fix int32 vs int64 issue for GPU test on windows (#1739)

Summary:
Part of the work towards #1586; I already mentioned this diff as necessary for a passing
GPU test suite on windows [here](#1586 (comment))

Pull Request resolved: #1739

Reviewed By: wickedfoo

Differential Revision: D26855633

Pulled By: mdouze

fbshipit-source-id: 96d3e627034ccbbe3a32cc4f2310a721e8ea0a69

faiss/gpu/test/test_gpu_index.py

@@ -61,8 +61,11 @@ def do_cpu_to_gpu(self, index_key):
         index.train(xt)
         ts.append(time.time())
 
-        # adding some ids because there was a bug in this case
-        index.add_with_ids(xb, np.arange(nb) * 3 + 12345)
+        # adding some ids because there was a bug in this case;
+        # those need to be cast to idx_t(= int64_t), because
+        # on windows the numpy int default is int32
+        ids = (np.arange(nb) * 3 + 12345).astype('int64')
+        index.add_with_ids(xb, ids)
         ts.append(time.time())
 
         index.nprobe = 4