Merge pull request #7715 from Yancey1989/split_selected_rows_to_multi_pserver

[WIP] Split SelectedRows to multiple pservers

paddle/operators/recv_op.cc

@@ -103,7 +103,7 @@ class RecvOp : public framework::OperatorBase {
 
     // TODO(typhoonzero): change this to a while_op for every cluster-batch.
     bool exit_flag = false;
-    int64_t barrier_size = param_count * fan_in;
+    size_t barrier_size = param_count * fan_in;
     while (!exit_flag) {
       // Get from multiple trainers, we don't care about the order in which
       // the gradients arrives, just add suffix 0~n and merge the gradient.

paddle/operators/split_selected_rows_op.cc

@@ -23,8 +23,6 @@ class SplitSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker {
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input SelectedRows.");
     AddOutput("Out", "The outputs of input SelectedRows.").AsDuplicable();
-    AddAttr<std::vector<int>>("rows_sections", "Rows section for output.")
-        .SetDefault(std::vector<int>({}));
     AddAttr<std::vector<int>>("height_sections",
                               "Height for each output SelectedRows.")
         .SetDefault(std::vector<int>({}));
@@ -35,16 +33,16 @@ height_sections is only needed when need to split the dims of the original tenso
 
 Example:
   Input:
-    X.rows = {0, 7, 5}
+    X.rows = {7, 5}
     X.height = 12
   Attr:
-    rows_sections = {1, 2}
-    height_sections = {}
+    height_sections = {4, 8}
   Out:
-    out0.rows = {0}
-    out0.height = 12
-    out1.rows = {7, 5}
-    out2.height = 12
+    out0.rows = {}
+    out0.height = 4
+
+    out1.rows = {5, 7}
+    out2.height = 8
 
 )DOC");
   }
@@ -61,11 +59,6 @@ class SplitSelectedRowsOp : public framework::OperatorWithKernel {
 
     std::vector<int> height_sections =
         ctx->Attrs().Get<std::vector<int>>("height_sections");
-    std::vector<int> rows_sections =
-        ctx->Attrs().Get<std::vector<int>>("rows_sections");
-    PADDLE_ENFORCE_EQ(
-        rows_sections.size(), ctx->Outputs("Out").size(),
-        "The size of rows section should be the same with Outputs size.");
     int64_t n = ctx->Outputs("Out").size();
 
     std::vector<framework::DDim> outs_dims;

paddle/operators/split_selected_rows_op.h

@@ -16,40 +16,70 @@ limitations under the License. */
 
 #include <vector>
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/selected_rows_functor.h"
 
 namespace paddle {
 namespace operators {
 
+static int FindOutIdx(int row, const std::vector<int>& height_sections) {
+  int offset = 0;
+  for (size_t i = 0; i < height_sections.size(); ++i) {
+    if (row >= offset && row < (offset + height_sections[i])) {
+      return i;
+    }
+    offset += height_sections[i];
+  }
+  return -1;
+}
+
 template <typename DeviceContext, typename T>
 class SplitSelectedRowsOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* x = ctx.Input<framework::SelectedRows>("X");
     auto outs = ctx.MultiOutput<framework::SelectedRows>("Out");
-
-    auto rows_sections = ctx.Attr<std::vector<int>>("rows_sections");
     auto height_sections = ctx.Attr<std::vector<int>>("height_sections");
 
-    int64_t n = outs.size();
-    int offset = 0;
+    auto x_rows = x->rows();
+    std::vector<std::vector<int>> outs_rows_idx;
+    outs_rows_idx.resize(outs.size());
+
+    auto row_numel = x->value().numel() / x->value().dims()[0];
+    auto src = x->value().data<T>();
 
-    for (int64_t i = 0; i < n; ++i) {
-      framework::Vector<int64_t> out_rows;
-      for (int64_t j = 0; j < rows_sections[i]; ++j) {
-        out_rows.push_back(x->rows()[offset + j]);
+    for (size_t i = 0; i < x_rows.size(); ++i) {
+      int out_idx = FindOutIdx(x_rows[i], height_sections);
+      outs_rows_idx[out_idx].push_back(i);
     }
+    auto place = ctx.GetPlace();
 
-      auto& out = outs[i];
-      auto x_dims = x->GetCompleteDims();
-      x_dims[0] = rows_sections[i];
-      out->mutable_value()->mutable_data<T>(x_dims, ctx.GetPlace());
-      framework::Copy(x->value().Slice(offset, rows_sections[i] + offset),
-                      x->place(), ctx.device_context(), out->mutable_value());
-      outs[i]->set_rows(out_rows);
-      if (height_sections.size()) {
-        outs[i]->set_height(height_sections[i]);
+    for (size_t i = 0; i < outs_rows_idx.size(); ++i) {
+      auto rows_idx = outs_rows_idx[i];
+      if (rows_idx.size() > 0) {
+        auto dims = x->GetCompleteDims();
+        dims[0] = rows_idx.size();
+        outs[i]->mutable_value()->mutable_data<T>(dims, x->place());
+        for (auto idx : rows_idx) {
+          outs[i]->mutable_rows()->push_back(x_rows[idx]);
+        }
+        auto dst = outs[i]->mutable_value()->mutable_data<T>(ctx.GetPlace());
+        for (size_t j = 0; j < rows_idx.size(); j++) {
+          if (platform::is_cpu_place(place)) {
+            memory::Copy(platform::CPUPlace(), dst + j * row_numel,
+                         platform::CPUPlace(), src + rows_idx[j] * row_numel,
+                         sizeof(T) * row_numel);
+          } else {
+#ifdef PADDLE_WITH_CUDA
+            auto stream = ctx.cuda_device_context().stream();
+            memory::Copy(platform::CUDAPlace(), dst + j * row_numel,
+                         platform::CUDAPlace(), src + rows_idx[j] * row_numel,
+                         sizeof(T) * row_numel, stream);
+#else
+            PADDLE_THROW("Paddle is not compiled with GPU");
+#endif
+          }
+        }
       }
-      offset += rows_sections[i];
     }
   }
 };

python/paddle/v2/fluid/distribute_transpiler.py

@@ -19,6 +19,7 @@ import optimizer
 from layer_helper import LayerHelper
 from distributed_spliter import *
 import math
+from . import core
 
 
 class VarBlock:
@@ -217,6 +218,16 @@ class DistributeTranspiler:
             if len(splited_vars) <= 1:
                 continue
             orig_var = program.global_block().vars[varname]
+            if orig_var == core.VarDesc.VarType.SELECTED_ROWS:
+                height_sections = []
+                for v in splited_vars:
+                    height_sections.append(v.shape[0])
+                program.global_block().append_op(
+                    type="split_selected_rows",
+                    inputs={"X": orig_var},
+                    outputs={"Out": splited_vars},
+                    attrs={"height_sections": height_sections})
+            elif orig_var == core.VarDesc.VarType.LOD_TENSOR:
                 sections = []
                 for v in splited_vars:
                     sections.append(v.shape[0])
@@ -226,6 +237,9 @@ class DistributeTranspiler:
                     outputs={"Out": splited_vars},
                     attrs={"sections": sections}  # assume split evenly
                 )
+            else:
+                AssertionError("Variable type should be in set "
+                               "[LOD_TENSOR, SELECTED_ROWS]")
         return var_mapping
 
     def get_trainer_program(self):

python/paddle/v2/fluid/tests/test_split_selected_rows_op.py

@@ -35,8 +35,8 @@ class TestSpliteSelectedRows(unittest.TestCase):
 
     def check_with_place(self, place):
         scope = core.Scope()
-        rows = [0, 5, 7, 4]
-        height = 10
+        rows = [0, 5, 7, 4, 20]
+        height = 20
         row_numel = 2
 
         # initialize input variable X
@@ -46,38 +46,41 @@ class TestSpliteSelectedRows(unittest.TestCase):
         np_array = np.ones((len(rows), row_numel)).astype("float32")
         np_array[0, 0] = 2.0
         np_array[2, 1] = 4.0
+        np_array[4, 1] = 8.0
         x_tensor = x.get_tensor()
         x_tensor.set(np_array, place)
 
-        rows_sections = [2, 2]
-        height_sections = []
+        height_sections = [5, 5, 5, 5, 3]
 
         # initialize output variables [out0, out1]
-        out0 = scope.var('out0').get_selected_rows()
-        out1 = scope.var('out1').get_selected_rows()
+        outs_name = ["out%d" % i for i in xrange(len(height_sections))]
+        outs = [
+            scope.var(var_name).get_selected_rows() for var_name in outs_name
+        ]
 
         # expected output selected rows
-        expected_out0_rows = [0, 5]
-        expected_out1_rows = [7, 4]
-        expected_height = height
+        expected_out0_rows = [0, 4]
+        expected_out1_rows = [5, 7]
+        expected_out4_rows = [20]
 
         op = Operator(
             "split_selected_rows",
             X="X",
-            Out=["out0", "out1"],
-            rows_sections=rows_sections,
+            Out=outs_name,
             height_sections=height_sections)
 
         op.run(scope, place)
 
-        self.assertEqual(out0.rows(), expected_out0_rows)
-        self.assertEqual(out1.rows(), expected_out1_rows)
+        self.assertEqual(outs[0].rows(), expected_out0_rows)
+        self.assertEqual(outs[1].rows(), expected_out1_rows)
+        self.assertEqual(outs[4].rows(), expected_out4_rows)
 
-        self.assertEqual(out0.height(), expected_height)
-        self.assertEqual(out1.height(), expected_height)
+        self.assertEqual(outs[0].height(), height_sections[0])
+        self.assertEqual(outs[4].height(), height_sections[4])
 
-        self.assertAlmostEqual(2.0, np.array(out0.get_tensor())[0, 0])
-        self.assertAlmostEqual(4.0, np.array(out1.get_tensor())[0, 1])
+        self.assertAlmostEqual(2.0, np.array(outs[0].get_tensor())[0, 0])
+        self.assertAlmostEqual(4.0, np.array(outs[1].get_tensor())[1, 1])
+        self.assertAlmostEqual(8.0, np.array(outs[4].get_tensor())[0, 1])
 
     def check_grad_with_place(self, place):
         scope = core.Scope()
@@ -85,8 +88,7 @@ class TestSpliteSelectedRows(unittest.TestCase):
         row_numel = 2
 
         # attr
-        rows_sections = [2, 2]
-        height_sections = []
+        height_sections = [5, 5]
 
         # initialize input variable X
         out0_grad = scope.var("out0@GRAD").get_selected_rows()
@@ -113,7 +115,6 @@ class TestSpliteSelectedRows(unittest.TestCase):
             "sum",
             X=["out0@GRAD", "out1@GRAD"],
             Out="X@GRAD",
-            rows_sections=rows_sections,
             height_sections=height_sections)
 
         grad_op.run(scope, place)