43 |
#include "../encoder.h" |
#include "../encoder.h" |
44 |
|
|
45 |
#include "../image/reduced.h" |
#include "../image/reduced.h" |
46 |
|
#include "../quant/quant_matrix.h" |
47 |
|
|
48 |
MBFIELDTEST_PTR MBFieldTest; |
MBFIELDTEST_PTR MBFieldTest; |
49 |
|
|
177 |
stop_iquant_timer(); |
stop_iquant_timer(); |
178 |
} |
} |
179 |
|
|
|
|
|
|
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); |
|
|
|
|
|
static int |
|
|
dct_quantize_trellis_h263_c(int16_t *const Out, |
|
|
const int16_t *const In, |
|
|
int Q, |
|
|
const uint16_t * const Zigzag, |
|
|
int Non_Zero); |
|
|
|
|
180 |
static int |
static int |
181 |
dct_quantize_trellis_mpeg_c(int16_t *const Out, |
dct_quantize_trellis_c(int16_t *const Out, |
182 |
const int16_t *const In, |
const int16_t *const In, |
183 |
int Q, |
int Q, |
184 |
const uint16_t * const Zigzag, |
const uint16_t * const Zigzag, |
185 |
|
const uint16_t * const QuantMatrix, |
186 |
int Non_Zero); |
int Non_Zero); |
187 |
|
|
188 |
/* Quantize all blocks -- Inter mode */ |
/* Quantize all blocks -- Inter mode */ |
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++) { |
217 |
sum = quant[mpeg](&qcoeff[i*64], &data[i*64], pMB->quant); |
sum = quant[mpeg](&qcoeff[i*64], &data[i*64], pMB->quant); |
218 |
|
|
219 |
if(sum && (frame->vop_flags & XVID_VOP_TRELLISQUANT)) { |
if(sum && (frame->vop_flags & XVID_VOP_TRELLISQUANT)) { |
220 |
sum = trellis[mpeg](&qcoeff[i*64], &data[i*64], pMB->quant, &scan_tables[0][0], 63); |
const static uint16_t h263matrix[] = |
221 |
|
{ |
222 |
|
16, 16, 16, 16, 16, 16, 16, 16, |
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 |
|
}; |
231 |
|
sum = dct_quantize_trellis_c(&qcoeff[i*64], &data[i*64], |
232 |
|
pMB->quant, &scan_tables[0][0], |
233 |
|
(mpeg)?(uint16_t*)get_inter_matrix():h263matrix, |
234 |
|
63); |
235 |
} |
} |
236 |
stop_quant_timer(); |
stop_quant_timer(); |
237 |
|
|
647 |
* IEEE Transactions on Image Processing, Vol.9, No.8, Aug. 2000. |
* IEEE Transactions on Image Processing, Vol.9, No.8, Aug. 2000. |
648 |
* |
* |
649 |
* we are at stake with a simplified Bellmand-Ford / Dijkstra Single |
* we are at stake with a simplified Bellmand-Ford / Dijkstra Single |
650 |
* Source Shorted Path algo. But due to the underlying graph structure |
* Source Shortest Path algo. But due to the underlying graph structure |
651 |
* ("Trellis"), it can be turned into a dynamic programming algo, |
* ("Trellis"), it can be turned into a dynamic programming algo, |
652 |
* partially saving the explicit graph's nodes representation. And |
* partially saving the explicit graph's nodes representation. And |
653 |
* 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 |
654 |
* known, and of fixed sized. |
* known, and of fixed size. |
655 |
*--------------------------------------------------------------------------*/ |
*--------------------------------------------------------------------------*/ |
656 |
|
|
657 |
|
|
751 |
Code_Len24,Code_Len23,Code_Len22,Code_Len21, Code_Len3, Code_Len1, |
Code_Len24,Code_Len23,Code_Len22,Code_Len21, Code_Len3, Code_Len1, |
752 |
}; |
}; |
753 |
|
|
754 |
#define TL(q) 0xfe00/(q*q) |
/* TL_SHIFT controls the precision of the RD optimizations in trellis |
755 |
|
* valid range is [10..16]. The bigger, the more trellis is vulnerable |
756 |
|
* to overflows in cost formulas. |
757 |
|
* - 10 allows ac values up to 2^11 == 2048 |
758 |
|
* - 16 allows ac values up to 2^8 == 256 |
759 |
|
*/ |
760 |
|
#define TL_SHIFT 11 |
761 |
|
#define TL(q) ((0xfe00>>(16-TL_SHIFT))/(q*q)) |
762 |
|
|
763 |
static const int Trellis_Lambda_Tabs[31] = { |
static const int Trellis_Lambda_Tabs[31] = { |
764 |
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), |
788 |
|
|
789 |
return(sum); |
return(sum); |
790 |
} |
} |
|
/* this routine has been strippen of all debug code */ |
|
791 |
|
|
792 |
|
/* this routine has been strippen of all debug code */ |
793 |
static int |
static int |
794 |
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, |
795 |
|
const int16_t *const In, |
796 |
|
int Q, |
797 |
|
const uint16_t * const Zigzag, |
798 |
|
const uint16_t * const QuantMatrix, |
799 |
|
int Non_Zero) |
800 |
{ |
{ |
801 |
|
|
802 |
/* |
/* Note: We should search last non-zero coeffs on *real* DCT input coeffs |
803 |
* 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 |
804 |
* not quantized one (Out[]). However, it only improves the result *very* |
* *very* slightly (~0.01dB), whereas speed drops to crawling level :) |
805 |
* slightly (~0.01dB), whereas speed drops to crawling level :) |
* Well, actually, taking 1 more coeff past Non_Zero into account sometimes |
806 |
* Well, actually, taking 1 more coeff past Non_Zero into account sometimes helps. |
* helps. */ |
|
*/ |
|
807 |
typedef struct { int16_t Run, Level; } NODE; |
typedef struct { int16_t Run, Level; } NODE; |
808 |
|
|
809 |
NODE Nodes[65], Last; |
NODE Nodes[65], Last; |
810 |
uint32_t Run_Costs0[64+1]; |
uint32_t Run_Costs0[64+1]; |
811 |
uint32_t * const Run_Costs = Run_Costs0 + 1; |
uint32_t * const Run_Costs = Run_Costs0 + 1; |
812 |
const int Mult = 2*Q; |
|
813 |
const int Bias = (Q-1) | 1; |
/* it's 1/lambda, actually */ |
814 |
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 */ |
|
815 |
|
|
816 |
int Run_Start = -1; |
int Run_Start = -1; |
817 |
uint32_t Min_Cost = 2<<16; |
uint32_t Min_Cost = 2<<TL_SHIFT; |
818 |
|
|
819 |
int Last_Node = -1; |
int Last_Node = -1; |
820 |
uint32_t Last_Cost = 0; |
uint32_t Last_Cost = 0; |
821 |
|
|
822 |
int i, j, sum; |
int i, j, sum; |
823 |
Run_Costs[-1] = 2<<16; /* source (w/ CBP penalty) */ |
|
824 |
|
/* source (w/ CBP penalty) */ |
825 |
|
Run_Costs[-1] = 2<<TL_SHIFT; |
826 |
|
|
827 |
Non_Zero = Find_Last(Out, Zigzag, Non_Zero); |
Non_Zero = Find_Last(Out, Zigzag, Non_Zero); |
828 |
if (Non_Zero<0) |
if (Non_Zero<0) |
829 |
return 0; /* Sum is zero if there are only zero coeffs */ |
return 0; /* Sum is zero if there are only zero coeffs */ |
830 |
|
|
831 |
for(i=0; i<=Non_Zero; i++) { |
for(i=0; i<=Non_Zero; i++) { |
832 |
|
const int q = ((Q*QuantMatrix[Zigzag[i]])>>4); |
833 |
|
const int Mult = 2*q; |
834 |
|
const int Bias = (q-1) | 1; |
835 |
|
const int Lev0 = Mult + Bias; |
836 |
|
|
837 |
const int AC = In[Zigzag[i]]; |
const int AC = In[Zigzag[i]]; |
838 |
const int Level1 = Out[Zigzag[i]]; |
const int Level1 = Out[Zigzag[i]]; |
839 |
const int Dist0 = Lambda* AC*AC; |
const unsigned int Dist0 = Lambda* AC*AC; |
840 |
uint32_t Best_Cost = 0xf0000000; |
uint32_t Best_Cost = 0xf0000000; |
841 |
Last_Cost += Dist0; |
Last_Cost += Dist0; |
842 |
|
|
856 |
Cost0 = Lambda*dQ*dQ; |
Cost0 = Lambda*dQ*dQ; |
857 |
|
|
858 |
Nodes[i].Run = 1; |
Nodes[i].Run = 1; |
859 |
Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0; |
Best_Cost = (Code_Len20[0]<<TL_SHIFT) + Run_Costs[i-1]+Cost0; |
860 |
for(Run=i-Run_Start; Run>0; --Run) { |
for(Run=i-Run_Start; Run>0; --Run) { |
861 |
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
862 |
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16); |
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<TL_SHIFT); |
863 |
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16); |
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<TL_SHIFT); |
864 |
|
|
865 |
/* |
/* TODO: what about tie-breaks? Should we favor short runs or |
|
* TODO: what about tie-breaks? Should we favor short runs or |
|
866 |
* long runs? Although the error is the same, it would not be |
* long runs? Although the error is the same, it would not be |
867 |
* spread the same way along high and low frequencies... |
* spread the same way along high and low frequencies... */ |
|
*/ |
|
868 |
|
|
869 |
/* (I'd say: favour short runs => hifreq errors (HVS) -- gruel ) */ |
/* Gruel: I'd say, favour short runs => hifreq errors (HVS) */ |
870 |
|
|
871 |
if (Cost<Best_Cost) { |
if (Cost<Best_Cost) { |
872 |
Best_Cost = Cost; |
Best_Cost = Cost; |
881 |
} |
} |
882 |
if (Last_Node==i) |
if (Last_Node==i) |
883 |
Last.Level = Nodes[i].Level; |
Last.Level = Nodes[i].Level; |
884 |
} else { /* "big" levels */ |
} else if (51U>(uint32_t)(Level1+25)) { |
885 |
|
/* "big" levels (not less than ESC3, though) */ |
886 |
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; |
887 |
int Level2; |
int Level2; |
888 |
int dQ1, dQ2; |
int dQ1, dQ2; |
918 |
uint32_t Cost1, Cost2; |
uint32_t Cost1, Cost2; |
919 |
int bLevel; |
int bLevel; |
920 |
|
|
921 |
/* |
/* for sub-optimal (but slightly worth it, speed-wise) search, |
922 |
* for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following: |
* uncomment the following: |
923 |
* if (Cost_Base>=Best_Cost) continue; |
* if (Cost_Base>=Best_Cost) continue; |
924 |
* (? doesn't seem to have any effect -- gruel ) |
* (? doesn't seem to have any effect -- gruel ) */ |
|
*/ |
|
925 |
|
|
926 |
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16); |
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<TL_SHIFT); |
927 |
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21; |
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<TL_SHIFT) + dDist21; |
928 |
|
|
929 |
if (Cost2<Cost1) { |
if (Cost2<Cost1) { |
930 |
Cost1 = Cost2; |
Cost1 = Cost2; |
939 |
Nodes[i].Level = bLevel; |
Nodes[i].Level = bLevel; |
940 |
} |
} |
941 |
|
|
942 |
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16); |
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<TL_SHIFT); |
943 |
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21; |
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<TL_SHIFT) + dDist21; |
944 |
|
|
945 |
if (Cost2<Cost1) { |
if (Cost2<Cost1) { |
946 |
Cost1 = Cost2; |
Cost1 = Cost2; |
956 |
Last_Node = i; |
Last_Node = i; |
957 |
} |
} |
958 |
} /* end of "for Run" */ |
} /* end of "for Run" */ |
959 |
|
} else { |
960 |
|
/* Very very high levels, with no chance of being optimizable |
961 |
|
* => Simply pick best Run. */ |
962 |
|
int Run; |
963 |
|
for(Run=i-Run_Start; Run>0; --Run) { |
964 |
|
/* 30 bits + no distortion */ |
965 |
|
const uint32_t Cost = (30<<TL_SHIFT) + Run_Costs[i-Run]; |
966 |
|
if (Cost<Best_Cost) { |
967 |
|
Best_Cost = Cost; |
968 |
|
Nodes[i].Run = Run; |
969 |
|
Nodes[i].Level = Level1; |
970 |
|
} |
971 |
|
|
972 |
|
if (Cost<Last_Cost) { |
973 |
|
Last_Cost = Cost; |
974 |
|
Last.Run = Run; |
975 |
|
Last.Level = Level1; |
976 |
|
Last_Node = i; |
977 |
|
} |
978 |
|
} |
979 |
} |
} |
980 |
|
|
981 |
|
|
982 |
Run_Costs[i] = Best_Cost; |
Run_Costs[i] = Best_Cost; |
983 |
|
|
984 |
if (Best_Cost < Min_Cost + Dist0) { |
if (Best_Cost < Min_Cost + Dist0) { |
985 |
Min_Cost = Best_Cost; |
Min_Cost = Best_Cost; |
986 |
Run_Start = i; |
Run_Start = i; |
987 |
} else { |
} else { |
988 |
/* |
/* as noticed by Michael Niedermayer (michaelni at gmx.at), |
989 |
* as noticed by Michael Niedermayer (michaelni at gmx.at), there's |
* there's a code shorter by 1 bit for a larger run (!), same |
990 |
* 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 |
991 |
* it a chance by not moving the left barrier too much. |
* much. */ |
992 |
*/ |
while( Run_Costs[Run_Start]>Min_Cost+(1<<TL_SHIFT) ) |
|
|
|
|
while( Run_Costs[Run_Start]>Min_Cost+(1<<16) ) |
|
993 |
Run_Start++; |
Run_Start++; |
994 |
|
|
995 |
/* spread on preceding coeffs the cost incurred by skipping this one */ |
/* spread on preceding coeffs the cost incurred by skipping this |
996 |
|
* one */ |
997 |
for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0; |
for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0; |
998 |
Min_Cost += Dist0; |
Min_Cost += Dist0; |
999 |
} |
} |
1000 |
} |
} |
1001 |
|
|
1002 |
/* It seems trellis doesn't give good results... just compute the Out sum and |
/* It seems trellis doesn't give good results... just compute the Out sum |
1003 |
* quit (even if we did not modify it, upperlayer relies on this data) */ |
* and quit */ |
1004 |
if (Last_Node<0) |
if (Last_Node<0) |
1005 |
return Compute_Sum(Out, Non_Zero); |
return Compute_Sum(Out, Non_Zero); |
1006 |
|
|
1018 |
return sum; |
return sum; |
1019 |
} |
} |
1020 |
|
|
|
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); |
|
|
} |
|
|
|
|
1021 |
/* original version including heavy debugging info */ |
/* original version including heavy debugging info */ |
1022 |
|
|
1023 |
#ifdef DBGTRELL |
#ifdef DBGTRELL |
1074 |
V -= Ref[Zigzag[i]]; |
V -= Ref[Zigzag[i]]; |
1075 |
Dist += V*V; |
Dist += V*V; |
1076 |
} |
} |
1077 |
Cost = Lambda*Dist + (Bits<<16); |
Cost = Lambda*Dist + (Bits<<TL_SHIFT); |
1078 |
if (DBG==1) |
if (DBG==1) |
1079 |
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 ); |
1080 |
return Cost; |
return Cost; |
1106 |
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 */ |
1107 |
|
|
1108 |
int Run_Start = -1; |
int Run_Start = -1; |
1109 |
Run_Costs[-1] = 2<<16; /* source (w/ CBP penalty) */ |
Run_Costs[-1] = 2<<TL_SHIFT; /* source (w/ CBP penalty) */ |
1110 |
uint32_t Min_Cost = 2<<16; |
uint32_t Min_Cost = 2<<TL_SHIFT; |
1111 |
|
|
1112 |
int Last_Node = -1; |
int Last_Node = -1; |
1113 |
uint32_t Last_Cost = 0; |
uint32_t Last_Cost = 0; |
1146 |
Cost0 = Lambda*dQ*dQ; |
Cost0 = Lambda*dQ*dQ; |
1147 |
|
|
1148 |
Nodes[i].Run = 1; |
Nodes[i].Run = 1; |
1149 |
Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0; |
Best_Cost = (Code_Len20[0]<<TL_SHIFT) + Run_Costs[i-1]+Cost0; |
1150 |
for(Run=i-Run_Start; Run>0; --Run) |
for(Run=i-Run_Start; Run>0; --Run) |
1151 |
{ |
{ |
1152 |
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
1153 |
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16); |
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<TL_SHIFT); |
1154 |
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16); |
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<TL_SHIFT); |
1155 |
|
|
1156 |
/* |
/* |
1157 |
* TODO: what about tie-breaks? Should we favor short runs or |
* TODO: what about tie-breaks? Should we favor short runs or |
1227 |
* 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: |
1228 |
* if (Cost_Base>=Best_Cost) continue; |
* if (Cost_Base>=Best_Cost) continue; |
1229 |
*/ |
*/ |
1230 |
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16); |
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<TL_SHIFT); |
1231 |
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21; |
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<TL_SHIFT) + dDist21; |
1232 |
|
|
1233 |
if (Cost2<Cost1) { |
if (Cost2<Cost1) { |
1234 |
Cost1 = Cost2; |
Cost1 = Cost2; |
1242 |
Nodes[i].Level = bLevel; |
Nodes[i].Level = bLevel; |
1243 |
} |
} |
1244 |
|
|
1245 |
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16); |
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<TL_SHIFT); |
1246 |
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21; |
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<TL_SHIFT) + dDist21; |
1247 |
|
|
1248 |
if (Cost2<Cost1) { |
if (Cost2<Cost1) { |
1249 |
Cost1 = Cost2; |
Cost1 = Cost2; |
1289 |
* it a chance by not moving the left barrier too much. |
* it a chance by not moving the left barrier too much. |
1290 |
*/ |
*/ |
1291 |
|
|
1292 |
while( Run_Costs[Run_Start]>Min_Cost+(1<<16) ) |
while( Run_Costs[Run_Start]>Min_Cost+(1<<TL_SHIFT) ) |
1293 |
Run_Start++; |
Run_Start++; |
1294 |
|
|
1295 |
/* spread on preceding coeffs the cost incurred by skipping this one */ |
/* spread on preceding coeffs the cost incurred by skipping this one */ |