost1, Cost2; |
887 |
int bLevel; |
int bLevel; |
888 |
|
|
889 |
/* |
/* for sub-optimal (but slightly worth it, speed-wise) search, |
890 |
* for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following: |
* uncomment the following: |
891 |
* if (Cost_Base>=Best_Cost) continue; |
* if (Cost_Base>=Best_Cost) continue; |
892 |
* (? doesn't seem to have any effect -- gruel ) |
* (? doesn't seem to have any effect -- gruel ) */ |
|
*/ |
|
893 |
|
|
894 |
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16); |
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<TL_SHIFT); |
895 |
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21; |
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<TL_SHIFT) + dDist21; |
896 |
|
|
897 |
if (Cost2<Cost1) { |
if (Cost2<Cost1) { |
898 |
Cost1 = Cost2; |
Cost1 = Cost2; |
907 |
Nodes[i].Level = bLevel; |
Nodes[i].Level = bLevel; |
908 |
} |
} |
909 |
|
|
910 |
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16); |
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<TL_SHIFT); |
911 |
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21; |
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<TL_SHIFT) + dDist21; |
912 |
|
|
913 |
if (Cost2<Cost1) { |
if (Cost2<Cost1) { |
914 |
Cost1 = Cost2; |
Cost1 = Cost2; |
924 |
Last_Node = i; |
Last_Node = i; |
925 |
} |
} |
926 |
} /* end of "for Run" */ |
} /* end of "for Run" */ |
927 |
|
} else { |
928 |
|
/* Very very high levels, with no chance of being optimizable |
929 |
|
* => Simply pick best Run. */ |
930 |
|
int Run; |
931 |
|
for(Run=i-Run_Start; Run>0; --Run) { |
932 |
|
/* 30 bits + no distortion */ |
933 |
|
const uint32_t Cost = (30<<TL_SHIFT) + Run_Costs[i-Run]; |
934 |
|
if (Cost<Best_Cost) { |
935 |
|
Best_Cost = Cost; |
936 |
|
Nodes[i].Run = Run; |
937 |
|
Nodes[i].Level = Level1; |
938 |
|
} |
939 |
|
|
940 |
|
if (Cost<Last_Cost) { |
941 |
|
Last_Cost = Cost; |
942 |
|
Last.Run = Run; |
943 |
|
Last.Level = Level1; |
944 |
|
Last_Node = i; |
945 |
|
} |
946 |
|
} |
947 |
} |
} |
948 |
|
|
949 |
|
|
950 |
Run_Costs[i] = Best_Cost; |
Run_Costs[i] = Best_Cost; |
951 |
|
|
952 |
if (Best_Cost < Min_Cost + Dist0) { |
if (Best_Cost < Min_Cost + Dist0) { |
953 |
Min_Cost = Best_Cost; |
Min_Cost = Best_Cost; |
954 |
Run_Start = i; |
Run_Start = i; |
955 |
} else { |
} else { |
956 |
/* |
/* as noticed by Michael Niedermayer (michaelni at gmx.at), |
957 |
* as noticed by Michael Niedermayer (michaelni at gmx.at), there's |
* there's a code shorter by 1 bit for a larger run (!), same |
958 |
* a code shorter by 1 bit for a larger run (!), same level. We give |
* level. We give it a chance by not moving the left barrier too |
959 |
* it a chance by not moving the left barrier too much. |
* much. */ |
960 |
*/ |
while( Run_Costs[Run_Start]>Min_Cost+(1<<TL_SHIFT) ) |
|
|
|
|
while( Run_Costs[Run_Start]>Min_Cost+(1<<16) ) |
|
961 |
Run_Start++; |
Run_Start++; |
962 |
|
|
963 |
/* spread on preceding coeffs the cost incurred by skipping this one */ |
/* spread on preceding coeffs the cost incurred by skipping this |
964 |
|
* one */ |
965 |
for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0; |
for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0; |
966 |
Min_Cost += Dist0; |
Min_Cost += Dist0; |
967 |
} |
} |
968 |
} |
} |
969 |
|
|
970 |
/* It seems trellis doesn't give good results... just compute the Out sum and |
/* It seems trellis doesn't give good results... just leave the block untouched |
971 |
* quit (even if we did not modify it, upperlayer relies on this data) */ |
* and return the original sum value */ |
972 |
if (Last_Node<0) |
if (Last_Node<0) |
973 |
return Compute_Sum(Out, Non_Zero); |
return Sum; |
974 |
|
|
975 |
/* reconstruct optimal sequence backward with surviving paths */ |
/* reconstruct optimal sequence backward with surviving paths */ |
976 |
memset(Out, 0x00, 64*sizeof(*Out)); |
memset(Out, 0x00, 64*sizeof(*Out)); |
977 |
Out[Zigzag[Last_Node]] = Last.Level; |
Out[Zigzag[Last_Node]] = Last.Level; |
978 |
i = Last_Node - Last.Run; |
i = Last_Node - Last.Run; |
979 |
sum = 0; |
Sum = abs(Last.Level); |
980 |
while(i>=0) { |
while(i>=0) { |
981 |
Out[Zigzag[i]] = Nodes[i].Level; |
Out[Zigzag[i]] = Nodes[i].Level; |
982 |
sum += abs(Nodes[i].Level); |
Sum += abs(Nodes[i].Level); |
983 |
i -= Nodes[i].Run; |
i -= Nodes[i].Run; |
984 |
} |
} |
985 |
|
|
986 |
return sum; |
return Sum; |
|
} |
|
|
|
|
|
static int |
|
|
dct_quantize_trellis_mpeg_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero) |
|
|
{ |
|
|
/* ToDo: Ok ok it's just a place holder for Gruel -- damn write this one :-) */ |
|
|
return Compute_Sum(Out, 63); |
|
987 |
} |
} |
988 |
|
|
989 |
/* original version including heavy debugging info */ |
/* original version including heavy debugging info */ |
1042 |
V -= Ref[Zigzag[i]]; |
V -= Ref[Zigzag[i]]; |
1043 |
Dist += V*V; |
Dist += V*V; |
1044 |
} |
} |
1045 |
Cost = Lambda*Dist + (Bits<<16); |
Cost = Lambda*Dist + (Bits<<TL_SHIFT); |
1046 |
if (DBG==1) |
if (DBG==1) |
1047 |
printf( " Last:%2d/%2d Cost = [(Bits=%5.0d) + Lambda*(Dist=%6.0d) = %d ] >>12= %d ", Last,Max, Bits, Dist, Cost, Cost>>12 ); |
printf( " Last:%2d/%2d Cost = [(Bits=%5.0d) + Lambda*(Dist=%6.0d) = %d ] >>12= %d ", Last,Max, Bits, Dist, Cost, Cost>>12 ); |
1048 |
return Cost; |
return Cost; |
1074 |
const int Lambda = Trellis_Lambda_Tabs[Q-1]; /* it's 1/lambda, actually */ |
const int Lambda = Trellis_Lambda_Tabs[Q-1]; /* it's 1/lambda, actually */ |
1075 |
|
|
1076 |
int Run_Start = -1; |
int Run_Start = -1; |
1077 |
Run_Costs[-1] = 2<<16; /* source (w/ CBP penalty) */ |
Run_Costs[-1] = 2<<TL_SHIFT; /* source (w/ CBP penalty) */ |
1078 |
uint32_t Min_Cost = 2<<16; |
uint32_t Min_Cost = 2<<TL_SHIFT; |
1079 |
|
|
1080 |
int Last_Node = -1; |
int Last_Node = -1; |
1081 |
uint32_t Last_Cost = 0; |
uint32_t Last_Cost = 0; |
1114 |
Cost0 = Lambda*dQ*dQ; |
Cost0 = Lambda*dQ*dQ; |
1115 |
|
|
1116 |
Nodes[i].Run = 1; |
Nodes[i].Run = 1; |
1117 |
Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0; |
Best_Cost = (Code_Len20[0]<<TL_SHIFT) + Run_Costs[i-1]+Cost0; |
1118 |
for(Run=i-Run_Start; Run>0; --Run) |
for(Run=i-Run_Start; Run>0; --Run) |
1119 |
{ |
{ |
1120 |
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
1121 |
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16); |
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<TL_SHIFT); |
1122 |
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16); |
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<TL_SHIFT); |
1123 |
|
|
1124 |
/* |
/* |
1125 |
* TODO: what about tie-breaks? Should we favor short runs or |
* TODO: what about tie-breaks? Should we favor short runs or |
1195 |
* for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following: |
* for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following: |
1196 |
* if (Cost_Base>=Best_Cost) continue; |
* if (Cost_Base>=Best_Cost) continue; |
1197 |
*/ |
*/ |
1198 |
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16); |
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<TL_SHIFT); |
1199 |
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21; |
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<TL_SHIFT) + dDist21; |
1200 |
|
|
1201 |
if (Cost2<Cost1) { |
if (Cost2<Cost1) { |
1202 |
Cost1 = Cost2; |
Cost1 = Cost2; |
1210 |
Nodes[i].Level = bLevel; |
Nodes[i].Level = bLevel; |
1211 |
} |
} |
1212 |
|
|
1213 |
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16); |
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<TL_SHIFT); |
1214 |
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21; |
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<TL_SHIFT) + dDist21; |
1215 |
|
|
1216 |
if (Cost2<Cost1) { |
if (Cost2<Cost1) { |
1217 |
Cost1 = Cost2; |
Cost1 = Cost2; |
1257 |
* it a chance by not moving the left barrier too much. |
* it a chance by not moving the left barrier too much. |
1258 |
*/ |
*/ |
1259 |
|
|
1260 |
while( Run_Costs[Run_Start]>Min_Cost+(1<<16) ) |
while( Run_Costs[Run_Start]>Min_Cost+(1<<TL_SHIFT) ) |
1261 |
Run_Start++; |
Run_Start++; |
1262 |
|
|
1263 |
/* spread on preceding coeffs the cost incurred by skipping this one */ |
/* spread on preceding coeffs the cost incurred by skipping this one */ |