[AutoParallel] dist reshape op (#41821)

* add dist reshape impl_idx=2

* fix cmakelist

python/paddle/distributed/auto_parallel/operators/dist_reshape.py

@@ -349,7 +349,159 @@ class DistributedReshapeImpl1(DistributedOperatorImpl):
         DistributedDefaultImpl0.backward(ctx, *args, **kwargs)
 
 
+class DistributedReshapeImpl2(DistributedOperatorImpl):
+    def __init__(self, name):
+        super(DistributedReshapeImpl2, self).__init__(name)
+        self._forward_implemented = True
+        self._backward_implemented = False
+
+    def is_input_compatible(self, dist_op):
+        op_desc = dist_op.serial_op.desc
+        op_dist_attr = dist_op.dist_attr
+        x_name = op_desc.input('X')[0]
+        out_name = op_desc.output('Out')[0]
+        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
+        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
+
+        if len(x_dims_mapping) != len(out_dims_mapping):
+            return False
+
+        return True
+
+    def is_output_compatible(self, dist_op):
+        op_desc = dist_op.serial_op.desc
+        op_dist_attr = dist_op.dist_attr
+        out_name = op_desc.output('Out')[0]
+        x_name = op_desc.input('X')[0]
+        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
+        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
+
+        if len(x_dims_mapping) != len(out_dims_mapping):
+            return False
+
+        return True
+
+    def is_auto_compatible(self, dist_op):
+        if (not self.is_input_compatible(dist_op)) or \
+            (not self.is_output_compatible(dist_op)):
+            return False
+
+        op_desc = dist_op.serial_op.desc
+        op_dist_attr = dist_op.dist_attr
+        x_name = op_desc.input('X')[0]
+        out_name = op_desc.output('Out')[0]
+        x_shape_name = op_desc.output('XShape')[0]
+        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
+        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
+        x_shape_dims_mapping = op_dist_attr.get_output_dims_mapping(
+            x_shape_name)
+
+        for idx, item in enumerate(x_dims_mapping[:-1]):
+            if out_dims_mapping[idx] != item:
+                return False
+
+        if x_shape_dims_mapping[0] != -1:
+            return False
+
+        if x_shape_dims_mapping[1:] != out_dims_mapping[:]:
+            return False
+
+        return True
+
+    def update_dims_mapping(self, dist_op):
+        changed = False
+        op_desc = dist_op.serial_op.desc
+        op_dist_attr = dist_op.dist_attr
+        x_name = op_desc.input('X')[0]
+        out_name = op_desc.output('Out')[0]
+        x_shape_name = op_desc.output('XShape')[0]
+        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
+        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
+        x_shape_dims_mapping = op_dist_attr.get_output_dims_mapping(
+            x_shape_name)
+
+        for i in range(len(out_dims_mapping) - 1):
+            dim_changed = compute_compatible_and_update_dim_mapping(
+                [x_dims_mapping, out_dims_mapping], [i, i])
+            if dim_changed:
+                changed = True
+
+        for i in range(len(out_dims_mapping)):
+            x_shape_dims_mapping[i + 1] = out_dims_mapping[i]
+
+        return changed
+
+    @staticmethod
+    def forward(ctx, *args, **kwargs):
+        """
+        kwargs: inputname_mapping & outputname_mapping
+        """
+
+        dist_op_context = ctx.dist_op_context
+        main_block = dist_op_context.work_block
+        src_op = dist_op_context.cur_src_op
+        op_dist_attr = ctx.get_op_dist_attr_for_program(src_op)
+        assert op_dist_attr is not None, "backward op [{}] don't have dist attribute !".format(
+            str(src_op))
+
+        # check validation of inputs / outputs
+        for input_name in src_op.desc.input_names():
+            assert input_name in kwargs, "input [{}] is not given".format(
+                input_name)
+            assert len(kwargs[input_name]) == len(
+                src_op.desc.input(input_name)
+            ), "number of tensor for input [{}] is not match".format(input_name)
+        for output_name in src_op.desc.output_names():
+            assert output_name in kwargs, "input [{}] is not given".format(
+                output_name)
+            assert len(kwargs[output_name]) == len(
+                src_op.desc.output(output_name)
+            ), "number of tensor for input [{}] is not match".format(
+                output_name)
+
+        X_var = main_block.var(kwargs['X'][0])
+        Out_var = main_block.var(kwargs['Out'][0])
+        XShape_var = main_block.var(kwargs['XShape'][0])
+        shape_list = src_op.desc.attr("shape")
+        ShapeTensor_var_list = []
+        for name in kwargs['ShapeTensor']:
+            ShapeTensor_var_list.append(name)
+        Shape_var_list = []
+        for name in kwargs['Shape']:
+            Shape_var_list.append(name)
+
+        # got dist attribute info
+        out_dim_mapping = op_dist_attr.get_output_dims_mapping(Out_var.name)
+        process_mesh_shape = op_dist_attr.process_mesh.topology
+
+        # modify target shape
+        for idx, axis in enumerate(out_dim_mapping):
+            if axis >= 0:
+                if len(shape_list) > idx:
+                    shape_list[idx] = shape_list[idx] // process_mesh_shape[
+                        axis]
+
+        # create op
+        new_op_desc = main_block.desc.append_op()
+        new_op_desc.copy_from(src_op.desc)
+        set_dist_op_desc_original_id(new_op_desc, src_op.desc, ctx)
+        new_op_desc.set_input('ShapeTensor', ShapeTensor_var_list)
+        new_op_desc.set_input('Shape', Shape_var_list)
+        new_op_desc.set_input('X', [X_var.name])
+        new_op_desc.set_output('XShape', [XShape_var.name])
+        new_op_desc.set_output('Out', [Out_var.name])
+        new_op_desc._set_attr('shape', shape_list)
+
+        main_block._sync_with_cpp()
+
+    @staticmethod
+    def backward(ctx, *args, **kwargs):
+        DistributedDefaultImpl0.backward(ctx, *args, **kwargs)
+
+
 register_distributed_operator_impl("reshape2",
                                    DistributedReshapeImpl0("add_one_dim_back"))
 register_distributed_operator_impl(
     "reshape2", DistributedReshapeImpl1("remove_one_dim_back"))
+register_distributed_operator_impl("reshape2",
+                                   DistributedReshapeImpl2("same_dim_shape"))

python/paddle/fluid/tests/unittests/auto_parallel/CMakeLists.txt

@@ -18,6 +18,7 @@ if(WITH_DISTRIBUTE AND WITH_GPU)
     py_test_modules(test_recorder MODULES test_recorder ENVS ${dist_ENVS})
     py_test_modules(test_trial MODULES test_trial ENVS ${dist_ENVS})
     py_test_modules(test_new_cost_model MODULES test_new_cost_model ENVS ${dist_ENVS})
+    py_test_modules(test_dist_reshape MODULES test_dist_reshape ENVS ${dist_ENVS})
     py_test_modules(test_dist_pnorm MODULES test_dist_pnorm ENVS ${dist_ENVS})
     py_test_modules(test_dist_slice MODULES test_dist_slice ENVS ${dist_ENVS})
     py_test_modules(test_cluster MODULES test_cluster ENVS ${dist_ENVS})

python/paddle/fluid/tests/unittests/auto_parallel/test_dist_reshape.py

+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import paddle
+import paddle.distributed.auto_parallel as auto
+
+from paddle.fluid import program_guard
+from paddle.fluid.backward import append_backward
+from paddle.distributed.auto_parallel.utils import print_program_with_dist_attr
+
+paddle.enable_static()
+
+
+def make_program_dp2():
+    main_program = paddle.fluid.Program()
+    start_program = paddle.fluid.Program()
+    with paddle.static.program_guard(main_program, start_program):
+        x = paddle.static.data(name='x', shape=[4, 4, 8], dtype='float32')
+        x.stop_gradient = False
+        auto.shard_tensor(
+            x,
+            dist_attr={
+                "process_mesh": auto.ProcessMesh([0, 1]),
+                "dims_mapping": [0, -1, -1]
+            })
+        tmp_0 = paddle.reshape(x, shape=[0, 0, 4, 2])
+        tmp_1 = paddle.reshape(tmp_0, shape=[0, 0, 8])
+        tmp_2 = tmp_1.reshape((tmp_1.shape[0], tmp_1.shape[1], -1))
+    return main_program, start_program
+
+
+def parallelizer(program_func, rank):
+    from paddle.distributed.auto_parallel.completion import Completer
+    from paddle.distributed.auto_parallel.partitioner import Partitioner
+    from paddle.distributed.auto_parallel.dist_context import DistributedContext
+
+    main_program, start_program = program_func()
+
+    dist_context = DistributedContext()
+    completer = Completer(dist_context)
+    completer.complete_forward_annotation(main_program)
+    dist_context.block_state.parse_forward_blocks(main_program)
+
+    partitioner = Partitioner(dist_context, rank)
+    dist_main_prog, _, _ = partitioner.partition(main_program, start_program,
+                                                 [])
+
+    return dist_main_prog, dist_context
+
+
+class TestDistReshape(unittest.TestCase):
+    def test_dist_reshape_mp2(self):
+
+        for rank in range(2):
+            dist_main_prog, dist_context = parallelizer(make_program_dp2, rank)
+            ops = dist_main_prog.global_block().ops
+            print_program_with_dist_attr(dist_main_prog, dist_context)
+            for idx, op in enumerate(ops):
+                op_dist_attr = dist_context.get_op_dist_attr_for_program(op)
+                assert op_dist_attr.impl_type == "reshape2"
+                assert op_dist_attr.impl_idx == idx
+
+                if op_dist_attr.impl_idx == 2:
+                    assert op.desc.attr('shape')[0] == 2
+
+
+if __name__ == "__main__":
+    unittest.main()