No modification in standard module so it can still be used the same way as before. Example imports:

import sparse_dot_topn.sparse_dot_topn as ct_original
import sparse_dot_topn.sparse_dot_topn_threaded as ct_thread

Example function calls:

ct_original.sparse_dot_topn(row_len, col_len, train_ptr, train_indices, train_data, g_ptr, g_indices, g_data, best_entries_len, probability_above, c_ptr, c_indicies, c_data)

ct_thread.sparse_dot_topn_threaded(row_len, col_len, train_ptr, train_indices, train_data, g_ptr, g_indices, g_data, best_entries_len, probability_above, c_ptr, c_indicies, c_data, n_jobs)

Speed comparison using two matrices:

• tf_idf_train_names is <100000x61134 sparse matrix of type '<class 'numpy.float64'>'
• tf_idf_g_names is <450256x61134 sparse matrix of type '<class 'numpy.float64'>'

Results using original function and threaded function with 1 thread:

Threaded function results using 8 threads:

Only common file with major changes is setup.py

It is now building two extensions and pack them into single python module. Also I modified extra_compile_args and used performance optimization instead of size.

Tested it on Linux / Windows - don't know how/if it will compile on Mac with these settings.

Also it's worth to mention that threading speeds up process significally only under certain conditions. If there is cpu -> memory bottleneck (like on my work laptop) it will slow down calculations if you run more than 1 thread.