【PaddlePaddle Hackathon 2】18、为 Paddle 新增 paddle.heaviside 和 paddle.Tensor.heaviside API

paddle/fluid/operators/elementwise/elementwise_heaviside_op.cc

@@ -0,0 +1,67 @@
+/* Copyright (c) 2022 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. */
+
+#include <string>
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
+
+namespace paddle {
+namespace operators {
+
+class ElementwiseHeavisideOpMaker : public ElementwiseOpMaker {
+ protected:
+  std::string GetName() const override { return "Heaviside"; }
+  std::string GetEquation() const override { return "Out = Heaviside(X, Y)"; }
+
+  void AddInputX() override {
+    AddInput("X",
+             "(Tensor), The input tensor of Heaviside step function. "
+             "Its dtype can be int32, int64, float32 and float64");
+  }
+
+  void AddInputY() override {
+    AddInput("Y",
+             "(Tensor), The tensor determining a Heaviside step function, "
+             "which is the value when X = 0. Its dtype should be same as X.");
+  }
+
+  std::string GetOpFuntionality() const override {
+    return "Computes the Heaviside step function determined by Y "
+           "for each element in X.";
+  }
+};
+
+template <typename T>
+class ElementwiseHeavisideGradOpMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+ protected:
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("elementwise_heaviside_grad");
+    op->SetInput("X", this->Input("X"));
+    op->SetInput("Y", this->Input("Y"));
+    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+    op->SetOutput(framework::GradVarName("Y"), this->InputGrad("Y"));
+    op->SetAttrMap(this->Attrs());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(
+    elementwise_heaviside, ops::ElementwiseOp, ops::ElementwiseHeavisideOpMaker,
+    ops::ElementwiseHeavisideGradOpMaker<paddle::framework::OpDesc>,
+    ops::ElementwiseHeavisideGradOpMaker<paddle::imperative::OpBase>);
+
+REGISTER_OPERATOR(elementwise_heaviside_grad, ops::ElementwiseOpGrad);

paddle/phi/kernels/cpu/elementwise_grad_kernel.cc

@@ -323,6 +323,16 @@ PD_REGISTER_KERNEL(minimum_grad,
                    int,
                    int64_t,
                    phi::dtype::bfloat16) {}
+
+PD_REGISTER_KERNEL(elementwise_heaviside_grad,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::ElementwiseHeavisideGradKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}
+
 PD_REGISTER_KERNEL(elementwise_pow_grad,
                    CPU,
                    ALL_LAYOUT,

paddle/phi/kernels/cpu/elementwise_kernel.cc

@@ -166,6 +166,15 @@ PD_REGISTER_KERNEL(
 PD_REGISTER_KERNEL(
     fmin, CPU, ALL_LAYOUT, phi::FMinKernel, float, double, int, int64_t) {}
 
+PD_REGISTER_KERNEL(elementwise_heaviside,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::ElementwiseHeavisideKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}
+
 PD_REGISTER_KERNEL(add_raw,
                    CPU,
                    ALL_LAYOUT,

paddle/phi/kernels/funcs/elementwise_functor.h

@@ -539,6 +539,28 @@ struct InverseModuloFunctor<
   }
 };
 
+// Heaviside
+template <typename T>
+struct HeavisideFunctor {
+  inline HOSTDEVICE T operator()(const T a, const T b) const {
+    return a == static_cast<T>(0) ? b : static_cast<T>(a > 0);
+  }
+};
+
+template <typename T>
+struct HeavisideGradDx {
+  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
+    return dout * static_cast<T>(0);
+  }
+};
+
+template <typename T>
+struct HeavisideGradDy {
+  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
+    return dout * static_cast<T>(x == static_cast<T>(0));
+  }
+};
+
 template <typename T>
 struct FloorDivideFunctor {
   inline HOSTDEVICE T operator()(const T a, const T b) const {

paddle/phi/kernels/gpu/elementwise_grad_kernel.cu

@@ -382,6 +382,16 @@ PD_REGISTER_KERNEL(minimum_grad,
                    int64_t,
                    phi::dtype::float16,
                    phi::dtype::bfloat16) {}
+
+PD_REGISTER_KERNEL(elementwise_heaviside_grad,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::ElementwiseHeavisideGradKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}
+
 PD_REGISTER_KERNEL(elementwise_pow_grad,
                    GPU,
                    ALL_LAYOUT,

paddle/phi/kernels/gpu/elementwise_kernel.cu

@@ -73,6 +73,15 @@ PD_REGISTER_KERNEL(
 PD_REGISTER_KERNEL(
     fmin, GPU, ALL_LAYOUT, phi::FMinKernel, float, double, int, int64_t) {}
 
+PD_REGISTER_KERNEL(elementwise_heaviside,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::ElementwiseHeavisideKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}
+
 PD_REGISTER_KERNEL(add_raw,
                    GPU,
                    ALL_LAYOUT,

paddle/phi/kernels/impl/elementwise_grad_kernel_impl.h

@@ -667,6 +667,55 @@ struct MinGradDy {
   }
 };
 
+template <typename T, typename Context>
+void ElementwiseHeavisideGradKernel(const Context& dev_ctx,
+                                    const DenseTensor& x,
+                                    const DenseTensor& y,
+                                    const DenseTensor& out_grad,
+                                    int axis,
+                                    DenseTensor* x_grad,
+                                    DenseTensor* y_grad) {
+  funcs::ElementwiseGradPreProcess(out_grad, x_grad);
+
+  auto out = out_grad;  // Fake out, not used
+  auto x_dim = x.dims();
+  auto y_dim = y.dims();
+  if (x.dims() == y.dims()) {
+    funcs::ElemwiseGradComputeNoBroadcast<Context,
+                                          T,
+                                          funcs::HeavisideGradDx<T>,
+                                          funcs::HeavisideGradDy<T>>(
+        dev_ctx,
+        x_dim,
+        y_dim,
+        x,
+        y,
+        out,
+        out_grad,
+        axis,
+        x_grad,
+        y_grad,
+        funcs::HeavisideGradDx<T>(),
+        funcs::HeavisideGradDy<T>());
+  } else {
+    funcs::ElemwiseGradComputeWithBroadcast<T,
+                                            funcs::HeavisideGradDx<T>,
+                                            funcs::HeavisideGradDy<T>>(
+        dev_ctx,
+        x_dim,
+        y_dim,
+        x,
+        y,
+        out,
+        out_grad,
+        axis,
+        x_grad,
+        y_grad,
+        funcs::HeavisideGradDx<T>(),
+        funcs::HeavisideGradDy<T>());
+  }
+}
+
 template <typename T>
 struct PowGradDX {
   HOSTDEVICE T operator()(T x, T y, T out, T dout) const {

paddle/phi/kernels/impl/elementwise_kernel_impl.h

@@ -44,4 +44,15 @@ void FMinKernel(const Context& dev_ctx,
       dev_ctx, x, y, axis, funcs::FMinFunctor<T>(), out);
 }
 
+template <typename T, typename Context>
+void ElementwiseHeavisideKernel(const Context& dev_ctx,
+                                const DenseTensor& x,
+                                const DenseTensor& y,
+                                int axis,
+                                DenseTensor* out) {
+  dev_ctx.template Alloc<T>(out);
+  funcs::ElementwiseCompute<funcs::HeavisideFunctor<T>, T, T>(
+      dev_ctx, x, y, axis, funcs::HeavisideFunctor<T>(), out);
+}
+
 }  // namespace phi

paddle/phi/ops/compat/elementwise_sig.cc

@@ -222,13 +222,23 @@ KernelSignature ElementwiseMinGradOpArgumentMapping(
                          {"axis"},
                          {GradVarName("X"), GradVarName("Y")});
 }
+
+KernelSignature ElementwiseHeavisideGradOpArgumentMapping(
+    const ArgumentMappingContext& ctx) {
+  return KernelSignature("elementwise_heaviside_grad",
+                         {"X", "Y", GradVarName("Out")},
+                         {"axis"},
+                         {GradVarName("X"), GradVarName("Y")});
+}
+
 KernelSignature ElementwisePowGradOpArgumentMapping(
     const ArgumentMappingContext& ctx) {
   return KernelSignature("elementwise_pow_grad",
                          {"X", "Y", GradVarName("Out")},
                          {"axis"},
                          {GradVarName("X"), GradVarName("Y")});
 }
+
 }  // namespace phi
 
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_add, add);
@@ -306,5 +316,7 @@ PD_REGISTER_ARG_MAPPING_FN(elementwise_max_grad,
                            phi::ElementwiseMaxGradOpArgumentMapping);
 PD_REGISTER_ARG_MAPPING_FN(elementwise_min_grad,
                            phi::ElementwiseMinGradOpArgumentMapping);
+PD_REGISTER_ARG_MAPPING_FN(elementwise_heaviside_grad,
+                           phi::ElementwiseHeavisideGradOpArgumentMapping);
 PD_REGISTER_ARG_MAPPING_FN(elementwise_pow_grad,
                            phi::ElementwisePowGradOpArgumentMapping);

python/paddle/__init__.py

@@ -268,6 +268,7 @@
 from .tensor.math import fmin  # noqa: F401
 from .tensor.math import inner  # noqa: F401
 from .tensor.math import outer  # noqa: F401
+from .tensor.math import heaviside  # noqa: F401
 from .tensor.math import frac  # noqa: F401
 
 from .tensor.random import bernoulli  # noqa: F401