| 1 | // Licensed to the Apache Software Foundation (ASF) under one |
| 2 | // or more contributor license agreements. See the NOTICE file |
| 3 | // distributed with this work for additional information |
| 4 | // regarding copyright ownership. The ASF licenses this file |
| 5 | // to you under the Apache License, Version 2.0 (the |
| 6 | // "License"); you may not use this file except in compliance |
| 7 | // with the License. You may obtain a copy of the License at |
| 8 | // |
| 9 | // http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | // |
| 11 | // Unless required by applicable law or agreed to in writing, |
| 12 | // software distributed under the License is distributed on an |
| 13 | // "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
| 14 | // KIND, either express or implied. See the License for the |
| 15 | // specific language governing permissions and limitations |
| 16 | // under the License. |
| 17 | |
| 18 | #pragma once |
| 19 | |
| 20 | #include <cstdint> |
| 21 | #include <cstring> |
| 22 | #include <memory> |
| 23 | #include <vector> |
| 24 | |
| 25 | #include "parquet/exception.h" |
| 26 | #include "parquet/platform.h" |
| 27 | #include "parquet/types.h" |
| 28 | |
| 29 | namespace arrow { |
| 30 | |
| 31 | class Array; |
| 32 | class ArrayBuilder; |
| 33 | class BinaryArray; |
| 34 | class BinaryBuilder; |
| 35 | class BinaryDictionary32Builder; |
| 36 | |
| 37 | namespace internal { |
| 38 | |
| 39 | class ChunkedBinaryBuilder; |
| 40 | |
| 41 | } // namespace internal |
| 42 | } // namespace arrow |
| 43 | |
| 44 | namespace parquet { |
| 45 | |
| 46 | class ColumnDescriptor; |
| 47 | |
| 48 | // Untyped base for all encoders |
| 49 | class Encoder { |
| 50 | public: |
| 51 | virtual ~Encoder() = default; |
| 52 | |
| 53 | virtual int64_t EstimatedDataEncodedSize() = 0; |
| 54 | virtual std::shared_ptr<Buffer> FlushValues() = 0; |
| 55 | virtual Encoding::type encoding() const = 0; |
| 56 | |
| 57 | virtual void Put(const ::arrow::Array& values) = 0; |
| 58 | |
| 59 | virtual MemoryPool* memory_pool() const = 0; |
| 60 | }; |
| 61 | |
| 62 | // Base class for value encoders. Since encoders may or not have state (e.g., |
| 63 | // dictionary encoding) we use a class instance to maintain any state. |
| 64 | // |
| 65 | // TODO(wesm): Encode interface API is temporary |
| 66 | template <typename DType> |
| 67 | class TypedEncoder : virtual public Encoder { |
| 68 | public: |
| 69 | typedef typename DType::c_type T; |
| 70 | |
| 71 | using Encoder::Put; |
| 72 | |
| 73 | virtual void Put(const T* src, int num_values) = 0; |
| 74 | |
| 75 | virtual void PutSpaced(const T* src, int num_values, const uint8_t* valid_bits, |
| 76 | int64_t valid_bits_offset) = 0; |
| 77 | }; |
| 78 | |
| 79 | // Base class for dictionary encoders |
| 80 | template <typename DType> |
| 81 | class DictEncoder : virtual public TypedEncoder<DType> { |
| 82 | public: |
| 83 | /// Writes out any buffered indices to buffer preceded by the bit width of this data. |
| 84 | /// Returns the number of bytes written. |
| 85 | /// If the supplied buffer is not big enough, returns -1. |
| 86 | /// buffer must be preallocated with buffer_len bytes. Use EstimatedDataEncodedSize() |
| 87 | /// to size buffer. |
| 88 | virtual int WriteIndices(uint8_t* buffer, int buffer_len) = 0; |
| 89 | |
| 90 | virtual int dict_encoded_size() = 0; |
| 91 | // virtual int dict_encoded_size() { return dict_encoded_size_; } |
| 92 | |
| 93 | virtual int bit_width() const = 0; |
| 94 | |
| 95 | /// Writes out the encoded dictionary to buffer. buffer must be preallocated to |
| 96 | /// dict_encoded_size() bytes. |
| 97 | virtual void WriteDict(uint8_t* buffer) = 0; |
| 98 | |
| 99 | virtual int num_entries() const = 0; |
| 100 | |
| 101 | /// \brief EXPERIMENTAL: Append dictionary indices into the encoder. It is |
| 102 | /// assumed (without any boundschecking) that the indices reference |
| 103 | /// pre-existing dictionary values |
| 104 | /// \param[in] indices the dictionary index values. Only Int32Array currently |
| 105 | /// supported |
| 106 | virtual void PutIndices(const ::arrow::Array& indices) = 0; |
| 107 | |
| 108 | /// \brief EXPERIMENTAL: Append dictionary into encoder, inserting indices |
| 109 | /// separately. Currently throws exception if the current dictionary memo is |
| 110 | /// non-empty |
| 111 | /// \param[in] values the dictionary values. Only valid for certain |
| 112 | /// Parquet/Arrow type combinations, like BYTE_ARRAY/BinaryArray |
| 113 | virtual void PutDictionary(const ::arrow::Array& values) = 0; |
| 114 | }; |
| 115 | |
| 116 | // ---------------------------------------------------------------------- |
| 117 | // Value decoding |
| 118 | |
| 119 | class Decoder { |
| 120 | public: |
| 121 | virtual ~Decoder() = default; |
| 122 | |
| 123 | // Sets the data for a new page. This will be called multiple times on the same |
| 124 | // decoder and should reset all internal state. |
| 125 | virtual void SetData(int num_values, const uint8_t* data, int len) = 0; |
| 126 | |
| 127 | // Returns the number of values left (for the last call to SetData()). This is |
| 128 | // the number of values left in this page. |
| 129 | virtual int values_left() const = 0; |
| 130 | virtual Encoding::type encoding() const = 0; |
| 131 | }; |
| 132 | |
| 133 | template <typename DType> |
| 134 | class TypedDecoder : virtual public Decoder { |
| 135 | public: |
| 136 | using T = typename DType::c_type; |
| 137 | |
| 138 | // Subclasses should override the ones they support. In each of these functions, |
| 139 | // the decoder would decode put to 'max_values', storing the result in 'buffer'. |
| 140 | // The function returns the number of values decoded, which should be max_values |
| 141 | // except for end of the current data page. |
| 142 | virtual int Decode(T* buffer, int max_values) = 0; |
| 143 | |
| 144 | // Decode the values in this data page but leave spaces for null entries. |
| 145 | // |
| 146 | // num_values is the size of the def_levels and buffer arrays including the number of |
| 147 | // null values. |
| 148 | virtual int DecodeSpaced(T* buffer, int num_values, int null_count, |
| 149 | const uint8_t* valid_bits, int64_t valid_bits_offset) { |
| 150 | int values_to_read = num_values - null_count; |
| 151 | int values_read = Decode(buffer, values_to_read); |
| 152 | if (values_read != values_to_read) { |
| 153 | throw ParquetException("Number of values / definition_levels read did not match" ); |
| 154 | } |
| 155 | |
| 156 | // Depending on the number of nulls, some of the value slots in buffer may |
| 157 | // be uninitialized, and this will cause valgrind warnings / potentially UB |
| 158 | memset(static_cast<void*>(buffer + values_read), 0, |
| 159 | (num_values - values_read) * sizeof(T)); |
| 160 | |
| 161 | // Add spacing for null entries. As we have filled the buffer from the front, |
| 162 | // we need to add the spacing from the back. |
| 163 | int values_to_move = values_read; |
| 164 | for (int i = num_values - 1; i >= 0; i--) { |
| 165 | if (BitUtil::GetBit(valid_bits, valid_bits_offset + i)) { |
| 166 | buffer[i] = buffer[--values_to_move]; |
| 167 | } |
| 168 | } |
| 169 | return num_values; |
| 170 | } |
| 171 | }; |
| 172 | |
| 173 | template <typename DType> |
| 174 | class DictDecoder : virtual public TypedDecoder<DType> { |
| 175 | public: |
| 176 | virtual void SetDict(TypedDecoder<DType>* dictionary) = 0; |
| 177 | |
| 178 | /// \brief Insert dictionary values into the Arrow dictionary builder's memo, |
| 179 | /// but do not append any indices |
| 180 | virtual void InsertDictionary(::arrow::ArrayBuilder* builder) = 0; |
| 181 | |
| 182 | /// \brief Decode only dictionary indices and append to dictionary |
| 183 | /// builder. The builder must have had the dictionary from this decoder |
| 184 | /// inserted already. |
| 185 | /// |
| 186 | /// Remember to reset the builder each time the dict decoder is initialized |
| 187 | /// with a new dictionary page |
| 188 | virtual int DecodeIndicesSpaced(int num_values, int null_count, |
| 189 | const uint8_t* valid_bits, int64_t valid_bits_offset, |
| 190 | ::arrow::ArrayBuilder* builder) = 0; |
| 191 | |
| 192 | /// \brief Decode only dictionary indices (no nulls) |
| 193 | /// |
| 194 | /// Remember to reset the builder each time the dict decoder is initialized |
| 195 | /// with a new dictionary page |
| 196 | virtual int DecodeIndices(int num_values, ::arrow::ArrayBuilder* builder) = 0; |
| 197 | }; |
| 198 | |
| 199 | // ---------------------------------------------------------------------- |
| 200 | // TypedEncoder specializations, traits, and factory functions |
| 201 | |
| 202 | class BooleanEncoder : virtual public TypedEncoder<BooleanType> { |
| 203 | public: |
| 204 | using TypedEncoder<BooleanType>::Put; |
| 205 | virtual void Put(const std::vector<bool>& src, int num_values) = 0; |
| 206 | }; |
| 207 | |
| 208 | using Int32Encoder = TypedEncoder<Int32Type>; |
| 209 | using Int64Encoder = TypedEncoder<Int64Type>; |
| 210 | using Int96Encoder = TypedEncoder<Int96Type>; |
| 211 | using FloatEncoder = TypedEncoder<FloatType>; |
| 212 | using DoubleEncoder = TypedEncoder<DoubleType>; |
| 213 | using ByteArrayEncoder = TypedEncoder<ByteArrayType>; |
| 214 | using FLBAEncoder = TypedEncoder<FLBAType>; |
| 215 | |
| 216 | class BooleanDecoder : virtual public TypedDecoder<BooleanType> { |
| 217 | public: |
| 218 | using TypedDecoder<BooleanType>::Decode; |
| 219 | virtual int Decode(uint8_t* buffer, int max_values) = 0; |
| 220 | }; |
| 221 | |
| 222 | using Int32Decoder = TypedDecoder<Int32Type>; |
| 223 | using Int64Decoder = TypedDecoder<Int64Type>; |
| 224 | using Int96Decoder = TypedDecoder<Int96Type>; |
| 225 | using FloatDecoder = TypedDecoder<FloatType>; |
| 226 | using DoubleDecoder = TypedDecoder<DoubleType>; |
| 227 | |
| 228 | /// \brief Internal helper class for decoding BYTE_ARRAY data where we can |
| 229 | /// overflow the capacity of a single arrow::BinaryArray |
| 230 | struct ArrowBinaryAccumulator { |
| 231 | std::unique_ptr<::arrow::BinaryBuilder> builder; |
| 232 | std::vector<std::shared_ptr<::arrow::Array>> chunks; |
| 233 | }; |
| 234 | |
| 235 | class ByteArrayDecoder : virtual public TypedDecoder<ByteArrayType> { |
| 236 | public: |
| 237 | using TypedDecoder<ByteArrayType>::DecodeSpaced; |
| 238 | |
| 239 | /// \brief Returns number of encoded values decoded |
| 240 | virtual int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits, |
| 241 | int64_t valid_bits_offset, |
| 242 | ::arrow::BinaryDictionary32Builder* builder) = 0; |
| 243 | |
| 244 | virtual int DecodeArrowNonNull(int num_values, |
| 245 | ::arrow::BinaryDictionary32Builder* builder) = 0; |
| 246 | |
| 247 | /// \brief Returns number of encoded values decoded |
| 248 | virtual int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits, |
| 249 | int64_t valid_bits_offset, ArrowBinaryAccumulator* out) = 0; |
| 250 | |
| 251 | virtual int DecodeArrowNonNull(int num_values, ArrowBinaryAccumulator* out) = 0; |
| 252 | }; |
| 253 | |
| 254 | class FLBADecoder : virtual public TypedDecoder<FLBAType> { |
| 255 | public: |
| 256 | using TypedDecoder<FLBAType>::DecodeSpaced; |
| 257 | |
| 258 | // TODO(wesm): As possible follow-up to PARQUET-1508, we should examine if |
| 259 | // there is value in adding specialized read methods for |
| 260 | // FIXED_LEN_BYTE_ARRAY. If only Decimal data can occur with this data type |
| 261 | // then perhaps not |
| 262 | }; |
| 263 | |
| 264 | template <typename T> |
| 265 | struct EncodingTraits {}; |
| 266 | |
| 267 | template <> |
| 268 | struct EncodingTraits<BooleanType> { |
| 269 | using Encoder = BooleanEncoder; |
| 270 | using Decoder = BooleanDecoder; |
| 271 | }; |
| 272 | |
| 273 | template <> |
| 274 | struct EncodingTraits<Int32Type> { |
| 275 | using Encoder = Int32Encoder; |
| 276 | using Decoder = Int32Decoder; |
| 277 | }; |
| 278 | |
| 279 | template <> |
| 280 | struct EncodingTraits<Int64Type> { |
| 281 | using Encoder = Int64Encoder; |
| 282 | using Decoder = Int64Decoder; |
| 283 | }; |
| 284 | |
| 285 | template <> |
| 286 | struct EncodingTraits<Int96Type> { |
| 287 | using Encoder = Int96Encoder; |
| 288 | using Decoder = Int96Decoder; |
| 289 | }; |
| 290 | |
| 291 | template <> |
| 292 | struct EncodingTraits<FloatType> { |
| 293 | using Encoder = FloatEncoder; |
| 294 | using Decoder = FloatDecoder; |
| 295 | }; |
| 296 | |
| 297 | template <> |
| 298 | struct EncodingTraits<DoubleType> { |
| 299 | using Encoder = DoubleEncoder; |
| 300 | using Decoder = DoubleDecoder; |
| 301 | }; |
| 302 | |
| 303 | template <> |
| 304 | struct EncodingTraits<ByteArrayType> { |
| 305 | using Encoder = ByteArrayEncoder; |
| 306 | using Decoder = ByteArrayDecoder; |
| 307 | }; |
| 308 | |
| 309 | template <> |
| 310 | struct EncodingTraits<FLBAType> { |
| 311 | using Encoder = FLBAEncoder; |
| 312 | using Decoder = FLBADecoder; |
| 313 | }; |
| 314 | |
| 315 | PARQUET_EXPORT |
| 316 | std::unique_ptr<Encoder> MakeEncoder( |
| 317 | Type::type type_num, Encoding::type encoding, bool use_dictionary = false, |
| 318 | const ColumnDescriptor* descr = NULLPTR, |
| 319 | ::arrow::MemoryPool* pool = ::arrow::default_memory_pool()); |
| 320 | |
| 321 | template <typename DType> |
| 322 | std::unique_ptr<typename EncodingTraits<DType>::Encoder> MakeTypedEncoder( |
| 323 | Encoding::type encoding, bool use_dictionary = false, |
| 324 | const ColumnDescriptor* descr = NULLPTR, |
| 325 | ::arrow::MemoryPool* pool = ::arrow::default_memory_pool()) { |
| 326 | using OutType = typename EncodingTraits<DType>::Encoder; |
| 327 | std::unique_ptr<Encoder> base = |
| 328 | MakeEncoder(DType::type_num, encoding, use_dictionary, descr, pool); |
| 329 | return std::unique_ptr<OutType>(dynamic_cast<OutType*>(base.release())); |
| 330 | } |
| 331 | |
| 332 | PARQUET_EXPORT |
| 333 | std::unique_ptr<Decoder> MakeDecoder(Type::type type_num, Encoding::type encoding, |
| 334 | const ColumnDescriptor* descr = NULLPTR); |
| 335 | |
| 336 | namespace detail { |
| 337 | |
| 338 | PARQUET_EXPORT |
| 339 | std::unique_ptr<Decoder> MakeDictDecoder(Type::type type_num, |
| 340 | const ColumnDescriptor* descr, |
| 341 | ::arrow::MemoryPool* pool); |
| 342 | |
| 343 | } // namespace detail |
| 344 | |
| 345 | template <typename DType> |
| 346 | std::unique_ptr<DictDecoder<DType>> MakeDictDecoder( |
| 347 | const ColumnDescriptor* descr = NULLPTR, |
| 348 | ::arrow::MemoryPool* pool = ::arrow::default_memory_pool()) { |
| 349 | using OutType = DictDecoder<DType>; |
| 350 | auto decoder = detail::MakeDictDecoder(DType::type_num, descr, pool); |
| 351 | return std::unique_ptr<OutType>(dynamic_cast<OutType*>(decoder.release())); |
| 352 | } |
| 353 | |
| 354 | template <typename DType> |
| 355 | std::unique_ptr<typename EncodingTraits<DType>::Decoder> MakeTypedDecoder( |
| 356 | Encoding::type encoding, const ColumnDescriptor* descr = NULLPTR) { |
| 357 | using OutType = typename EncodingTraits<DType>::Decoder; |
| 358 | std::unique_ptr<Decoder> base = MakeDecoder(DType::type_num, encoding, descr); |
| 359 | return std::unique_ptr<OutType>(dynamic_cast<OutType*>(base.release())); |
| 360 | } |
| 361 | |
| 362 | } // namespace parquet |
| 363 | |