[AutoParallel] dist slice op (#41780)

* add dist slice

* fix debug

* fix cmakelist

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

@@ -28,5 +28,6 @@ from . import dist_check_finite_and_unscale
 from . import dist_update_loss_scaling
 from . import dist_split
 from . import dist_fill_constant_batch_size_like
+from . import dist_slice
 from . import dist_fused_feedforward
 from . import dist_fused_attention

python/paddle/distributed/auto_parallel/operators/dist_slice.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.
+
+from .common import DistributedOperatorImplContainer
+from .common import DistributedOperatorImpl
+from .common import register_distributed_operator_impl_container
+from .common import register_distributed_operator_impl
+from ..utils import is_dim_shard
+from ..utils import compute_compatible_and_update_dim_mapping
+from .dist_default import DistributedDefaultImpl0
+
+
+class DistributedSlice(DistributedOperatorImplContainer):
+    def __init__(self, op_type):
+        super(DistributedSlice, self).__init__(op_type)
+
+
+register_distributed_operator_impl_container(DistributedSlice("slice"))
+
+
+class DistributedSliceImpl(DistributedOperatorImpl):
+    def __init__(self, name):
+        super(DistributedSliceImpl, self).__init__(name)
+        self._forward_implemented = True
+        self._backward_implemented = True
+
+    def is_input_compatible(self, dist_op):
+        op_desc = dist_op.serial_op.desc
+        op_dist_attr = dist_op.dist_attr
+        in_name = op_desc.input('Input')[0]
+        axes = op_desc.attr('axes')
+        in_dims_mapping = op_dist_attr.get_input_dims_mapping(in_name)
+        for axis in axes:
+            if is_dim_shard(in_dims_mapping[axis]):
+                return False
+        return True
+
+    def is_output_compatible(self, dist_op):
+        return True
+
+    def is_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
+        in_name = op_desc.input('Input')[0]
+        out_name = op_desc.output('Out')[0]
+        decrease_axis = op_desc.attr('decrease_axis')
+        in_dims_mapping = op_dist_attr.get_input_dims_mapping(in_name)
+        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
+        if len(in_dims_mapping) - len(decrease_axis) != 0 and len(
+                out_dims_mapping) != len(in_dims_mapping) - len(decrease_axis):
+            return False
+
+        new_out_dims_mapping = []
+        for i in range(len(in_dims_mapping)):
+            if i not in decrease_axis:
+                new_out_dims_mapping.append(in_dims_mapping[i])
+        if new_out_dims_mapping == []:
+            new_out_dims_mapping = [-1]
+        if new_out_dims_mapping != 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)) or \
+            (not self.is_compatible(dist_op)):
+            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
+        in_name = op_desc.input('Input')[0]
+        out_name = op_desc.output('Out')[0]
+        decrease_axis = op_desc.attr('decrease_axis')
+        in_dims_mapping = op_dist_attr.get_input_dims_mapping(in_name)
+        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
+
+        ref_dims_mapping = []
+        for i in range(len(in_dims_mapping)):
+            if i not in decrease_axis:
+                ref_dims_mapping.append(in_dims_mapping[i])
+        if ref_dims_mapping == []:
+            ref_dims_mapping = [-1]
+
+        assert len(ref_dims_mapping) == len(out_dims_mapping)
+        for i in range(len(out_dims_mapping)):
+            if out_dims_mapping[i] != ref_dims_mapping[i]:
+                out_dims_mapping[i] = ref_dims_mapping[i]
+                changed = True
+
+        return changed
+
+    @staticmethod
+    def forward(ctx, *args, **kwargs):
+        DistributedDefaultImpl0.forward(ctx, *args, **kwargs)
+
+    @staticmethod
+    def backward(ctx, *args, **kwargs):
+        DistributedDefaultImpl0.backward(ctx, *args, **kwargs)
+
+
+register_distributed_operator_impl("slice",
+                                   DistributedSliceImpl("decrease_in_axis"))

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

@@ -18,5 +18,6 @@ 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_slice MODULES test_dist_slice ENVS ${dist_ENVS})
     py_test_modules(test_cluster MODULES test_cluster ENVS ${dist_ENVS})
 endif()

python/paddle/fluid/tests/unittests/auto_parallel/test_dist_slice.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
+
+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, 5, 6], dtype='float32')
+        auto.shard_tensor(
+            x,
+            dist_attr={
+                "process_mesh": auto.ProcessMesh([0, 1]),
+                "dims_mapping": [0, -1, -1]
+            })
+        tmp_0 = x[0]
+        tmp_1 = x[:, 0, :]
+        tmp_2 = x[:, :, 1]
+        tmp_3 = x[:2, :2, :2]
+    return main_program, start_program
+
+
+def make_program_serial():
+    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, 5, 6], dtype='float32')
+        auto.shard_tensor(
+            x,
+            dist_attr={
+                "process_mesh": auto.ProcessMesh([0]),
+                "dims_mapping": [-1, -1, -1]
+            })
+        tmp_0 = x[0]
+        tmp_1 = x[:, 0, :]
+        tmp_2 = x[:, :, 1]
+        tmp_3 = x[2, 2, :]
+        tmp_4 = x[:2, :2, :2]
+        tmp_5 = x[0, 0, 0]
+    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 TestDistSlice(unittest.TestCase):
+    def test_dist_slice_dp2(self):
+
+        for rank in range(2):
+            dist_main_prog, dist_context = parallelizer(make_program_dp2, rank)
+            ops = dist_main_prog.global_block().ops
+            for op in ops:
+                axes = op.desc.attr('axes')
+                op_dist_attr = dist_context.get_op_dist_attr_for_program(op)
+                if axes[0] == 0:
+                    assert op_dist_attr.impl_type == "default"
+                else:
+                    assert op_dist_attr.impl_type == "slice"
+                    for out in op.output_arg_names:
+                        var_dims_mapping = op_dist_attr.get_output_dims_mapping(
+                            out)
+                        assert var_dims_mapping[0] == 0
+
+    def test_dist_slice_serial(self):
+        dist_main_prog, dist_context = parallelizer(make_program_serial, 0)
+        ops = dist_main_prog.global_block().ops
+        for op in ops:
+            op_dist_attr = dist_context.get_op_dist_attr_for_program(op)
+            assert op_dist_attr.impl_type == "slice"
+            for out in op.output_arg_names:
+                var_dims_mapping = op_dist_attr.get_output_dims_mapping(out)
+                ref_dims_mapping = [-1 for i in range(len(var_dims_mapping))]
+                assert ref_dims_mapping == ref_dims_mapping
+
+
+if __name__ == "__main__":
+    unittest.main()