Enabling large tensor support for binary broadcast operators (#16755)

* using dim_t as type for dimensions in BinaryBroadcastShape

* add arctan2 and hypot test

* large vector for both

* Revert "using dim_t as type for dimensions in BinaryBroadcastShape"

This reverts commit 3d12ed2.

tests/nightly/test_large_array.py

@@ -1678,6 +1678,17 @@ def test_gather():
     assert np.sum(arr[idx[0]] == 2) == SMALL_Y
 
 
+def test_binary_broadcast():
+    def check_correctness(mxnet_op, numpy_op, atol=1e-3):
+        a = mx.nd.ones((LARGE_X, SMALL_Y)).as_np_ndarray()
+        b = 2*mx.nd.ones((LARGE_X, SMALL_Y)).as_np_ndarray()
+        res = mxnet_op(a, b)
+        np_res = numpy_op(1, 2)
+        assert np.abs(res[-1][-1] - np_res) < atol
+    check_correctness(mx.np.arctan2, np.arctan2)
+    check_correctness(mx.np.hypot, np.hypot)
+
+
 if __name__ == '__main__':
     import nose
     nose.runmodule()

tests/nightly/test_large_vector.py

@@ -1065,6 +1065,17 @@ def test_infer_shape():
     assert out_shapes == [(LARGE_X,)]
 
 
+def test_binary_broadcast():
+    def check_correctness(mxnet_op, numpy_op, atol=1e-3):
+        a = mx.nd.ones(LARGE_X).as_np_ndarray()
+        b = 2*mx.nd.ones(LARGE_X).as_np_ndarray()
+        res = mxnet_op(a, b)
+        np_res = numpy_op(1, 2)
+        assert np.abs(res[-1] - np_res) < atol
+    check_correctness(mx.np.arctan2, np.arctan2)
+    check_correctness(mx.np.hypot, np.hypot)
+
+
 if __name__ == '__main__':
     import nose
     nose.runmodule()