Merge pull request #695 from stephenswat/refactor/fuse_ckf

Fuse inner loop kernels in device CKF

core/include/traccc/finding/finding_config.hpp

@@ -48,12 +48,6 @@ struct finding_config {
  /// Particle hypothesis
  detray::pdg_particle<traccc::scalar> ptc_hypothesis =
  detray::muon<traccc::scalar>();
-
- /****************************
- * GPU-specfic parameters
- ****************************/
- /// The number of measurements to be iterated per thread
- unsigned int n_measurements_per_thread = 8;
 };
 
 } // namespace traccc

device/common/CMakeLists.txt

@@ -63,17 +63,13 @@ traccc_add_library( traccc_device_common device_common TYPE SHARED
  # Track finding funtions(s).
  "include/traccc/finding/device/apply_interaction.hpp"
  "include/traccc/finding/device/build_tracks.hpp"
- "include/traccc/finding/device/count_measurements.hpp"
  "include/traccc/finding/device/find_tracks.hpp"
- "include/traccc/finding/device/add_links_for_holes.hpp"
  "include/traccc/finding/device/make_barcode_sequence.hpp"
  "include/traccc/finding/device/propagate_to_next_surface.hpp"
  "include/traccc/finding/device/prune_tracks.hpp"
  "include/traccc/finding/device/impl/apply_interaction.ipp"
  "include/traccc/finding/device/impl/build_tracks.ipp"
- "include/traccc/finding/device/impl/count_measurements.ipp"
  "include/traccc/finding/device/impl/find_tracks.ipp"
- "include/traccc/finding/device/impl/add_links_for_holes.ipp"
  "include/traccc/finding/device/impl/make_barcode_sequence.ipp"
  "include/traccc/finding/device/impl/propagate_to_next_surface.ipp"
  "include/traccc/finding/device/impl/prune_tracks.ipp"

device/common/include/traccc/edm/device/finding_global_counter.hpp

@@ -11,9 +11,6 @@ namespace traccc::device {
 
 struct finding_global_counter {
 
- // Sum of the number of measurements for every parameter
- unsigned int n_measurements_sum;
-
  // Number of found measurements for the current step
  unsigned int n_candidates;
 

device/common/include/traccc/finding/device/find_tracks.hpp

@@ -10,6 +10,8 @@
 // Project include(s).
 #include "traccc/definitions/primitives.hpp"
 #include "traccc/definitions/qualifiers.hpp"
+#include "traccc/device/concepts/barrier.hpp"
+#include "traccc/device/concepts/thread_id.hpp"
 #include "traccc/edm/measurement.hpp"
 #include "traccc/edm/track_parameters.hpp"
 
@@ -22,43 +24,47 @@ namespace traccc::device {
 /// If the chi2 of the measurement < chi2_max, its measurement index and the
 /// index of the link from the previous step are added to the link container.
 ///
-/// @param[in] globalIndex The index of the current thread
+/// @param[in] thread_id A thread identifier object
+/// @param[in] barrier A block-wide barrier
 /// @param[in] cfg Track finding config object
 /// @param[in] det_data Detector view object
 /// @param[in] measurements_view Measurements container view
+/// @param[in] in_params_view Input parameters
+/// @param[in] n_in_params The number of input params
+/// @param[in] barcodes_view View of a measurement -> barcode map
 /// @param[in] upper_bounds_view Upper bounds of measurements unique w.r.t
 /// barcode
-/// @param[in] in_params_view Input parameters
-/// @param[in] n_measurements_prefix_sum_view Prefix sum of the number of
-/// measurements per parameter
-/// @param[in] ref_meas_idx_view The first index of measurements per parameter
 /// @param[in] prev_links_view link container from the previous step
 /// @param[in] prev_param_to_link_view param_to_link container from the
 /// previous step
 /// @param[in] step Step index
 /// @param[in] n_max_candidates Number of maximum candidates
 /// @param[out] out_params_view Output parameters
-/// @param[out] n_candidates_view Number of candidates per input parameter
 /// @param[out] links_view link container for the current step
 /// @param[out] n_total_candidates The number of total candidates for the
 /// current step
+/// @param shared_num_candidates Shared memory scratch space
+/// @param shared_candidates Shared memory scratch space
+/// @param shared_candidates_size Shared memory scratch space
 ///
-template <typename detector_t, typename config_t>
+template <concepts::thread_id1 thread_id_t, concepts::barrier barrier_t,
+ typename detector_t, typename config_t>
 TRACCC_DEVICE inline void find_tracks(
- std::size_t globalIndex, const config_t cfg,
+ thread_id_t& thread_id, barrier_t& barrier, const config_t cfg,
  typename detector_t::view_type det_data,
  measurement_collection_types::const_view measurements_view,
  bound_track_parameters_collection_types::const_view in_params_view,
- vecmem::data::vector_view<const unsigned int>
-  n_measurements_prefix_sum_view,
- vecmem::data::vector_view<const unsigned int> ref_meas_idx_view,
+ const unsigned int n_in_params,
+ vecmem::data::vector_view<const detray::geometry::barcode> barcodes_view,
+ vecmem::data::vector_view<const unsigned int> upper_bounds_view,
  vecmem::data::vector_view<const candidate_link> prev_links_view,
  vecmem::data::vector_view<const unsigned int> prev_param_to_link_view,
  const unsigned int step, const unsigned int& n_max_candidates,
  bound_track_parameters_collection_types::view out_params_view,
- vecmem::data::vector_view<unsigned int> n_candidates_view,
  vecmem::data::vector_view<candidate_link> links_view,
- unsigned int& n_total_candidates);
+ unsigned int& n_total_candidates, unsigned int* shared_num_candidates,
+ std::pair<unsigned int, unsigned int>* shared_candidates,
+ unsigned int& shared_candidates_size);
 
 } // namespace traccc::device