diff options
Diffstat (limited to 'include/phat/representations/bit_tree_pivot_column.h')
| -rw-r--r-- | include/phat/representations/bit_tree_pivot_column.h | 117 |
1 files changed, 45 insertions, 72 deletions
diff --git a/include/phat/representations/bit_tree_pivot_column.h b/include/phat/representations/bit_tree_pivot_column.h index e8746f8..888e338 100644 --- a/include/phat/representations/bit_tree_pivot_column.h +++ b/include/phat/representations/bit_tree_pivot_column.h @@ -30,12 +30,14 @@ namespace phat { // 'add_index' is still the real bottleneck in practice.
class bit_tree_column
{
+ protected:
+
size_t offset; // data[i + offset] = ith block of the data-bitset
typedef uint64_t block_type;
std::vector< block_type > data;
// this static is not a problem with OMP, it's initialized just after program starts
- static const size_t debrujin_magic_table[64];
+ static const size_t debrujin_magic_table[ 64 ];
enum { block_size_in_bits = 64 };
enum { block_shift = 6 };
@@ -45,118 +47,90 @@ namespace phat { // (-x)&x isolates the rightmost bit.
// The whole method is much faster than calling log2i, and very comparable to using ScanBitForward/Reverse intrinsic,
// which should be one CPU instruction, but is not portable.
- size_t rightmost_pos(const block_type value) const
- {
- return 64 - 1 - debrujin_magic_table[((value & (-(int64_t)value))*0x07EDD5E59A4E28C2) >> 58];
+ size_t rightmost_pos( const block_type value ) const {
+ return 64 - 1 - debrujin_magic_table[ ( (value & (-(int64_t)value) ) * 0x07EDD5E59A4E28C2 ) >> 58 ];
}
- public:
+ public:
- void init(index num_cols)
- {
+ void init( index num_cols ) {
int64_t n = 1; // in case of overflow
- int64_t bottom_blocks_needed = (num_cols+block_size_in_bits-1)/block_size_in_bits;
+ int64_t bottom_blocks_needed = ( num_cols + block_size_in_bits - 1 ) / block_size_in_bits;
int64_t upper_blocks = 1;
// How many blocks/nodes of index needed to index the whole bitset?
- while(n * block_size_in_bits < bottom_blocks_needed)
- {
+ while( n * block_size_in_bits < bottom_blocks_needed ) {
n *= block_size_in_bits;
upper_blocks += n;
}
offset = upper_blocks;
- data.resize(upper_blocks + bottom_blocks_needed, 0);
+ data.resize( upper_blocks + bottom_blocks_needed, 0 );
}
- index get_max_index() const
- {
- if (!data[0])
+ index get_max_index() const {
+ if( !data[ 0 ] )
return -1;
- const size_t size = data.size();
size_t n = 0;
-
- while(true)
- {
- const block_type val = data[n];
- const size_t index = rightmost_pos(val);
- const size_t newn = (n << block_shift) + index + 1;
-
- if (newn >= size)
- {
- const size_t bottom_index = n - offset;
- return (bottom_index << block_shift) + index;
- }
-
+ size_t newn = 0;
+ size_t index = 0;
+ while( newn < data.size() ) {
n = newn;
- }
+ index = rightmost_pos( data[ n ] );
+ newn = ( n << block_shift ) + index + 1;
+ }
- return -1;
+ return ( ( n - offset ) << block_shift ) + index;
}
- bool is_empty() const
- {
- return data[0] == 0;
+ bool is_empty() const {
+ return data[ 0 ] == 0;
}
- void add_index(const size_t entry)
- {
+ void add_index( const size_t entry ) {
const block_type ONE = 1;
- const block_type block_modulo_mask = ((ONE << block_shift) - 1);
+ const block_type block_modulo_mask = ( ONE << block_shift ) - 1;
size_t index_in_level = entry >> block_shift;
size_t address = index_in_level + offset;
size_t index_in_block = entry & block_modulo_mask;
- block_type mask = (ONE << (block_size_in_bits - index_in_block - 1));
-
- while(true)
- {
- data[address]^=mask;
+ block_type mask = ( ONE << ( block_size_in_bits - index_in_block - 1 ) );
- // First we check if we reached the root.
- // Also, if anyone else was in this block, we don't need to update the path up.
- if (!address || (data[address] & ~mask))
- return;
+ data[ address ] ^= mask;
+ // Check if we reached the root. Also, if anyone else was in this block, we don't need to update the path up.
+ while( address && !( data[ address ] & ~mask ) ) {
index_in_block = index_in_level & block_modulo_mask;
index_in_level >>= block_shift;
--address;
address >>= block_shift;
- mask = (ONE << (block_size_in_bits - index_in_block - 1));
+ mask = ( ONE << ( block_size_in_bits - index_in_block - 1 ) );
+ data[ address ] ^= mask;
}
}
- void get_col_and_clear(column &out)
- {
- out.clear();
- while(true)
- {
- index mx = this->get_max_index();
- if (mx == -1)
- break;
- out.push_back(mx);
- add_index(mx);
+ void get_col_and_clear( column &out ) {
+ index mx = this->get_max_index();
+ while( mx != -1 ) {
+ out.push_back( mx );
+ add_index( mx );
+ mx = this->get_max_index();
}
- std::reverse(out.begin(), out.end());
+ std::reverse( out.begin(), out.end() );
}
- void add_col(const column &col)
- {
- for (size_t i = 0; i < col.size(); ++i)
- {
+ void add_col(const column &col) {
+ for( size_t i = 0; i < col.size(); ++i )
add_index(col[i]);
- }
}
void clear() {
- while(true)
- {
- index mx = this->get_max_index();
- if (mx == -1)
- break;
- add_index(mx);
+ index mx = this->get_max_index();
+ while( mx != -1 ) {
+ add_index( mx );
+ mx = this->get_max_index();
}
}
@@ -164,19 +138,18 @@ namespace phat { add_index( get_max_index() );
}
- void set_col( const column& col ) {
+ void set_col( const column& col ) {
clear();
add_col( col );
}
- void get_col( column& col ) {
- col.clear();
+ void get_col( column& col ) {
get_col_and_clear( col );
add_col( col );
}
};
- const size_t bit_tree_column::debrujin_magic_table[64] = {
+ const size_t bit_tree_column::debrujin_magic_table[ 64 ] = {
63, 0, 58, 1, 59, 47, 53, 2,
60, 39, 48, 27, 54, 33, 42, 3,
61, 51, 37, 40, 49, 18, 28, 20,
|