Skip to content

Commit

Permalink
Fix SIMD unit tests for transforms
Browse files Browse the repository at this point in the history
  • Loading branch information
@adamjw24
adamjw24 committed Nov 18, 2024
1 parent cb94be7 commit b442364
Showing 1 changed file with 79 additions and 38 deletions.
117 changes: 79 additions & 38 deletions test/vvenc_unit_test/vvenc_unit_test.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -97,6 +97,23 @@ class InputGenerator
unsigned m_bits;
};

template<typename T>
class TrafoGenerator
{
public:
explicit TrafoGenerator( unsigned bits ) : m_bits( bits )
{
}

T operator()() const
{
return ( rand() & ( ( 1 << m_bits ) - 1 ) ) - ( 1 << m_bits >> 1 );
}

private:
unsigned m_bits;
};

class DimensionGenerator
{
public:
Expand All @@ -108,78 +125,101 @@ class DimensionGenerator
}
};

template<typename G>
static bool check_one_fastInvCore( TCoeffOps* ref, TCoeffOps* opt, int idx, int trSize, unsigned lines,
unsigned reducedLines, unsigned rows, G input_generator )
template<typename G, typename T>
static bool check_one_fastInvCore( TCoeffOps* ref, TCoeffOps* opt, unsigned idx, unsigned trSize, unsigned lines,
unsigned reducedLines, unsigned cutoff, G input_generator, T trafo_generator )
{
CHECK( lines == 0, "Lines must be non-zero" );
CHECK( reducedLines > lines, "ReducedLines must be less than or equal to lines" );
CHECK( rows == 0, "Rows must be non-zero" );
CHECK( rows % 4, "Rows must be a multiple of four" );
CHECK( cutoff == 0, "Cutoff must be non-zero" );
CHECK( cutoff % 4, "Cutoff must be a multiple of four" );
CHECK( cutoff > trSize, "Cutoff must not be larger than transformation size" );

std::ostringstream sstm;
sstm << "fastInvCore trSize=" << trSize << " lines=" << lines << " reducedLines=" << reducedLines << " rows=" << rows;
sstm << "fastInvCore trSize=" << trSize << " lines=" << lines << " reducedLines=" << reducedLines << " cutoff=" << cutoff;

std::vector<TMatrixCoeff> it( rows * trSize );
std::vector<TCoeff> src( rows * lines );
std::vector<TCoeff> dst0( lines * trSize );
std::vector<TCoeff> dst1 = dst0;
TMatrixCoeff *it = ( TMatrixCoeff* ) xMalloc( TMatrixCoeff, trSize * trSize );
TCoeff *src = ( TCoeff* ) xMalloc( TCoeff, trSize * lines );
TCoeff *dst0 = ( TCoeff* ) xMalloc( TCoeff, trSize * lines );
TCoeff *dst1 = ( TCoeff* ) xMalloc( TCoeff, trSize * lines );

// Initialize source buffers.
std::generate( it.begin(), it.end(), input_generator );
std::generate( src.begin(), src.end(), input_generator );
std::generate( it, it + trSize * trSize, trafo_generator );
std::generate( src, src + trSize * lines, input_generator );
memset( dst0, 0, trSize * lines * sizeof( TCoeff ) );
memset( dst1, 0, trSize * lines * sizeof( TCoeff ) );

ref->fastInvCore[ idx ]( it, src, dst0, lines, reducedLines, cutoff );
opt->fastInvCore[ idx ]( it, src, dst1, lines, reducedLines, cutoff );

ref->fastInvCore[ idx ]( it.data(), src.data(), dst0.data(), lines, reducedLines, rows );
opt->fastInvCore[ idx ]( it.data(), src.data(), dst1.data(), lines, reducedLines, rows );
return compare_values_2d( sstm.str(), dst0.data(), dst1.data(), reducedLines, trSize );
auto ret = compare_values_2d( sstm.str(), dst0, dst1, trSize, lines );

xFree( it );
xFree( src );
xFree( dst0 );
xFree( dst1 );

return ret;
}

template<typename G>
static bool check_one_fastFwdCore_2D( TCoeffOps* ref, TCoeffOps* opt, int idx, int trSize, unsigned line,
unsigned reducedLine, unsigned cutoff, unsigned shift, G input_generator )
template<typename G, typename T>
static bool check_one_fastFwdCore_2D( TCoeffOps* ref, TCoeffOps* opt, unsigned idx, unsigned trSize, unsigned line,
unsigned reducedLine, unsigned cutoff, unsigned shift, G input_generator, T trafo_generator )
{
CHECK( line == 0, "Line must be non-zero" );
CHECK( reducedLine > line, "ReducedLine must be less than or equal to line" );
CHECK( cutoff == 0, "Cutoff must be non-zero" );
CHECK( cutoff % 4, "Cutoff must be a multiple of four" );
CHECK( cutoff > trSize, "Cutoff must not be larger than transformation size" );
CHECK( shift == 0, "Shift must be at least one" );

std::ostringstream sstm;
sstm << "fastFwdCore_2D trSize=" << trSize << " line=" << line << " reducedLine=" << reducedLine
<< " cutoff=" << cutoff << " shift=" << shift;

std::vector<TMatrixCoeff> tc( trSize * cutoff );
std::vector<TCoeff> src( trSize * reducedLine );
std::vector<TCoeff> dst0( line * cutoff );
std::vector<TCoeff> dst1 = dst0;
TMatrixCoeff *tc = ( TMatrixCoeff* ) xMalloc( TMatrixCoeff, trSize * trSize );
TCoeff *src = ( TCoeff* ) xMalloc( TCoeff, trSize * line );
TCoeff *dst0 = ( TCoeff* ) xMalloc( TCoeff, trSize * line );
TCoeff *dst1 = ( TCoeff* ) xMalloc( TCoeff, trSize * line );

// Initialize source and destination buffers, make sure that destination
// buffers match in elements that are not written to by the kernel being
// tested.
std::generate( tc.begin(), tc.end(), input_generator );
std::generate( src.begin(), src.end(), input_generator );
std::generate( tc, tc + trSize * trSize, trafo_generator );
std::generate( src, src + trSize * line, input_generator );
memset( dst0, 0, trSize * line * sizeof( TCoeff ) );
memset( dst1, 0, trSize * line * sizeof( TCoeff ) );

ref->fastFwdCore_2D[ idx ]( tc, src, dst0, line, reducedLine, cutoff, shift );
opt->fastFwdCore_2D[ idx ]( tc, src, dst1, line, reducedLine, cutoff, shift );

ref->fastFwdCore_2D[ idx ]( tc.data(), src.data(), dst0.data(), line, reducedLine, cutoff, shift );
opt->fastFwdCore_2D[ idx ]( tc.data(), src.data(), dst1.data(), line, reducedLine, cutoff, shift );
// Don't check for over-writes past reducedLine columns here, since the
// existing x86 implementations would fail.
return compare_values_2d( sstm.str(), dst0.data(), dst1.data(), cutoff, reducedLine, line );
auto ret = compare_values_2d( sstm.str(), dst0, dst1, trSize, reducedLine, line );

xFree( tc );
xFree( src );
xFree( dst0 );
xFree( dst1 );

return ret;
}

static bool check_fastInvCore( TCoeffOps* ref, TCoeffOps* opt, unsigned num_cases, int idx, int trSize )
static bool check_fastInvCore( TCoeffOps* ref, TCoeffOps* opt, unsigned num_cases, unsigned idx, unsigned trSize )
{
printf( "Testing TCoeffOps::fastInvCore trSize=%d\n", trSize );
InputGenerator<TCoeff> g{ 10 };
TrafoGenerator<TMatrixCoeff> t{ 8 };
DimensionGenerator rng;

for( unsigned i = 0; i < num_cases; ++i )
{
// Clamp lines down to the next multiple of four when generating
// reducedLines to avoid existing x86 implementations over-writing.
unsigned lines = rng.get( 4, 1024 );
unsigned reducedLines = rng.get( 4, lines & ~3U );
unsigned rows = rng.get( 4, 128, 4 ); // Rows must be a non-zero multiple of four.
if( !check_one_fastInvCore( ref, opt, idx, trSize, lines, reducedLines, rows, g ) )
unsigned lines = 1 << rng.get( 2, 6 );
unsigned reducedLines = std::min( 32u, rng.get( 4u, lines, 4 ) );
unsigned cutoff = rng.get( 4, trSize, 4 ); // Cutoff must be a non-zero multiple of four.
if( !check_one_fastInvCore( ref, opt, idx, trSize, lines, reducedLines, cutoff, g, t ) )
{
return false;
}
Expand All @@ -188,21 +228,22 @@ static bool check_fastInvCore( TCoeffOps* ref, TCoeffOps* opt, unsigned num_case
return true;
}

static bool check_fastFwdCore_2D( TCoeffOps* ref, TCoeffOps* opt, unsigned num_cases, int idx, int trSize )
static bool check_fastFwdCore_2D( TCoeffOps* ref, TCoeffOps* opt, unsigned num_cases, unsigned idx, unsigned trSize )
{
printf( "Testing TCoeffOps::fastFwdCore_2D trSize=%d\n", trSize );
InputGenerator<TCoeff> g{ 10 };
TrafoGenerator<TMatrixCoeff> t{ 8 };
DimensionGenerator rng;

for( unsigned i = 0; i < num_cases; ++i )
{
// Clamp line down to the next multiple of four when generating reducedLine
// to avoid existing x86 implementations over-writing.
unsigned line = rng.get( 1, 1024 );
unsigned reducedLine = rng.get( 0, line & ~3U );
unsigned cutoff = rng.get( 4, 1024, 4 ); // Cutoff must be a non-zero multiple of four.
unsigned shift = rng.get( 1, 16 ); // Shift must be at least one to avoid UB.
if( !check_one_fastFwdCore_2D( ref, opt, idx, trSize, line, reducedLine, cutoff, shift, g ) )
unsigned line = 1 << rng.get( 2, 6 );
unsigned reducedLine = std::min( 32u, rng.get( 4u, line, 4 ) );
unsigned cutoff = rng.get( 4, trSize, 4 ); // Cutoff must be a non-zero multiple of four.
unsigned shift = rng.get( 1, 16 ); // Shift must be at least one to avoid UB.
if( !check_one_fastFwdCore_2D( ref, opt, idx, trSize, line, reducedLine, cutoff, shift, g, t ) )
{
return false;
}
Expand Down

0 comments on commit b442364

Please sign in to comment.