/* Copyright (c) 2021 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 #include #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/operators/npu_op_runner.h" #include "paddle/fluid/operators/optimizers/adam_op.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; template class AdamNPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { const auto* param_var = ctx.InputVar("Param"); PADDLE_ENFORCE_EQ(param_var->IsType(), true, platform::errors::InvalidArgument( "The Var(%s)'s type should be LoDTensor, " "but the received is %s", ctx.InputNames("Param").front(), framework::ToTypeName(param_var->Type()))); auto* param = ctx.Input("Param"); auto* grad_var = ctx.InputVar("Grad"); PADDLE_ENFORCE_EQ(grad_var->IsType(), true, platform::errors::InvalidArgument( "The Grad(%s)'s type should be LoDTensor, " "but the received is %s", ctx.InputNames("Grad").front(), framework::ToTypeName(param_var->Type()))); auto* grad = ctx.Input("Grad"); auto* mom1 = ctx.Input("Moment1"); auto* mom2 = ctx.Input("Moment2"); auto* lr = ctx.Input("LearningRate"); auto* beta1_pow = ctx.Input("Beta1Pow"); auto* beta2_pow = ctx.Input("Beta2Pow"); auto* param_out = ctx.Output("ParamOut"); auto* mom1_out = ctx.Output("Moment1Out"); auto* mom2_out = ctx.Output("Moment2Out"); auto* beta1_pow_out = ctx.Output("Beta1PowOut"); auto* beta2_pow_out = ctx.Output("Beta2PowOut"); param_out->mutable_data(ctx.GetPlace()); mom1_out->mutable_data(ctx.GetPlace()); mom2_out->mutable_data(ctx.GetPlace()); // NOTE(zhiqiu): beta1_pow and beta2_pow may on CPU and not transform place. if (beta1_pow->place() == platform::CPUPlace()) { T beta1 = *beta1_pow->data(); // `mutable_data` operation needs to be done after getting data beta1_pow_out->mutable_data(ctx.GetPlace()); FillNpuTensorWithConstant(beta1_pow_out, beta1); } else { beta1_pow_out->mutable_data(ctx.GetPlace()); } if (beta2_pow->place() == platform::CPUPlace()) { T beta2 = *beta2_pow->data(); beta2_pow_out->mutable_data(ctx.GetPlace()); FillNpuTensorWithConstant(beta2_pow_out, beta2); } else { beta2_pow_out->mutable_data(ctx.GetPlace()); } const Tensor* beta1_tensor = nullptr; const Tensor* beta2_tensor = nullptr; const Tensor* epsilon_tensor = nullptr; Tensor beta1_tmp(framework::proto::VarType::FP32); Tensor beta2_tmp(framework::proto::VarType::FP32); Tensor epsilon_tmp(framework::proto::VarType::FP32); if (ctx.HasInput("Beta1Tensor")) { beta1_tensor = ctx.Input("Beta1Tensor"); PADDLE_ENFORCE_EQ(beta1_tensor->numel(), 1, platform::errors::InvalidArgument( "Input(Beta1Tensor) size must be 1, but get %d", beta1_tensor->numel())); } else { T beta1 = static_cast(ctx.Attr("beta1")); beta1_tmp.mutable_data({1}, ctx.GetPlace()); FillNpuTensorWithConstant(&beta1_tmp, beta1); beta1_tensor = &beta1_tmp; } if (ctx.HasInput("Beta2Tensor")) { beta2_tensor = ctx.Input("Beta2Tensor"); PADDLE_ENFORCE_EQ(beta1_tensor->numel(), 1, platform::errors::InvalidArgument( "Input(Beta2Tensor) size must be 1, but get %d", beta2_tensor->numel())); } else { T beta2 = static_cast(ctx.Attr("beta2")); beta2_tmp.mutable_data({1}, ctx.GetPlace()); FillNpuTensorWithConstant(&beta2_tmp, beta2); beta2_tensor = &beta2_tmp; } if (ctx.HasInput("EpsilonTensor")) { epsilon_tensor = ctx.Input("EpsilonTensor"); PADDLE_ENFORCE_EQ(epsilon_tensor->numel(), 1, platform::errors::InvalidArgument( "Input(EpsilonTensor) size must be 1, but get %d", epsilon_tensor->numel())); } else { T epsilon = static_cast(ctx.Attr("epsilon")); epsilon_tmp.mutable_data({1}, ctx.GetPlace()); FillNpuTensorWithConstant(&epsilon_tmp, epsilon); epsilon_tensor = &epsilon_tmp; } VLOG(3) << "beta1_pow.numel() : " << beta1_pow->numel() << "beta2_pow.numel() : " << beta2_pow->numel(); VLOG(3) << "param.numel(): " << param->numel(); PADDLE_ENFORCE_EQ(beta1_pow_out->numel(), 1, platform::errors::InvalidArgument( "beta1 pow output size should be 1, but received " "value is:%d.", beta1_pow_out->numel())); PADDLE_ENFORCE_EQ(beta2_pow_out->numel(), 1, platform::errors::InvalidArgument( "beta2 pow output size should be 1, but received " "value is:%d.", beta2_pow_out->numel())); auto stream = ctx.template device_context() .stream(); auto runner = NpuOpRunner("ApplyAdamD", { *param, *mom1, *mom2, *beta1_pow, *beta2_pow, *lr, *beta1_tensor, *beta2_tensor, *epsilon_tensor, *grad, }, { *param_out, *mom1_out, *mom2_out, }, {}); runner.Run(stream); // NOTE(zhiqiu): ApplyAdamD updates params inplace, so // if param and param_out is not same, we need to do copy. if (param_out->data() != param->data()) { framework::TensorCopy( *param, ctx.GetPlace(), ctx.template device_context(), param_out); } if (mom1_out->data() != mom1->data()) { framework::TensorCopy( *mom1, ctx.GetPlace(), ctx.template device_context(), mom1_out); } if (mom2_out->data() != mom2->data()) { framework::TensorCopy( *mom2, ctx.GetPlace(), ctx.template device_context(), mom2_out); } auto runner_m1 = NpuOpRunner("Mul", {*beta1_pow, *beta1_tensor}, {*beta1_pow_out}, {}); runner_m1.Run(stream); auto runner_m2 = NpuOpRunner("Mul", {*beta2_pow, *beta2_tensor}, {*beta2_pow_out}, {}); runner_m2.Run(stream); } }; } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OP_NPU_KERNEL( adam, ops::AdamNPUKernel, ops::AdamNPUKernel);