Use broadcasted output shape in local_useless_switch optimization

Closes #270

Author: brandonwillard

tests/tensor/test_opt.py

@@ -6163,8 +6163,7 @@ class TestLocalUselessSwitch:
     def setup_method(self):
         self.mode = mode_opt.excluding("constant_folding")
 
-    def test_const0(self):
-
+    def test_const_0(self):
         for dtype1 in ["int32", "int64"]:
             for dtype2 in ["int32", "int64"]:
                 x = tt.matrix("x", dtype=dtype1)
@@ -6186,10 +6185,15 @@ def test_const0(self):
                 )
                 vx = np.array([[1, 2, 3], [4, 5, 6]], dtype=dtype1)
                 vy = np.array([[7, 8, 9], [10, 11, 12]], dtype=dtype2)
-                assert np.all(f(vx, vy) == vy)
+                np_res = np.where(0, vx, vy)
+                assert np.array_equal(f(vx, vy), np_res)
 
-    def test_const1(self):
+        res_non_bool_np = np.where(np.ones(10), 0, 1)
+        non_bool_graph = tt.switch(np.ones(10), 0, 1)
+        non_bool_fn = function([], non_bool_graph, mode=self.mode)
+        assert np.array_equal(non_bool_fn(), res_non_bool_np)
 
+    def test_const_1(self):
         for dtype1 in ["int32", "int64"]:
             for dtype2 in ["int32", "int64"]:
                 x = tt.matrix("x", dtype=dtype1)
@@ -6211,10 +6215,10 @@ def test_const1(self):
                 )
                 vx = np.array([[1, 2, 3], [4, 5, 6]], dtype=dtype1)
                 vy = np.array([[7, 8, 9], [10, 11, 12]], dtype=dtype2)
-                assert np.all(f(vx, vy) == vx)
+                np_res = np.where(1, vx, vy)
+                assert np.array_equal(f(vx, vy), np_res)
 
     def test_left_is_right(self):
-
         for dtype1 in ["int32", "int64"]:
             x = tt.matrix("x", dtype=dtype1)
             varc = tt.matrix("varc", dtype=dtype1)
@@ -6239,12 +6243,11 @@ def test_left_is_right(self):
 
             vx = np.array([[1, 2, 3], [4, 5, 6]], dtype=dtype1)
             vc = np.array([[1, 2, 3], [4, 5, 6]], dtype=dtype1)
-            assert np.all(f1(vx) == vx)
-            assert np.all(f0(vx) == vx)
-            assert np.all(f2(vx, vc) == vx)
+            assert np.array_equal(f1(vx), vx)
+            assert np.array_equal(f0(vx), vx)
+            assert np.array_equal(f2(vx, vc), vx)
 
     def test_shape_le_0(self):
-
         for dtype1 in ["float32", "float64"]:
             x = tt.matrix("x", dtype=dtype1)
             z0 = tt.switch(tt.le(x.shape[0], 0), 0, x.shape[0])
@@ -6259,84 +6262,63 @@ def test_shape_le_0(self):
             assert f0(vx) == 0
             assert f1(vx) == 5
 
-    def test_broadcast1(self):
+    def test_broadcasting_1(self):
         # test switch(cst, matrix, row)
         x = tt.matrix("x", dtype="int32")
         y = tt.vector("y", dtype="int64")
 
         z = tt.switch(1, x, y)
         f = function([x, y], z, mode=self.mode)
-        assert (
-            len(
-                [
-                    node.op
-                    for node in f.maker.fgraph.toposort()
-                    if isinstance(node.op, tt.Elemwise)
-                    and not isinstance(node.op.scalar_op, scal.basic.Cast)
-                ]
-            )
-            == 0
-        )
+
+        assert isinstance(f.maker.fgraph.outputs[0].owner.op, tt.Elemwise)
+        assert isinstance(f.maker.fgraph.outputs[0].owner.op.scalar_op, scal.basic.Cast)
+        assert not any(node.op == tt.switch for node in f.maker.fgraph.toposort())
+
         vx = np.array([[1, 2, 3], [4, 5, 6]], dtype="int32")
         vy = np.array([10, 11, 12], dtype="int64")
-        assert np.all(f(vx, vy) == vx)
+        np_res = np.where(1, vx, vy)
+        assert np.array_equal(f(vx, vy), np_res)
 
         z = tt.switch(0, x, y)
         f = function([x, y], z, mode=self.mode)
-        assert (
-            len(
-                [
-                    node.op
-                    for node in f.maker.fgraph.toposort()
-                    if isinstance(node.op, tt.Elemwise)
-                ]
-            )
-            == 0
-        )
+
+        assert isinstance(f.maker.fgraph.outputs[0].owner.op, tt.Alloc)
+        assert f.maker.fgraph.inputs[1] == f.maker.fgraph.outputs[0].owner.inputs[0]
+        assert not any(node.op == tt.switch for node in f.maker.fgraph.toposort())
+
         vx = np.array([[1, 2, 3], [4, 5, 6]], dtype="int32")
         vy = np.array([10, 11, 12], dtype="int64")
-        assert np.all(f(vx, vy) == vy)
+        np_res = np.where(0, vx, vy)
+        assert np.array_equal(f(vx, vy), np_res)
 
-    def test_broadcast2(self):
+    def test_broadcasting_2(self):
         # test switch(cst, vector, matrix)
 
-        # This case is not optimized for now.
         x = tt.vector("x", dtype="int32")
         y = tt.matrix("y", dtype="int64")
         z = tt.switch(1, x, y)
         f = function([x, y], z, mode=self.mode)
-        assert (
-            len(
-                [
-                    node.op
-                    for node in f.maker.fgraph.toposort()
-                    if isinstance(node.op, tt.Elemwise)
-                    and not isinstance(node.op.scalar_op, scal.basic.Cast)
-                ]
-            )
-            == 0
-        )
+
+        assert isinstance(f.maker.fgraph.outputs[0].owner.op, tt.Alloc)
+        assert not any(node.op == tt.switch for node in f.maker.fgraph.toposort())
+
         vx = np.array([4, 5, 6], dtype="int32")
         vy = np.array([[7, 8, 9], [10, 11, 12]], dtype="int64")
-        assert np.all(f(vx, vy) == vx)
+        np_res = np.where(1, vx, vy)
+        assert np.array_equal(f(vx, vy), np_res)
 
         z = tt.switch(0, x, y)
         f = function([x, y], z, mode=self.mode)
-        assert (
-            len(
-                [
-                    node.op
-                    for node in f.maker.fgraph.toposort()
-                    if isinstance(node.op, tt.Elemwise)
-                ]
-            )
-            == 0
-        )
+
+        assert isinstance(f.maker.fgraph.outputs[0].owner.op, DeepCopyOp)
+        assert not any(node.op == tt.switch for node in f.maker.fgraph.toposort())
+
         vx = np.array([4, 5, 6], dtype="int32")
         vy = np.array([[7, 8, 9], [10, 11, 12]], dtype="int64")
-        assert np.all(f(vx, vy) == vy)
+        np_res = np.where(0, vx, vy)
+        assert np.array_equal(f(vx, vy), np_res)
 
-    def test_broadcast3(self):
+    def test_broadcasting_3(self):
         # test switch(matrix, same_vector, same_vector)
 
         x = tt.matrix("x", dtype="int32")
@@ -6346,16 +6328,9 @@ def test_broadcast3(self):
         vx = np.array([[0, 1], [1, 0]], dtype="int32")
         vy = np.array([7, 8], dtype="int64")
         utt.assert_allclose(f(vx, vy), np.where(vx, vy, vy))
-        assert (
-            len(
-                [
-                    node.op
-                    for node in f.maker.fgraph.toposort()
-                    if isinstance(node.op, tt.Elemwise)
-                ]
-            )
-            == 0
-        )
+
+        assert isinstance(f.maker.fgraph.outputs[0].owner.op, tt.Alloc)
+        assert not any(node.op == tt.switch for node in f.maker.fgraph.toposort())
 
 
 class TestLocalMergeSwitchSameCond:

theano/tensor/opt.py

@@ -97,6 +97,7 @@
     ProdWithoutZeros,
     Sum,
 )
+from theano.tensor.extra_ops import broadcast_shape
 from theano.tensor.sort import TopKOp
 from theano.tensor.subtensor import (
     AdvancedIncSubtensor,
@@ -4327,7 +4328,9 @@ def local_useless_switch(fgraph, node):
         T.switch(le(shape_i{id}(X), 0), 0, shape_i{id}(X)) -> shape_i{id}(X)
     """
     if isinstance(node.op, Elemwise) and isinstance(node.op.scalar_op, ts.Switch):
+
         cond = tt.extract_constant(node.inputs[0], only_process_constants=True)
+
         if (isinstance(cond, np.ndarray) and cond.ndim == 0) or isinstance(
             cond, np.number
         ):
@@ -4336,37 +4339,18 @@ def local_useless_switch(fgraph, node):
             else:
                 correct_out = node.inputs[1]
 
-            if correct_out.ndim != node.outputs[0].ndim:
-                # TODO: broadcast?
-                return False
             if correct_out.dtype != node.outputs[0].dtype:
                 out = tt.cast(correct_out, node.outputs[0].dtype)
             else:
                 out = correct_out
 
-            if out.type.broadcastable != node.outputs[0].type.broadcastable:
-                # We need to copy data to the new dimensions during execution
-
-                # We should not depend on node.outputs as this would
-                # make the new node depend on the old one that will
-                # get optimized again. So this create a cycle.
-                shps = []
-                for idx, (b1, b2), in enumerate(
-                    zip(out.type.broadcastable, node.outputs[0].type.broadcastable)
-                ):
-                    if b1 == b2:
-                        shps.append(out.shape[idx])
-                    elif not node.inputs[1].type.broadcastable[idx]:
-                        shps.append(node.inputs[1].shape[idx])
-                    else:
-                        shps.append(node.inputs[2].shape[idx])
-                out = alloc(out, *shps)
-            else:
-                out = out
+            out_shape = broadcast_shape(*node.inputs)
+            out = alloc(out, *out_shape)
 
             # Copy over stacktrace from selected output to new output
             copy_stack_trace(node.outputs + correct_out, out)
             return [out]
+
         # if left is right -> left
         if node.inputs[1] is node.inputs[2]:
             # Note: No need to copy over stacktrace, because the input node