【PaddlePaddle Hackathon 2】18、为 Paddle 新增 paddle.heaviside 和 paddle.Te…

…nsor.heaviside API (#41872) * Create elementwise_heaviside_op.cc * add ElementwiseHeavisideFunctor * Create test_elementwise_heaviside_op.py * 增加heaviside的python接口 * add heaviside in white list * 增加heaviside的签名 * 增加heaviside的核函数 * 增加heaviside梯度的核函数 * 增加heaviside梯度的注册 * 调整代码格式 * Update elementwise_sig.cc * add heaviside in __all__ * Update heaviside docs * Update math.py * Update math.py * Update math.py
PaddlePaddle · May 10, 2022 · 4892d59 · 4892d59
1 parent 8164414
commit 4892d59
Show file tree

Hide file tree

Showing 17 changed files with 468 additions and 0 deletions.
diff --git a/paddle/fluid/operators/elementwise/elementwise_heaviside_op.cc b/paddle/fluid/operators/elementwise/elementwise_heaviside_op.cc
@@ -0,0 +1,70 @@
+// 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);
diff --git a/paddle/phi/kernels/cpu/elementwise_grad_kernel.cc b/paddle/phi/kernels/cpu/elementwise_grad_kernel.cc
@@ -88,6 +88,16 @@ 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,

diff --git a/paddle/phi/kernels/cpu/elementwise_kernel.cc b/paddle/phi/kernels/cpu/elementwise_kernel.cc
@@ -95,6 +95,18 @@ void ElementwisePowRawKernel(const Context& dev_ctx,
       dev_ctx, x, y, axis, funcs::ElementwisePowFunctor<T>(), out);
 }
 
+template <typename T, typename Context>
+void ElementwiseHeavisideRawKernel(const Context& dev_ctx,
+                                   const DenseTensor& x,
+                                   const DenseTensor& y,
+                                   int axis,
+                                   DenseTensor* out) {
+  // allocate memory for out
+  dev_ctx.template Alloc<T>(out);
+  funcs::ElementwiseCompute<funcs::ElementwiseHeavisideFunctor<T>, T>(
+      dev_ctx, x, y, axis, funcs::ElementwiseHeavisideFunctor<T>(), out);
+}
+
 }  // namespace phi
 
 using complex64 = ::phi::dtype::complex<float>;
@@ -149,3 +161,11 @@ PD_REGISTER_KERNEL(elementwise_pow_raw,
                    double,
                    int,
                    int64_t) {}
+PD_REGISTER_KERNEL(elementwise_heaviside_raw,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::ElementwiseHeavisideRawKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}
diff --git a/paddle/phi/kernels/elementwise_grad_kernel.h b/paddle/phi/kernels/elementwise_grad_kernel.h
@@ -55,6 +55,15 @@ void MinimumGradKernel(const Context& dev_ctx,
                        DenseTensor* dx,
                        DenseTensor* dy);
 
+template <typename T, typename Context>
+void ElementwiseHeavisideGradKernel(const Context& dev_ctx,
+                                    const DenseTensor& x,
+                                    const DenseTensor& y,
+                                    const DenseTensor& dout,
+                                    int axis,
+                                    DenseTensor* dx,
+                                    DenseTensor* dy);
+
 template <typename T, typename Context>
 void ElementwisePowGradKernel(const Context& dev_ctx,
                               const DenseTensor& x,

diff --git a/paddle/phi/kernels/elementwise_kernel.cc b/paddle/phi/kernels/elementwise_kernel.cc
@@ -64,6 +64,15 @@ void ElementwisePowKernel(const Context& dev_ctx,
   ElementwisePowRawKernel<T>(dev_ctx, x, y, axis, out);
 }
 
+template <typename T, typename Context>
+void ElementwiseHeavisideKernel(const Context& dev_ctx,
+                                const DenseTensor& x,
+                                const DenseTensor& y,
+                                DenseTensor* out) {
+  int axis = -1;
+  ElementwiseHeavisideRawKernel<T>(dev_ctx, x, y, axis, out);
+}
+
 }  // namespace phi
 
 using complex64 = ::phi::dtype::complex<float>;
@@ -91,6 +100,14 @@ PD_REGISTER_KERNEL(
     modulo, CPU, ALL_LAYOUT, phi::ModuloKernel, float, double, int, int64_t) {}
 PD_REGISTER_KERNEL(
     floor_divide, CPU, ALL_LAYOUT, phi::FloorDivideKernel, int, int64_t) {}
+PD_REGISTER_KERNEL(elementwise_heaviside,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::ElementwiseHeavisideKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}
 PD_REGISTER_KERNEL(elementwise_pow,
                    CPU,
                    ALL_LAYOUT,
@@ -126,6 +143,14 @@ PD_REGISTER_KERNEL(
     modulo, GPU, ALL_LAYOUT, phi::ModuloKernel, float, double, int, int64_t) {}
 PD_REGISTER_KERNEL(
     floor_divide, KPS, ALL_LAYOUT, phi::FloorDivideKernel, int, int64_t) {}
+PD_REGISTER_KERNEL(elementwise_heaviside,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::ElementwiseHeavisideKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}
 PD_REGISTER_KERNEL(elementwise_pow,
                    KPS,
                    ALL_LAYOUT,

diff --git a/paddle/phi/kernels/elementwise_kernel.h b/paddle/phi/kernels/elementwise_kernel.h
@@ -98,6 +98,19 @@ void ElementwisePowKernel(const Context& dev_ctx,
                           const DenseTensor& y,
                           DenseTensor* out);
 
+template <typename T, typename Context>
+void ElementwiseHeavisideRawKernel(const Context& dev_ctx,
+                                   const DenseTensor& x,
+                                   const DenseTensor& y,
+                                   int axis,
+                                   DenseTensor* out);
+
+template <typename T, typename Context>
+void ElementwiseHeavisideKernel(const Context& dev_ctx,
+                                const DenseTensor& x,
+                                const DenseTensor& y,
+                                DenseTensor* out);
+
 template <typename T, typename Context>
 DenseTensor Maximum(const Context& dev_ctx,
                     const DenseTensor& x,
@@ -142,6 +155,17 @@ DenseTensor FloorDivide(const Context& dev_ctx,
   return dense_out;
 }
 
+template <typename T, typename Context>
+DenseTensor ElementwiseHeaviside(const Context& dev_ctx,
+                                 const DenseTensor& x,
+                                 const DenseTensor& y) {
+  DenseTensor dense_out;
+  MetaTensor meta_out(&dense_out);
+  ElementwiseInferMeta(x, y, &meta_out);
+  ElementwiseHeavisideKernel<T, Context>(dev_ctx, x, y, &dense_out);
+  return dense_out;
+}
+
 template <typename T, typename Context>
 DenseTensor ElementwisePow(const Context& dev_ctx,
                            const DenseTensor& x,

diff --git a/paddle/phi/kernels/funcs/elementwise_functor.h b/paddle/phi/kernels/funcs/elementwise_functor.h
@@ -543,6 +543,13 @@ struct InverseModuloFunctor<
   }
 };
 
+template <typename T>
+struct ElementwiseHeavisideFunctor {
+  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 FloorDivideFunctor {
   inline HOSTDEVICE T operator()(const T a, const T b) const {

diff --git a/paddle/phi/kernels/gpu/elementwise_grad_kernel.cu b/paddle/phi/kernels/gpu/elementwise_grad_kernel.cu
@@ -128,6 +128,16 @@ 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,

diff --git a/paddle/phi/kernels/impl/elementwise_grad_kernel_impl.h b/paddle/phi/kernels/impl/elementwise_grad_kernel_impl.h
@@ -683,6 +683,43 @@ struct MinGradDy {
   }
 };
 
+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, typename Context>
+void ElementwiseHeavisideGradKernel(const Context& dev_ctx,
+                                    const DenseTensor& x,
+                                    const DenseTensor& y,
+                                    const DenseTensor& dout,
+                                    int axis,
+                                    DenseTensor* dx,
+                                    DenseTensor* dy) {
+  funcs::ElementwiseGradPreProcess(dout, dx);
+  phi::funcs::
+      ElemwiseGradCompute<Context, T, HeavisideGradDx<T>, HeavisideGradDy<T>>(
+          dev_ctx,
+          x,
+          y,
+          dout,
+          dout,
+          axis,
+          dx,
+          dy,
+          HeavisideGradDx<T>(),
+          HeavisideGradDy<T>());
+}
+
 template <typename T>
 struct PowGradDX {
   HOSTDEVICE T operator()(T x, T y, T out, T dout) const {

diff --git a/paddle/phi/kernels/kps/elementwise_kernel.cu b/paddle/phi/kernels/kps/elementwise_kernel.cu
@@ -54,6 +54,8 @@ void FloorDivideKernel(const Context& dev_ctx,
   int axis = -1;
   FloorDivideRawKernel<T>(dev_ctx, x, y, axis, out);
 }
+// Create the definition of Heaviside
+DEFINE_CUDA_ELEMENTWISE_OP(ElementwiseHeaviside)
 // Create the definition of Pow
 DEFINE_CUDA_ELEMENTWISE_OP(ElementwisePow)
 template <typename T, typename Context>
@@ -130,6 +132,14 @@ PD_REGISTER_KERNEL(floor_divide_raw,
                    phi::FloorDivideRawKernel,
                    int,
                    int64_t) {}
+PD_REGISTER_KERNEL(elementwise_heaviside_raw,
+                   KPS,
+                   ALL_LAYOUT,
+                   phi::ElementwiseHeavisideRawKernel,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}
 PD_REGISTER_KERNEL(elementwise_pow_raw,
                    KPS,
                    ALL_LAYOUT,

diff --git a/paddle/phi/ops/compat/elementwise_sig.cc b/paddle/phi/ops/compat/elementwise_sig.cc
@@ -95,6 +95,16 @@ KernelSignature ElementwiseFloorDivOpArgumentMapping(
   return KernelSignature("floor_divide_raw", {"X", "Y"}, {"axis"}, {"Out"});
 }
 
+KernelSignature ElementwiseHeavisideOpArgumentMapping(
+    const ArgumentMappingContext& ctx) {
+  int axis = paddle::any_cast<int>(ctx.Attr("axis"));
+  if (axis == -1) {
+    return KernelSignature("elementwise_heaviside", {"X", "Y"}, {}, {"Out"});
+  }
+  return KernelSignature(
+      "elementwise_heaviside_raw", {"X", "Y"}, {"axis"}, {"Out"});
+}
+
 KernelSignature ElementwisePowOpArgumentMapping(
     const ArgumentMappingContext& ctx) {
   int axis = paddle::any_cast<int>(ctx.Attr("axis"));
@@ -208,6 +218,15 @@ KernelSignature ElementwiseMinGradOpArgumentMapping(
   return KernelSignature(
       "minimum_grad", {"X", "Y", "Out@GRAD"}, {"axis"}, {"X@GRAD", "Y@GRAD"});
 }
+
+KernelSignature ElementwiseHeavisideGradOpArgumentMapping(
+    const ArgumentMappingContext& ctx) {
+  return KernelSignature("elementwise_heaviside_grad",
+                         {"X", "Y", "Out@GRAD"},
+                         {"axis"},
+                         {"X@GRAD", "Y@GRAD"});
+}
+
 KernelSignature ElementwisePowGradOpArgumentMapping(
     const ArgumentMappingContext& ctx) {
   return KernelSignature("elementwise_pow_grad",
@@ -258,6 +277,8 @@ PD_REGISTER_ARG_MAPPING_FN(elementwise_mod,
                            phi::ElementwiseModOpArgumentMapping);
 PD_REGISTER_ARG_MAPPING_FN(elementwise_floordiv,
                            phi::ElementwiseFloorDivOpArgumentMapping);
+PD_REGISTER_ARG_MAPPING_FN(elementwise_heaviside,
+                           phi::ElementwiseHeavisideOpArgumentMapping);
 PD_REGISTER_ARG_MAPPING_FN(elementwise_pow,
                            phi::ElementwisePowOpArgumentMapping);
 PD_REGISTER_ARG_MAPPING_FN(elementwise_add_grad,
@@ -292,5 +313,7 @@ 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);
diff --git a/python/paddle/__init__.py b/python/paddle/__init__.py
@@ -269,6 +269,7 @@
 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
@@ -635,4 +636,5 @@
            'renorm',
            'take_along_axis',
            'put_along_axis',
+           'heaviside',
 ]