Merge pull request scipy#18833 from ThomasLoke/lsap-gil

ENH: optimize: release the GIL while computing the LSAP solution

[skip ci]

benchmarks/benchmarks/optimize_lap.py

@@ -1,3 +1,5 @@
+from concurrent.futures import ThreadPoolExecutor, wait
+
 import numpy as np
 from .common import Benchmark, safe_import
 
@@ -50,3 +52,17 @@ def setup(self, shape, cost_type):
 
     def time_evaluation(self, *args):
         linear_sum_assignment(self.cost_matrix)
+
+
+class ParallelLinearAssignment(Benchmark):
+    shape = (100, 100)
+    param_names = ['threads']
+    params = [[1, 2, 4]]
+
+    def setup(self, threads):
+        self.cost_matrices = [random_uniform(self.shape) for _ in range(20)]
+
+    def time_evaluation(self, threads):
+        with ThreadPoolExecutor(max_workers=threads) as pool:
+            wait({pool.submit(linear_sum_assignment, cost_matrix)
+                  for cost_matrix in self.cost_matrices})

scipy/optimize/_lsap.c

@@ -78,10 +78,17 @@ linear_sum_assignment(PyObject* self, PyObject* args, PyObject* kwargs)
     if (!b)
         goto cleanup;
 
-    int ret = solve_rectangular_linear_sum_assignment(
-      num_rows, num_cols, cost_matrix, maximize,
-      PyArray_DATA((PyArrayObject*)a),
-      PyArray_DATA((PyArrayObject*)b));
+    int64_t* a_ptr = PyArray_DATA((PyArrayObject*)a);
+    int64_t* b_ptr = PyArray_DATA((PyArrayObject*)b);
+    int ret;
+
+    Py_BEGIN_ALLOW_THREADS
+
+    ret = solve_rectangular_linear_sum_assignment(
+      num_rows, num_cols, cost_matrix, maximize, a_ptr, b_ptr);
+
+    Py_END_ALLOW_THREADS
+
     if (ret == RECTANGULAR_LSAP_INFEASIBLE) {
         PyErr_SetString(PyExc_ValueError, "cost matrix is infeasible");
         goto cleanup;