Merge pull request #12295 from jacquesqiao/speedup-reduce-sum-grad-op

Speedup reduce sum grad op

paddle/fluid/operators/reduce_sum_op.cc

@@ -23,12 +23,13 @@ REGISTER_OP_CPU_KERNEL(
     ops::ReduceKernel<paddle::platform::CPUDeviceContext, int, ops::SumFunctor>,
     ops::ReduceKernel<paddle::platform::CPUDeviceContext, int64_t,
                       ops::SumFunctor>);
-REGISTER_OP_CPU_KERNEL(reduce_sum_grad,
-                       ops::ReduceGradKernel<paddle::platform::CPUDeviceContext,
-                                             float, ops::SumGradFunctor>,
-                       ops::ReduceGradKernel<paddle::platform::CPUDeviceContext,
-                                             double, ops::SumGradFunctor>,
-                       ops::ReduceGradKernel<paddle::platform::CPUDeviceContext,
-                                             int, ops::SumGradFunctor>,
-                       ops::ReduceGradKernel<paddle::platform::CPUDeviceContext,
-                                             int64_t, ops::SumGradFunctor>);
+REGISTER_OP_CPU_KERNEL(
+    reduce_sum_grad,
+    ops::ReduceSumGradKernel<paddle::platform::CPUDeviceContext, float,
+                             ops::SumGradFunctor>,
+    ops::ReduceSumGradKernel<paddle::platform::CPUDeviceContext, double,
+                             ops::SumGradFunctor>,
+    ops::ReduceSumGradKernel<paddle::platform::CPUDeviceContext, int,
+                             ops::SumGradFunctor>,
+    ops::ReduceSumGradKernel<paddle::platform::CPUDeviceContext, int64_t,
+                             ops::SumGradFunctor>);

paddle/fluid/operators/reduce_sum_op.h

@@ -14,11 +14,69 @@
 
 #pragma once
 
+#include <vector>
+
 #include "paddle/fluid/operators/reduce_op.h"
 
 namespace paddle {
 namespace operators {
 
+// use for loop to speed up Eigen broadcast. 4 timer faster then broadcast
+template <typename DeviceContext, typename T, typename Functor>
+class ReduceSumGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto dims = context.Attr<std::vector<int>>("dim");
+    if (context.GetPlace().type() == typeid(platform::CPUPlace) &&
+        dims.size() == 1) {
+      auto* input0 = context.Input<Tensor>("X");
+      auto* input2 = context.Input<Tensor>(framework::GradVarName("Out"));
+      auto* output = context.Output<Tensor>(framework::GradVarName("X"));
+      output->mutable_data<T>(context.GetPlace());
+      const auto* input2_d = input2->data<T>();
+      auto* output_d = output->data<T>();
+
+      // handle reduce_all
+      if (input2->dims().size() == 1 && input2->dims()[0] == 1) {
+        for (int64_t i = 0; i < framework::product(input0->dims()); ++i) {
+          output_d[i] = input2_d[0];
+        }
+        return;
+      }
+
+      // handle reduce by one dimension
+      int reduce_dim_index = dims[0];
+      if (reduce_dim_index < 0) {
+        reduce_dim_index += input0->dims().size();
+      }
+
+      auto& input_dim = input0->dims();
+      int64_t before_dim = 1;
+      for (int i = 0; i < reduce_dim_index; ++i) {
+        before_dim *= input_dim[i];
+      }
+      int64_t reduce_dim = input_dim[reduce_dim_index];
+      int64_t after_dim = 1;
+      for (int i = reduce_dim_index + 1; i < input_dim.size(); ++i) {
+        after_dim *= input_dim[i];
+      }
+      for (int64_t i = 0; i < before_dim; ++i) {
+        for (int64_t j = 0; j < reduce_dim; ++j) {
+          for (int64_t k = 0; k < after_dim; ++k) {
+            output_d[i * reduce_dim * after_dim + j * after_dim + k] =
+                input2_d[i * after_dim + k];
+          }
+        }
+      }
+      return;
+    }
+
+    // default use Eigen broadcast
+    ReduceGradKernel<DeviceContext, T, Functor> kernel;
+    kernel.Compute(context);
+  }
+};
+
 struct SumFunctor {
   template <typename DeviceContext, typename X, typename Y, typename Dim>
   void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) {
@@ -31,7 +89,7 @@ struct SumGradFunctor {
             typename DY, typename Dim>
   void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy,
                   const Dim& dim, int size) {
-    dx->device(place) = dy->broadcast(dim);
+    dx->device(place) = dy->eval().broadcast(dim);
   }
 };
 

python/paddle/fluid/layers/nn.py

@@ -2961,7 +2961,7 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None):
             # x is a Tensor variable with following elements:
             #    [[0.2, 0.3, 0.5, 0.9]
             #     [0.1, 0.2, 0.6, 0.7]]
-            # Each example is followed by the correspending output tensor.
+            # Each example is followed by the corresponding output tensor.
             fluid.layers.reduce_sum(x)  # [3.5]
             fluid.layers.reduce_sum(x, dim=0)  # [0.3, 0.5, 1.1, 1.6]
             fluid.layers.reduce_sum(x, dim=-1)  # [1.9, 1.6]
@@ -2970,7 +2970,7 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None):
             # x is a Tensor variable with shape [2, 2, 2] and elements as below:
             #      [[[1, 2], [3, 4]],
             #      [[5, 6], [7, 8]]]
-            # Each example is followed by the correspending output tensor.
+            # Each example is followed by the corresponding output tensor.
             fluid.layers.reduce_sum(x, dim=[1, 2]) # [10, 26]
             fluid.layers.reduce_sum(x, dim=[0, 1]) # [16, 20]
 

python/paddle/fluid/tests/unittests/test_reduce_op.py

@@ -89,15 +89,11 @@ class TestProdOp(OpTest):
         self.check_grad(['X'], 'Out')
 
 
-class TestKeepDimReduce(OpTest):
+class Test1DReduce(OpTest):
     def setUp(self):
         self.op_type = "reduce_sum"
-        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
-        self.attrs = {'dim': [-2], 'keep_dim': True}
-        self.outputs = {
-            'Out':
-            self.inputs['X'].sum(axis=tuple(self.attrs['dim']), keepdims=True)
-        }
+        self.inputs = {'X': np.random.random(20).astype("float64")}
+        self.outputs = {'Out': self.inputs['X'].sum(axis=0)}
 
     def test_check_output(self):
         self.check_output()
@@ -106,32 +102,82 @@ class TestKeepDimReduce(OpTest):
         self.check_grad(['X'], 'Out')
 
 
-class Test1DReduce(OpTest):
+class Test2DReduce0(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_sum"
-        self.inputs = {'X': np.random.random(20).astype("float64")}
+        self.attrs = {'dim': [0]}
+        self.inputs = {'X': np.random.random((20, 10)).astype("float64")}
         self.outputs = {'Out': self.inputs['X'].sum(axis=0)}
 
-    def test_check_output(self):
-        self.check_output()
 
-    def test_check_grad(self):
-        self.check_grad(['X'], 'Out')
+class Test2DReduce1(Test1DReduce):
+    def setUp(self):
+        self.op_type = "reduce_sum"
+        self.attrs = {'dim': [1]}
+        self.inputs = {'X': np.random.random((20, 10)).astype("float64")}
+        self.outputs = {
+            'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim']))
+        }
 
 
-class TestReduceAll(OpTest):
+class Test3DReduce0(Test1DReduce):
+    def setUp(self):
+        self.op_type = "reduce_sum"
+        self.attrs = {'dim': [1]}
+        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")}
+        self.outputs = {
+            'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim']))
+        }
+
+
+class Test3DReduce1(Test1DReduce):
+    def setUp(self):
+        self.op_type = "reduce_sum"
+        self.attrs = {'dim': [2]}
+        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")}
+        self.outputs = {
+            'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim']))
+        }
+
+
+class Test3DReduce2(Test1DReduce):
+    def setUp(self):
+        self.op_type = "reduce_sum"
+        self.attrs = {'dim': [-2]}
+        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")}
+        self.outputs = {
+            'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim']))
+        }
+
+
+class Test3DReduce3(Test1DReduce):
+    def setUp(self):
+        self.op_type = "reduce_sum"
+        self.attrs = {'dim': [1, 2]}
+        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")}
+        self.outputs = {
+            'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim']))
+        }
+
+
+class TestKeepDimReduce(Test1DReduce):
+    def setUp(self):
+        self.op_type = "reduce_sum"
+        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
+        self.attrs = {'dim': [1], 'keep_dim': True}
+        self.outputs = {
+            'Out': self.inputs['X'].sum(axis=tuple(self.attrs['dim']),
+                                        keepdims=self.attrs['keep_dim'])
+        }
+
+
+class TestReduceAll(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_sum"
         self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float64")}
         self.attrs = {'reduce_all': True}
         self.outputs = {'Out': self.inputs['X'].sum()}
 
-    def test_check_output(self):
-        self.check_output()
-
-    def test_check_grad(self):
-        self.check_grad(['X'], 'Out')
-
 
 ## reduction in multi dims
 class TestReduceMeanOpMultiAxises(OpTest):