[Feature] Efficient DTW in C++

.gitignore

@@ -25,3 +25,7 @@ docs_env
 .venv
 .env
 uv.lock
+
+# C++ build artifacts
+*.so
+_ext.pyi

CMakeLists.txt

@@ -0,0 +1,86 @@
+cmake_minimum_required(VERSION 3.15...3.26)
+
+project(darts LANGUAGES CXX)
+
+if (NOT SKBUILD)
+  message(WARNING "\
+  This CMake file is meant to be executed using 'scikit-build'. Running
+  it directly will almost certainly not produce the desired result. If
+  you are a user trying to install this package, please use the command
+  below, which will install all necessary build dependencies, compile
+  the package in an isolated environment, and then install it.
+  =====================================================================
+   $ pip install .
+  =====================================================================
+  If you are a software developer, and this is your own package, then
+  it is usually much more efficient to install the build dependencies
+  in your environment once and use the following command that avoids
+  a costly creation of a new virtual environment at every compilation:
+  =====================================================================
+   $ pip install nanobind scikit-build-core[pyproject]
+   $ pip install --no-build-isolation -ve .
+  =====================================================================
+  You may optionally add -Ceditable.rebuild=true to auto-rebuild when
+  the package is imported. Otherwise, you need to re-run the above
+  after editing C++ files.")
+endif()
+
+# Try to import all Python components potentially needed by nanobind
+find_package(Python 3.10
+  REQUIRED COMPONENTS Interpreter Development.Module
+  OPTIONAL_COMPONENTS Development.SABIModule)
+
+# Import nanobind through CMake's find_package mechanism
+find_package(nanobind CONFIG REQUIRED)
+
+include_directories(include)
+
+# We are now ready to compile the actual extension module
+nanobind_add_module(
+  # Name of the extension
+  _ext
+
+  # This extension is free-threaded (only applies to Python 3.14t+)
+  FREE_THREADED
+
+  # Target the stable ABI for Python 3.12+, which reduces
+  # the number of binary wheels that must be built. This
+  # does nothing on older Python versions
+  STABLE_ABI
+
+  # Build libnanobind statically and merge it into the
+  # extension (which itself remains a shared library)
+  #
+  # If your project builds multiple extensions, you can
+  # replace this flag by NB_SHARED to conserve space by
+  # reusing a shared libnanobind across libraries
+  NB_STATIC
+
+  # Source code goes here
+  src/bind.cpp
+  src/dtw.cpp
+)
+
+nanobind_add_stub(
+  _ext_stub
+  MODULE _ext
+  OUTPUT _ext.pyi
+  PYTHON_PATH $<TARGET_FILE_DIR:_ext>
+  MARKER_FILE py.typed
+  DEPENDS _ext
+)
+
+# Keep generated typing artifacts visible to language servers during local/editable
+# development. VS Code/Pylance resolves this workspace package from source.
+add_custom_command(
+  TARGET _ext_stub POST_BUILD
+  COMMAND ${CMAKE_COMMAND} -E copy_if_different
+    ${CMAKE_CURRENT_BINARY_DIR}/_ext.pyi
+    ${CMAKE_CURRENT_SOURCE_DIR}/darts/_ext.pyi
+)
+
+# Install directive for scikit-build-core
+install(TARGETS _ext LIBRARY DESTINATION darts)
+
+# Install the generated stub file as well, so that it can be included in the wheel
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/py.typed ${CMAKE_CURRENT_BINARY_DIR}/_ext.pyi DESTINATION darts)

MANIFEST.in

@@ -0,0 +1 @@
+include include/*.h

darts/dataprocessing/dtw/cost_matrix.py

@@ -5,7 +5,9 @@
 
 import array
 from abc import ABC, abstractmethod
+from collections.abc import Generator
 from itertools import repeat
+from typing import Any
 
 import numpy as np
 
@@ -30,15 +32,15 @@ def fill(self, value: float):
         pass
 
     @abstractmethod
-    def __getitem__(self, item):
+    def __getitem__(self, item) -> Any:
         pass
 
     @abstractmethod
     def __setitem__(self, key, value):
         pass
 
     @abstractmethod
-    def __iter__(self):
+    def __iter__(self) -> Generator:
         pass
 
     @abstractmethod
@@ -77,14 +79,24 @@ def _from_window(window: Window):
             return DenseCostMatrix(window.n, window.m)
 
 
-class DenseCostMatrix(np.ndarray, CostMatrix):
-    def __new__(self, n, m):
+class DenseCostMatrix(CostMatrix):
+    def __init__(self, n: int, m: int, dense: np.ndarray | None = None):
         self.n = n
         self.m = m
-        return super().__new__(self, (n + 1, m + 1), float)
+        self.dense = np.ndarray((n + 1, m + 1), float) if dense is None else dense
 
-    def to_dense(self) -> np.ndarray:
-        return self[1:, 1:]
+    def to_dense(self, copy: bool = False) -> np.ndarray:
+        data = self.dense[1:, 1:]
+        return data.copy() if copy else data
+
+    def fill(self, value: float):
+        self.dense.fill(value)
+
+    def __getitem__(self, item):
+        return self.dense[item]
+
+    def __setitem__(self, key, value):
+        self.dense[key] = value
 
     def __iter__(self):
         for n in range(1, self.n):

darts/dataprocessing/dtw/dtw.py

@@ -10,7 +10,8 @@
 import pandas as pd
 
 from darts import TimeSeries
-from darts.dataprocessing.dtw.cost_matrix import CostMatrix
+from darts._ext import dtw_cost_matrix_no_window
+from darts.dataprocessing.dtw.cost_matrix import CostMatrix, DenseCostMatrix
 from darts.dataprocessing.dtw.window import CRWindow, NoWindow, Window
 from darts.logging import get_logger, raise_if, raise_if_not
 
@@ -362,6 +363,9 @@ def dtw(
             logger,
         )
         cost_matrix = _fast_dtw(values_x, values_y, distance, multi_grid_radius)
+    elif isinstance(window, NoWindow):
+        cost_array = dtw_cost_matrix_no_window(values_x, values_y)
+        cost_matrix = DenseCostMatrix(window.n, window.m, cost_array)
     else:
         cost_matrix = _dtw_cost_matrix(values_x, values_y, distance, window)