optimize elementwise_max_grad using new interfaces (#37906)

* init elem_max_grad op * optimize code and reply review comments * ternary functors * apply new reduce func * move functor to .h * multi-outputs init * rearrange code * modifed functors * optimizer code * pass nullptr * revert the last change as seg fault occurs * optimize code * remove inplace * remove comments
PaddlePaddle · Jan 12, 2022 · 4a64ca1 · 4a64ca1
1 parent 5f5f626
commit 4a64ca1
Show file tree

Hide file tree

Showing 3 changed files with 80 additions and 6 deletions.
diff --git a/paddle/fluid/operators/elementwise/elementwise_functor.h b/paddle/fluid/operators/elementwise/elementwise_functor.h
@@ -301,5 +301,32 @@ struct MulGradXYFunctor<Complex<InT>, Complex<OutT>> {
   }
 };
 
+// Ternary compare
+template <typename T>
+struct MaxGradXFunctor {
+  inline HOSTDEVICE T operator()(const T& x, const T& y, const T& dout) const {
+    return dout * static_cast<T>(x > y);
+  }
+};
+template <typename T>
+struct MaxGradYFunctor {
+  inline HOSTDEVICE T operator()(const T& x, const T& y, const T& dout) const {
+    return dout * static_cast<T>(x <= y);
+  }
+};
+
+template <typename InT, typename OutT>
+struct MaxGradXYFunctor {
+  inline HOSTDEVICE paddle::framework::Array<OutT, 2> operator()(
+      const InT& x, const InT& y, const InT& dout) {
+    paddle::framework::Array<OutT, 2> outs;
+    // dx = dout * (x > y)
+    outs[0] = static_cast<OutT>(dout * static_cast<InT>(x > y));
+    // dy = dout * (x <= y)
+    outs[1] = static_cast<OutT>(dout * static_cast<InT>(x <= y));
+    return outs;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/fluid/operators/elementwise/elementwise_max_op.cu b/paddle/fluid/operators/elementwise/elementwise_max_op.cu
@@ -24,15 +24,41 @@ class ElementwiseMaxKernel<platform::CUDADeviceContext, T>
   void Compute(const framework::ExecutionContext& ctx) const override {
     std::vector<const framework::Tensor*> ins;
     std::vector<framework::Tensor*> outs;
-    const auto& cuda_ctx =
+    const auto& dev_ctx =
         ctx.template device_context<platform::CUDADeviceContext>();
 
     int axis = PackTensorsIntoVector<T>(ctx, &ins, &outs);
     LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
-        cuda_ctx, ins, &outs, axis, MaxFunctor<T>());
+        dev_ctx, ins, &outs, axis, MaxFunctor<T>());
   }
 };
 
+template <typename DeviceContext, typename T>
+typename std::enable_if<
+    std::is_same<DeviceContext, platform::CUDADeviceContext>::value>::type
+ElementwiseMaxGrad(const framework::ExecutionContext& ctx,
+                   const framework::Tensor* x, const framework::Tensor* y,
+                   const framework::Tensor* out, const framework::Tensor* dout,
+                   framework::Tensor* dx, framework::Tensor* dy) {
+  int axis = ctx.Attr<int>("axis");
+  const auto& dev_ctx =
+      ctx.template device_context<platform::CUDADeviceContext>();
+  const auto place = ctx.GetPlace();
+  if (dx != nullptr && dy != nullptr) {
+    std::vector<const framework::Tensor*> ins = {x, y, dout};
+    GetGradXAndYOut<ElementwiseType::kTernary, T>(
+        dev_ctx, place, axis, ins, dout, dx, dy, MaxGradXYFunctor<T, T>());
+  } else if (dx != nullptr && dy == nullptr) {
+    std::vector<const framework::Tensor*> ins = {x, y, dout};
+    GetGradXOrYOut<ElementwiseType::kTernary, T>(
+        dev_ctx, place, axis, ins, dout, dx, MaxGradXFunctor<T>());
+  } else if (dx == nullptr && dy != nullptr) {
+    std::vector<const framework::Tensor*> ins = {x, y, dout};
+    GetGradXOrYOut<ElementwiseType::kTernary, T>(
+        dev_ctx, place, axis, ins, dout, dy, MaxGradYFunctor<T>());
+  }
+}
+
 }  // namespace operators
 }  // namespace paddle
 

diff --git a/paddle/fluid/operators/elementwise/elementwise_max_op.h b/paddle/fluid/operators/elementwise/elementwise_max_op.h
@@ -64,6 +64,28 @@ struct MaxGradDy {
   }
 };
 
+template <typename DeviceContext, typename T>
+typename std::enable_if<
+    std::is_same<DeviceContext, platform::CPUDeviceContext>::value>::type
+ElementwiseMaxGrad(const framework::ExecutionContext& ctx,
+                   const framework::Tensor* x, const framework::Tensor* y,
+                   const framework::Tensor* out, const framework::Tensor* dout,
+                   framework::Tensor* dx, framework::Tensor* dy) {
+  int axis = ctx.Attr<int>("axis");
+  ElemwiseGradCompute<DeviceContext, T, MaxGradDx<T>, MaxGradDy<T>>(
+      ctx, *x, *y, *out, *dout, axis, dx, dy, MaxGradDx<T>(), MaxGradDy<T>());
+}
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+template <typename DeviceContext, typename T>
+typename std::enable_if<
+    std::is_same<DeviceContext, platform::CUDADeviceContext>::value>::type
+ElementwiseMaxGrad(const framework::ExecutionContext& ctx,
+                   const framework::Tensor* x, const framework::Tensor* y,
+                   const framework::Tensor* out, const framework::Tensor* dout,
+                   framework::Tensor* dx, framework::Tensor* dy);
+#endif
+
 template <typename DeviceContext, typename T>
 class ElementwiseMaxGradKernel : public ElemwiseGradKernel<T> {
  public:
@@ -74,12 +96,11 @@ class ElementwiseMaxGradKernel : public ElemwiseGradKernel<T> {
     auto* x = ctx.Input<Tensor>("X");
     auto* y = ctx.Input<Tensor>("Y");
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* out = dout;  // out is not necessary
     auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
-    auto* out = dout;  // Fake out, not used
-    int axis = ctx.Attr<int>("axis");
-    ElemwiseGradCompute<DeviceContext, T, MaxGradDx<T>, MaxGradDy<T>>(
-        ctx, *x, *y, *out, *dout, axis, dx, dy, MaxGradDx<T>(), MaxGradDy<T>());
+
+    ElementwiseMaxGrad<DeviceContext, T>(ctx, x, y, out, dout, dx, dy);
   }
 };