generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

StereoSGBM.cpp - use SSE2 for pass 2 using MODE_HH

Browse Source 
With a test image set of 2800x1400 bytes on a Intel Core i7 5960X this improves runtime of MODE_HH with about 10%. (this particular replaced code segment is approx 3 times faster than the non-SSE2 variant). I was able to reduce runtime by 130 ms by this simple fix.

The second part of the SSE2 optimized part could probably be optimized further by using shift SSE2 operations, but I imagine this would improve performance 10-20 ms at best.
This commit is contained in:
Kai Hugo Hustoft Endresen 2016-01-08 00:32:52 +01:00
parent 345678770b
commit 551b5d3e1a
1 changed files with 43 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

50
modules/calib3d/src/stereosgbm.cpp
View File

@ -759,14 +759,50 @@ static void computeDisparitySGBM( const Mat& img1, const Mat& img2,
}
else
{
for( d = 0; d < D; d++ )
#if CV_SSE2
if( useSIMD )
{
int Sval = Sp[d];
if( Sval < minS )
{
minS = Sval;
bestDisp = d;
}
__m128i _minS = _mm_set1_epi16(MAX_COST), _bestDisp = _mm_set1_epi16(-1);
__m128i _d8 = _mm_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7), _8 = _mm_set1_epi16(8);
for( d = 0; d < D; d+= 8 )
{
__m128i L0 = _mm_load_si128((const __m128i*)( Sp + d ));
__m128i mask = _mm_cmplt_epi16( L0, _minS );
_minS = _mm_min_epi16( L0, _minS );
_bestDisp = _mm_xor_si128(_bestDisp, _mm_and_si128(_mm_xor_si128( _bestDisp, _d8), mask));
_d8 = _mm_adds_epi16(_d8, _8 );
}
short CV_DECL_ALIGNED(16) bestDispBuf[8];
_mm_store_si128((__m128i*)bestDispBuf, _bestDisp);
short CV_DECL_ALIGNED(16) minSBuf[8];
_mm_store_si128((__m128i*)minSBuf, _minS );
for( int i = 0; i < 8; i++ )
{
int Sval = minSBuf[ i ];
if( Sval <= minS )
{
if( ( Sval < minS ) || ( bestDispBuf[i] < bestDisp ) )
{
bestDisp = bestDispBuf[i];
}
minS = Sval;
}
}
}
else
#endif
{
for( d = 0; d < D; d++ )
{
int Sval = Sp[d];
if( Sval < minS )
{
minS = Sval;
bestDisp = d;
}
}
}
}