Move some activation to phi (#40727)

* update * add forward case * update * update; test=develop * add some grad kernel; test=develop * move gpu kernel; test=develop * update * update; * update test; * fix selected rows bug; * add mix vector include ; * add mixed vector depen; test=develop * add logit grad signature; * polish code * fix bug; * add namespace for abs * revert code * not move softsign * revmove duplate register; * fix softsign bug * polish code * format * format * fix bug * remove cmake dep * add square sqrt selected rows support * update * remove clip norm * add standalone executor sqrt dep * standalone exec denp sqrt * remove sqrt op in cmkaelist * open some case

Author: phlrain

paddle/fluid/framework/new_executor/standalone_executor_test.cc

@@ -53,7 +53,7 @@ USE_OP_ITSELF(tanh_grad);
 USE_OP(sum);
 USE_OP_ITSELF(slice_grad);
 USE_OP_ITSELF(lookup_table_grad);
-USE_OP(sqrt);
+USE_OP_ITSELF(sqrt);
 USE_OP_ITSELF(elementwise_max);
 USE_OP_ITSELF(elementwise_div);
 USE_OP_ITSELF(sgd);
@@ -83,6 +83,7 @@ PD_DECLARE_KERNEL(max_raw, GPU, ALL_LAYOUT);
 PD_DECLARE_KERNEL(sgd, GPU, ALL_LAYOUT);
 PD_DECLARE_KERNEL(slice, GPU, ALL_LAYOUT);
 PD_DECLARE_KERNEL(slice_grad, GPU, ALL_LAYOUT);
+PD_DECLARE_KERNEL(sqrt, GPU, ALL_LAYOUT);
 
 DECLARE_double(eager_delete_tensor_gb);
 

paddle/fluid/operators/activation_op.cc

@@ -1496,6 +1496,14 @@ REGISTER_ACTIVATION_OP(hard_sigmoid, HardSigmoid, HardSigmoidFunctor,
  HardSigmoidGradFunctor);
 REGISTER_ACTIVATION_OP(logsigmoid, LogSigmoid, LogSigmoidFunctor,
  LogSigmoidGradFunctor);
+REGISTER_ACTIVATION_OP(expm1, Expm1, Expm1Functor, Expm1GradFunctor);
+REGISTER_ACTIVATION_OP(softplus, Softplus, SoftplusFunctor,
+ SoftplusGradFunctor);
+REGISTER_ACTIVATION_OP(mish, Mish, MishFunctor, MishGradFunctor);
+REGISTER_ACTIVATION_OP(stanh, STanh, STanhFunctor, STanhGradFunctor);
+REGISTER_ACTIVATION_OP(reciprocal, Reciprocal, ReciprocalFunctor,
+ ReciprocalGradFunctor);
+
 REGISTER_ACTIVATION_OP(log2, Log2, Log2Functor, Log2GradFunctor);
 REGISTER_ACTIVATION_OP(log10, Log10, Log10Functor, Log10GradFunctor);
 REGISTER_ACTIVATION_OP(log1p, Log1p, Log1pFunctor, Log1pGradFunctor);
@@ -1630,12 +1638,7 @@ REGISTER_OPERATOR(logit, ops::LogitOp, ops::LogitOpMaker,
  ops::LogitGradOpMaker<paddle::framework::OpDesc>,
  ops::LogitGradOpMaker<paddle::imperative::OpBase>);
 REGISTER_OPERATOR(logit_grad, ops::LogitGradOp);
-REGISTER_OP_CPU_KERNEL(
- logit, ops::LogitKernel<paddle::platform::CPUDeviceContext, float>,
- ops::LogitKernel<paddle::platform::CPUDeviceContext, double>);
-REGISTER_OP_CPU_KERNEL(
- logit_grad, ops::LogitGradKernel<paddle::platform::CPUDeviceContext, float>,
- ops::LogitGradKernel<paddle::platform::CPUDeviceContext, double>);
+
 /* ========================================================================== */
 
 /* ======================== celu register ============================
@@ -1684,7 +1687,6 @@ REGISTER_OPERATOR(
  ops::ActivationOpDoubleGrad<ops::SqrtGradGradFunctor<float>::FwdDeps()>,
  ops::ActivationDoubleGradOpInplaceInferer);
 
-REGISTER_ACTIVATION_CPU_KERNEL(sqrt, Sqrt, SqrtFunctor, SqrtGradFunctor);
 REGISTER_OP_CPU_KERNEL(
  sqrt_grad_grad, ops::SqrtDoubleGradKernel<plat::CPUDeviceContext,
  ops::SqrtGradGradFunctor<float>>,
@@ -1712,7 +1714,6 @@ REGISTER_OPERATOR(
  ops::ActivationOpDoubleGrad<ops::RsqrtGradGradFunctor<float>::FwdDeps()>,
  ops::ActivationDoubleGradOpInplaceInferer);
 
-REGISTER_ACTIVATION_CPU_KERNEL(rsqrt, Rsqrt, RsqrtFunctor, RsqrtGradFunctor);
 REGISTER_OP_CPU_KERNEL(
  rsqrt_grad_grad,
  ops::RsqrtDoubleGradKernel<plat::CPUDeviceContext,
@@ -1741,25 +1742,6 @@ REGISTER_OPERATOR(
  ops::ActivationOpDoubleGrad<ops::SquareGradGradFunctor<float>::FwdDeps()>,
  ops::ActivationDoubleGradOpInplaceInferer);
 
-REGISTER_OP_CPU_KERNEL(square,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::SquareFunctor<float>>,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::SquareFunctor<double>>,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::SquareFunctor<int>>,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::SquareFunctor<int64_t>>);
-REGISTER_OP_CPU_KERNEL(
- square_grad, ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::SquareGradFunctor<float>>,
- ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::SquareGradFunctor<double>>,
- ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::SquareGradFunctor<int>>,
- ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::SquareGradFunctor<int64_t>>);
-
 REGISTER_OP_CPU_KERNEL(
  square_grad_grad,
  ops::SquareDoubleGradKernel<plat::CPUDeviceContext,
@@ -1798,54 +1780,6 @@ REGISTER_OPERATOR(
 REGISTER_OPERATOR(exp_grad, ops::ActivationOpGrad,
  ops::ActivationGradOpInplaceInferer);
 
-REGISTER_OP_CPU_KERNEL(exp,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::ExpFunctor<float>>,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::ExpFunctor<double>>,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::ExpFunctor<int>>,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::ExpFunctor<int64_t>>);
-REGISTER_OP_CPU_KERNEL(
- exp_grad, ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::ExpGradFunctor<float>>,
- ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::ExpGradFunctor<double>>,
- ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::ExpGradFunctor<int>>,
- ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::ExpGradFunctor<int64_t>>);
-/* ========================================================================== */
-
-/* ========================== expm1 register ============================ */
-REGISTER_OPERATOR(
- expm1, ops::ActivationOp, ops::Expm1OpMaker, ops::ActivationOpInferVarType,
- ops::ActivationGradOpMaker<ops::Expm1GradFunctor<float>::FwdDeps(),
- paddle::framework::OpDesc>,
- ops::ActivationGradOpMaker<ops::Expm1GradFunctor<float>::FwdDeps(),
- paddle::imperative::OpBase>,
- std::conditional<ops::CanInplaceAct<ops::Expm1GradFunctor<float>>(),
- ops::ActFwdInplaceInferer, void>::type);
-REGISTER_OPERATOR(expm1_grad, ops::ActivationOpGrad,
- ops::ActivationGradOpInplaceInferer);
-
-REGISTER_OP_CPU_KERNEL(expm1,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::Expm1Functor<float>>,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::Expm1Functor<double>>,
- ops::ActivationKernel<paddle::platform::CPUDeviceContext,
- ops::Expm1Functor<plat::float16>>);
-REGISTER_OP_CPU_KERNEL(
- expm1_grad, ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::Expm1GradFunctor<float>>,
- ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::Expm1GradFunctor<double>>,
- ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
- ops::Expm1GradFunctor<plat::float16>>);
-/* ========================================================================== */
-
 /* ========================== Log register ==================================*/
 REGISTER_OPERATOR(
  log, ops::ActivationOp, ops::LogOpMaker, ops::ActivationOpInferVarType,
@@ -1864,8 +1798,6 @@ REGISTER_OPERATOR(
  ops::ActivationOpDoubleGrad<ops::LogGradGradFunctor<float>::FwdDeps()>,
  ops::ActivationDoubleGradOpInplaceInferer);
 
-/* ========================================================================== */
-
 /* ========================== register checkpoint ===========================*/
 REGISTER_OP_VERSION(leaky_relu)
  .AddCheckpoint(