Merge branch 'synchro' into 'main'

Add querying for KNN

See merge request jjerphan/cython_plus_experiments!1

kdtree/Makefile

-INCLUDE_DIRS = -I/usr/include/python3.8
+SHELL = /bin/bash
+PROJECT = cython+
+
+VENV_PATH=`conda info --base`/envs/${PROJECT}
+PIP_EXECUTABLE=${VENV_PATH}/bin/pip
+PYTHON_EXECUTABLE=${VENV_PATH}/bin/python
+PYTEST_EXECUTABLE=${VENV_PATH}/bin/pytest
+
+# Used when not using the python runtime
+INCLUDE_DIRS = -I/usr/include/python3.9
 EXE = kdtree
 CXX = g++
-CPPFLAGS = -O2 -g -Wno-unused-result -Wsign-compare -pthread $(INCLUDE_DIRS)
+CPPFLAGS = -O2 -g -Wno-unused-result -Wsign-compare -pthread $(INCLUDE_DIRS) -fopenmp
 LDFLAGS += -Wl,--unresolved-symbols=ignore-all
 MACROS = -DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION
-LDLIBS = -lcrypto -lfmt
-EXT_SUFFIX := $(shell python3 -c "import sysconfig; print(sysconfig.get_config_var('EXT_SUFFIX'))")
+EXT_SUFFIX := $(shell python -c "import sysconfig; print(sysconfig.get_config_var('EXT_SUFFIX'))")
 EXT = $(EXE)$(EXT_SUFFIX)
 
-# Build with Python runtime
-all: $(EXT)
+.DEFAULT_GOAL := all
 
-$(EXT): setup.py
-	@echo "[Cython Compiling $^ -> $@]"
-	python3 setup.py build_ext --inplace
+## help: Display list of commands
+.PHONY: help
+help: Makefile
+	@sed -n 's|^##||p' $< | column -t -s ':' | sed -e 's|^| |'
 
-# Run with Python runtime
-run: $(EXT)
-	python3 -c "import $(EXE); $(EXE).python_main()" 2>/dev/null
+## all: Run the main targets
+.PHONY: all
+all: install benchmark
 
-# Build without Python runtime
+## install: Install conda env.
+.PHONY: install
+install: clean
+	conda env create --force -f environment.yml
+	${PIP_EXECUTABLE} install -e . -v
+
+# nopython: Build without the Python runtime
+.PHONY: nopython
 nopython: $(EXE)
 
 %.cpp: %.pyx
 	@echo "[Cython Compiling $^ -> $@]"
-	python3 -c "from Cython.Compiler.Main import main; main(command_line=1)" $^ --cplus -3
+	${PYTEST_EXECUTABLE} -c "from Cython.Compiler.Main import main; main(command_line=1)" $^ --cplus -3
 	@rm -f $(subst .cpp,.h,$@)
 
 %: %.cpp
 	@echo "[C++ Compiling $^ -> $@]"
-	$(LINK.cpp) $^ $(LOADLIBES) $(LDLIBS) $(MACROS) -o $@
+	$(LINK.cpp) $^ $(MACROS) -o $@
 
-# Run without Python runtime
+## runnopython: Run without Python runtime
+.PHONY: runnopython
 runnopython: $(EXE)
 	# Information of the runtime are currently redirected to stderr.
 	# This is just a simple way to mute them.
 	./$(EXE) 2>/dev/null
 
+## clean: Remove generated files from Cython and C/C++ compilation
+.PHONY: clean
 clean:
 	-rm -f *.c *.cpp *.html
 	-rm -f *.h
@@ -46,5 +64,17 @@ clean:
 	-rm -f -r build
 	-rm -f *.json
 
-.PHONY: all run nopython runnopython clean
 .PRECIOUS: %.cpp
+
+## benchmark: Run benchmarks
+# Uses taskset to cap to a cpu solely
+.PHONY: benchmark
+benchmark:
+	for i in {0..5}; do \
+		taskset -c 0-$$((2**i-1)) ${PYTHON_EXECUTABLE} benchmarks/benchmark.py `git rev-parse --short HEAD`_$$((2**i-1))_thread ;\
+	done
+
+## test: Launch all the test.
+.PHONY: test
+test:
+	${PYTEST_EXECUTABLE} tests

kdtree/benchmarks/benchmark.py

+import argparse
+import glob
+import importlib
+import json
+import os
+import subprocess
+import time
+import kdtree
+import numpy as np
+import pandas as pd
+import seaborn as sns
+import threadpoolctl
+import yaml
+
+from pprint import pprint
+from matplotlib import pyplot as plt
+from memory_profiler import memory_usage
+from sklearn import set_config
+from sklearn.neighbors import KDTree
+
+
+# Be gentle with eyes
+plt.rcParams["figure.dpi"] = 200
+
+
+def benchmark(config, results_folder, bench_name):
+    datasets = config["datasets"]
+    estimators = config["estimators"]
+    leaf_sizes = config["leaf_sizes"]
+    n_neighbors = config.get("n_neighbors", [])
+
+    n_trials = config.get("n_trials", 3)
+    return_distance = config.get("return_distance", False)
+    one_GiB = 1e9
+    benchmarks = pd.DataFrame()
+
+    env_specs_file = f"{results_folder}/{bench_name}.json"
+
+    # TODO: This is ugly, but I haven't found something better.
+    commit = (
+        str(subprocess.check_output(["git", "rev-parse", "--short", "HEAD"]))
+        .replace("b'", "")
+        .replace("\\n'", "")
+    )
+
+    env_specs = dict(
+        threadpool_info=threadpoolctl.threadpool_info(),
+        commit=commit,
+        config=config,
+    )
+
+    set_config(assume_finite=True)
+
+    with open(env_specs_file, "w") as outfile:
+        json.dump(env_specs, outfile)
+
+    for dataset in datasets:
+        for leaf_size in leaf_sizes:
+            for trial in range(n_trials):
+                dataset = {k: int(float(v)) for k, v in dataset.items()}
+                ns_train, ns_test, n_features = dataset.values()
+                X_train = np.random.rand(ns_train, n_features)
+                X_test = np.random.rand(ns_test, n_features)
+                bytes_processed_data_init = X_train.nbytes
+                bytes_processed_data_query = X_test.nbytes
+
+                t0_ = time.perf_counter()
+                sk_tree = KDTree(X_train, leaf_size=leaf_size)
+                t1_ = time.perf_counter()
+                time_elapsed = round(t1_ - t0_, 5)
+
+                row = dict(
+                    trial=trial,
+                    func="init",
+                    implementation="sklearn",
+                    leaf_size=leaf_size,
+                    n_samples_train=ns_train,
+                    n_samples_test=ns_test,
+                    n_features=n_features,
+                    n_neighbors=np.nan,
+                    time_elapsed=time_elapsed,
+                    throughput=bytes_processed_data_init / time_elapsed / one_GiB,
+                )
+
+                benchmarks = benchmarks.append(row, ignore_index=True)
+                pprint(row)
+                print("---")
+
+                t0_ = time.perf_counter()
+                tree = kdtree.KDTree(X_train, leaf_size=leaf_size)
+                t1_ = time.perf_counter()
+                time_elapsed = round(t1_ - t0_, 5)
+
+                row = dict(
+                    trial=trial,
+                    func="init",
+                    implementation="kdtree",
+                    leaf_size=leaf_size,
+                    n_samples_train=ns_train,
+                    n_samples_test=ns_test,
+                    n_features=n_features,
+                    n_neighbors=np.nan,
+                    time_elapsed=time_elapsed,
+                    throughput=bytes_processed_data_init / time_elapsed / one_GiB,
+                )
+
+                benchmarks = benchmarks.append(row, ignore_index=True)
+                pprint(row)
+                print("---")                
+
+                benchmarks.to_csv(
+                    f"{results_folder}/{bench_name}.csv",
+                    mode="w+",
+                    index=False,
+                )
+                
+                for k in n_neighbors:
+                    
+                    t0_ = time.perf_counter()
+                    sk_tree.query(X_test, k=k, return_distance=False)
+                    t1_ = time.perf_counter()
+                    time_elapsed = round(t1_ - t0_, 5)
+                    
+                    row = dict(
+                        trial=trial,
+                        func="query",
+                        implementation="sklearn",
+                        leaf_size=leaf_size,
+                        n_samples_train=ns_train,
+                        n_samples_test=ns_test,
+                        n_features=n_features,
+                        n_neighbors=k,
+                        time_elapsed=time_elapsed,
+                        throughput=bytes_processed_data_query / time_elapsed / one_GiB,
+                    )
+
+                    benchmarks = benchmarks.append(row, ignore_index=True)
+                    pprint(row)
+                    print("---")
+                        
+                    closests = np.zeros((ns_test, k), dtype=np.int32) 
+                    t0_ = time.perf_counter()
+                    tree.query(X_test, closests)
+                    t1_ = time.perf_counter()
+                    time_elapsed = round(t1_ - t0_, 5)
+
+                    row = dict(
+                        trial=trial,
+                        func="query",
+                        implementation="kdtree",
+                        leaf_size=leaf_size,
+                        n_samples_train=ns_train,
+                        n_samples_test=ns_test,
+                        n_features=n_features,
+                        n_neighbors=k,
+                        time_elapsed=time_elapsed,
+                        throughput=bytes_processed_data_query / time_elapsed / one_GiB,
+                    )
+
+                    benchmarks = benchmarks.append(row, ignore_index=True)
+                    pprint(row)
+                    print("---")                
+
+                    benchmarks.to_csv(
+                        f"{results_folder}/{bench_name}.csv",
+                        mode="w+",
+                        index=False,
+                    )
+
+                
+
+    # Overriding again now that all the dyn. lib. have been loaded
+    env_specs["threadpool_info"] = threadpoolctl.threadpool_info()
+
+    with open(env_specs_file, "w") as outfile:
+        json.dump(env_specs, outfile)
+
+
+def report(results_folder, bench_name):
+    df = pd.read_csv(glob.glob(f"{results_folder}/*.csv")[0])
+    with open(glob.glob(f"{results_folder}/*.json")[0], "r") as json_file:
+        env_specs = json.load(json_file)
+
+    cols = [
+        "n_samples_train",
+        "n_features",
+        "leaf_size",
+    ]
+
+    df[cols] = df[cols].astype(np.uint32)
+    df['d'] = df.n_features.apply(str)
+    df['leaf'] = df.leaf_size.apply(str)
+
+    df_grouped = df.groupby(cols)
+
+    for i, (vals, df) in enumerate(df_grouped):
+        # 16:9 ratio
+        fig = plt.figure(figsize=(24, 13.5))
+        ax = plt.gca()
+        splot = sns.barplot(
+            y="leaf", x="throughput", hue="implementation", data=df, ax=ax
+        )
+        _ = ax.set_xlabel("Throughput (in GB/s)")
+        _ = ax.set_ylabel("Leaf Size")
+        _ = ax.tick_params(labelrotation=45)
+
+        # Adding the numerical values of "x" to bar
+        for p in splot.patches:
+            _ = splot.annotate(
+                f"{p.get_width():.4e}",
+                (p.get_width(), p.get_y() + p.get_height() / 2),
+                ha="center",
+                va="center",
+                size=10,
+                xytext=(0, -12),
+                textcoords="offset points",
+            )
+
+        title = (
+            f"KDTree@{env_specs['commit']} - "
+            f"Euclidean Distance, dtype=np.float64, {df.trial.max() + 1} trials - Bench. Name: {bench_name}\n"
+        )
+        title += (
+            "n_samples_train=%s - n_features=%s - leaf_size=%s" % vals
+        )
+        _ = fig.suptitle(title, fontsize=16)
+        plt.savefig(f"{results_folder}/{bench_name}_{i}.pdf", bbox_inches="tight")
+
+    # Unifying pdf files into one
+    pdf_files = sorted(glob.glob(f"{results_folder}/{bench_name}*.pdf"))
+    subprocess.check_output(
+        ["pdfunite", *pdf_files, f"{results_folder}/{bench_name}.pdf"]
+    )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser("benchmark")
+
+    parser.add_argument("bench_name")
+
+    args = parser.parse_args()
+
+    bench_name = args.bench_name
+    with open("benchmarks/config.yml", "r") as f:
+        config = yaml.full_load(f)
+
+    results_folder = f"benchmarks/results/{bench_name}"
+    os.makedirs(results_folder, exist_ok=True)
+
+    print(f"Benchmarking {bench_name}")
+    benchmark(config, results_folder, bench_name)
+    print(f"Benchmark results wrote in {results_folder}")
+
+    print(f"Reporting results for {bench_name}")
+    report(results_folder, bench_name)
+    print(f"Reporting results wrote in {results_folder}")
\ No newline at end of file

kdtree/benchmarks/config.yml

+estimators:
+  - name: sklearn
+    estimator: sklearn.neighbors.KDTree
+  - name: cython+
+    estimator: kdtree.KDTree
+
+n_trials: 3
+
+datasets:
+  - n_samples_train: 1e6
+    n_samples_test: 1e6
+    n_features: 5
+  - n_samples_train: 1e6
+    n_samples_test: 1e6
+    n_features: 10
+  - n_samples_train: 1e6
+    n_samples_test: 1e6
+    n_features: 50
+  - n_samples_train: 1e6
+    n_samples_test: 1e6
+    n_features: 100
+
+leaf_sizes:
+  - 64
+  - 128
+  - 256
+  - 512
+  - 1024
+  - 2048
+  - 4096

kdtree/environment.yml

+name: cython+
+channels:
+  - conda-forge
+dependencies:
+  - python=3.9
+  - compilers
+  - jupyter
+  - numpy
+  - matplotlib
+  - seaborn
+  - pandas
+  - pyaml
+  - pip
+  - threadpoolctl
+  - pytest
+  - scikit-learn
+  - memory_profiler
+  - pip:
+      # Install cython+ from upstream directly
+    - -e git+https://gitlab.inria.fr/jjerphan/cython.git@b30eafec6a7b174afdc4f023b45b21f85104e2fe#egg=Cython
+      # The installation of the 'kdtree' module is made then

kdtree/kdtree.pyx

@@ -5,16 +5,23 @@ import numpy as np
 
 np.import_array()
 
-from runtime.runtime cimport BatchMailBox, NullResult, Scheduler
+from runtime.runtime cimport BatchMailBox, NullResult, Scheduler, WaitResult
 from libc.stdio cimport printf
 from libc.stdlib cimport malloc, free
+from openmp cimport omp_get_max_threads
+
+
+from cython.parallel import prange
 
 ## Types declaration
 ctypedef int I_t
 ctypedef double D_t
 
 cdef lock Scheduler scheduler
+cdef D_t INF = 1e38
 
+# NOTE: The following extern definition is used to interface
+# std::nth_element, a robust partitioning algorithm, in Cython 
 cdef extern from *:
     """
     #include <algorithm>
@@ -63,53 +70,384 @@ cdef extern from *:
                 I n_features
     ) nogil except +
 
+
+cdef cypclass Counter activable:
+    """ A simple Counter.
+    
+    This can be useful for synchronisation for the caller after
+    triggering the actors logic as it wait for the value of
+    the Coutner to reach a given one before moving on.
+    """
+
+    I_t _n
+
+    __init__(self):
+        self._n = 0
+        self._active_result_class = WaitResult.construct
+        self._active_queue_class = consume BatchMailBox(scheduler)
+
+    void add(self, I_t value):
+        self._n += value
+
+    void reset(self):
+        self._n = 0
+
+    I_t value(self):
+        return self._n
+
+
+cdef cypclass Container activable:
+    """ A simple wrapper of an array.
+    
+    The wrapped array is passed by the initial caller which then
+    trigger the actors logic modifying it.
+
+    The initial caller can wait for specific number of update
+    before proceeding.
+    """
+
+    I_t * _array
+    I_t _size
+    I_t _n_updates
+
+
+    __init__(self, I_t *array, I_t size):
+        self._array = array
+        self._size = size
+        self._active_result_class = WaitResult.construct
+        self._active_queue_class = consume BatchMailBox(scheduler)
+        self._n_updates = 0
+
+    void update(self, I_t idx, I_t value):
+        self._n_updates += 1
+        self._array[idx] = value
+
+    int n_updates(self):
+        return self._n_updates
+
+
+cdef inline void dual_swap(D_t* darr, I_t* iarr, I_t i1, I_t i2) nogil:
+    """swap the values at inex i1 and i2 of both darr and iarr"""
+    cdef D_t dtmp = darr[i1]
+    darr[i1] = darr[i2]
+    darr[i2] = dtmp
+
+    cdef I_t itmp = iarr[i1]
+    iarr[i1] = iarr[i2]
+    iarr[i2] = itmp
+
+
+cdef void _simultaneous_sort(
+    D_t* dist,
+    I_t* idx,
+    I_t size
+) nogil:
+    """
+    Perform a recursive quicksort on the dist array, simultaneously
+    performing the same swaps on the idx array.  The equivalent in
+    numpy (though quite a bit slower) is
+
+    def simultaneous_sort(dist, idx):
+        i = np.argsort(dist)
+        return dist[i], idx[i]
+    """
+    cdef I_t pivot_idx, i, store_idx
+    cdef D_t pivot_val
+
+    # in the small-array case, do things efficiently
+    if size <= 1:
+        pass
+    elif size == 2:
+        if dist[0] > dist[1]:
+            dual_swap(dist, idx, 0, 1)
+    elif size == 3:
+        if dist[0] > dist[1]:
+            dual_swap(dist, idx, 0, 1)
+        if dist[1] > dist[2]:
+            dual_swap(dist, idx, 1, 2)
+            if dist[0] > dist[1]:
+                dual_swap(dist, idx, 0, 1)
+    else:
+        # Determine the pivot using the median-of-three rule.
+        # The smallest of the three is moved to the beginning of the array,
+        # the middle (the pivot value) is moved to the end, and the largest
+        # is moved to the pivot index.
+        pivot_idx = size // 2
+        if dist[0] > dist[size - 1]:
+            dual_swap(dist, idx, 0, size - 1)
+        if dist[size - 1] > dist[pivot_idx]:
+            dual_swap(dist, idx, size - 1, pivot_idx)
+            if dist[0] > dist[size - 1]:
+                dual_swap(dist, idx, 0, size - 1)
+        pivot_val = dist[size - 1]
+
+        # partition indices about pivot.  At the end of this operation,
+        # pivot_idx will contain the pivot value, everything to the left
+        # will be smaller, and everything to the right will be larger.
+        store_idx = 0
+        for i in range(size - 1):
+            if dist[i] < pivot_val:
+                dual_swap(dist, idx, i, store_idx)
+                store_idx += 1
+        dual_swap(dist, idx, store_idx, size - 1)
+        pivot_idx = store_idx
+
+        # recursively sort each side of the pivot
+        if pivot_idx > 1:
+            _simultaneous_sort(dist, idx, pivot_idx)
+        if pivot_idx + 2 < size:
+            _simultaneous_sort(dist + pivot_idx + 1,
+                                    idx + pivot_idx + 1,
+                                    size - pivot_idx - 1)
+
+
+cdef cypclass NeighborsHeap activable:
+    """A max-heap structure to keep track of distances/indices of neighbors
+
+    This implements an efficient pre-allocated set of fixed-size heaps
+    for chasing neighbors, holding both an index and a distance.
+    When any row of the heap is full, adding an additional point will push
+    the furthest point off the heap.
+
+    Taken and adapted from:
+    https://github.com/scikit-learn/scikit-learn/blob/e4bb9fa86b0df873ad750b6d59090843d9d23d50/sklearn/neighbors/_binary_tree.pxi#L513
+
+    The initial caller is responsible for providing the array of indices
+    which will be modified in place by the actors logic.
+
+    n_pushes and is_sorted can be used by the initial caller to know
+    when to pursue.
+
+    Parameters
+    ----------
+    n_pts : int
+        the number of heaps to use
+    n_nbrs : int
+        the size of each heap.
+    """
+    D_t *_distances
+    I_t *_indices
+
+    I_t _n_pts
+    I_t _n_nbrs
+    I_t _n_pushes
+    bint _sorted
+
+    __init__(self, I_t * indices, I_t n_pts, I_t n_nbrs):
+        cdef I_t i
+        self._n_pts = n_pts
+        self._n_nbrs = n_nbrs
+        self._active_result_class = WaitResult.construct
+        self._active_queue_class = consume BatchMailBox(scheduler)
+        self._distances = <D_t *> malloc(n_pts * n_nbrs * sizeof(D_t))
+        self._indices = indices
+        self._n_pushes = 0
+        self._sorted = False
+
+        # We can't use memset here
+        for i in range(n_pts * n_nbrs):
+            self._distances[i] = INF
+
+    void __dealloc__(self):
+        free(self._distances)
+
+    void push(self, I_t row, D_t val, I_t i_val):
+        """push (val, i_val) into the given row"""
+        cdef:
+            I_t i, left_child_idx, right_child_idx, swap_idx
+
+            # Getting the heap to use
+            I_t *indices = self._indices + row * self._n_nbrs
+            D_t *distances = self._distances + row * self._n_nbrs
+
+        self._n_pushes += 1
+
+        # check if val should be in heap
+        if val > distances[0]:
+            return
+
+        # insert val at position zero
+        distances[0] = val
+        indices[0] = i_val
+
+        # descend the heap, swapping values until the max heap criterion is met
+        i = 0
+        while True:
+            left_child_idx = 2 * i + 1
+            right_child_idx = left_child_idx + 1
+
+            if left_child_idx >= self._n_nbrs:
+                break
+            elif right_child_idx >= self._n_nbrs:
+                if distances[left_child_idx] > val:
+                    swap_idx = left_child_idx
+                else:
+                    break
+            elif distances[left_child_idx] >= distances[right_child_idx]:
+                if val < distances[left_child_idx]:
+                    swap_idx = left_child_idx
+                else:
+                    break
+            else:
+                if val < distances[right_child_idx]:
+                    swap_idx = right_child_idx
+                else:
+                    break
+
+            distances[i] = distances[swap_idx]
+            indices[i] = indices[swap_idx]
+
+            i = swap_idx
+
+        distances[i] = val
+        indices[i] = i_val
+
+
+    void sort(self):
+        # NOTE: Ideally we could sort results in parallel, but
+        # OpenMP threadpool and this runtime's aren't working
+        # nicely together (using prange here would create OpenMP
+        # threads within workers, causing unexpected behavior.)
+        cdef I_t row
+
+        for row in range(self._n_pts):
+            # We use a function here to be able to recurse
+            _simultaneous_sort(
+                self._distances + row * self._n_nbrs,
+                self._indices + row * self._n_nbrs,
+                self._n_nbrs)
+
+        self._sorted = True
+
+    int n_pushes(self):
+        return self._n_pushes
+
+    int is_sorted(self):
+        # TODO: is there a support for returning bool via
+        # promises? As of now returning a int for making
+        # use of getIntResult
+        return 1 if self._sorted else 0
+
+
 cdef cypclass Node activable:
-    """A KDTree Node"""
+    """A KDTree Node
+    
+    Node delegate tasks to their children Nodes.
+
+    Some Nodes are set as Leaves when they are associated
+    to ``leaf_size`` or less points.
+    
+    Leafs are terminal Nodes and do not have children.
+    """
+
+    # Reference to the head of the allocated arrays
+    # data gets not modified via _data_ptr 
+    D_t * _data_ptr
+    I_t * _indices_ptr
+
+    # The point the Node split on
+    D_t * _point
+    I_t _n_dims
+    I_t _dim
+
+    # Portion of _indices covered by the Node is:
+    # _indices_ptr[_start:_end]
+    I_t _start
+    I_t _end
+
+    bint _is_leaf
 
-    D_t * point
-    I_t n_dims
-    active Node left
-    active Node right
+    active Node _left
+    active Node _right
 
     __init__(self):
-        self._active_result_class = NullResult
+        self._active_result_class = WaitResult.construct
         self._active_queue_class = consume BatchMailBox(scheduler)
-        self.left = NULL
-        self.right = NULL
+        self._left = NULL
+        self._right = NULL
+        self._is_leaf = False
 
+
+    # We use this to allow using actors for initialisation
+    # because __init__ can't be reified. 
     void build_node(
             self,
             D_t * data_ptr,
             I_t * indices_ptr,
-            I_t depth,
+            I_t leaf_size,
             I_t n_dims,
             I_t dim,
             I_t start,
             I_t end,
+            active Counter counter,
     ):
         cdef I_t i
         cdef I_t next_dim = (dim + 1) % n_dims
         cdef I_t mid = (start + end) // 2
 
-        self.n_dims = n_dims
+        self._data_ptr = data_ptr
+        self._indices_ptr = indices_ptr
+
+        self._dim = dim
+        self._n_dims = n_dims
+        self._start = start
+        self._end = end
 
-        if (depth < 0) or (end - start <= 1):
+        if (end - start <= leaf_size):
+            self._is_leaf = True
+            # Adding to the global counter the number
+            # of samples the leaf is responsible of
+            counter.add(NULL, end - start)
             return
 
+        # We partition the samples in two nodes on a given dimension,
+        # with the middle point as a pivot
         partition_node_indices(data_ptr, indices_ptr, start, mid, end, dim, n_dims)
-        self.point = data_ptr + mid
+        self._point = data_ptr + mid
 
-        self.left = consume Node()
-        self.right = consume Node()
+        self._left = consume Node()
+        self._right = consume Node()
 
-        self.left.build_node(NULL,
+        # Recursing on both partition
+        self._left.build_node(NULL,
                               data_ptr, indices_ptr,
-                             depth - 1, n_dims, next_dim,
-                             start, mid)
-        self.right.build_node(NULL,
+                              leaf_size, n_dims, next_dim,
+                              start, mid, counter)
+        self._right.build_node(NULL,
                                data_ptr, indices_ptr,
-                              depth - 1, n_dims, next_dim,
-                              mid, end)
+                               leaf_size, n_dims, next_dim,
+                               mid, end, counter)
+
+    void query(self,
+                D_t * query_points,
+                I_t i,
+                active NeighborsHeap heaps,
+    ):
+        cdef:
+            I_t j, k
+            D_t dist
+            D_t tmp
+
+        if self._is_leaf:
+            # Computing all the euclideans distances here
+            for j in range(self._start, self._end):
+                dist = 0
+                for k in range(self._n_dims):
+                    tmp = (
+                        query_points[i * self._n_dims + k] - 
+                        self._data_ptr[self._indices_ptr[j] * self._n_dims + k]
+                        )
+                    dist += tmp * tmp
+                
+                # The heap is doing the smart work of keeping
+                # the closest points for each query point i
+                heaps.push(NULL, i, dist, self._indices_ptr[j])
+         
+            return
+
+        # TODO: one can implement a pruning strategy here
+        self._left.query(NULL, query_points, i, heaps)
+        self._right.query(NULL, query_points, i, heaps)
 
 
 cdef cypclass KDTree:
@@ -124,66 +462,105 @@ cdef cypclass KDTree:
     This relies on a Cython+ runtime using actors.
     """
 
-    I_t n # number of data
-    I_t d # number of dimensions / features
-    I_t depth # max_depth of the tree (to be unified with leaf_size)
+    I_t _n # number of examples
+    I_t _d # number of dimensions / features
+    I_t _leaf_size # maximum number of vectors at leaf
+
+    I_t _n_leafs
 
-    active Node root
+    active Node _root
 
-    D_t *data_ptr
-    I_t *indices_ptr
+    D_t *_data_ptr
+    I_t *_indices_ptr
 
     __init__(self,
-             np.ndarray data,
-             I_t depth,
+             np.ndarray X,
+             I_t leaf_size,
              ):
         cdef I_t i
-        cdef I_t n = data.shape[0]
-        cdef I_t d = data.shape[1]
-
-        self.n = n
-        self.d = d
-        self.depth = depth
-
-        self.data_ptr = <D_t *> data.data
-        self.indices_ptr = <I_t *> malloc(n * sizeof(I_t))
+        cdef I_t n = X.shape[0]
+        cdef I_t d = X.shape[1]
+        cdef I_t initialised = 0
+        # Accessing _SC_NPROCESSORS_ONLN does not return the 
+        # effective number of threads which were assigned to
+        # this task using `taskset(1)` for instance. 
+        # 
+        # This OpenMP API is a workable way to access it.
+        cdef I_t num_workers = omp_get_max_threads()
+
+        self._n = n
+        self._d = d
+        self._leaf_size = leaf_size
+
+        self._data_ptr = <D_t *> X.data
+        self._indices_ptr = <I_t *> malloc(n * sizeof(I_t))
         for i in range(n):
-            self.indices_ptr[i] = i
+            self._indices_ptr[i] = i
 
-        self._recursive_build()
-
-    void _recursive_build(self):
-        # TODO: introducing a context manager for the runtime
-        # would be nice here:
-        #   ```
-        #   with scheduler:
-        #        self.root = ...
-        #   ```
+        # Recurvisely building the tree here
         global scheduler
-        scheduler = Scheduler()
+        scheduler = Scheduler(num_workers)
+        cdef active Counter counter = consume Counter()
 
-        self.root = consume Node()
-        if self.root is NULL:
+        self._root = consume Node()
+        if self._root is NULL:
             printf("Error consuming node\n")
 
         # When object are activated (set as Actors), methods
         # are reified. When using those reified methods
         # a new argument is prepredend for a predicate,
         # which we aren't using using here, hence the extra NULL.
-        self.root.build_node(NULL,
-                             self.data_ptr,
-                             self.indices_ptr,
-                             depth, n_dims=d, dim=0, start=0, end=n)
+        #
+        # Also using this separate method allowing using actors
+        # because __init__ can't be reified. 
+        self._root.build_node(NULL,
+                              self._data_ptr,
+                              self._indices_ptr,
+                              self._leaf_size, n_dims=self._d,
+                              dim=0, start=0, end=self._n,
+                              counter=counter)
 
-        scheduler.finish()
+        # Waiting for the tree construction to end
+        # Somewhat similar to a thread barrier
+        while(initialised < self._n):
+            initialised = counter.value(NULL).getIntResult()
+
+        counter.reset(NULL)
         
 
     void __dealloc__(self):
-        printf("Deallocating KDTree datastructures\n")
-        free(self.data_ptr)
-        free(self.indices_ptr)
-        printf("Done deallocating KDTree datastructures\n")
-        # XXX It sometimes blocks over here and then segfaults
+        scheduler.finish()
+        free(self._indices_ptr)
+
+
+    void query(self,
+               np.ndarray query_points,  # IN
+               np.ndarray closests,      # OUT
+               ):
+        cdef:
+            I_t completed_queries = 0
+            I_t i
+            I_t n_query = query_points.shape[0]
+            I_t n_neighbors = closests.shape[1]
+            I_t total_n_pushes = n_query * self._n
+
+            active NeighborsHeap heaps
+
+        heaps = consume NeighborsHeap(<I_t *> closests.data,
+                                      n_query,
+                                      n_neighbors)
+
+        for i in range(n_query):
+            self._root.query(NULL, <D_t *> query_points.data, i, heaps)
+
+        while(completed_queries < total_n_pushes):
+            completed_queries = heaps.n_pushes(NULL).getIntResult()
+
+        heaps.sort(NULL)
+        
+        while not(heaps.is_sorted(NULL).getIntResult()):
+            pass
+
 
 cdef public int main() nogil:
     # Entry point for the compiled binary file

kdtree/main.py

-import numpy as np
-import kdtree
-
-if __name__ == "__main__":
-    n, d = 1000000, 2
-    golden_ratio = (1 + 5 ** 0.5) / 2
-    X = np.zeros((n, d))
-    for i in range(n):
-        X[i, 0] = (i / golden_ratio) % 1
-        X[i, 1] = i / n
-    tree = kdtree.KDTree(X, depth=10)
-    del tree

kdtree/runtime/runtime.pxd

@@ -217,3 +217,47 @@ cdef cypclass BatchMailBox(SequentialMailBox):
 
 cdef inline ActhonResultInterface NullResult() nogil:
     return NULL
+
+
+# Taken from:
+# https://lab.nexedi.com/nexedi/cython/blob/3.0a6-cypclass/tests/run/cypclass_acthon.pyx#L66
+cdef cypclass WaitResult(ActhonResultInterface):
+    union result_t:
+        int int_val
+        void* ptr
+    result_t result
+    sem_t semaphore
+
+    __init__(self):
+        self.result.ptr = NULL
+        sem_init(&self.semaphore, 0, 0)
+
+    __dealloc__(self):
+        sem_destroy(&self.semaphore)
+
+    @staticmethod
+    ActhonResultInterface construct():
+        return WaitResult()
+
+    void pushVoidStarResult(self, void* result):
+        self.result.ptr = result
+        sem_post(&self.semaphore)
+
+    void pushIntResult(self, int result):
+        self.result.int_val = result
+        sem_post(&self.semaphore)
+
+    result_t _getRawResult(const self):
+        # We must ensure a result exists, but we can let others access it immediately
+        # The cast here is a way of const-casting (we're modifying the semaphore in a const method)
+        sem_wait(<sem_t*> &self.semaphore)
+        sem_post(<sem_t*> &self.semaphore)
+        return self.result
+
+    void* getVoidStarResult(const self):
+        res = self._getRawResult()
+        return res.ptr
+
+    int getIntResult(const self):
+        res = self._getRawResult()
+        return res.int_val

kdtree/setup.py

 from distutils.core import setup
 from distutils.extension import Extension
+
+import numpy
 from Cython.Build import cythonize
 
 extensions = [
@@ -7,11 +9,15 @@ extensions = [
         "kdtree",
         language="c++",
         define_macros=[("NPY_NO_DEPRECATED_API", "NPY_1_7_API_VERSION")],
+        include_dirs=[numpy.get_include()],
+        extra_compile_args=["-fopenmp"],
+        extra_link_args=["-fopenmp"],
         sources=["kdtree.pyx"],
-        libraries=["crypto", "fmt"],
     ),
 ]
 
 setup(
-    ext_modules=cythonize(extensions)
+    ext_modules=cythonize(extensions),
+    name="kdtree",
+    version="0.1",
 )

kdtree/tests/test_conf.py

+import numpy as np
+import pytest
+import kdtree
+from sklearn.neighbors import KDTree
+
+@pytest.mark.parametrize("n", [10, 100, 1000, 10000])
+@pytest.mark.parametrize("d", [10, 100])
+@pytest.mark.parametrize("k", [1, 2, 5, 10])
+@pytest.mark.parametrize("leaf_size", [256, 1024])
+def test_against_sklearn(n, d, k, leaf_size):
+    np.random.seed(1)
+    X = np.random.rand(n, d)
+    query_points = np.random.rand(n, d)
+    
+    tree = kdtree.KDTree(X, leaf_size=256)
+    skl_tree = KDTree(X, leaf_size=256)
+    
+    closests = np.zeros((n, k), dtype=np.int32)
+    tree.query(query_points, closests)
+    skl_closests = skl_tree.query(query_points, k=k, return_distance=False).astype(np.int32)
+    
+    np.testing.assert_equal(closests, skl_closests)
\ No newline at end of file

requirements.txt

-ipython==7.22.0
-jupyter==1.0.0
-line-profiler==3.1.0
-matplotlib==3.4.1
-numpy==1.20.2
--e git+https://lab.nexedi.com/nexedi/cython.git@608c29a9ab7b7803c900621424279a9e71fac106#egg=Cython&subdirectory=../../cython