Don't fuse take with dynamic inputs (apache#6979)

* add a regression test for fusing dynamic take

* add legalize for take that stops fusion on dynamic inputs

* fix lint

* fix typo

python/tvm/relay/op/_transform.py

@@ -341,6 +341,25 @@ def take_shape_func(attrs, inputs, out_ndims):
     return [_take_with_axis_shape_func(*inputs, convert(axis), out_ndims[0])]
 
 
+@_reg.register_legalize("take")
+def legalize_dyn_topk(attrs, inputs, types):
+    """Legalize take op.
+    Parameters
+    ----------
+    attrs : tvm.ir.Attrs
+        Attributes of current op
+    inputs : list of tvm.relay.Expr
+        The args of the Relay expr to be legalized
+    types : list of types
+        List of input and output types
+    Returns
+    -------
+    result : tvm.relay.Expr
+        The legalized expr
+    """
+    return topi.take_legalize(attrs, inputs, types)
+
+
 @script
 def _argwhere_shape_func_1d(condition):
     out = output_tensor((2,), "int64")

python/tvm/topi/transform.py

@@ -426,6 +426,28 @@ def take(a, indices, axis=None, mode="clip"):
     return cpp.take(a, indices, int(axis), mode)
 
 
+@tvm.target.generic_func
+def take_legalize(attrs, inputs, types):
+    """Legalizes dyn.topk op.
+
+    Parameters
+    ----------
+    attrs : tvm.ir.Attrs
+        Attributes of current op
+    inputs : list of tvm.relay.Expr
+        The args of the Relay expr to be legalized
+    types : list of types
+        List of input and output types
+    Returns
+    -------
+    result : tvm.relay.Expr
+        The legalized expr
+    """
+    if tvm.relay.ty.is_dynamic(types[0]):
+        return tvm.relay.take(tvm.relay.annotation.stop_fusion(inputs[0]), inputs[1], **attrs)
+    return None
+
+
 def gather(data, axis, indices):
     """Gather values along given axis from given indices.
 

tests/python/relay/test_pass_fuse_ops.py

@@ -14,6 +14,8 @@
 # KIND, either express or implied.  See the License for the
 # specific language governing permissions and limitations
 # under the License.
+import numpy as np
+
 import tvm
 from tvm import relay
 from tvm.relay import transform
@@ -757,6 +759,31 @@ def create_diamond_func(inp):
     assert tvm.ir.structural_equal(fused, expected)
 
 
+def test_fuse_dynamic_squeeze_slice_take():
+    input_data = [
+        np.random.random([1, 2, 4]).astype("float32"),
+        np.array([0]).astype("int64"),
+    ]
+
+    x = relay.var("p0107", shape=(relay.Any(), relay.Any(), 4), dtype="float32")
+    take_val = relay.var("p166", shape=(relay.Any(),), dtype="int64")
+
+    squeeze = relay.op.squeeze(x, axis=[0])
+    strided_slice = relay.op.strided_slice(
+        squeeze, begin=[0, 0], end=[15130, 9223372036854775807], strides=[1, 1]
+    )
+    take = relay.op.take(strided_slice, take_val, axis=0)
+
+    mod = tvm.IRModule.from_expr(take)
+    ex = relay.create_executor("vm", mod=mod, ctx=tvm.cpu(), target="llvm")
+
+    result = ex.evaluate()(*input_data)
+
+    np_result = np.squeeze(input_data[0][:, input_data[1][0], :], axis=0)
+
+    assert np.allclose(result.asnumpy(), np_result)
+
+
 if __name__ == "__main__":
     test_fuse_simple()
     test_conv2d_fuse()