/* 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 "paddle/fluid/operators/truncated_gaussian_random_op.h" #include #include #include "paddle/fluid/framework/convert_utils.h" #include "paddle/fluid/platform/device/npu/npu_op_runner.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; template class TruncatedGaussianRandomNPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { // TODO(zhiqiu): support dynamic shape and call ParameterizedTruncatedNormal std::vector shape = ctx.Attr>("shape"); Tensor shape_tensor(experimental::DataType::INT32); shape_tensor.mutable_data({static_cast(shape.size())}, ctx.GetPlace()); paddle::framework::TensorFromVector(shape, ctx.device_context(), &shape_tensor); float mean = ctx.Attr("mean"); Tensor mean_tensor(experimental::DataType::FLOAT32); mean_tensor.mutable_data({1}, ctx.GetPlace()); FillNpuTensorWithConstant(&mean_tensor, mean); float std = ctx.Attr("std"); Tensor std_tensor(experimental::DataType::FLOAT32); std_tensor.mutable_data({1}, ctx.GetPlace()); FillNpuTensorWithConstant(&std_tensor, std); int32_t seed_var = ctx.Attr("seed"); Tensor min_tensor(experimental::DataType::FLOAT32); min_tensor.mutable_data({1}, ctx.GetPlace()); float min_value = mean - std * 2.0; FillNpuTensorWithConstant(&min_tensor, min_value); Tensor max_tensor(experimental::DataType::FLOAT32); max_tensor.mutable_data({1}, ctx.GetPlace()); float max_value = mean + std * 2.0; FillNpuTensorWithConstant(&max_tensor, max_value); auto* out = ctx.Output("Out"); out->mutable_data(ctx.GetPlace()); auto stream = ctx.template device_context() .stream(); const auto& runner = NpuOpRunner( "ParameterizedTruncatedNormal", {shape_tensor, mean_tensor, std_tensor, min_tensor, max_tensor}, {*out}, {{"seed", seed_var}}); runner.Run(stream); } }; // NOTE(zhiqiu): actually, this is cpu version kernel, and we need to make the // above // npu version work in the future. template class NPUTruncatedGaussianRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { float mean = context.Attr("mean"); float std = context.Attr("std"); auto* tensor = context.Output("Out"); tensor->mutable_data(context.GetPlace()); Tensor cpu_tensor(tensor->dtype()); cpu_tensor.Resize(tensor->dims()); T* cpu_data = cpu_tensor.mutable_data(platform::CPUPlace()); auto normal_cdf = [](float x) { return (1.0 + std::erf(x / std::sqrt(2.0))) / 2.0; }; float a_normal_cdf = normal_cdf((-2.0 - mean) / std); float b_normal_cdf = normal_cdf((2.0 - mean) / std); std::uniform_real_distribution dist(2.0 * a_normal_cdf - 1.0, 2.0 * b_normal_cdf - 1.0); TruncatedNormal truncated_normal(mean, std); int64_t size = tensor->numel(); unsigned int seed = static_cast(context.Attr("seed")); auto engine = framework::GetCPURandomEngine(seed); for (int64_t i = 0; i < size; ++i) { cpu_data[i] = truncated_normal(dist(*engine)); } framework::TensorCopy( cpu_tensor, context.GetPlace(), context.template device_context(), tensor); context.template device_context() .Wait(); } }; } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OP_NPU_KERNEL(truncated_gaussian_random, ops::NPUTruncatedGaussianRandomKernel);