Merge pull request #530 from vlnistor/I-356/batch-search

Batch search inside stack search

README.md

@@ -140,7 +140,7 @@ candidates = [trj for trj in gen]
 
 # Do the actual search.
 search = kb.StackSearch(stack)
-search.search(
+search.search_all(
     strategy,
     7,  # The minimum number of observations
 )

notebooks/Kbmod_Reference.ipynb

@@ -406,7 +406,7 @@
     "candidates = [trj for trj in gen]\n",
     "print(f\"Created {len(candidates)} candidate trajectories per pixel.\")\n",
     "\n",
-    "search.search(candidates, 2)"
+    "search.search_all(candidates, 2)"
    ]
   },
   {

src/kbmod/batch_search.py

@@ -0,0 +1,31 @@
+class BatchSearchManager:
+    def __init__(self, stack_search, search_list, min_observations):
+        """Manages a batch search over a list of Trajectory instances using a StackSearch instance.
+
+        Parameters
+        ----------
+        stack_search: `StackSearch`
+            StackSearch instance to use for the batch search.
+        search_list: `list[Trajectory]`
+            List of Trajectory instances to search along the stack.
+        min_observations: `int`
+            Minimum number of observations required to consider a candidate.
+        """
+        self.stack_search = stack_search
+        self.search_list = search_list
+        self.min_observations = min_observations
+
+    def __enter__(self):
+        # Prepare memory for the batch search.
+        self.stack_search.prepare_search(self.search_list, self.min_observations)
+        return self.stack_search
+
+    def __exit__(self, *_):
+        """
+        This method is called when exiting the context.
+        It cleans up resources by calling `finish_search` on the StackSearch instance.
+        We return False to indicate that we do not want to suppress any exceptions that may have been raised.
+        """
+        # Clean up
+        self.stack_search.finish_search()
+        return False

src/kbmod/run_search.py

@@ -197,7 +197,7 @@ def do_gpu_search(self, config, stack, trj_generator):
 
         # Do the actual search.
         candidates = [trj for trj in trj_generator]
-        search.search(candidates, int(config["num_obs"]))
+        search.search_all(candidates, int(config["num_obs"]))
         search_timer.stop()
 
         # Load the results.

src/kbmod/search/pydocs/stack_search_docs.h

@@ -177,6 +177,46 @@ static const auto DOC_StackSearch_get_results = R"doc(
   ``RunTimeError`` if start < 0 or count <= 0.
   )doc";
 
+static const auto DOC_StackSearch_prepare_batch_search = R"doc(
+  Prepare the search for a batch of trajectories.
+
+  Parameters
+  ----------
+  search_list : `List`
+      A list of ``Trajectory`` objects to search.
+  min_observations : `int`
+      The minimum number of observations for a trajectory to be considered.
+  )doc";
+
+static const auto DOC_StackSearch_compute_max_results = R"doc(
+  Compute the maximum number of results according to the x, y bounds and the RESULTS_PER_PIXEL constant
+
+  Returns
+  -------
+  max_results : `int`
+      The maximum number of results that a search will return according to the current bounds and the RESULTS_PER_PIXEL constant.
+  )doc";
+
+static const auto DOC_StackSearch_search_single_batch = R"doc(
+  Perform a search on the given trajectories for the current batch.
+  Batch is defined by the parameters set `set_start_bounds_x` & `set_start_bounds_y`.
+
+  Returns
+  -------
+  results : `List`
+      A list of ``Trajectory`` search results
+  )doc";
+
+static const auto DOC_StackSearch_finish_search = R"doc(
+  Clears memory used for the batch search.
+
+  This method should be called after a batch search is completed to ensure that any resources allocated during the search are properly freed.
+
+  Returns
+  -------
+  None
+  )doc";
+
 static const auto DOC_StackSearch_set_results = R"doc(
   Set the cached results. Used for testing.
 

src/kbmod/search/stack_search.cpp

@@ -15,7 +15,7 @@ extern "C" void evaluateTrajectory(PsiPhiArrayMeta psi_phi_meta, void* psi_phi_v
 // I'd imaging...
 auto rs_logger = logging::getLogger("kbmod.search.run_search");
 
-StackSearch::StackSearch(ImageStack& imstack) : stack(imstack), results(0) {
+StackSearch::StackSearch(ImageStack& imstack) : stack(imstack), results(0), gpu_search_list(0) {
     debug_info = false;
     psi_phi_generated = false;
 
@@ -155,63 +155,81 @@ Trajectory StackSearch::search_linear_trajectory(short x, short y, float vx, flo
     return result;
 }
 
-void StackSearch::search(std::vector<Trajectory>& search_list, int min_observations) {
-    DebugTimer core_timer = DebugTimer("core search", rs_logger);
+void StackSearch::finish_search(){
+    psi_phi_array.clear_from_gpu();
+    gpu_search_list.move_to_cpu();
+}
 
-    DebugTimer psi_phi_timer = DebugTimer("creating psi/phi buffers", rs_logger);
+void StackSearch::prepare_search(std::vector<Trajectory>& search_list, int min_observations){
+    DebugTimer psi_phi_timer = DebugTimer("Creating psi/phi buffers", rs_logger);
     prepare_psi_phi();
     psi_phi_array.move_to_gpu();
     psi_phi_timer.stop();
 
-    // Allocate a vector for the results and move it onto the GPU.
-    int search_width = params.x_start_max - params.x_start_min;
-    int search_height = params.y_start_max - params.y_start_min;
-    int num_search_pixels = search_width * search_height;
-    int max_results = num_search_pixels * RESULTS_PER_PIXEL;
+
+    int num_to_search = search_list.size();
+    if (debug_info) std::cout << "Preparing to search " << num_to_search << " trajectories... \n" << std::flush;
+    gpu_search_list.set_trajectories(search_list);
+    gpu_search_list.move_to_gpu();
+
+    params.min_observations = min_observations;
+}
+
+void StackSearch::search_all(std::vector<Trajectory>& search_list, int min_observations) {
+    prepare_search(search_list, min_observations);
+    search_batch();
+    finish_search();
+}
+
+void StackSearch::search_batch(){
+    if(!psi_phi_array.gpu_array_allocated()){
+        throw std::runtime_error("PsiPhiArray array not allocated on GPU. Did you forget to call prepare_search?");
+    }
+
+    DebugTimer core_timer = DebugTimer("Running batch search", rs_logger);
+    int max_results = compute_max_results();
+
     // staple C++
     std::stringstream logmsg;
     logmsg << "Searching X=[" << params.x_start_min << ", " << params.x_start_max << "] "
            << "Y=[" << params.y_start_min << ", " << params.y_start_max << "]\n"
            << "Allocating space for " << max_results << " results.";
     rs_logger->info(logmsg.str());
 
+
     results.resize(max_results);
     results.move_to_gpu();
 
-    // Allocate space for the search list and move that to the GPU.
-    int num_to_search = search_list.size();
-
-    logmsg.str("");
-    logmsg << search_list.size() << " trajectories...";
-    rs_logger->info(logmsg.str());
-
-    TrajectoryList gpu_search_list(search_list);
-    gpu_search_list.move_to_gpu();
-
-    // Set the minimum number of observations.
-    params.min_observations = min_observations;
-
     // Do the actual search on the GPU.
-    DebugTimer search_timer = DebugTimer("search execution", rs_logger);
+    DebugTimer search_timer = DebugTimer("Running search", rs_logger);
 #ifdef HAVE_CUDA
     deviceSearchFilter(psi_phi_array, params, gpu_search_list, results);
 #else
     throw std::runtime_error("Non-GPU search is not implemented.");
 #endif
     search_timer.stop();
 
-    // Move data back to CPU to unallocate GPU space (this will happen automatically
-    // for gpu_search_list when the object goes out of scope, but we do it explicitly here).
-    psi_phi_array.clear_from_gpu();
     results.move_to_cpu();
-    gpu_search_list.move_to_cpu();
-
     DebugTimer sort_timer = DebugTimer("Sorting results", rs_logger);
     results.sort_by_likelihood();
     sort_timer.stop();
     core_timer.stop();
 }
 
+std::vector<Trajectory> StackSearch::search_single_batch(){
+    int max_results = compute_max_results();
+    search_batch();
+    return results.get_batch(0, max_results);
+}
+
+
+int StackSearch::compute_max_results(){
+    int search_width = params.x_start_max - params.x_start_min;
+    int search_height = params.y_start_max - params.y_start_min;
+    int num_search_pixels = search_width * search_height;
+    return num_search_pixels * RESULTS_PER_PIXEL;
+}
+
 std::vector<float> StackSearch::extract_psi_or_phi_curve(Trajectory& trj, bool extract_psi) {
     prepare_psi_phi();
 
@@ -261,7 +279,7 @@ static void stack_search_bindings(py::module& m) {
 
     py::class_<ks>(m, "StackSearch", pydocs::DOC_StackSearch)
             .def(py::init<is&>())
-            .def("search", &ks::search, pydocs::DOC_StackSearch_search)
+            .def("search_all", &ks::search_all, pydocs::DOC_StackSearch_search)
             .def("evaluate_single_trajectory", &ks::evaluate_single_trajectory,
                  pydocs::DOC_StackSearch_evaluate_single_trajectory)
             .def("search_linear_trajectory", &ks::search_linear_trajectory,
@@ -288,7 +306,11 @@ static void stack_search_bindings(py::module& m) {
             .def("prepare_psi_phi", &ks::prepare_psi_phi, pydocs::DOC_StackSearch_prepare_psi_phi)
             .def("clear_psi_phi", &ks::clear_psi_phi, pydocs::DOC_StackSearch_clear_psi_phi)
             .def("get_results", &ks::get_results, pydocs::DOC_StackSearch_get_results)
-            .def("set_results", &ks::set_results, pydocs::DOC_StackSearch_set_results);
+            .def("set_results", &ks::set_results, pydocs::DOC_StackSearch_set_results)
+            .def("compute_max_results", &ks::compute_max_results, pydocs::DOC_StackSearch_compute_max_results)
+            .def("search_single_batch", &ks::search_single_batch, pydocs::DOC_StackSearch_search_single_batch)
+            .def("prepare_search", &ks::prepare_search, pydocs::DOC_StackSearch_prepare_batch_search)
+            .def("finish_search", &ks::finish_search, pydocs::DOC_StackSearch_finish_search);
 }
 #endif /* Py_PYTHON_H */
 

src/kbmod/search/stack_search.h

@@ -31,7 +31,7 @@ using Image = search::Image;
 class StackSearch {
 public:
     StackSearch(ImageStack& imstack);
-
+    int compute_max_results();
     int num_images() const { return stack.img_count(); }
     int get_image_width() const { return stack.get_width(); }
     int get_image_height() const { return stack.get_height(); }
@@ -50,7 +50,11 @@ class StackSearch {
     // The primary search functions
     void evaluate_single_trajectory(Trajectory& trj);
     Trajectory search_linear_trajectory(short x, short y, float vx, float vy);
-    void search(std::vector<Trajectory>& search_list, int min_observations);
+    void prepare_search(std::vector<Trajectory>& search_list, int min_observations);
+    std::vector<Trajectory> search_single_batch();
+    void search_batch();
+    void search_all(std::vector<Trajectory>& search_list, int min_observations);
+    void finish_search();
 
     // Gets the vector of result trajectories from the grid search.
     std::vector<Trajectory> get_results(int start, int end);
@@ -82,6 +86,9 @@ class StackSearch {
 
     // Results from the grid search.
     TrajectoryList results;
+
+    // Trajectories that are being searched.
+    TrajectoryList gpu_search_list;
 };
 
 } /* namespace search */

tests/test_readme_example.py

@@ -53,7 +53,7 @@ def test_make_and_copy(self):
 
         # Do the actual search.
         search = kb.StackSearch(stack)
-        search.search(
+        search.search_all(
             candidates,
             7,  # The minimum number of observations
         )