void genRefs()
{
gen( TPointX, TPoint, x );
gen( TPointY, TPoint, y );
genx( TViewSizeX, TView, size );
geny( TViewSizeY, TView, size );
gen( TViewState, TView, state );
gen( TViewOwner, TView, owner );
geny( TViewOriginY, TView, origin );
genx( TViewOriginX, TView, origin );
geny( TViewCursorY, TView, cursor );
genx( TViewCursorX, TView, cursor );
gen( TViewNext, TView, next );
gen( TViewOptions, TView, options );
genay( TGroupClipAY, TGroup, clip );
genax( TGroupClipAX, TGroup, clip );
genby( TGroupClipBY, TGroup, clip );
genbx( TGroupClipBX, TGroup, clip );
gen( TGroupLast, TGroup, last );
gen( TGroupBuffer, TGroup, buffer );
gen( TGroupLockFlag, TGroup, lockFlag );
genx( MsEventWhereX, MouseEventType, where );
geny( MsEventWhereY, MouseEventType, where );
genx( TFrameSizeX, TFrame, size );
gen( TFrameOwner, TFrame, owner );
gen( TDrawBufferData, TDrawBuffer, data );
gen( TEditorCurPtr, TEditor, curPtr );
gen( TEditorGapLen, TEditor, gapLen );
gen( TEditorBuffer, TEditor, buffer );
gen( TEditorSelStart, TEditor, selStart );
gen( TEditorSelEnd, TEditor, selEnd );
gen( TEditorBufSize, TEditor, bufSize );
gen( TEditorBufLen, TEditor, bufLen );
gen( TTerminalBuffer, TTerminal, buffer );
gen( TTerminalBufSize, TTerminal, bufSize );
gen( TTerminalQueBack, TTerminal, queBack );
genConst( sfVisible );
genConst( sfCursorVis );
genConst( sfCursorIns );
genConst( sfFocused );
genConst( sfShadow );
genConst( sfExposed );
genConst( ofFramed );
}
void time_hog( const std::vector<carp::record_t>& pool, const std::vector<float>& sizes, int num_positions, int repeat )
{
carp::Timing timing("HOG");
for (;repeat>0; --repeat) {
for ( auto & size : sizes ) {
for ( auto & item : pool ) {
std::mt19937 rng(0); //uses same seed, reseed for all iteration
cv::Mat cpu_gray;
cv::cvtColor( item.cpuimg(), cpu_gray, CV_RGB2GRAY );
cv::Mat_<float> locations(num_positions, 2);
cv::Mat_<float> blocksizes(num_positions, 2);
size_t max_blocksize_x = std::ceil(size);
size_t max_blocksize_y = std::ceil(size);
//fill locations and blocksizes
std::uniform_real_distribution<float> genx(size/2+1, cpu_gray.rows-1-size/2-1);
std::uniform_real_distribution<float> geny(size/2+1, cpu_gray.cols-1-size/2-1);
for( int i = 0; i < num_positions; ++i) {
locations(i, 0) = genx(rng);
locations(i, 1) = geny(rng);
blocksizes(i, 0) = size;
blocksizes(i, 1) = size;
}
const int HISTOGRAM_BINS = NUMBER_OF_CELLS * NUMBER_OF_CELLS * NUMBER_OF_BINS;
std::vector<float> cpu_result(num_positions * HISTOGRAM_BINS), gpu_result(num_positions * HISTOGRAM_BINS), pen_result(num_positions * HISTOGRAM_BINS);
std::chrono::duration<double> elapsed_time_cpu, elapsed_time_gpu_p_copy, elapsed_time_gpu_nocopy, elapsed_time_pencil;
{
//CPU implement
static nel::HOGDescriptorCPP descriptor( NUMBER_OF_CELLS
, NUMBER_OF_BINS
, GAUSSIAN_WEIGHTS
, SPARTIAL_WEIGHTS
, SIGNED_HOG
);
const auto cpu_start = std::chrono::high_resolution_clock::now();
const auto result = descriptor.compute(cpu_gray, locations, blocksizes);
const auto cpu_end = std::chrono::high_resolution_clock::now();
std::copy(result.begin(), result.end(), cpu_result.begin());
elapsed_time_cpu = cpu_end - cpu_start;
//Free up resources
}
{
//GPU implement
static nel::HOGDescriptorOCL descriptor( NUMBER_OF_CELLS
, NUMBER_OF_BINS
, GAUSSIAN_WEIGHTS
, SPARTIAL_WEIGHTS
, SIGNED_HOG
);
const auto gpu_start = std::chrono::high_resolution_clock::now();
const auto result = descriptor.compute(cpu_gray, locations, blocksizes, max_blocksize_x, max_blocksize_y, elapsed_time_gpu_nocopy);
const auto gpu_end = std::chrono::high_resolution_clock::now();
std::copy(result.begin(), result.end(), gpu_result.begin());
elapsed_time_gpu_p_copy = gpu_end - gpu_start;
//Free up resources
}
{
pen_result.resize(num_positions * HISTOGRAM_BINS, 0.0f);
const auto pencil_start = std::chrono::high_resolution_clock::now();
pencil_hog( NUMBER_OF_CELLS, NUMBER_OF_BINS, GAUSSIAN_WEIGHTS, SPARTIAL_WEIGHTS, SIGNED_HOG
, cpu_gray.rows, cpu_gray.cols, cpu_gray.step1(), cpu_gray.ptr<uint8_t>()
, num_positions
, reinterpret_cast<const float (*)[2]>(locations.data)
, reinterpret_cast<const float (*)[2]>(blocksizes.data)
, pen_result.data()
);
const auto pencil_end = std::chrono::high_resolution_clock::now();
elapsed_time_pencil = pencil_end - pencil_start;
//Free up resources
}
// Verifying the results
if ( cv::norm( cpu_result, gpu_result, cv::NORM_INF) > cv::norm( gpu_result, cv::NORM_INF)*1e-5
|| cv::norm( cpu_result, pen_result, cv::NORM_INF) > cv::norm( cpu_result, cv::NORM_INF)*1e-5
)
{
std::vector<float> diff;
std::transform(cpu_result.begin(), cpu_result.end(), gpu_result.begin(), std::back_inserter(diff), std::minus<float>());
std::cerr << "ERROR: Results don't match." << std::endl;
std::cerr << "CPU norm:" << cv::norm(cpu_result, cv::NORM_INF) << std::endl;
std::cerr << "GPU norm:" << cv::norm(gpu_result, cv::NORM_INF) << std::endl;
std::cerr << "PEN norm:" << cv::norm(pen_result, cv::NORM_INF) << std::endl;
std::cerr << "GPU-CPU norm:" << cv::norm(gpu_result, cpu_result, cv::NORM_INF) << std::endl;
std::cerr << "PEN-CPU norm:" << cv::norm(pen_result, cpu_result, cv::NORM_INF) << std::endl;
throw std::runtime_error("The OpenCL or PENCIL results are not equivalent with the C++ results.");
}
timing.print( elapsed_time_cpu, elapsed_time_gpu_p_copy, elapsed_time_gpu_nocopy, elapsed_time_pencil );
}
}
}
}