Merge pull request #14337 from jczaja/prv-dam-softmax

Softmax op optimization for inference 

paddle/fluid/operators/math/softmax.cc

@@ -19,8 +19,10 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template class SoftmaxFunctor<platform::CPUDeviceContext, float>;
-template class SoftmaxFunctor<platform::CPUDeviceContext, double>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, float, true>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, float, false>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, double, true>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, double, false>;
 template class SoftmaxGradFunctor<platform::CPUDeviceContext, float>;
 template class SoftmaxGradFunctor<platform::CPUDeviceContext, double>;
 

paddle/fluid/operators/math/softmax.cu

@@ -98,9 +98,14 @@ template class SoftmaxGradCUDNNFunctor<float>;
 template class SoftmaxGradCUDNNFunctor<double>;
 template class SoftmaxGradCUDNNFunctor<platform::float16>;
 
-template class SoftmaxFunctor<platform::CUDADeviceContext, platform::float16>;
-template class SoftmaxFunctor<platform::CUDADeviceContext, float>;
-template class SoftmaxFunctor<platform::CUDADeviceContext, double>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, platform::float16,
+                              false>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, platform::float16,
+                              true>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, float, false>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, double, false>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, float, true>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, double, true>;
 template class SoftmaxGradFunctor<platform::CUDADeviceContext, float>;
 template class SoftmaxGradFunctor<platform::CUDADeviceContext, double>;
 template class SoftmaxGradFunctor<platform::CUDADeviceContext,

paddle/fluid/operators/math/softmax.h

@@ -19,7 +19,7 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template <typename DeviceContext, typename T>
+template <typename DeviceContext, typename T, bool is_test>
 class SoftmaxFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor* X,

paddle/fluid/operators/math/softmax_impl.h

@@ -32,9 +32,9 @@ struct ValueClip {
   }
 };
 
-template <typename DeviceContext, typename T>
-void SoftmaxFunctor<DeviceContext, T>::operator()(const DeviceContext& context,
-                                                  const framework::Tensor* X,
+template <typename DeviceContext, typename T, bool is_test>
+void SoftmaxFunctor<DeviceContext, T, is_test>::operator()(
+    const DeviceContext& context, const framework::Tensor* X,
     framework::Tensor* Y) {
   auto logits = EigenMatrix<T>::From(*X);
   auto softmax = EigenMatrix<T>::From(*Y);
@@ -65,6 +65,39 @@ void SoftmaxFunctor<DeviceContext, T>::operator()(const DeviceContext& context,
                                                  .broadcast(one_by_class));
 }
 
+template <typename DeviceContext, typename T>
+class SoftmaxFunctor<DeviceContext, T, true> {
+  void operator()(const DeviceContext& context, const framework::Tensor* X,
+                  framework::Tensor* Y) {
+    auto logits = EigenMatrix<T>::From(*X);
+    auto softmax = EigenMatrix<T>::From(*Y);
+
+    const int kBatchDim = 0;
+    const int kClassDim = 1;
+
+    const int batch_size = logits.dimension(kBatchDim);
+    const int num_classes = logits.dimension(kClassDim);
+
+    Eigen::DSizes<int, 1> along_class(kClassDim);
+    Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
+    Eigen::DSizes<int, 2> one_by_class(1, num_classes);
+
+    auto shifted_logits = (logits -
+                           logits.maximum(along_class)
+                               .eval()
+                               .reshape(batch_by_one)
+                               .broadcast(one_by_class));
+
+    softmax.device(*context.eigen_device()) = shifted_logits.exp();
+    softmax.device(*context.eigen_device()) = (softmax *
+                                               softmax.sum(along_class)
+                                                   .inverse()
+                                                   .eval()
+                                                   .reshape(batch_by_one)
+                                                   .broadcast(one_by_class));
+  }
+};
+
 template <typename DeviceContext, typename T>
 void SoftmaxGradFunctor<DeviceContext, T>::operator()(
     const DeviceContext& context, const framework::Tensor* y,

paddle/fluid/operators/softmax_op.h

@@ -35,8 +35,14 @@ class SoftmaxKernel : public framework::OpKernel<T> {
     Tensor X_2d = framework::ReshapeToMatrix(*X, rank - 1);
     Tensor Out_2d = framework::ReshapeToMatrix(*Out, rank - 1);
 
-    math::SoftmaxFunctor<DeviceContext, T>()(
+    const bool is_test = context.Attr<bool>("is_test");
+    if (is_test == true) {
+      math::SoftmaxFunctor<DeviceContext, T, true>()(
           context.template device_context<DeviceContext>(), &X_2d, &Out_2d);
+    } else {
+      math::SoftmaxFunctor<DeviceContext, T, false>()(
+          context.template device_context<DeviceContext>(), &X_2d, &Out_2d);
+    }
   }
 };
 

paddle/fluid/operators/softmax_with_cross_entropy_op.h

@@ -42,8 +42,8 @@ class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
 
     auto& dev_ctx =
         context.template device_context<platform::CPUDeviceContext>();
-    math::SoftmaxFunctor<platform::CPUDeviceContext, T>()(dev_ctx, logits,
-                                                          softmax);
+    math::SoftmaxFunctor<platform::CPUDeviceContext, T, false>()(
+        dev_ctx, logits, softmax);
     math::CrossEntropyFunctor<platform::CPUDeviceContext, T>()(
         dev_ctx, loss, softmax, labels, context.Attr<bool>("soft_label"),
         context.Attr<int>("ignore_index"));

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

@@ -35,6 +35,7 @@ class TestSoftmaxOp(OpTest):
         self.op_type = "softmax"
         self.use_cudnn = False
         self.use_mkldnn = False
+        self.is_test = False
         self.dtype = np.float32
         self.init_kernel_type()
         self.shape = self.get_x_shape()
@@ -48,7 +49,8 @@ class TestSoftmaxOp(OpTest):
         self.outputs = {'Out': out}
         self.attrs = {
             'use_cudnn': self.use_cudnn,
-            'use_mkldnn': self.use_mkldnn
+            'use_mkldnn': self.use_mkldnn,
+            'is_test': self.is_test
         }
 
     def init_kernel_type(self):
@@ -144,6 +146,11 @@ class TestSoftmaxFP16CUDNNOp2(TestSoftmaxFP16CUDNNOp):
         return [2, 3, 4, 5]
 
 
+class TestSoftmaxInference(TestSoftmaxOp):
+    def init_kernel_type(self):
+        self.is_test = True
+
+
 class TestSoftmaxMKLDNNOp(TestSoftmaxOp):
     def init_kernel_type(self):
         self.use_mkldnn = True