use threadpoolctl for detecting BLAS threadpools (#692)

We were using numpy.__config__.get_info to figure out what BLAS library was being used,
in an effort to warn people if they hadn't disabled the internal threadpool for the
BLAS they were using. This method is removed in numpy 1.26, causing errors.

Fix by using the threadpoolctl library to detect both the BLAS library and the number
of threads its configured for.

While we could automatically configure the BLAS library to reduce the threadpool size
to 1, this is process wide - and would have side effects for our users. Instead just
warn here, and give instructions for people on how to configure themselves.

Author: benfred

implicit/utils.py

@@ -1,9 +1,9 @@
-import os
 import time
 import warnings
 
 import numpy as np
 import scipy.sparse
+import threadpoolctl
 
 
 def nonzeros(m, row):
@@ -22,19 +22,44 @@ def check_blas_config():
     global _checked_blas_config  # pylint: disable=global-statement
     if _checked_blas_config:
         return
-    _checked_blas_config = True
 
-    if np.__config__.get_info("openblas_info") and os.environ.get("OPENBLAS_NUM_THREADS") != "1":
-        warnings.warn(
-            "OpenBLAS detected. Its highly recommend to set the environment variable "
-            "'export OPENBLAS_NUM_THREADS=1' to disable its internal multithreading"
-        )
-    if np.__config__.get_info("blas_mkl_info") and os.environ.get("MKL_NUM_THREADS") != "1":
-        warnings.warn(
-            "Intel MKL BLAS detected. Its highly recommend to set the environment "
-            "variable 'export MKL_NUM_THREADS=1' to disable its internal "
-            "multithreading"
-        )
+    for api in threadpoolctl.threadpool_info():
+        num_threads = api["num_threads"]
+        if api["user_api"] != "blas" or num_threads == 1:
+            continue
+
+        internal_api = api["internal_api"]
+        if internal_api == "openblas":
+            warnings.warn(
+                f"OpenBLAS is configured to use {num_threads} threads. It is highly recommended"
+                " to disable its internal threadpool by setting the environment variable"
+                " 'OPENBLAS_NUM_THREADS=1' or by calling 'threadpoolctl.threadpool_limits(1,"
+                ' "blas")\'. Having OpenBLAS use a threadpool can lead to severe performance'
+                " issues here.",
+                RuntimeWarning,
+                stacklevel=2,
+            )
+        elif internal_api == "mkl":
+            warnings.warn(
+                f"Intel MKL BLAS is configured to use {num_threads} threads. It is highly"
+                " recommended to disable its internal threadpool by setting the environment"
+                " variable 'MKL_NUM_THREADS=1' or by callng 'threadpoolctl.threadpool_limits(1,"
+                ' "blas")\'. Having MKL use a threadpool can lead to severe performance issues',
+                RuntimeWarning,
+                stacklevel=2,
+            )
+        else:
+            # probably using blis, which is by default single threaded. warn anyways
+            # for when threadpoolctl gets support for VecLib/Accelerate etc
+            warnings.warn(
+                f"BLAS library {internal_api} is configured to use {num_threads} threads."
+                " It is highly recommended to disable its internal threadpool by calling"
+                " 'threadpoolctl.threadpool_limits(1, \"blas\")'.",
+                RuntimeWarning,
+                stacklevel=2,
+            )
+
+    _checked_blas_config = True
 
 
 def check_random_state(random_state):

requirements.txt

@@ -1,3 +1,4 @@
 scipy>=0.16.0
 Cython>=0.24.0
 tqdm>=4.27.0
+threadpoolctl

setup.py

@@ -44,6 +44,6 @@ def exclude_non_implicit_cmake_files(cmake_manifest):
         "Collaborative Filtering, Recommender Systems"
     ),
     packages=find_packages(),
-    install_requires=["numpy", "scipy>=0.16", "tqdm>=4.27"],
+    install_requires=["numpy", "scipy>=0.16", "tqdm>=4.27", "threadpoolctl"],
     cmake_process_manifest_hook=exclude_non_implicit_cmake_files,
 )