Skip to content

Commit 238dfef

Browse files
krzysgkrzysg
committed
Little cleanup
1 parent cde5865 commit 238dfef

File tree

3 files changed

+58
-43
lines changed

3 files changed

+58
-43
lines changed

‎src/algorithm/APRConverter.hpp‎

Lines changed: 58 additions & 37 deletions
Original file line numberDiff line numberDiff line change
@@ -230,9 +230,34 @@ inline bool APRConverter<ImageType>::get_apr_method(APR<ImageType> &aAPR, PixelD
230230
method_timer.start_timer("compute_levels");
231231
computeLevels(grad_temp, local_scale_temp, (*apr).level_max(), par.rel_error, par.dx, par.dy, par.dz);
232232
method_timer.stop_timer();
233+
234+
method_timer.start_timer("initialize_particle_cell_tree");
235+
iPullingScheme.initialize_particle_cell_tree(aAPR.apr_access);
236+
method_timer.stop_timer();
237+
238+
method_timer.start_timer("compute_local_particle_set");
239+
get_local_particle_cell_set(local_scale_temp, local_scale_temp2);
240+
method_timer.stop_timer();
241+
242+
method_timer.start_timer("compute_pulling_scheme");
243+
iPullingScheme.pulling_scheme_main();
244+
method_timer.stop_timer();
245+
246+
method_timer.start_timer("downsample_pyramid");
247+
std::vector<PixelData<T>> downsampled_img;
248+
//Down-sample the image for particle intensity estimation
249+
downsamplePyrmaid(input_image, downsampled_img, aAPR.level_max(), aAPR.level_min());
250+
method_timer.stop_timer();
251+
252+
method_timer.start_timer("compute_apr_datastructure");
253+
aAPR.apr_access.initialize_structure_from_particle_cell_tree(aAPR.parameters, iPullingScheme.getParticleCellTree());
254+
method_timer.stop_timer();
255+
256+
method_timer.start_timer("sample_particles");
257+
aAPR.get_parts_from_img(downsampled_img,aAPR.particles_intensities);
258+
method_timer.stop_timer();
233259
#else
234260
method_timer.start_timer("compute_gradient_magnitude_using_bsplines and local instensity scale CUDA");
235-
// getFullPipeline(image_temp, grad_temp, local_scale_temp, local_scale_temp2, bspline_offset, par, (*apr).level_max());
236261
APRTimer t(true);
237262
APRTimer d(true);
238263
t.start_timer(" =========== ALL");
@@ -241,35 +266,57 @@ inline bool APRConverter<ImageType>::get_apr_method(APR<ImageType> &aAPR, PixelD
241266

242267
std::vector<GpuProcessingTask<ImageType>> gpts;
243268

244-
int n = 3;
245-
int rep = 5;
246-
for (int i = 0; i < n; ++i) {
269+
int numOfStreams = 1;
270+
int repetitionsPerStream = 1;
271+
272+
// Create streams and send initial task to do
273+
for (int i = 0; i < numOfStreams; ++i) {
247274
gpts.emplace_back(GpuProcessingTask<ImageType>(image_temp, local_scale_temp, par, bspline_offset, (*apr).level_max()));
248275
gpts.back().sendDataToGpu();
249276
gpts.back().processOnGpu();
250277
}
251278

252-
for (int i = 0; i < n * rep; ++i) {
253-
int c = i % n;
279+
for (int i = 0; i < numOfStreams * repetitionsPerStream; ++i) {
280+
int c = i % numOfStreams;
281+
282+
// get data from previous task
254283
gpts[c].getDataFromGpu();
255284

256285
// in theory we get new data and send them to task
257-
if (i < n * (rep - 1)) { gpts[c].sendDataToGpu();
258-
gpts[c].processOnGpu();
259-
}
260-
std::cout << "--------- start CPU processing ---------- " << i << std::endl;
261-
d.start_timer("CPU processing");
286+
if (i < numOfStreams * (repetitionsPerStream - 1)) {
287+
gpts[c].sendDataToGpu();
288+
gpts[c].processOnGpu();
289+
}
290+
291+
// Postprocess on CPU
292+
std::cout << "--------- start CPU processing ---------- " << i << std::endl;
262293
init_apr(aAPR, input_image);
294+
d.start_timer("1");
263295
iPullingScheme.initialize_particle_cell_tree(aAPR.apr_access);
296+
d.stop_timer();
297+
d.start_timer("2");
264298
PixelData<float> lst(local_scale_temp, true);
299+
d.stop_timer();
300+
d.start_timer("3");
265301
get_local_particle_cell_set(lst, local_scale_temp2);
302+
d.stop_timer();
303+
d.start_timer("4");
266304
iPullingScheme.pulling_scheme_main();
305+
d.stop_timer();
306+
d.start_timer("5");
267307
PixelData<T> inImg(input_image, true);
308+
d.stop_timer();
309+
d.start_timer("6");
268310
std::vector<PixelData<T>> downsampled_img;
269311
downsamplePyrmaid(inImg, downsampled_img, aAPR.level_max(), aAPR.level_min());
312+
d.stop_timer();
313+
d.start_timer("7");
270314
aAPR.apr_access.initialize_structure_from_particle_cell_tree(aAPR.parameters, iPullingScheme.getParticleCellTree());
315+
d.stop_timer();
316+
d.start_timer("8");
271317
aAPR.get_parts_from_img(downsampled_img, aAPR.particles_intensities);
272318
d.stop_timer();
319+
273320
}
274321
std::cout << "Total n ENDED" << std::endl;
275322

@@ -278,32 +325,6 @@ inline bool APRConverter<ImageType>::get_apr_method(APR<ImageType> &aAPR, PixelD
278325
method_timer.stop_timer();
279326
#endif
280327

281-
// method_timer.start_timer("initialize_particle_cell_tree");
282-
// iPullingScheme.initialize_particle_cell_tree(aAPR.apr_access);
283-
// method_timer.stop_timer();
284-
//
285-
// method_timer.start_timer("compute_local_particle_set");
286-
// get_local_particle_cell_set(local_scale_temp, local_scale_temp2);
287-
// method_timer.stop_timer();
288-
//
289-
// method_timer.start_timer("compute_pulling_scheme");
290-
// iPullingScheme.pulling_scheme_main();
291-
// method_timer.stop_timer();
292-
//
293-
// method_timer.start_timer("downsample_pyramid");
294-
// std::vector<PixelData<T>> downsampled_img;
295-
// //Down-sample the image for particle intensity estimation
296-
// downsamplePyrmaid(input_image, downsampled_img, aAPR.level_max(), aAPR.level_min());
297-
// method_timer.stop_timer();
298-
//
299-
// method_timer.start_timer("compute_apr_datastructure");
300-
// aAPR.apr_access.initialize_structure_from_particle_cell_tree(aAPR.parameters, iPullingScheme.getParticleCellTree());
301-
// method_timer.stop_timer();
302-
//
303-
// method_timer.start_timer("sample_particles");
304-
// aAPR.get_parts_from_img(downsampled_img,aAPR.particles_intensities);
305-
// method_timer.stop_timer();
306-
307328
computation_timer.stop_timer();
308329

309330
total_timer.stop_timer();

‎src/algorithm/ComputeGradientCuda.hpp‎

Lines changed: 0 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -9,9 +9,6 @@
99
#include "algorithm/APRParameters.hpp"
1010

1111

12-
template <typename ImgType>
13-
void getFullPipeline(PixelData<ImgType> &image, PixelData<ImgType> &grad_temp, PixelData<float> &local_scale_temp, PixelData<float> &local_scale_temp2, float bspline_offset, const APRParameters &par, int maxLevel);
14-
1512
// Test helpers and definitions
1613
using TypeOfRecBsplineFlags = uint16_t;
1714
constexpr TypeOfRecBsplineFlags BSPLINE_Y_DIR = 0x01;
@@ -41,8 +38,6 @@ template <typename T>
4138
void thresholdGradient(PixelData<float> &output, const PixelData<T> &input, const float Ip_th);
4239
void cudaDownsampledGradient(PixelData<float> &input, PixelData<float> &grad, const float hx, const float hy, const float hz);
4340

44-
// -------------------------------------------------
45-
4641
template <typename ImgType>
4742
class GpuProcessingTask {
4843
// Using PIMPL to seperate GPU code from CPU (this file is to be included in regular code not compiled with nvcc)

‎src/algorithm/PullingScheme.hpp‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -175,7 +175,6 @@ inline void PullingScheme::set_ascendant_neighbours(int level) {
175175
// type is EMPTY
176176
particle_cell_tree[level].mesh[neighbour_index] = ASCENDANTNEIGHBOUR;
177177
}
178-
179178
if (particle_cell_tree[level].mesh[neighbour_index] == SEED_TYPE) {
180179
// type is SEED
181180
particle_cell_tree[level].mesh[neighbour_index] = PROPOGATE;

0 commit comments

Comments
 (0)