void masking_show_stats
(arg_dont_complain(FILE* f))
{
#ifdef collect_stats
if (f == NULL) return;
fprintf (f, " masked bases: %s\n", commatize(maskingStats.maskedBases));
fprintf (f, "-------------------\n");
#endif // collect_stats
}
void
clsparseDeviceTimer::Print( cl_ulong flopCount, std::string unit )
{
const int tableWidth = 60;
const int tableHalf = tableWidth / 2;
const int tableThird = tableWidth / 3;
const int tableFourth = tableWidth / 4;
const int tableFifth = tableWidth / 5;
for( cl_uint id = 0; id < labelID.size( ); ++id )
{
size_t halfString = labelID[ id ].first.size( ) / 2;
// Print label of timer, in a header
std::cout << std::endl << std::setw( tableHalf + halfString ) << std::setfill( '=' ) << labelID[ id ].first
<< std::setw( tableHalf - halfString ) << "=" << std::endl;
std::cout << std::setfill( ' ' );
std::vector< StatData > mean = getMean( id );
// Print each individual dimension
std::stringstream catLengths;
for( cl_uint t = 0; t < mean.size( ); ++t )
{
cl_double time = mean[ t ].doubleNanoSec;
cl_double gFlops = flopCount / time;
if( mean[ t ].outEvents.size( ) != 0 )
{
std::cout << std::setw( tableFourth ) << "OutEvents:" << std::setw( tableThird );
for( size_t i = 0; i < mean[ t ].outEvents.size( ); ++i )
{
std::cout << mean[ t ].outEvents[ i ]( );
if( i < ( mean[ t ].outEvents.size( ) - 1 ) )
{
std::cout << "," << std::endl;
std::cout << std::setw( tableFourth + tableThird );
}
}
std::cout << std::endl;
}
std::cout << std::setw( tableFourth ) << unit << ":"
<< std::setw( 2 * tableFourth ) << gFlops << std::endl;
std::cout << std::setw( tableFourth ) << "Time (ns):"
<< std::setw( 3 * tableFourth ) << commatize( static_cast<cl_ulong>( time ) ) << std::endl;
std::cout << std::endl;
}
}
}
segtable* new_segment_table
(u32 size,
unspos coverageLimit)
{
segtable* st;
size_t bytesNeeded;
// sanity check
if (size < 1)
suicidef ("in new_segment_table(), size can't be %d", size);
// allocate
bytesNeeded = segtable_bytes (size);
if (bytesNeeded > mallocLimit) goto overflow;
st = (segtable*) malloc_or_die ("new_segment_table", bytesNeeded);
// initialize
st->size = size;
st->len = 0;
st->haveScores = false;
st->coverageLimit = coverageLimit;
st->coverage = 0;
st->lowScore = worstPossibleScore;
return st;
// failure exits
overflow:
suicidef ("internal error, in new_segment_table()\n"
"table size (%s) exceeds allocation limit of %s;",
commatize(bytesNeeded),
commatize(mallocLimit));
return NULL; // (doesn't get here)
}
void* realloc_or_die
(char* id,
void* _p,
size_t size)
{
void* p;
// make sure size is legit
if (size > mallocLimit)
{
if (id == NULL)
suicidef ("realloc_or_die blocked large request, for %s bytes (max is %s)",
commatize(size), commatize(mallocLimit));
else
suicidef ("realloc_or_die blocked large request, for %s bytes (max is %s), for %s",
commatize(size), commatize(mallocLimit), id);
}
if (size == 0) size = 1;
// allocate the memory
p = realloc (_p, size);
if (p == NULL)
{
if (id == NULL)
suicidef ("call to realloc failed to allocate %lu bytes",
commatize(size));
else
suicidef ("call to realloc failed to allocate %lu bytes, for %s",
commatize(size), id);
}
reportRealloc (id, _p, p, size);
return p;
}
void
GpuStatTimer::Print( )
{
const int tableWidth = 60;
const int tableHalf = tableWidth / 2;
const int tableThird = tableWidth / 3;
const int tableFourth = tableWidth / 4;
const int tableFifth = tableWidth / 5;
for( cl_uint id = 0; id < labelID.size( ); ++id )
{
size_t halfString = labelID[ id ].first.size( ) / 2;
// Print label of timer, in a header
std::cout << std::endl << std::setw( tableHalf + halfString ) << std::setfill( '=' ) << labelID[ id ].first
<< std::setw( tableHalf - halfString ) << "=" << std::endl;
tout << std::setfill( _T( ' ' ) );
std::vector< StatData > mean = getMean( id );
// Print each individual dimension
tstringstream catLengths;
for( cl_uint t = 0; t < mean.size( ); ++t )
{
cl_double time = 0;
if( mean[ t ].kernel == NULL )
{
for( cl_uint m = 0; m < t; ++m )
{
if( mean[ m ].plHandle == mean[ t ].planX ||
mean[ m ].plHandle == mean[ t ].planY ||
mean[ m ].plHandle == mean[ t ].planZ ||
mean[ m ].plHandle == mean[ t ].planTX ||
mean[ m ].plHandle == mean[ t ].planTY ||
mean[ m ].plHandle == mean[ t ].planTZ ||
mean[ m ].plHandle == mean[ t ].planRCcopy ||
mean[ m ].plHandle == mean[ t ].planCopy )
{
time += mean[ m ].doubleNanoSec;
}
}
mean[ t ].doubleNanoSec = time;
}
else
{
time = mean[ t ].doubleNanoSec;
}
double gFlops = mean[ t ].calcFlops( ) / time;
tout << std::setw( tableFourth ) << _T( "Handle:" )
<< std::setw( tableThird ) << mean[ t ].plHandle << std::endl;
if( mean[ t ].kernel != 0 )
{
tout << std::setw( tableFourth ) << _T( "Kernel:" )
<< std::setw( tableThird ) << reinterpret_cast<void*>( mean[ t ].kernel ) << std::endl;
}
if( ( mean[ t ].planX + mean[ t ].planY + mean[ t ].planZ ) > 0 ||
( mean[ t ].planTX + mean[ t ].planTY + mean[ t ].planTZ ) > 0 ||
( mean[ t ].planRCcopy + mean[ t ].planCopy ) > 0 )
{
tout << std::setw( tableFourth ) << _T( "Child Handles:" );
catLengths.str( _T( "" ) );
catLengths << _T( "(" );
if( mean[ t ].planX != 0 )
catLengths << mean[ t ].planX;
if( mean[ t ].planTX != 0 )
{
catLengths << _T( "," );
catLengths << mean[ t ].planTX;
}
if( mean[ t ].planY != 0 )
{
catLengths << _T( "," );
catLengths << mean[ t ].planY;
}
if( mean[ t ].planTY != 0 )
{
catLengths << _T( "," );
catLengths << mean[ t ].planTY;
}
if( mean[ t ].planZ != 0 )
{
catLengths << _T( "," );
catLengths << mean[ t ].planZ;
}
if( mean[ t ].planTZ != 0 )
{
catLengths << _T( "," );
catLengths << mean[ t ].planTZ;
}
if( mean[ t ].planRCcopy != 0 )
{
catLengths << _T( "," );
catLengths << mean[ t ].planRCcopy;
}
if( mean[ t ].planCopy != 0 )
{
catLengths << _T( "," );
catLengths << mean[ t ].planCopy;
}
catLengths << _T( ")" );
tout << std::setw( tableThird ) << catLengths.str( ) << std::endl;
}
if( mean[ t ].outEvents.size( ) != 0 )
{
tout << std::setw( tableFourth ) << _T( "OutEvents:" ) << std::setw( tableThird );
for( size_t i = 0; i < mean[ t ].outEvents.size( ); ++i )
{
tout << mean[ t ].outEvents[ i ];
if( i < (mean[ t ].outEvents.size( )-1) )
{
tout << _T( "," ) << std::endl;
tout << std::setw( tableFourth+tableThird );
}
}
tout << std::endl;
}
tout << std::setw( tableFourth ) << _T( "Length:" );
catLengths.str( _T( "" ) );
catLengths << _T( "(" );
for( size_t i = 0; i < mean[ t ].lengths.size( ); ++i )
{
catLengths << mean[ t ].lengths.at( i );
if( i < (mean[ t ].lengths.size( )-1) )
catLengths << _T( "," );
}
catLengths << _T( ")" );
tout << std::setw( tableThird ) << catLengths.str( ) << std::endl;
if( mean[ t ].batchSize > 1 )
{
tout << std::setw( tableFourth ) << _T( "Batch:" )
<< std::setw( tableThird ) << mean[ t ].batchSize << std::endl;
}
tout << std::setw( tableFourth ) << _T( "Input Stride:" );
catLengths.str( _T( "" ) );
catLengths << _T( "(" );
for( size_t i = 0; i < mean[ t ].inStride.size( ); ++i )
{
catLengths << mean[ t ].inStride.at( i );
if( i < (mean[ t ].inStride.size( )-1) )
catLengths << _T( "," );
}
catLengths << _T( ")" );
tout << std::setw( tableThird ) << catLengths.str( ) << std::endl;
tout << std::setw( tableFourth ) << _T( "Output Stride:" );
catLengths.str( _T( "" ) );
catLengths << _T( "(" );
for( size_t i = 0; i < mean[ t ].outStride.size( ); ++i )
{
catLengths << mean[ t ].outStride.at( i );
if( i < (mean[ t ].outStride.size( )-1) )
catLengths << _T( "," );
}
catLengths << _T( ")" );
tout << std::setw( tableThird ) << catLengths.str( ) << std::endl;
if( mean[ t ].enqueueWorkSize.size( ) != 0 )
{
tout << std::setw( tableFourth ) << _T( "Global Work:" );
catLengths.str( _T( "" ) );
catLengths << _T( "(" );
for( size_t i = 0; i < mean[ t ].enqueueWorkSize.size( ); ++i )
{
catLengths << mean[ t ].enqueueWorkSize.at( i );
if( i < (mean[ t ].enqueueWorkSize.size( )-1) )
catLengths << _T( "," );
}
catLengths << _T( ")" );
tout << std::setw( tableThird ) << catLengths.str( ) << std::endl;
}
tout << std::setw( tableFourth ) << _T( "Gflops:" )
<< std::setw( 2*tableFourth ) << gFlops << std::endl;
tout << std::setw( tableFourth ) << _T( "Time (ns):" )
<< std::setw( 3*tableFourth ) << commatize( static_cast< cl_ulong >( time ) ) << std::endl;
tout << std::endl;
}
}
}
segtable* add_segment
(segtable* st,
unspos pos1,
unspos pos2,
unspos length,
score s,
int id)
{
u32 newSize;
size_t bytesNeeded;
segment* seg, *parent;
segment tempSeg;
int ix, pIx;
int tied, stopped;
// fprintf (stderr, "add " unsposSlashSFmt " " unsposFmt " " scoreFmtSimple "; id %d\n",
// pos1+1, "+",
// pos2+1, ((id & rcf_rev) != 0)? "-" : "+",
// length, s, id);
//////////
// add the segment to the table, enlarging the table if needed, but
// discarding the segment if it is low-scoring and the table has met its
// coverage limit
//////////
// if the table is already full and this segment scores less than the
// lowest score in the table, discard it
if ((st->len > 0)
&& (st->coverageLimit != 0)
&& (st->coverage >= st->coverageLimit)
&& (s < st->lowScore))
return st;
// if there's no room for the new segment, re-allocate
if (st->len >= st->size)
{
newSize = st->size + 100 + (st->size / 3);
bytesNeeded = segtable_bytes (newSize);
if (bytesNeeded > mallocLimit) goto overflow;
st = (segtable*) realloc_or_die ("add_segment", st, bytesNeeded);
st->size = newSize;
}
// add the segment, by appending it at the end
seg = &st->seg[st->len++];
seg->pos1 = pos1;
seg->pos2 = pos2;
seg->length = length;
seg->s = s;
seg->id = id;
seg->filter = false;
seg->scoreCov = (possum) length;
st->coverage += length;
if ((st->len == 1) || (s < st->lowScore)) st->lowScore = s;
//////////
// handle the transition between the two table states
// below-the-coverage-limit: table is kept as a simple list
// met-the-coverage-limit: table is kept as a proper min-heap
//////////
// if this segment leaves us below the limit, we're done
if ((st->coverageLimit == 0)
|| (st->coverage < st->coverageLimit))
return st;
// if this is the first time we've reached the limit, sort the segments to
// create a proper min-heap, and add the tied-score information
// nota bene: if we reach here, st->coverageLimit > 0 and
// st->coverage >= st->coverageLimit
if (st->coverage - length < st->coverageLimit)
{
sort_segments (st, qSegmentsByIncreasingScore);
record_tie_scores (st);
#ifdef debugBinaryHeap
fprintf (stderr, "\nafter sort:\n");
dump_segments (stderr, st, NULL, NULL);
validate_heap (st, "after sort");
#endif // debugBinaryHeap
goto prune;
}
//////////
// maintain the min-heap property
//////////
#ifdef debugBinaryHeap
//fprintf (stderr, "\nbefore percolation:\n");
//dump_segments (stderr, st, NULL, NULL);
#endif // debugBinaryHeap
// the rest of the list is a proper min-heap, so percolate the new segment
// up the tree, while maintaining the tied-score information
// nota bene: if we reach here, length >= 2
tied = false;
for (ix=st->len-1 ; ix>0 ; )
{
pIx = (ix-1) / 2;
seg = &st->seg[ix];
parent = &st->seg[pIx];
if (seg->s >= parent->s)
{ tied = (seg->s == parent->s); break; }
// swap this segment with its parent, and adjust old parent's tied-score
// subheap
tempSeg = *seg; *seg = *parent; *parent = tempSeg;
record_tie_score (st, ix);
ix = pIx;
}
record_tie_score (st, ix);
// if the new segment tied an existing score, we must continue to percolate
// the tied-score info up the tree
if (tied)
{
stopped = false;
for (ix=(ix-1)/2 ; ix>0 ; ix=(ix-1)/2)
{
if (!record_tie_score (st, ix))
{ stopped = true; break; }
}
if (!stopped) record_tie_score (st, 0);
}
#ifdef debugBinaryHeap
fprintf (stderr, "\nafter percolation:\n");
dump_segments (stderr, st, NULL, NULL);
validate_heap (st, "after percolation");
#endif // debugBinaryHeap
//////////
// remove low-scoring segments
//////////
prune:
// if removing the minimum scoring subheap would bring us below the
// limit, no pruning is necessary
if (st->coverage - st->seg[0].scoreCov < st->coverageLimit)
return st;
// otherwise, we must remove subheaps as long as doing so leaves us at or
// above the limit
while (st->coverage - st->seg[0].scoreCov >= st->coverageLimit)
{
s = st->seg[0].s;
while (st->seg[0].s == s)
{
remove_root (st);
#ifdef debugBinaryHeap
fprintf (stderr, "\nafter a pruning:\n");
dump_segments (stderr, st, NULL, NULL);
validate_heap (st, "after pruning");
#endif // debugBinaryHeap
}
}
st->lowScore = st->seg[0].s;
#ifdef debugBinaryHeap
fprintf (stderr, "\nafter pruning:\n");
dump_segments (stderr, st, NULL, NULL);
validate_heap (st, "after pruning");
#endif // debugBinaryHeap
return st;
// failure exits
#define suggestions " consider using lastz_m40," \
" or setting max_malloc_index for a special build," \
" or raising scoring threshold (--hspthresh or --exact)," \
" or break your target sequence into smaller pieces"
overflow:
suicidef ("in add_segment()\n"
"table size (%s for %s segments) exceeds allocation limit of %s;\n"
suggestions,
commatize(bytesNeeded),
commatize(newSize),
commatize(mallocLimit));
return NULL; // (doesn't get here)
}