/* Copyright (c) 2016 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. */ #pragma once #include #include #include "lite/core/context.h" #include "lite/core/tensor.h" #include "lite/fluid/lod.h" #include "lite/utils/cp_logging.h" namespace paddle { namespace lite { namespace x86 { namespace math { enum PadLayout { kBatchLengthWidth = 0, kLengthBatchWidth }; enum CopyType { kSeqToPad, kPadToSeq }; inline static uint64_t MaximumSequenceLength( const std::vector& seq_offset) { uint64_t seq_num = seq_offset.size() - 1; uint64_t max_seq_len = 0; for (size_t i = 0; i < seq_num; ++i) { max_seq_len = (std::max)(max_seq_len, seq_offset[i + 1] - seq_offset[i]); } return max_seq_len; } inline static void CheckDims(const lite::DDim& seq_tensor_dims, const lite::DDim& pad_tensor_dims, const std::vector& seq_offset, int64_t padded_seq_len, int64_t step_width, const PadLayout& layout) { CHECK_EQ(static_cast(seq_tensor_dims[0]), seq_offset.back()) << "Value of 1st dimension of the sequence tensor should be " "equal to sum of lengths of all sequences."; CHECK(seq_tensor_dims.size() + 1 == pad_tensor_dims.size() || seq_tensor_dims.size() == pad_tensor_dims.size()) << "pad_tensor's rank should be 1 greater than seq_tensor's " "rank, or be equal with it."; } /* * \brief Padding/Unpadding LoDTensor to/from normal Tensor of the shape * [max_sequence_length, num_sequences, sequence_width]. * * Padding sequence: * padding[i] = seq[lod[level][i]] * Unpadding sequence: * seq[lod[level][i]] = padding[i] * * All sequences will be padded to the same length and stored in a transposed * shape. * Example: * seq (s0, s0, s0, s0; s1, s1; s2, s2, s2; s3) * padding (s0, s1, s2, s3; s0, s1, s2, 0; s0, 0, s2, 0; s0, 0, 0, 0) * * \param context device context of this functor. * \param seq LoDTensor which is stored in sequence format, the shape * is [total_sequence_length, sequence_width] where * total_sequence_length is the sum of all sequences' * length. * \param padding Tensor which is padded to the same length, the shape is * [max_sequence_length, num_sequences, sequence_width]. * \param norm_by_times whether dividing sequence's length. * * \note transposition is also done in this functor. */ template class PaddingLoDTensorFunctor { public: void operator()(const lite::Context& context, const lite::Tensor& seq_tensor, lite::Tensor* pad_tensor, const lite::Tensor& pad_value, int pad_seq_len = -1, int lod_level = 0, bool norm_by_times = false, const PadLayout layout = kBatchLengthWidth); }; template class UnpaddingLoDTensorFunctor { public: void operator()(const lite::Context& context, const lite::Tensor& pad_tensor, lite::Tensor* seq_tensor, int pad_seq_len = -1, int lod_level = 0, bool norm_by_times = false, const PadLayout layout = kBatchLengthWidth); }; } // namespace math } // namespace x86 } // namespace lite } // namespace paddle