【Hackathon 5th No.123】 Support r to p on cross mesh

paddle/fluid/pybind/auto_parallel_py.cc

@@ -207,6 +207,10 @@ void BindAutoParallel(py::module *m) {
       *m, "RToPReshardFunction", ReshardFunction)
       .def(py::init<>());
 
+  py::class_<phi::distributed::RToPReshardFunctionCrossMesh>(
+      *m, "RToPReshardFunctionCrossMesh", ReshardFunction)
+      .def(py::init<>());
+
   py::class_<phi::distributed::PToRReshardFunction>(
       *m, "PToRReshardFunction", ReshardFunction)
       .def(py::init<>());

paddle/phi/core/distributed/auto_parallel/reshard/r_to_p_reshard_function.cc

@@ -19,6 +19,7 @@
 #include "paddle/phi/core/distributed/auto_parallel/dist_attr.h"
 #include "paddle/phi/core/distributed/auto_parallel/dist_tensor.h"
 #include "paddle/phi/core/distributed/auto_parallel/reshard/reshard_utils.h"
+#include "paddle/phi/core/distributed/auto_parallel/reshard/same_status_reshard_function.h"
 #include "paddle/phi/kernels/assign_kernel.h"
 #include "paddle/phi/kernels/full_kernel.h"
 
@@ -70,5 +71,48 @@ void RToPReshardFunction::Eval(phi::DeviceContext* dev_ctx,
   SetDistProps(out, in.dims(), out_dist_attr);
 }
 
+bool RToPReshardFunctionCrossMesh::IsSuitable(
+    const DistTensor& in, const TensorDistAttr& out_dist_attr) {
+  const auto& in_dist_attr = in.dist_attr();
+
+  RESHARD_SHORTCUT_IF_FALSE(in_dist_attr.is_replicated());
+  RESHARD_SHORTCUT_IF_FALSE(out_dist_attr.is_partial());
+
+  const auto& in_process_mesh = in_dist_attr.process_mesh();
+  const auto& out_process_mesh = out_dist_attr.process_mesh();
+
+  RESHARD_SHORTCUT_IF_FALSE(in_process_mesh.ndim() == 1);
+  RESHARD_SHORTCUT_IF_FALSE(out_process_mesh.ndim() == 1);
+  RESHARD_SHORTCUT_IF_FALSE(in_process_mesh.shape() ==
+                            out_process_mesh.shape());
+  RESHARD_SHORTCUT_IF_FALSE(in_process_mesh != out_process_mesh);
+
+  return true;
+}
+
+void RToPReshardFunctionCrossMesh::Eval(phi::DeviceContext* dev_ctx,
+                                        const DistTensor& in,
+                                        const TensorDistAttr& out_dist_attr,
+                                        DistTensor* out) {
+  VLOG(3) << "Call RToPReshardFunctionCrossMesh Eval";
+  const auto& out_process_mesh = out_dist_attr.process_mesh();
+
+  DistTensor tmp_result;
+
+  SameStatusReshardFunction same_status_func;
+  TensorDistAttr tmp_dist_attr = in.dist_attr();
+  tmp_dist_attr.set_process_mesh(out_process_mesh);
+  same_status_func.Eval(dev_ctx, in, tmp_dist_attr, &tmp_result);
+
+  RToPReshardFunction r_to_p_func;
+  PADDLE_ENFORCE(
+      r_to_p_func.IsSuitable(tmp_result, out_dist_attr),
+      phi::errors::InvalidArgument(
+          "Invoke the r to p reshard function is not valid from %s to %s.",
+          tmp_result.dist_attr(),
+          out_dist_attr));
+  r_to_p_func.Eval(dev_ctx, tmp_result, out_dist_attr, out);
+}
+
 }  // namespace distributed
 }  // namespace phi

paddle/phi/core/distributed/auto_parallel/reshard/r_to_p_reshard_function.h

@@ -32,5 +32,18 @@ class RToPReshardFunction final : public ReshardFunction {
   std::string Name() override { return "RToPReshard"; }
 };
 
+class RToPReshardFunctionCrossMesh final : public ReshardFunction {
+ public:
+  bool IsSuitable(const DistTensor& in,
+                  const TensorDistAttr& out_dist_attr) override;
+
+  void Eval(DeviceContext* dev_ctx,
+            const DistTensor& in,
+            const TensorDistAttr& out_dist_attr,
+            DistTensor* out) override;
+
+  std::string Name() override { return "RToPReshardCrossMesh"; }
+};
+
 }  // namespace distributed
 }  // namespace phi

test/auto_parallel/reshard_r_to_p_cross_mesh.py

@@ -0,0 +1,75 @@
+# Copyright (c) 2023 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 os
+
+import numpy as np
+
+import paddle
+import paddle.distributed as dist
+from paddle.base import core
+
+
+class TestReshardRToPCrossMesh:
+    def __init__(self):
+        self._shape = eval(os.getenv("shape"))
+        self._dtype = os.getenv("dtype")
+        self._seeds = eval(os.getenv("seeds"))
+        self._shard = eval(os.getenv("shard"))
+        self._backend = os.getenv("backend")
+        self._in_mesh = dist.ProcessMesh([0, 1], dim_names=["x"])
+        self._out_mesh = dist.ProcessMesh([1, 0], dim_names=["x"])
+
+    def run_test_case(self):
+        # cpu does not support send/recv
+        if self._backend == "cpu":
+            return
+        elif self._backend == "gpu":
+            place = paddle.CUDAPlace(dist.get_rank())
+
+        dev_ctx = core.DeviceContext.create(place)
+        a = paddle.ones(self._shape)
+
+        in_shard_specs = [None for i in range(len(self._shape))]
+        out_shard_specs = [None for i in range(len(self._shape))]
+
+        dist_attr = dist.DistAttr(
+            mesh=self._in_mesh, sharding_specs=in_shard_specs
+        )
+        out_dist_attr = dist.DistAttr(
+            mesh=self._out_mesh, sharding_specs=out_shard_specs
+        )
+        out_dist_attr._set_partial_dims([0])
+
+        input_tensor = dist.shard_tensor(a, dist_attr=dist_attr)
+
+        reshard_func = core.RToPReshardFunctionCrossMesh()
+        assert reshard_func.is_suitable(input_tensor, out_dist_attr)
+
+        out = reshard_func.eval(dev_ctx, input_tensor, out_dist_attr)
+
+        if dist.get_rank() == 1:
+            np.testing.assert_equal(
+                out._local_value().numpy(), input_tensor.numpy()
+            )
+        else:
+            zeros = paddle.zeros(self._shape)
+            np.testing.assert_equal(out._local_value().numpy(), zeros.numpy())
+
+        assert np.equal(out.shape, input_tensor.shape).all()
+        assert np.equal(out._local_shape, input_tensor._local_shape).all()
+
+
+if __name__ == '__main__':
+    TestReshardRToPCrossMesh().run_test_case()

test/auto_parallel/test_reshard_r_to_p.py

@@ -23,10 +23,11 @@ def setUp(self):
         self._default_envs = {
             "shape": "(10, 20)",
             "dtype": "float32",
-            "seeds": str(self._seeds),
+            "seeds": "2023",
         }
         self._changeable_envs = {
             "shape": ["(10, 20)"],
+            "shard": ["0", "1"],
             "backend": ["cpu", "gpu"],
         }
 
@@ -40,6 +41,17 @@ def test_reshard_r_to_p(self):
                 user_defined_envs=envs,
             )
 
+    def test_reshard_r_to_p_cross_mesh(self):
+        envs_list = test_base.gen_product_envs_list(
+            self._default_envs, self._changeable_envs
+        )
+        for envs in envs_list:
+            if envs["backend"] != "cpu":
+                self.run_test_case(
+                    "reshard_r_to_p_cross_mesh.py",
+                    user_defined_envs=envs,
+                )
+
 
 if __name__ == "__main__":
     unittest.main()