Commit ff53c41f authored by lorenm's avatar lorenm

13% faster decode_residual (cavlc).

patch by diane_cartman at gmx dot de.


git-svn-id: file:///var/local/repositories/ffmpeg/trunk@4617 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b
parent 13a51b81
...@@ -4437,8 +4437,8 @@ static inline int get_dct8x8_allowed(H264Context *h){ ...@@ -4437,8 +4437,8 @@ static inline int get_dct8x8_allowed(H264Context *h){
static int decode_residual(H264Context *h, GetBitContext *gb, DCTELEM *block, int n, const uint8_t *scantable, const uint16_t *qmul, int max_coeff){ static int decode_residual(H264Context *h, GetBitContext *gb, DCTELEM *block, int n, const uint8_t *scantable, const uint16_t *qmul, int max_coeff){
MpegEncContext * const s = &h->s; MpegEncContext * const s = &h->s;
static const int coeff_token_table_index[17]= {0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3}; static const int coeff_token_table_index[17]= {0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3};
int level[16], run[16]; int level[16];
int suffix_length, zeros_left, coeff_num, coeff_token, total_coeff, i, trailing_ones; int zeros_left, coeff_num, coeff_token, total_coeff, i, j, trailing_ones, run_before;
//FIXME put trailing_onex into the context //FIXME put trailing_onex into the context
...@@ -4471,12 +4471,12 @@ static int decode_residual(H264Context *h, GetBitContext *gb, DCTELEM *block, in ...@@ -4471,12 +4471,12 @@ static int decode_residual(H264Context *h, GetBitContext *gb, DCTELEM *block, in
level[i]= 1 - 2*get_bits1(gb); level[i]= 1 - 2*get_bits1(gb);
} }
suffix_length= total_coeff > 10 && trailing_ones < 3; if(i<total_coeff) {
for(; i<total_coeff; i++){
const int prefix= get_level_prefix(gb);
int level_code, mask; int level_code, mask;
int suffix_length = total_coeff > 10 && trailing_ones < 3;
int prefix= get_level_prefix(gb);
//first coefficient has suffix_length equal to 0 or 1
if(prefix<14){ //FIXME try to build a large unified VLC table for all this if(prefix<14){ //FIXME try to build a large unified VLC table for all this
if(suffix_length) if(suffix_length)
level_code= (prefix<<suffix_length) + get_bits(gb, suffix_length); //part level_code= (prefix<<suffix_length) + get_bits(gb, suffix_length); //part
...@@ -4495,20 +4495,32 @@ static int decode_residual(H264Context *h, GetBitContext *gb, DCTELEM *block, in ...@@ -4495,20 +4495,32 @@ static int decode_residual(H264Context *h, GetBitContext *gb, DCTELEM *block, in
return -1; return -1;
} }
if(i==trailing_ones && i<3) level_code+= 2; //FIXME split first iteration if(trailing_ones < 3) level_code += 2;
suffix_length = 1;
if(level_code > 5)
suffix_length++;
mask= -(level_code&1); mask= -(level_code&1);
level[i]= (((2+level_code)>>1) ^ mask) - mask; level[i]= (((2+level_code)>>1) ^ mask) - mask;
i++;
if(suffix_length==0) suffix_length=1; //FIXME split first iteration
//remaining coefficients have suffix_length > 0
#if 1 for(;i<total_coeff;i++) {
if(ABS(level[i]) > (3<<(suffix_length-1)) && suffix_length<6) suffix_length++; static const int suffix_limit[7] = {0,5,11,23,47,95,INT_MAX };
#else prefix = get_level_prefix(gb);
if((2+level_code)>>1) > (3<<(suffix_length-1)) && suffix_length<6) suffix_length++; if(prefix<15){
/* ? == prefix > 2 or sth */ level_code = (prefix<<suffix_length) + get_bits(gb, suffix_length);
#endif }else if(prefix==15){
tprintf("level: %d suffix_length:%d\n", level[i], suffix_length); level_code = (prefix<<suffix_length) + get_bits(gb, 12);
}else{
av_log(h->s.avctx, AV_LOG_ERROR, "prefix too large at %d %d\n", s->mb_x, s->mb_y);
return -1;
}
mask= -(level_code&1);
level[i]= (((2+level_code)>>1) ^ mask) - mask;
if(level_code > suffix_limit[suffix_length])
suffix_length++;
}
} }
if(total_coeff == max_coeff) if(total_coeff == max_coeff)
...@@ -4520,49 +4532,48 @@ static int decode_residual(H264Context *h, GetBitContext *gb, DCTELEM *block, in ...@@ -4520,49 +4532,48 @@ static int decode_residual(H264Context *h, GetBitContext *gb, DCTELEM *block, in
zeros_left= get_vlc2(gb, total_zeros_vlc[ total_coeff-1 ].table, TOTAL_ZEROS_VLC_BITS, 1); zeros_left= get_vlc2(gb, total_zeros_vlc[ total_coeff-1 ].table, TOTAL_ZEROS_VLC_BITS, 1);
} }
for(i=0; i<total_coeff-1; i++){ coeff_num = zeros_left + total_coeff - 1;
if(zeros_left <=0) j = scantable[coeff_num];
break; if(n > 24){
block[j] = level[0];
for(i=1;i<total_coeff;i++) {
if(zeros_left <= 0)
run_before = 0;
else if(zeros_left < 7){ else if(zeros_left < 7){
run[i]= get_vlc2(gb, run_vlc[zeros_left-1].table, RUN_VLC_BITS, 1); run_before= get_vlc2(gb, run_vlc[zeros_left-1].table, RUN_VLC_BITS, 1);
}else{ }else{
run[i]= get_vlc2(gb, run7_vlc.table, RUN7_VLC_BITS, 2); run_before= get_vlc2(gb, run7_vlc.table, RUN7_VLC_BITS, 2);
}
zeros_left -= run[i];
}
if(zeros_left<0){
av_log(h->s.avctx, AV_LOG_ERROR, "negative number of zero coeffs at %d %d\n", s->mb_x, s->mb_y);
return -1;
} }
zeros_left -= run_before;
for(; i<total_coeff-1; i++){ coeff_num -= 1 + run_before;
run[i]= 0;
}
run[i]= zeros_left;
coeff_num=-1;
if(n > 24){
for(i=total_coeff-1; i>=0; i--){ //FIXME merge into rundecode?
int j;
coeff_num += run[i] + 1; //FIXME add 1 earlier ?
j= scantable[ coeff_num ]; j= scantable[ coeff_num ];
block[j]= level[i]; block[j]= level[i];
} }
}else{ }else{
for(i=total_coeff-1; i>=0; i--){ //FIXME merge into rundecode? block[j] = level[0] * qmul[j];
int j; for(i=1;i<total_coeff;i++) {
if(zeros_left <= 0)
coeff_num += run[i] + 1; //FIXME add 1 earlier ? run_before = 0;
else if(zeros_left < 7){
run_before= get_vlc2(gb, run_vlc[zeros_left-1].table, RUN_VLC_BITS, 1);
}else{
run_before= get_vlc2(gb, run7_vlc.table, RUN7_VLC_BITS, 2);
}
zeros_left -= run_before;
coeff_num -= 1 + run_before;
j= scantable[ coeff_num ]; j= scantable[ coeff_num ];
block[j]= level[i] * qmul[j]; block[j]= level[i] * qmul[j];
// printf("%d %d ", block[j], qmul[j]); // printf("%d %d ", block[j], qmul[j]);
} }
} }
if(zeros_left<0){
av_log(h->s.avctx, AV_LOG_ERROR, "negative number of zero coeffs at %d %d\n", s->mb_x, s->mb_y);
return -1;
}
return 0; return 0;
} }
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment