support elementwise backward rule (PaddlePaddle#57813)

paddle/phi/api/lib/data_transform.cc

@@ -652,18 +652,9 @@ ReshardApiInputToReplicatedKernelInput(
     if (ReshardIsNeeded(dist_tensor->dist_attr(), dist_attr)) {
       VLOG(6) << "ApiIn to Replicated KernelIn - "
               << ReshardDebugInfo(*dist_tensor, dist_attr);
-      if (dist_tensor->initialized()) {
-        auto* func = phi::distributed::ChooseProperReshardFunction(*dist_tensor,
-                                                                   dist_attr);
-        return func->Eval(dev_ctx, *dist_tensor, dist_attr);
-      } else {
-        // when no tensor data need to be reshard, we still need to set correct
-        // replicated dist attr and local dims for output
-        dist_tensor->unsafe_set_dist_attr(dist_attr);
-        auto dense_tensor_meta = dist_tensor->value().meta();
-        dense_tensor_meta.dims = dist_tensor->dims();
-        dist_tensor->unsafe_mutable_value()->set_meta(dense_tensor_meta);
-      }
+      auto* func = phi::distributed::ChooseProperReshardFunction(*dist_tensor,
+                                                                 dist_attr);
+      return func->Eval(dev_ctx, *dist_tensor, dist_attr);
     }
     return std::static_pointer_cast<phi::distributed::DistTensor>(tensor_in);
   }

paddle/phi/api/yaml/legacy_backward.yaml

@@ -19,6 +19,7 @@
   infer_meta :
     func : GeneralBinaryGradInferMeta
     param : [x, y]
+    spmd_rule : ElementwiseBinaryGradInferSpmd
   kernel :
     func : add_grad
   no_need_buffer : x, y
@@ -680,6 +681,7 @@
   infer_meta :
     func : GeneralBinaryGradInferMeta
     param : [x, y]
+    spmd_rule : ElementwiseBinaryGradInferSpmd
   kernel :
     func : subtract_grad
   no_need_buffer : x, y

paddle/phi/api/yaml/legacy_ops.yaml

@@ -17,6 +17,7 @@
   output : Tensor(out)
   infer_meta :
     func : ElementwiseInferMeta
+    spmd_rule : ElementwiseBinaryInferSpmd
   kernel :
     func : add
   inplace : (x -> out)
@@ -1003,6 +1004,7 @@
   output : Tensor(out)
   infer_meta :
     func : ElementwiseInferMeta
+    spmd_rule : ElementwiseBinaryInferSpmd
   kernel :
     func : subtract
   inplace : (x -> out)

paddle/phi/infermeta/spmd_rules/elementwise.cc

@@ -309,5 +309,72 @@ SpmdInfo ElementwiseBinaryInferSpmdReverse(const DistMetaTensor& x,
   return {{x_dist_attr_dst, y_dist_attr_dst}, {out_dist_attr}};
 }
 
+SpmdInfo ElementwiseBinaryGradInferSpmd(const DistMetaTensor& x,
+                                        const DistMetaTensor& y,
+                                        const DistMetaTensor& out_grad,
+                                        int64_t axis) {
+  TensorDistAttr x_dist_attr = out_grad.dist_attr();
+  TensorDistAttr y_dist_attr = out_grad.dist_attr();
+  TensorDistAttr x_grad_dist_attr = out_grad.dist_attr();
+  TensorDistAttr y_grad_dist_attr = out_grad.dist_attr();
+
+  PADDLE_ENFORCE_GE(
+      out_grad.dims().size(),
+      x.dims().size(),
+      phi::errors::InvalidArgument("If being broadcast, the dims of out_grad "
+                                   "must larger or equal to the inputs."
+                                   "But we get the rank of output as [%d] and "
+                                   "the rank of input as [%d].",
+                                   out_grad.dims().size(),
+                                   x.dims().size()));
+
+  PADDLE_ENFORCE_GE(
+      out_grad.dims().size(),
+      y.dims().size(),
+      phi::errors::InvalidArgument("If being broadcast, the dims of out_grad "
+                                   "must larger or equal to the inputs."
+                                   "But we get the rank of output as [%d] and "
+                                   "the rank of input as [%d].",
+                                   out_grad.dims().size(),
+                                   y.dims().size()));
+
+  // The backward rule of elementwise follows the princple: the dist_attr
+  // of input should equal to out_grad.
+  // Caution the special case when the inputs calculate together with different
+  // shape it means one of the input is broadcast to same shape with the other
+  // first. When doing backward the input_grad with broadcast input is in
+  // partial status, which need to do communicate and get the right result.
+  if (x.dims() != out_grad.dims()) {
+    int64_t diff = out_grad.dims().size() - x.dims().size();
+    auto dims_mapping = x_dist_attr.dims_mapping();
+    dims_mapping.erase(dims_mapping.begin(), dims_mapping.begin() + diff);
+    x_dist_attr.set_dims_mapping(dims_mapping);
+    x_grad_dist_attr.set_dims_mapping(dims_mapping);
+    for (int64_t i = 0; i < diff; ++i) {
+      if (out_grad.dist_attr().dims_mapping()[i] != -1) {
+        x_grad_dist_attr.set_partial_status(
+            std::vector<int64_t>{out_grad.dist_attr().dims_mapping()[i]});
+      }
+    }
+  }
+
+  if (y.dims() != out_grad.dims()) {
+    int64_t diff = out_grad.dims().size() - y.dims().size();
+    auto dims_mapping = y_dist_attr.dims_mapping();
+    dims_mapping.erase(dims_mapping.begin(), dims_mapping.begin() + diff);
+    y_dist_attr.set_dims_mapping(dims_mapping);
+    y_grad_dist_attr.set_dims_mapping(dims_mapping);
+    for (int64_t i = 0; i < diff; ++i) {
+      if (out_grad.dist_attr().dims_mapping()[i] != -1) {
+        y_grad_dist_attr.set_partial_status(
+            std::vector<int64_t>{out_grad.dist_attr().dims_mapping()[i]});
+      }
+    }
+  }
+
+  return {{x_dist_attr, y_dist_attr, out_grad.dist_attr()},
+          {x_grad_dist_attr, y_grad_dist_attr}};
+}
+
 }  // namespace distributed
 }  // namespace phi

paddle/phi/infermeta/spmd_rules/elementwise.h

@@ -34,5 +34,10 @@ SpmdInfo ElementwiseBinaryInferSpmdReverse(const DistMetaTensor& x,
                                            const DistMetaTensor& y,
                                            const DistMetaTensor& out);
 
+SpmdInfo ElementwiseBinaryGradInferSpmd(const DistMetaTensor& x,
+                                        const DistMetaTensor& y,
+                                        const DistMetaTensor& out_grad,
+                                        int64_t axis);
+
 }  // namespace distributed
 }  // namespace phi

paddle/phi/kernels/all_gather_kernel.h

@@ -34,7 +34,9 @@ void AllGather(const Context& dev_ctx,
   MetaTensor* out_meta_ptr = &out_meta;
 
   AllGatherInferMeta(phi::MetaTensor(x), nranks, out_meta_ptr);
-  AllGatherKernel<T, Context>(dev_ctx, x, nranks, out);
+  if (x.initialized()) {
+    AllGatherKernel<T, Context>(dev_ctx, x, nranks, out);
+  }
 }
 
 }  // namespace phi

paddle/phi/kernels/all_reduce_kernel.h

@@ -35,7 +35,9 @@ void AllReduce(const Context& dev_ctx,
   MetaTensor* out_meta_ptr = &out_meta;
 
   AllReduceInferMeta(phi::MetaTensor(x), out_meta_ptr);
-  AllReduceKernel<T, Context>(dev_ctx, x, reduce_type, out);
+  if (x.initialized()) {
+    AllReduceKernel<T, Context>(dev_ctx, x, reduce_type, out);
+  }
 }
 
 }  // namespace phi

paddle/phi/kernels/all_to_all_kernel.h

@@ -30,7 +30,9 @@ void AllToAll(const Context& dev_ctx, const DenseTensor& x, DenseTensor* out) {
   MetaTensor* out_meta_ptr = &out_meta;
 
   AllToAllInferMeta(phi::MetaTensor(x), out_meta_ptr);
-  AllToAllKernel<T, Context>(dev_ctx, x, out);
+  if (x.initialized()) {
+    AllToAllKernel<T, Context>(dev_ctx, x, out);
+  }
 }
 
 }  // namespace phi

paddle/phi/kernels/concat_kernel.h

@@ -41,7 +41,10 @@ void Concat(const Context& dev_ctx,
 
   MetaTensor meta_out(dense_out);
   ConcatInferMeta(meta_x_ptr, axis.to<int>(), &meta_out);
-  ConcatKernel<T, Context>(dev_ctx, x, axis, dense_out);
+
+  if (x[0]->initialized()) {
+    ConcatKernel<T, Context>(dev_ctx, x, axis, dense_out);
+  }
 }
 
 template <typename T, typename Context>

paddle/phi/kernels/cpu/svd_kernel.cc

@@ -98,7 +98,6 @@ void SvdKernel(const Context& dev_ctx,
   /*Create Tensors and output, set the dim ...*/
   auto numel = X.numel();
   DenseTensor trans_x = ::phi::TransposeLast2Dim<T>(dev_ctx, X);
-  auto* x_data = trans_x.data<T>();
   auto x_dims = X.dims();
   int rows = static_cast<int>(x_dims[x_dims.size() - 2]);
   int cols = static_cast<int>(x_dims[x_dims.size() - 1]);
@@ -113,6 +112,7 @@ void SvdKernel(const Context& dev_ctx,
       0,
       cols,
       errors::InvalidArgument("The col of Input(X) should be greater than 0."));
+  auto* x_data = trans_x.data<T>();
   int batches = static_cast<int>(numel / (rows * cols));
   auto* U_out = dev_ctx.template Alloc<phi::dtype::Real<T>>(U);
   auto* VH_out = dev_ctx.template Alloc<phi::dtype::Real<T>>(VH);

paddle/phi/kernels/reshape_kernel.h

@@ -48,7 +48,9 @@ void Reshape(const Context& dev_ctx,
              DenseTensor* out) {
   MetaTensor meta_out(out);
   InferMetaFromVecValue(x, shape, &meta_out);
-  ReshapeInferKernel<Context>(dev_ctx, x, IntArray(shape), out);
+  if (x.initialized()) {
+    ReshapeInferKernel<Context>(dev_ctx, x, IntArray(shape), out);
+  }
 }
 
 template <typename T, typename Context>

paddle/phi/kernels/split_kernel.h

@@ -74,7 +74,9 @@ void Split(const Context& dev_ctx,
     outs.push_back(&result->at(i));
   }
 
-  SplitKernel<T, Context>(dev_ctx, x, sections, axis, outs);
+  if (x.initialized()) {
+    SplitKernel<T, Context>(dev_ctx, x, sections, axis, outs);
+  }
 }
 
 template <typename T, typename Context>

paddle/phi/kernels/transpose_kernel.h

@@ -19,6 +19,7 @@
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/infermeta/unary.h"
 #include "paddle/phi/kernels/empty_kernel.h"
+
 namespace phi {
 
 template <typename T, typename Context>
@@ -43,7 +44,9 @@ void Transpose(const Context& dev_ctx,
 
   // do not call TransposeStridedKernel, because some other kernels call
   // Transpose directly
-  TransposeKernel<T, Context>(dev_ctx, x, axis, dense_out);
+  if (x.initialized()) {
+    TransposeKernel<T, Context>(dev_ctx, x, axis, dense_out);
+  }
 }
 
 template <typename T, typename Context>

test/auto_parallel/semi_auto_parallel_for_elementwise.py

@@ -0,0 +1,147 @@
+# 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
+
+
+class TestElementwiseApiForSemiAutoParallel:
+    def __init__(self):
+        self._dtype = os.getenv("dtype")
+        self._backend = os.getenv("backend")
+        self._seed = eval(os.getenv("seed"))
+        self._mesh = dist.ProcessMesh([0, 1], dim_names=["x"])
+
+    def check_tensor_eq(self, a, b):
+        np1 = a.numpy()
+        np2 = b.numpy()
+        np.testing.assert_allclose(np1, np2, rtol=1e-05, verbose=True)
+
+    def test_binary_body(
+        self, x_shape, y_shape, out_shape, x_specs, y_specs, binary_func
+    ):
+        paddle.seed(self._seed)
+        np.random.seed(self._seed)
+
+        x = paddle.randn(x_shape, self._dtype)
+        y = paddle.randn(y_shape, self._dtype)
+        x.stop_gradient = False
+        y.stop_gradient = False
+
+        x_dist_attr = dist.DistAttr(mesh=self._mesh, sharding_specs=x_specs)
+        y_dist_attr = dist.DistAttr(mesh=self._mesh, sharding_specs=y_specs)
+
+        dist_x = dist.shard_tensor(x, dist_attr=x_dist_attr)
+        dist_y = dist.shard_tensor(y, dist_attr=y_dist_attr)
+        dist_x.stop_gradient = False
+        dist_y.stop_gradient = False
+
+        dist_out = binary_func(dist_x, dist_y)
+        out = binary_func(x, y)
+        self.check_tensor_eq(out, dist_out)
+
+        dist_out.backward()
+        out.backward()
+        self.check_tensor_eq(x.grad, dist_x.grad)
+        self.check_tensor_eq(y.grad, dist_y.grad)
+
+    def test_add_x_shard(self):
+        self.test_binary_body(
+            x_shape=[16, 32],
+            y_shape=[16, 32],
+            out_shape=[16, 32],
+            x_specs=['x', None],
+            y_specs=[None, None],
+            binary_func=paddle.add,
+        )
+
+    def test_sub_x_shard(self):
+        self.test_binary_body(
+            x_shape=[16, 32],
+            y_shape=[16, 32],
+            out_shape=[16, 32],
+            x_specs=['x', None],
+            y_specs=[None, None],
+            binary_func=paddle.subtract,
+        )
+
+    def test_add_x_shard_broadcast(self):
+        self.test_binary_body(
+            x_shape=[16, 32],
+            y_shape=[2, 16, 32],
+            out_shape=[2, 16, 32],
+            x_specs=['x', None],
+            y_specs=[None, None, None],
+            binary_func=paddle.add,
+        )
+
+    def test_add_x_y_shard(self):
+        if self._backend == "cpu":
+            return
+
+        self.test_binary_body(
+            x_shape=[16, 32],
+            y_shape=[16, 32],
+            out_shape=[16, 32],
+            x_specs=['x', None],
+            y_specs=[None, 'x'],
+            binary_func=paddle.add,
+        )
+
+    def test_add_x_y_shard_broadcast(self):
+        if self._backend == "cpu":
+            return
+
+        self.test_binary_body(
+            x_shape=[4, 16, 32],
+            y_shape=[16, 32],
+            out_shape=[4, 16, 32],
+            x_specs=['x', None, None],
+            y_specs=[None, None],
+            binary_func=paddle.add,
+        )
+
+    def test_sub_x_y_shard_broadcast(self):
+        if self._backend == "cpu":
+            return
+
+        self.test_binary_body(
+            x_shape=[4, 16, 32],
+            y_shape=[16, 32],
+            out_shape=[4, 16, 32],
+            x_specs=['x', None, None],
+            y_specs=[None, None],
+            binary_func=paddle.subtract,
+        )
+
+    def run_test_case(self):
+        if self._backend == "cpu":
+            paddle.set_device("cpu")
+        elif self._backend == "gpu":
+            paddle.set_device("gpu:" + str(dist.get_rank()))
+        else:
+            raise ValueError("Only support cpu or gpu backend.")
+
+        self.test_add_x_shard()
+        self.test_add_x_shard_broadcast()
+        self.test_add_x_y_shard()
+        self.test_add_x_y_shard_broadcast()
+
+
+if __name__ == '__main__':
+    TestElementwiseApiForSemiAutoParallel().run_test_case()