Rename elementwise_heaviside to heaviside (#50821)

* rename elementwise_heaviside to heaviside

* delete __init__.py

* fix bug

paddle/phi/api/yaml/legacy_backward.yaml

@@ -413,16 +413,6 @@
   kernel :
     func : einsum_grad
 
-- backward_op : elementwise_heaviside_grad
-  forward : heaviside (Tensor x, Tensor y) -> Tensor(out)
-  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
-  output : Tensor(x_grad), Tensor(y_grad)
-  infer_meta :
-    func : GeneralBinaryGradInferMeta
-    param : [x, y]
-  kernel :
-    func : elementwise_heaviside_grad
-
 - backward_op : elementwise_pow_grad
   forward : elementwise_pow(Tensor x, Tensor y) -> Tensor(out)
   args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
@@ -570,6 +560,16 @@
     func : hardswish_grad
   inplace : (out_grad -> x_grad)
 
+- backward_op : heaviside_grad
+  forward : heaviside (Tensor x, Tensor y) -> Tensor(out)
+  args : (Tensor x, Tensor y, Tensor out_grad)
+  output : Tensor(x_grad), Tensor(y_grad)
+  infer_meta :
+    func : GeneralBinaryGradInferMeta
+    param : [x, y]
+  kernel :
+    func : heaviside_grad
+
 - backward_op : hsigmoid_loss_grad
   forward : hsigmoid_loss (Tensor x, Tensor label, Tensor w, Tensor bias, Tensor path, Tensor code, int num_classes, bool remote_prefetch, bool is_sparse) -> Tensor(out), Tensor(pre_out), Tensor(w_out)
   args : (Tensor x, Tensor w, Tensor label, Tensor path, Tensor code, Tensor bias, Tensor pre_out, Tensor out_grad, int num_classes, bool remote_prefetch, bool is_sparse)

paddle/phi/api/yaml/legacy_ops.yaml

@@ -532,15 +532,6 @@
     func : einsum
   backward : einsum_grad
 
-- op : elementwise_heaviside
-  args : (Tensor x, Tensor y)
-  output : Tensor
-  infer_meta :
-    func : ElementwiseInferMeta
-  kernel :
-    func : elementwise_heaviside
-  backward : elementwise_heaviside_grad
-
 - op : elementwise_pow
   args : (Tensor x, Tensor y)
   output : Tensor(out)
@@ -827,6 +818,15 @@
     func : hardswish
   backward : hardswish_grad
 
+- op : heaviside
+  args : (Tensor x, Tensor y)
+  output : Tensor
+  infer_meta :
+    func : ElementwiseInferMeta
+  kernel :
+    func : heaviside
+  backward : heaviside_grad
+
 - op : hsigmoid_loss
   args : (Tensor x, Tensor label, Tensor w, Tensor bias, Tensor path, Tensor code, int num_classes, bool remote_prefetch, bool is_sparse)
   output : Tensor(out), Tensor(pre_out), Tensor(w_out)

paddle/phi/kernels/cpu/elementwise_grad_kernel.cc

@@ -89,10 +89,10 @@ PD_REGISTER_KERNEL(minimum_grad,
                    int64_t,
                    phi::dtype::bfloat16) {}
 
-PD_REGISTER_KERNEL(elementwise_heaviside_grad,
+PD_REGISTER_KERNEL(heaviside_grad,
                    CPU,
                    ALL_LAYOUT,
-                   phi::ElementwiseHeavisideGradKernel,
+                   phi::HeavisideGradKernel,
                    float,
                    double,
                    int,

paddle/phi/kernels/cpu/elementwise_kernel.cc

@@ -103,15 +103,14 @@ void ElementwisePowRawKernel(const Context& dev_ctx,
 }
 
 template <typename T, typename Context>
-void ElementwiseHeavisideRawKernel(const Context& dev_ctx,
+void HeavisideKernel(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);
+      dev_ctx, x, y, -1, funcs::ElementwiseHeavisideFunctor<T>(), out);
 }
 
 }  // namespace phi
@@ -174,10 +173,11 @@ PD_REGISTER_KERNEL(elementwise_pow_raw,
                    double,
                    int,
                    int64_t) {}
-PD_REGISTER_KERNEL(elementwise_heaviside_raw,
+
+PD_REGISTER_KERNEL(heaviside,
                    CPU,
                    ALL_LAYOUT,
-                   phi::ElementwiseHeavisideRawKernel,
+                   phi::HeavisideKernel,
                    float,
                    double,
                    int,

paddle/phi/kernels/elementwise_grad_kernel.h

@@ -55,11 +55,10 @@ void MinimumGradKernel(const Context& dev_ctx,
                        DenseTensor* dy);
 
 template <typename T, typename Context>
-void ElementwiseHeavisideGradKernel(const Context& dev_ctx,
+void HeavisideGradKernel(const Context& dev_ctx,
                          const DenseTensor& x,
                          const DenseTensor& y,
                          const DenseTensor& dout,
-                                    int axis,
                          DenseTensor* dx,
                          DenseTensor* dy);
 

paddle/phi/kernels/elementwise_kernel.cc

@@ -68,15 +68,6 @@ 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);
-}
-
 template <typename T, typename Context>
 void DivideKernel(const Context& dev_ctx,
                   const DenseTensor& x,
@@ -153,14 +144,6 @@ PD_REGISTER_KERNEL(remainder,
                    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,
@@ -262,15 +245,6 @@ PD_REGISTER_KERNEL(remainder,
                    phi::dtype::float16) {}
 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,
-                   phi::dtype::float16) {}
 PD_REGISTER_KERNEL(elementwise_pow,
                    KPS,
                    ALL_LAYOUT,

paddle/phi/kernels/elementwise_kernel.h

@@ -104,14 +104,7 @@ void ElementwisePowKernel(const Context& dev_ctx,
                           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,
+void HeavisideKernel(const Context& dev_ctx,
                      const DenseTensor& x,
                      const DenseTensor& y,
                      DenseTensor* out);
@@ -161,13 +154,13 @@ DenseTensor FloorDivide(const Context& dev_ctx,
 }
 
 template <typename T, typename Context>
-DenseTensor ElementwiseHeaviside(const Context& dev_ctx,
+DenseTensor Heaviside(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);
+  HeavisideKernel<T, Context>(dev_ctx, x, y, &dense_out);
   return dense_out;
 }
 

paddle/phi/kernels/gpu/elementwise_grad_kernel.cu

@@ -132,10 +132,10 @@ PD_REGISTER_KERNEL(minimum_grad,
                    phi::dtype::float16,
                    phi::dtype::bfloat16) {}
 
-PD_REGISTER_KERNEL(elementwise_heaviside_grad,
+PD_REGISTER_KERNEL(heaviside_grad,
                    GPU,
                    ALL_LAYOUT,
-                   phi::ElementwiseHeavisideGradKernel,
+                   phi::HeavisideGradKernel,
                    float,
                    double,
                    int,

paddle/phi/kernels/impl/elementwise_grad_kernel_impl.h

@@ -830,11 +830,10 @@ struct HeavisideGradDy {
 };
 
 template <typename T, typename Context>
-void ElementwiseHeavisideGradKernel(const Context& dev_ctx,
+void HeavisideGradKernel(const Context& dev_ctx,
                          const DenseTensor& x,
                          const DenseTensor& y,
                          const DenseTensor& dout,
-                                    int axis,
                          DenseTensor* dx,
                          DenseTensor* dy) {
   funcs::ElementwiseGradPreProcess(dout, dx);
@@ -845,7 +844,7 @@ void ElementwiseHeavisideGradKernel(const Context& dev_ctx,
           y,
           dout,
           dout,
-          axis,
+          -1,
           dx,
           dy,
           HeavisideGradDx<T>(),

paddle/phi/kernels/kps/elementwise_kernel.cu

@@ -56,7 +56,23 @@ void FloorDivideKernel(const Context& dev_ctx,
   FloorDivideRawKernel<T>(dev_ctx, x, y, axis, out);
 }
 // Create the definition of Heaviside
-DEFINE_CUDA_ELEMENTWISE_OP(ElementwiseHeaviside)
+template <typename T, typename Context>
+void HeavisideKernel(const Context& dev_ctx,
+                     const DenseTensor& x,
+                     const DenseTensor& y,
+                     DenseTensor* out) {
+  std::vector<const DenseTensor*> inputs;
+  inputs.reserve(2);
+  std::vector<DenseTensor*> outputs;
+  outputs.reserve(1);
+  inputs.emplace_back(&x);
+  inputs.emplace_back(&y);
+  outputs.emplace_back(out);
+  dev_ctx.template Alloc<T>(out);
+  funcs::BroadcastKernel<ElementwiseType::kBinary, T, T>(
+      dev_ctx, inputs, &outputs, -1, funcs::ElementwiseHeavisideFunctor<T>());
+}
+
 // Create the definition of Pow
 DEFINE_CUDA_ELEMENTWISE_OP(ElementwisePow)
 template <typename T, typename Context>
@@ -148,10 +164,10 @@ PD_REGISTER_KERNEL(floor_divide_raw,
                    phi::FloorDivideRawKernel,
                    int,
                    int64_t) {}
-PD_REGISTER_KERNEL(elementwise_heaviside_raw,
+PD_REGISTER_KERNEL(heaviside,
                    KPS,
                    ALL_LAYOUT,
-                   phi::ElementwiseHeavisideRawKernel,
+                   phi::HeavisideKernel,
                    float,
                    double,
                    int,

paddle/phi/ops/compat/elementwise_sig.cc

@@ -102,12 +102,7 @@ KernelSignature ElementwiseFloorDivOpArgumentMapping(
 
 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"});
+  return KernelSignature("heaviside", {"X", "Y"}, {}, {"Out"});
 }
 
 KernelSignature ElementwisePowOpArgumentMapping(
@@ -226,10 +221,8 @@ KernelSignature ElementwiseMinGradOpArgumentMapping(
 
 KernelSignature ElementwiseHeavisideGradOpArgumentMapping(
     const ArgumentMappingContext& ctx) {
-  return KernelSignature("elementwise_heaviside_grad",
-                         {"X", "Y", "Out@GRAD"},
-                         {"axis"},
-                         {"X@GRAD", "Y@GRAD"});
+  return KernelSignature(
+      "heaviside_grad", {"X", "Y", "Out@GRAD"}, {}, {"X@GRAD", "Y@GRAD"});
 }
 
 KernelSignature ElementwisePowGradOpArgumentMapping(
@@ -249,6 +242,7 @@ PD_REGISTER_BASE_KERNEL_NAME(elementwise_max, maximum);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_min, minimum);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_mod, remainder);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_floordiv, floor_divide);
+PD_REGISTER_BASE_KERNEL_NAME(elementwise_heaviside, heaviside);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_add_grad, add_grad);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_add_grad_grad, add_double_grad);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_add_triple_grad, add_triple_grad);
@@ -265,6 +259,7 @@ PD_REGISTER_BASE_KERNEL_NAME(elementwise_fmax_grad, fmax_grad);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_fmin_grad, fmin_grad);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_max_grad, maximum_grad);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_min_grad, minimum_grad);
+PD_REGISTER_BASE_KERNEL_NAME(elementwise_heaviside_grad, heaviside_grad);
 
 PD_REGISTER_ARG_MAPPING_FN(elementwise_add,
                            phi::ElementwiseAddOpArgumentMapping);

python/paddle/tensor/math.py

@@ -4797,7 +4797,7 @@ def heaviside(x, y, name=None):
             #     [0.        , 1.        , 0.30000001]]
     """
     if in_dygraph_mode():
-        return _C_ops.elementwise_heaviside(x, y)
+        return _C_ops.heaviside(x, y)
     else:
         op_type = 'elementwise_heaviside'
         return _elementwise_op(LayerHelper(op_type, **locals()))