| 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 | #include "parquet/encoding.h" |
| 19 | |
| 20 | #include <algorithm> |
| 21 | #include <cstdint> |
| 22 | #include <cstdlib> |
| 23 | #include <memory> |
| 24 | #include <utility> |
| 25 | #include <vector> |
| 26 | |
| 27 | #include "arrow/array.h" |
| 28 | #include "arrow/builder.h" |
| 29 | #include "arrow/util/bit_stream_utils.h" |
| 30 | #include "arrow/util/checked_cast.h" |
| 31 | #include "arrow/util/hashing.h" |
| 32 | #include "arrow/util/logging.h" |
| 33 | #include "arrow/util/rle_encoding.h" |
| 34 | #include "arrow/util/ubsan.h" |
| 35 | |
| 36 | #include "parquet/exception.h" |
| 37 | #include "parquet/platform.h" |
| 38 | #include "parquet/schema.h" |
| 39 | #include "parquet/types.h" |
| 40 | |
| 41 | using arrow::Status; |
| 42 | using arrow::internal::checked_cast; |
| 43 | |
| 44 | namespace parquet { |
| 45 | |
| 46 | constexpr int64_t kInMemoryDefaultCapacity = 1024; |
| 47 | |
| 48 | class EncoderImpl : virtual public Encoder { |
| 49 | public: |
| 50 | EncoderImpl(const ColumnDescriptor* descr, Encoding::type encoding, MemoryPool* pool) |
| 51 | : descr_(descr), |
| 52 | encoding_(encoding), |
| 53 | pool_(pool), |
| 54 | type_length_(descr ? descr->type_length() : -1) {} |
| 55 | |
| 56 | Encoding::type encoding() const override { return encoding_; } |
| 57 | |
| 58 | MemoryPool* memory_pool() const override { return pool_; } |
| 59 | |
| 60 | protected: |
| 61 | // For accessing type-specific metadata, like FIXED_LEN_BYTE_ARRAY |
| 62 | const ColumnDescriptor* descr_; |
| 63 | const Encoding::type encoding_; |
| 64 | MemoryPool* pool_; |
| 65 | |
| 66 | /// Type length from descr |
| 67 | int type_length_; |
| 68 | }; |
| 69 | |
| 70 | // ---------------------------------------------------------------------- |
| 71 | // Plain encoder implementation |
| 72 | |
| 73 | template <typename DType> |
| 74 | class PlainEncoder : public EncoderImpl, virtual public TypedEncoder<DType> { |
| 75 | public: |
| 76 | using T = typename DType::c_type; |
| 77 | |
| 78 | explicit PlainEncoder(const ColumnDescriptor* descr, MemoryPool* pool) |
| 79 | : EncoderImpl(descr, Encoding::PLAIN, pool), sink_(pool) {} |
| 80 | |
| 81 | int64_t EstimatedDataEncodedSize() override { return sink_.length(); } |
| 82 | |
| 83 | std::shared_ptr<Buffer> FlushValues() override { |
| 84 | std::shared_ptr<Buffer> buffer; |
| 85 | PARQUET_THROW_NOT_OK(sink_.Finish(&buffer)); |
| 86 | return buffer; |
| 87 | } |
| 88 | |
| 89 | void Put(const T* buffer, int num_values) override; |
| 90 | |
| 91 | void Put(const arrow::Array& values) override; |
| 92 | |
| 93 | void PutSpaced(const T* src, int num_values, const uint8_t* valid_bits, |
| 94 | int64_t valid_bits_offset) override { |
| 95 | std::shared_ptr<ResizableBuffer> buffer; |
| 96 | PARQUET_THROW_NOT_OK(arrow::AllocateResizableBuffer(this->memory_pool(), |
| 97 | num_values * sizeof(T), &buffer)); |
| 98 | int32_t num_valid_values = 0; |
| 99 | arrow::internal::BitmapReader valid_bits_reader(valid_bits, valid_bits_offset, |
| 100 | num_values); |
| 101 | T* data = reinterpret_cast<T*>(buffer->mutable_data()); |
| 102 | for (int32_t i = 0; i < num_values; i++) { |
| 103 | if (valid_bits_reader.IsSet()) { |
| 104 | data[num_valid_values++] = src[i]; |
| 105 | } |
| 106 | valid_bits_reader.Next(); |
| 107 | } |
| 108 | Put(data, num_valid_values); |
| 109 | } |
| 110 | |
| 111 | void UnsafePutByteArray(const void* data, uint32_t length) { |
| 112 | DCHECK(length == 0 || data != nullptr) << "Value ptr cannot be NULL" ; |
| 113 | sink_.UnsafeAppend(&length, sizeof(uint32_t)); |
| 114 | sink_.UnsafeAppend(data, static_cast<int64_t>(length)); |
| 115 | } |
| 116 | |
| 117 | void Put(const ByteArray& val) { |
| 118 | // Write the result to the output stream |
| 119 | const int64_t increment = static_cast<int64_t>(val.len + sizeof(uint32_t)); |
| 120 | if (ARROW_PREDICT_FALSE(sink_.length() + increment > sink_.capacity())) { |
| 121 | PARQUET_THROW_NOT_OK(sink_.Reserve(increment)); |
| 122 | } |
| 123 | UnsafePutByteArray(val.ptr, val.len); |
| 124 | } |
| 125 | |
| 126 | protected: |
| 127 | arrow::BufferBuilder sink_; |
| 128 | }; |
| 129 | |
| 130 | template <typename DType> |
| 131 | void PlainEncoder<DType>::Put(const T* buffer, int num_values) { |
| 132 | if (num_values > 0) { |
| 133 | PARQUET_THROW_NOT_OK(sink_.Append(buffer, num_values * sizeof(T))); |
| 134 | } |
| 135 | } |
| 136 | |
| 137 | template <> |
| 138 | inline void PlainEncoder<ByteArrayType>::Put(const ByteArray* src, int num_values) { |
| 139 | for (int i = 0; i < num_values; ++i) { |
| 140 | Put(src[i]); |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | template <typename DType> |
| 145 | void PlainEncoder<DType>::Put(const arrow::Array& values) { |
| 146 | ParquetException::NYI(values.type()->ToString()); |
| 147 | } |
| 148 | |
| 149 | void AssertBinary(const arrow::Array& values) { |
| 150 | if (values.type_id() != arrow::Type::BINARY && |
| 151 | values.type_id() != arrow::Type::STRING) { |
| 152 | throw ParquetException("Only BinaryArray and subclasses supported" ); |
| 153 | } |
| 154 | } |
| 155 | |
| 156 | template <> |
| 157 | void PlainEncoder<ByteArrayType>::Put(const arrow::Array& values) { |
| 158 | AssertBinary(values); |
| 159 | const auto& data = checked_cast<const arrow::BinaryArray&>(values); |
| 160 | const int64_t total_bytes = data.value_offset(data.length()) - data.value_offset(0); |
| 161 | PARQUET_THROW_NOT_OK(sink_.Reserve(total_bytes + data.length() * sizeof(uint32_t))); |
| 162 | |
| 163 | if (data.null_count() == 0) { |
| 164 | // no nulls, just dump the data |
| 165 | for (int64_t i = 0; i < data.length(); i++) { |
| 166 | auto view = data.GetView(i); |
| 167 | UnsafePutByteArray(view.data(), static_cast<uint32_t>(view.size())); |
| 168 | } |
| 169 | } else { |
| 170 | for (int64_t i = 0; i < data.length(); i++) { |
| 171 | if (data.IsValid(i)) { |
| 172 | auto view = data.GetView(i); |
| 173 | UnsafePutByteArray(view.data(), static_cast<uint32_t>(view.size())); |
| 174 | } |
| 175 | } |
| 176 | } |
| 177 | } |
| 178 | |
| 179 | template <> |
| 180 | inline void PlainEncoder<FLBAType>::Put(const FixedLenByteArray* src, int num_values) { |
| 181 | if (descr_->type_length() == 0) { |
| 182 | return; |
| 183 | } |
| 184 | for (int i = 0; i < num_values; ++i) { |
| 185 | // Write the result to the output stream |
| 186 | DCHECK(src[i].ptr != nullptr) << "Value ptr cannot be NULL" ; |
| 187 | PARQUET_THROW_NOT_OK(sink_.Append(src[i].ptr, descr_->type_length())); |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | class PlainFLBAEncoder : public PlainEncoder<FLBAType>, virtual public FLBAEncoder { |
| 192 | public: |
| 193 | using BASE = PlainEncoder<FLBAType>; |
| 194 | using BASE::PlainEncoder; |
| 195 | }; |
| 196 | |
| 197 | class PlainBooleanEncoder : public EncoderImpl, |
| 198 | virtual public TypedEncoder<BooleanType>, |
| 199 | virtual public BooleanEncoder { |
| 200 | public: |
| 201 | explicit PlainBooleanEncoder(const ColumnDescriptor* descr, MemoryPool* pool) |
| 202 | : EncoderImpl(descr, Encoding::PLAIN, pool), |
| 203 | bits_available_(kInMemoryDefaultCapacity * 8), |
| 204 | bits_buffer_(AllocateBuffer(pool, kInMemoryDefaultCapacity)), |
| 205 | sink_(pool), |
| 206 | bit_writer_(bits_buffer_->mutable_data(), |
| 207 | static_cast<int>(bits_buffer_->size())) {} |
| 208 | |
| 209 | int64_t EstimatedDataEncodedSize() override; |
| 210 | std::shared_ptr<Buffer> FlushValues() override; |
| 211 | |
| 212 | void Put(const bool* src, int num_values) override; |
| 213 | void Put(const std::vector<bool>& src, int num_values) override; |
| 214 | |
| 215 | void PutSpaced(const bool* src, int num_values, const uint8_t* valid_bits, |
| 216 | int64_t valid_bits_offset) override { |
| 217 | std::shared_ptr<ResizableBuffer> buffer; |
| 218 | PARQUET_THROW_NOT_OK(arrow::AllocateResizableBuffer(this->memory_pool(), |
| 219 | num_values * sizeof(T), &buffer)); |
| 220 | int32_t num_valid_values = 0; |
| 221 | arrow::internal::BitmapReader valid_bits_reader(valid_bits, valid_bits_offset, |
| 222 | num_values); |
| 223 | T* data = reinterpret_cast<T*>(buffer->mutable_data()); |
| 224 | for (int32_t i = 0; i < num_values; i++) { |
| 225 | if (valid_bits_reader.IsSet()) { |
| 226 | data[num_valid_values++] = src[i]; |
| 227 | } |
| 228 | valid_bits_reader.Next(); |
| 229 | } |
| 230 | Put(data, num_valid_values); |
| 231 | } |
| 232 | |
| 233 | void Put(const arrow::Array& values) override { |
| 234 | ParquetException::NYI("Direct Arrow to Boolean writes not implemented" ); |
| 235 | } |
| 236 | |
| 237 | private: |
| 238 | int bits_available_; |
| 239 | std::shared_ptr<ResizableBuffer> bits_buffer_; |
| 240 | arrow::BufferBuilder sink_; |
| 241 | arrow::BitUtil::BitWriter bit_writer_; |
| 242 | |
| 243 | template <typename SequenceType> |
| 244 | void PutImpl(const SequenceType& src, int num_values); |
| 245 | }; |
| 246 | |
| 247 | template <typename SequenceType> |
| 248 | void PlainBooleanEncoder::PutImpl(const SequenceType& src, int num_values) { |
| 249 | int bit_offset = 0; |
| 250 | if (bits_available_ > 0) { |
| 251 | int bits_to_write = std::min(bits_available_, num_values); |
| 252 | for (int i = 0; i < bits_to_write; i++) { |
| 253 | bit_writer_.PutValue(src[i], 1); |
| 254 | } |
| 255 | bits_available_ -= bits_to_write; |
| 256 | bit_offset = bits_to_write; |
| 257 | |
| 258 | if (bits_available_ == 0) { |
| 259 | bit_writer_.Flush(); |
| 260 | PARQUET_THROW_NOT_OK( |
| 261 | sink_.Append(bit_writer_.buffer(), bit_writer_.bytes_written())); |
| 262 | bit_writer_.Clear(); |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | int bits_remaining = num_values - bit_offset; |
| 267 | while (bit_offset < num_values) { |
| 268 | bits_available_ = static_cast<int>(bits_buffer_->size()) * 8; |
| 269 | |
| 270 | int bits_to_write = std::min(bits_available_, bits_remaining); |
| 271 | for (int i = bit_offset; i < bit_offset + bits_to_write; i++) { |
| 272 | bit_writer_.PutValue(src[i], 1); |
| 273 | } |
| 274 | bit_offset += bits_to_write; |
| 275 | bits_available_ -= bits_to_write; |
| 276 | bits_remaining -= bits_to_write; |
| 277 | |
| 278 | if (bits_available_ == 0) { |
| 279 | bit_writer_.Flush(); |
| 280 | PARQUET_THROW_NOT_OK( |
| 281 | sink_.Append(bit_writer_.buffer(), bit_writer_.bytes_written())); |
| 282 | bit_writer_.Clear(); |
| 283 | } |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | int64_t PlainBooleanEncoder::EstimatedDataEncodedSize() { |
| 288 | int64_t position = sink_.length(); |
| 289 | return position + bit_writer_.bytes_written(); |
| 290 | } |
| 291 | |
| 292 | std::shared_ptr<Buffer> PlainBooleanEncoder::FlushValues() { |
| 293 | if (bits_available_ > 0) { |
| 294 | bit_writer_.Flush(); |
| 295 | PARQUET_THROW_NOT_OK(sink_.Append(bit_writer_.buffer(), bit_writer_.bytes_written())); |
| 296 | bit_writer_.Clear(); |
| 297 | bits_available_ = static_cast<int>(bits_buffer_->size()) * 8; |
| 298 | } |
| 299 | |
| 300 | std::shared_ptr<Buffer> buffer; |
| 301 | PARQUET_THROW_NOT_OK(sink_.Finish(&buffer)); |
| 302 | return buffer; |
| 303 | } |
| 304 | |
| 305 | void PlainBooleanEncoder::Put(const bool* src, int num_values) { |
| 306 | PutImpl(src, num_values); |
| 307 | } |
| 308 | |
| 309 | void PlainBooleanEncoder::Put(const std::vector<bool>& src, int num_values) { |
| 310 | PutImpl(src, num_values); |
| 311 | } |
| 312 | |
| 313 | // ---------------------------------------------------------------------- |
| 314 | // DictEncoder<T> implementations |
| 315 | |
| 316 | template <typename DType> |
| 317 | struct DictEncoderTraits { |
| 318 | using c_type = typename DType::c_type; |
| 319 | using MemoTableType = arrow::internal::ScalarMemoTable<c_type>; |
| 320 | }; |
| 321 | |
| 322 | template <> |
| 323 | struct DictEncoderTraits<ByteArrayType> { |
| 324 | using MemoTableType = arrow::internal::BinaryMemoTable; |
| 325 | }; |
| 326 | |
| 327 | template <> |
| 328 | struct DictEncoderTraits<FLBAType> { |
| 329 | using MemoTableType = arrow::internal::BinaryMemoTable; |
| 330 | }; |
| 331 | |
| 332 | // Initially 1024 elements |
| 333 | static constexpr int32_t kInitialHashTableSize = 1 << 10; |
| 334 | |
| 335 | /// See the dictionary encoding section of |
| 336 | /// https://github.com/Parquet/parquet-format. The encoding supports |
| 337 | /// streaming encoding. Values are encoded as they are added while the |
| 338 | /// dictionary is being constructed. At any time, the buffered values |
| 339 | /// can be written out with the current dictionary size. More values |
| 340 | /// can then be added to the encoder, including new dictionary |
| 341 | /// entries. |
| 342 | template <typename DType> |
| 343 | class DictEncoderImpl : public EncoderImpl, virtual public DictEncoder<DType> { |
| 344 | using MemoTableType = typename DictEncoderTraits<DType>::MemoTableType; |
| 345 | |
| 346 | public: |
| 347 | typedef typename DType::c_type T; |
| 348 | |
| 349 | explicit DictEncoderImpl(const ColumnDescriptor* desc, MemoryPool* pool) |
| 350 | : EncoderImpl(desc, Encoding::PLAIN_DICTIONARY, pool), |
| 351 | dict_encoded_size_(0), |
| 352 | memo_table_(pool, kInitialHashTableSize) {} |
| 353 | |
| 354 | ~DictEncoderImpl() override { DCHECK(buffered_indices_.empty()); } |
| 355 | |
| 356 | int dict_encoded_size() override { return dict_encoded_size_; } |
| 357 | |
| 358 | int WriteIndices(uint8_t* buffer, int buffer_len) override { |
| 359 | // Write bit width in first byte |
| 360 | *buffer = static_cast<uint8_t>(bit_width()); |
| 361 | ++buffer; |
| 362 | --buffer_len; |
| 363 | |
| 364 | arrow::util::RleEncoder encoder(buffer, buffer_len, bit_width()); |
| 365 | |
| 366 | for (int32_t index : buffered_indices_) { |
| 367 | if (!encoder.Put(index)) return -1; |
| 368 | } |
| 369 | encoder.Flush(); |
| 370 | |
| 371 | ClearIndices(); |
| 372 | return 1 + encoder.len(); |
| 373 | } |
| 374 | |
| 375 | void set_type_length(int type_length) { this->type_length_ = type_length; } |
| 376 | |
| 377 | /// Returns a conservative estimate of the number of bytes needed to encode the buffered |
| 378 | /// indices. Used to size the buffer passed to WriteIndices(). |
| 379 | int64_t EstimatedDataEncodedSize() override { |
| 380 | // Note: because of the way RleEncoder::CheckBufferFull() is called, we have to |
| 381 | // reserve |
| 382 | // an extra "RleEncoder::MinBufferSize" bytes. These extra bytes won't be used |
| 383 | // but not reserving them would cause the encoder to fail. |
| 384 | return 1 + |
| 385 | arrow::util::RleEncoder::MaxBufferSize( |
| 386 | bit_width(), static_cast<int>(buffered_indices_.size())) + |
| 387 | arrow::util::RleEncoder::MinBufferSize(bit_width()); |
| 388 | } |
| 389 | |
| 390 | /// The minimum bit width required to encode the currently buffered indices. |
| 391 | int bit_width() const override { |
| 392 | if (ARROW_PREDICT_FALSE(num_entries() == 0)) return 0; |
| 393 | if (ARROW_PREDICT_FALSE(num_entries() == 1)) return 1; |
| 394 | return BitUtil::Log2(num_entries()); |
| 395 | } |
| 396 | |
| 397 | /// Encode value. Note that this does not actually write any data, just |
| 398 | /// buffers the value's index to be written later. |
| 399 | inline void Put(const T& value); |
| 400 | |
| 401 | // Not implemented for other data types |
| 402 | inline void PutByteArray(const void* ptr, int32_t length); |
| 403 | |
| 404 | void Put(const T* src, int num_values) override { |
| 405 | for (int32_t i = 0; i < num_values; i++) { |
| 406 | Put(src[i]); |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | void PutSpaced(const T* src, int num_values, const uint8_t* valid_bits, |
| 411 | int64_t valid_bits_offset) override { |
| 412 | arrow::internal::BitmapReader valid_bits_reader(valid_bits, valid_bits_offset, |
| 413 | num_values); |
| 414 | for (int32_t i = 0; i < num_values; i++) { |
| 415 | if (valid_bits_reader.IsSet()) { |
| 416 | Put(src[i]); |
| 417 | } |
| 418 | valid_bits_reader.Next(); |
| 419 | } |
| 420 | } |
| 421 | |
| 422 | void Put(const arrow::Array& values) override; |
| 423 | void PutDictionary(const arrow::Array& values) override; |
| 424 | |
| 425 | template <typename ArrowType> |
| 426 | void PutIndicesTyped(const arrow::Array& data) { |
| 427 | using ArrayType = typename arrow::TypeTraits<ArrowType>::ArrayType; |
| 428 | const auto& indices = checked_cast<const ArrayType&>(data); |
| 429 | auto values = indices.raw_values(); |
| 430 | |
| 431 | size_t buffer_position = buffered_indices_.size(); |
| 432 | buffered_indices_.resize( |
| 433 | buffer_position + static_cast<size_t>(indices.length() - indices.null_count())); |
| 434 | if (indices.null_count() > 0) { |
| 435 | arrow::internal::BitmapReader valid_bits_reader(indices.null_bitmap_data(), |
| 436 | indices.offset(), indices.length()); |
| 437 | for (int64_t i = 0; i < indices.length(); ++i) { |
| 438 | if (valid_bits_reader.IsSet()) { |
| 439 | buffered_indices_[buffer_position++] = static_cast<int32_t>(values[i]); |
| 440 | } |
| 441 | valid_bits_reader.Next(); |
| 442 | } |
| 443 | } else { |
| 444 | for (int64_t i = 0; i < indices.length(); ++i) { |
| 445 | buffered_indices_[buffer_position++] = static_cast<int32_t>(values[i]); |
| 446 | } |
| 447 | } |
| 448 | } |
| 449 | |
| 450 | void PutIndices(const arrow::Array& data) override { |
| 451 | switch (data.type()->id()) { |
| 452 | case arrow::Type::INT8: |
| 453 | return PutIndicesTyped<arrow::Int8Type>(data); |
| 454 | case arrow::Type::INT16: |
| 455 | return PutIndicesTyped<arrow::Int16Type>(data); |
| 456 | case arrow::Type::INT32: |
| 457 | return PutIndicesTyped<arrow::Int32Type>(data); |
| 458 | case arrow::Type::INT64: |
| 459 | return PutIndicesTyped<arrow::Int64Type>(data); |
| 460 | default: |
| 461 | throw ParquetException("Dictionary indices were not signed integer" ); |
| 462 | } |
| 463 | } |
| 464 | |
| 465 | std::shared_ptr<Buffer> FlushValues() override { |
| 466 | std::shared_ptr<ResizableBuffer> buffer = |
| 467 | AllocateBuffer(this->pool_, EstimatedDataEncodedSize()); |
| 468 | int result_size = WriteIndices(buffer->mutable_data(), |
| 469 | static_cast<int>(EstimatedDataEncodedSize())); |
| 470 | PARQUET_THROW_NOT_OK(buffer->Resize(result_size, false)); |
| 471 | return std::move(buffer); |
| 472 | } |
| 473 | |
| 474 | /// Writes out the encoded dictionary to buffer. buffer must be preallocated to |
| 475 | /// dict_encoded_size() bytes. |
| 476 | void WriteDict(uint8_t* buffer) override; |
| 477 | |
| 478 | /// The number of entries in the dictionary. |
| 479 | int num_entries() const override { return memo_table_.size(); } |
| 480 | |
| 481 | private: |
| 482 | /// Clears all the indices (but leaves the dictionary). |
| 483 | void ClearIndices() { buffered_indices_.clear(); } |
| 484 | |
| 485 | /// Indices that have not yet be written out by WriteIndices(). |
| 486 | std::vector<int32_t> buffered_indices_; |
| 487 | |
| 488 | /// The number of bytes needed to encode the dictionary. |
| 489 | int dict_encoded_size_; |
| 490 | |
| 491 | MemoTableType memo_table_; |
| 492 | }; |
| 493 | |
| 494 | template <typename DType> |
| 495 | void DictEncoderImpl<DType>::WriteDict(uint8_t* buffer) { |
| 496 | // For primitive types, only a memcpy |
| 497 | DCHECK_EQ(static_cast<size_t>(dict_encoded_size_), sizeof(T) * memo_table_.size()); |
| 498 | memo_table_.CopyValues(0 /* start_pos */, reinterpret_cast<T*>(buffer)); |
| 499 | } |
| 500 | |
| 501 | // ByteArray and FLBA already have the dictionary encoded in their data heaps |
| 502 | template <> |
| 503 | void DictEncoderImpl<ByteArrayType>::WriteDict(uint8_t* buffer) { |
| 504 | memo_table_.VisitValues(0, [&buffer](const arrow::util::string_view& v) { |
| 505 | uint32_t len = static_cast<uint32_t>(v.length()); |
| 506 | memcpy(buffer, &len, sizeof(len)); |
| 507 | buffer += sizeof(len); |
| 508 | memcpy(buffer, v.data(), len); |
| 509 | buffer += len; |
| 510 | }); |
| 511 | } |
| 512 | |
| 513 | template <> |
| 514 | void DictEncoderImpl<FLBAType>::WriteDict(uint8_t* buffer) { |
| 515 | memo_table_.VisitValues(0, [&](const arrow::util::string_view& v) { |
| 516 | DCHECK_EQ(v.length(), static_cast<size_t>(type_length_)); |
| 517 | memcpy(buffer, v.data(), type_length_); |
| 518 | buffer += type_length_; |
| 519 | }); |
| 520 | } |
| 521 | |
| 522 | template <typename DType> |
| 523 | inline void DictEncoderImpl<DType>::Put(const T& v) { |
| 524 | // Put() implementation for primitive types |
| 525 | auto on_found = [](int32_t memo_index) {}; |
| 526 | auto on_not_found = [this](int32_t memo_index) { |
| 527 | dict_encoded_size_ += static_cast<int>(sizeof(T)); |
| 528 | }; |
| 529 | |
| 530 | auto memo_index = memo_table_.GetOrInsert(v, on_found, on_not_found); |
| 531 | buffered_indices_.push_back(memo_index); |
| 532 | } |
| 533 | |
| 534 | template <typename DType> |
| 535 | inline void DictEncoderImpl<DType>::PutByteArray(const void* ptr, int32_t length) { |
| 536 | DCHECK(false); |
| 537 | } |
| 538 | |
| 539 | template <> |
| 540 | inline void DictEncoderImpl<ByteArrayType>::PutByteArray(const void* ptr, |
| 541 | int32_t length) { |
| 542 | static const uint8_t empty[] = {0}; |
| 543 | |
| 544 | auto on_found = [](int32_t memo_index) {}; |
| 545 | auto on_not_found = [&](int32_t memo_index) { |
| 546 | dict_encoded_size_ += static_cast<int>(length + sizeof(uint32_t)); |
| 547 | }; |
| 548 | |
| 549 | DCHECK(ptr != nullptr || length == 0); |
| 550 | ptr = (ptr != nullptr) ? ptr : empty; |
| 551 | auto memo_index = memo_table_.GetOrInsert(ptr, length, on_found, on_not_found); |
| 552 | buffered_indices_.push_back(memo_index); |
| 553 | } |
| 554 | |
| 555 | template <> |
| 556 | inline void DictEncoderImpl<ByteArrayType>::Put(const ByteArray& val) { |
| 557 | return PutByteArray(val.ptr, static_cast<int32_t>(val.len)); |
| 558 | } |
| 559 | |
| 560 | template <> |
| 561 | inline void DictEncoderImpl<FLBAType>::Put(const FixedLenByteArray& v) { |
| 562 | static const uint8_t empty[] = {0}; |
| 563 | |
| 564 | auto on_found = [](int32_t memo_index) {}; |
| 565 | auto on_not_found = [this](int32_t memo_index) { dict_encoded_size_ += type_length_; }; |
| 566 | |
| 567 | DCHECK(v.ptr != nullptr || type_length_ == 0); |
| 568 | const void* ptr = (v.ptr != nullptr) ? v.ptr : empty; |
| 569 | auto memo_index = memo_table_.GetOrInsert(ptr, type_length_, on_found, on_not_found); |
| 570 | buffered_indices_.push_back(memo_index); |
| 571 | } |
| 572 | |
| 573 | template <typename DType> |
| 574 | void DictEncoderImpl<DType>::Put(const arrow::Array& values) { |
| 575 | ParquetException::NYI(values.type()->ToString()); |
| 576 | } |
| 577 | |
| 578 | template <> |
| 579 | void DictEncoderImpl<ByteArrayType>::Put(const arrow::Array& values) { |
| 580 | AssertBinary(values); |
| 581 | const auto& data = checked_cast<const arrow::BinaryArray&>(values); |
| 582 | if (data.null_count() == 0) { |
| 583 | // no nulls, just dump the data |
| 584 | for (int64_t i = 0; i < data.length(); i++) { |
| 585 | auto view = data.GetView(i); |
| 586 | PutByteArray(view.data(), static_cast<int32_t>(view.size())); |
| 587 | } |
| 588 | } else { |
| 589 | for (int64_t i = 0; i < data.length(); i++) { |
| 590 | if (data.IsValid(i)) { |
| 591 | auto view = data.GetView(i); |
| 592 | PutByteArray(view.data(), static_cast<int32_t>(view.size())); |
| 593 | } |
| 594 | } |
| 595 | } |
| 596 | } |
| 597 | |
| 598 | template <typename DType> |
| 599 | void DictEncoderImpl<DType>::PutDictionary(const arrow::Array& values) { |
| 600 | ParquetException::NYI(values.type()->ToString()); |
| 601 | } |
| 602 | |
| 603 | template <> |
| 604 | void DictEncoderImpl<ByteArrayType>::PutDictionary(const arrow::Array& values) { |
| 605 | AssertBinary(values); |
| 606 | if (this->num_entries() > 0) { |
| 607 | throw ParquetException("Can only call PutDictionary on an empty DictEncoder" ); |
| 608 | } |
| 609 | |
| 610 | const auto& data = checked_cast<const arrow::BinaryArray&>(values); |
| 611 | if (data.null_count() > 0) { |
| 612 | throw ParquetException("Inserted binary dictionary cannot cannot contain nulls" ); |
| 613 | } |
| 614 | for (int64_t i = 0; i < data.length(); i++) { |
| 615 | auto v = data.GetView(i); |
| 616 | dict_encoded_size_ += static_cast<int>(v.size() + sizeof(uint32_t)); |
| 617 | ARROW_IGNORE_EXPR( |
| 618 | memo_table_.GetOrInsert(v.data(), static_cast<int32_t>(v.size()), |
| 619 | /*on_found=*/[](int32_t memo_index) {}, |
| 620 | /*on_not_found=*/[](int32_t memo_index) {})); |
| 621 | } |
| 622 | } |
| 623 | |
| 624 | // ---------------------------------------------------------------------- |
| 625 | // Encoder and decoder factory functions |
| 626 | |
| 627 | std::unique_ptr<Encoder> MakeEncoder(Type::type type_num, Encoding::type encoding, |
| 628 | bool use_dictionary, const ColumnDescriptor* descr, |
| 629 | MemoryPool* pool) { |
| 630 | if (use_dictionary) { |
| 631 | switch (type_num) { |
| 632 | case Type::INT32: |
| 633 | return std::unique_ptr<Encoder>(new DictEncoderImpl<Int32Type>(descr, pool)); |
| 634 | case Type::INT64: |
| 635 | return std::unique_ptr<Encoder>(new DictEncoderImpl<Int64Type>(descr, pool)); |
| 636 | case Type::INT96: |
| 637 | return std::unique_ptr<Encoder>(new DictEncoderImpl<Int96Type>(descr, pool)); |
| 638 | case Type::FLOAT: |
| 639 | return std::unique_ptr<Encoder>(new DictEncoderImpl<FloatType>(descr, pool)); |
| 640 | case Type::DOUBLE: |
| 641 | return std::unique_ptr<Encoder>(new DictEncoderImpl<DoubleType>(descr, pool)); |
| 642 | case Type::BYTE_ARRAY: |
| 643 | return std::unique_ptr<Encoder>(new DictEncoderImpl<ByteArrayType>(descr, pool)); |
| 644 | case Type::FIXED_LEN_BYTE_ARRAY: |
| 645 | return std::unique_ptr<Encoder>(new DictEncoderImpl<FLBAType>(descr, pool)); |
| 646 | default: |
| 647 | DCHECK(false) << "Encoder not implemented" ; |
| 648 | break; |
| 649 | } |
| 650 | } else if (encoding == Encoding::PLAIN) { |
| 651 | switch (type_num) { |
| 652 | case Type::BOOLEAN: |
| 653 | return std::unique_ptr<Encoder>(new PlainBooleanEncoder(descr, pool)); |
| 654 | case Type::INT32: |
| 655 | return std::unique_ptr<Encoder>(new PlainEncoder<Int32Type>(descr, pool)); |
| 656 | case Type::INT64: |
| 657 | return std::unique_ptr<Encoder>(new PlainEncoder<Int64Type>(descr, pool)); |
| 658 | case Type::INT96: |
| 659 | return std::unique_ptr<Encoder>(new PlainEncoder<Int96Type>(descr, pool)); |
| 660 | case Type::FLOAT: |
| 661 | return std::unique_ptr<Encoder>(new PlainEncoder<FloatType>(descr, pool)); |
| 662 | case Type::DOUBLE: |
| 663 | return std::unique_ptr<Encoder>(new PlainEncoder<DoubleType>(descr, pool)); |
| 664 | case Type::BYTE_ARRAY: |
| 665 | return std::unique_ptr<Encoder>(new PlainEncoder<ByteArrayType>(descr, pool)); |
| 666 | case Type::FIXED_LEN_BYTE_ARRAY: |
| 667 | return std::unique_ptr<Encoder>(new PlainEncoder<FLBAType>(descr, pool)); |
| 668 | default: |
| 669 | DCHECK(false) << "Encoder not implemented" ; |
| 670 | break; |
| 671 | } |
| 672 | } else { |
| 673 | ParquetException::NYI("Selected encoding is not supported" ); |
| 674 | } |
| 675 | DCHECK(false) << "Should not be able to reach this code" ; |
| 676 | return nullptr; |
| 677 | } |
| 678 | |
| 679 | class DecoderImpl : virtual public Decoder { |
| 680 | public: |
| 681 | void SetData(int num_values, const uint8_t* data, int len) override { |
| 682 | num_values_ = num_values; |
| 683 | data_ = data; |
| 684 | len_ = len; |
| 685 | } |
| 686 | |
| 687 | int values_left() const override { return num_values_; } |
| 688 | Encoding::type encoding() const override { return encoding_; } |
| 689 | |
| 690 | protected: |
| 691 | explicit DecoderImpl(const ColumnDescriptor* descr, Encoding::type encoding) |
| 692 | : descr_(descr), encoding_(encoding), num_values_(0), data_(NULLPTR), len_(0) {} |
| 693 | |
| 694 | // For accessing type-specific metadata, like FIXED_LEN_BYTE_ARRAY |
| 695 | const ColumnDescriptor* descr_; |
| 696 | |
| 697 | const Encoding::type encoding_; |
| 698 | int num_values_; |
| 699 | const uint8_t* data_; |
| 700 | int len_; |
| 701 | int type_length_; |
| 702 | }; |
| 703 | |
| 704 | template <typename DType> |
| 705 | class PlainDecoder : public DecoderImpl, virtual public TypedDecoder<DType> { |
| 706 | public: |
| 707 | using T = typename DType::c_type; |
| 708 | explicit PlainDecoder(const ColumnDescriptor* descr); |
| 709 | |
| 710 | int Decode(T* buffer, int max_values) override; |
| 711 | }; |
| 712 | |
| 713 | template <typename DType> |
| 714 | PlainDecoder<DType>::PlainDecoder(const ColumnDescriptor* descr) |
| 715 | : DecoderImpl(descr, Encoding::PLAIN) { |
| 716 | if (descr_ && descr_->physical_type() == Type::FIXED_LEN_BYTE_ARRAY) { |
| 717 | type_length_ = descr_->type_length(); |
| 718 | } else { |
| 719 | type_length_ = -1; |
| 720 | } |
| 721 | } |
| 722 | |
| 723 | // Decode routine templated on C++ type rather than type enum |
| 724 | template <typename T> |
| 725 | inline int DecodePlain(const uint8_t* data, int64_t data_size, int num_values, |
| 726 | int type_length, T* out) { |
| 727 | int bytes_to_decode = num_values * static_cast<int>(sizeof(T)); |
| 728 | if (data_size < bytes_to_decode) { |
| 729 | ParquetException::EofException(); |
| 730 | } |
| 731 | // If bytes_to_decode == 0, data could be null |
| 732 | if (bytes_to_decode > 0) { |
| 733 | memcpy(out, data, bytes_to_decode); |
| 734 | } |
| 735 | return bytes_to_decode; |
| 736 | } |
| 737 | |
| 738 | // Template specialization for BYTE_ARRAY. The written values do not own their |
| 739 | // own data. |
| 740 | template <> |
| 741 | inline int DecodePlain<ByteArray>(const uint8_t* data, int64_t data_size, int num_values, |
| 742 | int type_length, ByteArray* out) { |
| 743 | int bytes_decoded = 0; |
| 744 | int increment; |
| 745 | for (int i = 0; i < num_values; ++i) { |
| 746 | uint32_t len = out[i].len = arrow::util::SafeLoadAs<uint32_t>(data); |
| 747 | increment = static_cast<int>(sizeof(uint32_t) + len); |
| 748 | if (data_size < increment) ParquetException::EofException(); |
| 749 | out[i].ptr = data + sizeof(uint32_t); |
| 750 | data += increment; |
| 751 | data_size -= increment; |
| 752 | bytes_decoded += increment; |
| 753 | } |
| 754 | return bytes_decoded; |
| 755 | } |
| 756 | |
| 757 | // Template specialization for FIXED_LEN_BYTE_ARRAY. The written values do not |
| 758 | // own their own data. |
| 759 | template <> |
| 760 | inline int DecodePlain<FixedLenByteArray>(const uint8_t* data, int64_t data_size, |
| 761 | int num_values, int type_length, |
| 762 | FixedLenByteArray* out) { |
| 763 | int bytes_to_decode = type_length * num_values; |
| 764 | if (data_size < bytes_to_decode) { |
| 765 | ParquetException::EofException(); |
| 766 | } |
| 767 | for (int i = 0; i < num_values; ++i) { |
| 768 | out[i].ptr = data; |
| 769 | data += type_length; |
| 770 | data_size -= type_length; |
| 771 | } |
| 772 | return bytes_to_decode; |
| 773 | } |
| 774 | |
| 775 | template <typename DType> |
| 776 | int PlainDecoder<DType>::Decode(T* buffer, int max_values) { |
| 777 | max_values = std::min(max_values, num_values_); |
| 778 | int bytes_consumed = DecodePlain<T>(data_, len_, max_values, type_length_, buffer); |
| 779 | data_ += bytes_consumed; |
| 780 | len_ -= bytes_consumed; |
| 781 | num_values_ -= max_values; |
| 782 | return max_values; |
| 783 | } |
| 784 | |
| 785 | class PlainBooleanDecoder : public DecoderImpl, |
| 786 | virtual public TypedDecoder<BooleanType>, |
| 787 | virtual public BooleanDecoder { |
| 788 | public: |
| 789 | explicit PlainBooleanDecoder(const ColumnDescriptor* descr); |
| 790 | void SetData(int num_values, const uint8_t* data, int len) override; |
| 791 | |
| 792 | // Two flavors of bool decoding |
| 793 | int Decode(uint8_t* buffer, int max_values) override; |
| 794 | int Decode(bool* buffer, int max_values) override; |
| 795 | |
| 796 | private: |
| 797 | std::unique_ptr<arrow::BitUtil::BitReader> bit_reader_; |
| 798 | }; |
| 799 | |
| 800 | PlainBooleanDecoder::PlainBooleanDecoder(const ColumnDescriptor* descr) |
| 801 | : DecoderImpl(descr, Encoding::PLAIN) {} |
| 802 | |
| 803 | void PlainBooleanDecoder::SetData(int num_values, const uint8_t* data, int len) { |
| 804 | num_values_ = num_values; |
| 805 | bit_reader_.reset(new BitUtil::BitReader(data, len)); |
| 806 | } |
| 807 | |
| 808 | int PlainBooleanDecoder::Decode(uint8_t* buffer, int max_values) { |
| 809 | max_values = std::min(max_values, num_values_); |
| 810 | bool val; |
| 811 | arrow::internal::BitmapWriter bit_writer(buffer, 0, max_values); |
| 812 | for (int i = 0; i < max_values; ++i) { |
| 813 | if (!bit_reader_->GetValue(1, &val)) { |
| 814 | ParquetException::EofException(); |
| 815 | } |
| 816 | if (val) { |
| 817 | bit_writer.Set(); |
| 818 | } |
| 819 | bit_writer.Next(); |
| 820 | } |
| 821 | bit_writer.Finish(); |
| 822 | num_values_ -= max_values; |
| 823 | return max_values; |
| 824 | } |
| 825 | |
| 826 | int PlainBooleanDecoder::Decode(bool* buffer, int max_values) { |
| 827 | max_values = std::min(max_values, num_values_); |
| 828 | if (bit_reader_->GetBatch(1, buffer, max_values) != max_values) { |
| 829 | ParquetException::EofException(); |
| 830 | } |
| 831 | num_values_ -= max_values; |
| 832 | return max_values; |
| 833 | } |
| 834 | |
| 835 | struct ArrowBinaryHelper { |
| 836 | explicit ArrowBinaryHelper(ArrowBinaryAccumulator* out) { |
| 837 | this->out = out; |
| 838 | this->builder = out->builder.get(); |
| 839 | this->chunk_space_remaining = |
| 840 | ::arrow::kBinaryMemoryLimit - this->builder->value_data_length(); |
| 841 | } |
| 842 | |
| 843 | Status PushChunk() { |
| 844 | std::shared_ptr<::arrow::Array> result; |
| 845 | RETURN_NOT_OK(builder->Finish(&result)); |
| 846 | out->chunks.push_back(result); |
| 847 | chunk_space_remaining = ::arrow::kBinaryMemoryLimit; |
| 848 | return Status::OK(); |
| 849 | } |
| 850 | |
| 851 | bool CanFit(int64_t length) const { return length <= chunk_space_remaining; } |
| 852 | |
| 853 | void UnsafeAppend(const uint8_t* data, int32_t length) { |
| 854 | chunk_space_remaining -= length; |
| 855 | builder->UnsafeAppend(data, length); |
| 856 | } |
| 857 | |
| 858 | void UnsafeAppendNull() { builder->UnsafeAppendNull(); } |
| 859 | |
| 860 | Status Append(const uint8_t* data, int32_t length) { |
| 861 | chunk_space_remaining -= length; |
| 862 | return builder->Append(data, length); |
| 863 | } |
| 864 | |
| 865 | Status AppendNull() { return builder->AppendNull(); } |
| 866 | |
| 867 | ArrowBinaryAccumulator* out; |
| 868 | arrow::BinaryBuilder* builder; |
| 869 | int64_t chunk_space_remaining; |
| 870 | }; |
| 871 | |
| 872 | class PlainByteArrayDecoder : public PlainDecoder<ByteArrayType>, |
| 873 | virtual public ByteArrayDecoder { |
| 874 | public: |
| 875 | using Base = PlainDecoder<ByteArrayType>; |
| 876 | using Base::DecodeSpaced; |
| 877 | using Base::PlainDecoder; |
| 878 | |
| 879 | // ---------------------------------------------------------------------- |
| 880 | // Dictionary read paths |
| 881 | |
| 882 | int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits, |
| 883 | int64_t valid_bits_offset, |
| 884 | arrow::BinaryDictionary32Builder* builder) override { |
| 885 | int result = 0; |
| 886 | PARQUET_THROW_NOT_OK(DecodeArrow(num_values, null_count, valid_bits, |
| 887 | valid_bits_offset, builder, &result)); |
| 888 | return result; |
| 889 | } |
| 890 | |
| 891 | int DecodeArrowNonNull(int num_values, |
| 892 | arrow::BinaryDictionary32Builder* builder) override { |
| 893 | int result = 0; |
| 894 | PARQUET_THROW_NOT_OK(DecodeArrowNonNull(num_values, builder, &result)); |
| 895 | return result; |
| 896 | } |
| 897 | |
| 898 | // ---------------------------------------------------------------------- |
| 899 | // Optimized dense binary read paths |
| 900 | |
| 901 | int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits, |
| 902 | int64_t valid_bits_offset, ArrowBinaryAccumulator* out) override { |
| 903 | int result = 0; |
| 904 | PARQUET_THROW_NOT_OK(DecodeArrowDense(num_values, null_count, valid_bits, |
| 905 | valid_bits_offset, out, &result)); |
| 906 | return result; |
| 907 | } |
| 908 | |
| 909 | int DecodeArrowNonNull(int num_values, ArrowBinaryAccumulator* out) override { |
| 910 | int result = 0; |
| 911 | PARQUET_THROW_NOT_OK(DecodeArrowDenseNonNull(num_values, out, &result)); |
| 912 | return result; |
| 913 | } |
| 914 | |
| 915 | private: |
| 916 | Status DecodeArrowDense(int num_values, int null_count, const uint8_t* valid_bits, |
| 917 | int64_t valid_bits_offset, ArrowBinaryAccumulator* out, |
| 918 | int* out_values_decoded) { |
| 919 | ArrowBinaryHelper helper(out); |
| 920 | arrow::internal::BitmapReader bit_reader(valid_bits, valid_bits_offset, num_values); |
| 921 | int values_decoded = 0; |
| 922 | |
| 923 | RETURN_NOT_OK(helper.builder->Reserve(num_values)); |
| 924 | RETURN_NOT_OK(helper.builder->ReserveData( |
| 925 | std::min<int64_t>(len_, helper.chunk_space_remaining))); |
| 926 | for (int i = 0; i < num_values; ++i) { |
| 927 | if (bit_reader.IsSet()) { |
| 928 | auto value_len = static_cast<int32_t>(arrow::util::SafeLoadAs<uint32_t>(data_)); |
| 929 | int increment = static_cast<int>(sizeof(uint32_t) + value_len); |
| 930 | if (ARROW_PREDICT_FALSE(len_ < increment)) ParquetException::EofException(); |
| 931 | if (ARROW_PREDICT_FALSE(!helper.CanFit(value_len))) { |
| 932 | // This element would exceed the capacity of a chunk |
| 933 | RETURN_NOT_OK(helper.PushChunk()); |
| 934 | RETURN_NOT_OK(helper.builder->Reserve(num_values - i)); |
| 935 | RETURN_NOT_OK(helper.builder->ReserveData( |
| 936 | std::min<int64_t>(len_, helper.chunk_space_remaining))); |
| 937 | } |
| 938 | helper.UnsafeAppend(data_ + sizeof(uint32_t), value_len); |
| 939 | data_ += increment; |
| 940 | len_ -= increment; |
| 941 | ++values_decoded; |
| 942 | } else { |
| 943 | helper.UnsafeAppendNull(); |
| 944 | } |
| 945 | bit_reader.Next(); |
| 946 | } |
| 947 | |
| 948 | num_values_ -= values_decoded; |
| 949 | *out_values_decoded = values_decoded; |
| 950 | return Status::OK(); |
| 951 | } |
| 952 | |
| 953 | Status DecodeArrowDenseNonNull(int num_values, ArrowBinaryAccumulator* out, |
| 954 | int* values_decoded) { |
| 955 | ArrowBinaryHelper helper(out); |
| 956 | num_values = std::min(num_values, num_values_); |
| 957 | RETURN_NOT_OK(helper.builder->Reserve(num_values)); |
| 958 | RETURN_NOT_OK(helper.builder->ReserveData( |
| 959 | std::min<int64_t>(len_, helper.chunk_space_remaining))); |
| 960 | for (int i = 0; i < num_values; ++i) { |
| 961 | int32_t value_len = static_cast<int32_t>(arrow::util::SafeLoadAs<uint32_t>(data_)); |
| 962 | int increment = static_cast<int>(sizeof(uint32_t) + value_len); |
| 963 | if (ARROW_PREDICT_FALSE(len_ < increment)) ParquetException::EofException(); |
| 964 | if (ARROW_PREDICT_FALSE(!helper.CanFit(value_len))) { |
| 965 | // This element would exceed the capacity of a chunk |
| 966 | RETURN_NOT_OK(helper.PushChunk()); |
| 967 | RETURN_NOT_OK(helper.builder->Reserve(num_values - i)); |
| 968 | RETURN_NOT_OK(helper.builder->ReserveData( |
| 969 | std::min<int64_t>(len_, helper.chunk_space_remaining))); |
| 970 | } |
| 971 | helper.UnsafeAppend(data_ + sizeof(uint32_t), value_len); |
| 972 | data_ += increment; |
| 973 | len_ -= increment; |
| 974 | } |
| 975 | |
| 976 | num_values_ -= num_values; |
| 977 | *values_decoded = num_values; |
| 978 | return Status::OK(); |
| 979 | } |
| 980 | |
| 981 | template <typename BuilderType> |
| 982 | Status DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits, |
| 983 | int64_t valid_bits_offset, BuilderType* builder, |
| 984 | int* out_values_decoded) { |
| 985 | RETURN_NOT_OK(builder->Reserve(num_values)); |
| 986 | arrow::internal::BitmapReader bit_reader(valid_bits, valid_bits_offset, num_values); |
| 987 | int values_decoded = 0; |
| 988 | for (int i = 0; i < num_values; ++i) { |
| 989 | if (bit_reader.IsSet()) { |
| 990 | uint32_t value_len = arrow::util::SafeLoadAs<uint32_t>(data_); |
| 991 | int increment = static_cast<int>(sizeof(uint32_t) + value_len); |
| 992 | if (len_ < increment) { |
| 993 | ParquetException::EofException(); |
| 994 | } |
| 995 | RETURN_NOT_OK(builder->Append(data_ + sizeof(uint32_t), value_len)); |
| 996 | data_ += increment; |
| 997 | len_ -= increment; |
| 998 | ++values_decoded; |
| 999 | } else { |
| 1000 | RETURN_NOT_OK(builder->AppendNull()); |
| 1001 | } |
| 1002 | bit_reader.Next(); |
| 1003 | } |
| 1004 | num_values_ -= values_decoded; |
| 1005 | *out_values_decoded = values_decoded; |
| 1006 | return Status::OK(); |
| 1007 | } |
| 1008 | |
| 1009 | template <typename BuilderType> |
| 1010 | Status DecodeArrowNonNull(int num_values, BuilderType* builder, int* values_decoded) { |
| 1011 | num_values = std::min(num_values, num_values_); |
| 1012 | RETURN_NOT_OK(builder->Reserve(num_values)); |
| 1013 | for (int i = 0; i < num_values; ++i) { |
| 1014 | uint32_t value_len = arrow::util::SafeLoadAs<uint32_t>(data_); |
| 1015 | int increment = static_cast<int>(sizeof(uint32_t) + value_len); |
| 1016 | if (len_ < increment) ParquetException::EofException(); |
| 1017 | RETURN_NOT_OK(builder->Append(data_ + sizeof(uint32_t), value_len)); |
| 1018 | data_ += increment; |
| 1019 | len_ -= increment; |
| 1020 | } |
| 1021 | num_values_ -= num_values; |
| 1022 | *values_decoded = num_values; |
| 1023 | return Status::OK(); |
| 1024 | } |
| 1025 | }; |
| 1026 | |
| 1027 | class PlainFLBADecoder : public PlainDecoder<FLBAType>, virtual public FLBADecoder { |
| 1028 | public: |
| 1029 | using Base = PlainDecoder<FLBAType>; |
| 1030 | using Base::PlainDecoder; |
| 1031 | }; |
| 1032 | |
| 1033 | // ---------------------------------------------------------------------- |
| 1034 | // Dictionary encoding and decoding |
| 1035 | |
| 1036 | template <typename Type> |
| 1037 | class DictDecoderImpl : public DecoderImpl, virtual public DictDecoder<Type> { |
| 1038 | public: |
| 1039 | typedef typename Type::c_type T; |
| 1040 | |
| 1041 | // Initializes the dictionary with values from 'dictionary'. The data in |
| 1042 | // dictionary is not guaranteed to persist in memory after this call so the |
| 1043 | // dictionary decoder needs to copy the data out if necessary. |
| 1044 | explicit DictDecoderImpl(const ColumnDescriptor* descr, |
| 1045 | MemoryPool* pool = arrow::default_memory_pool()) |
| 1046 | : DecoderImpl(descr, Encoding::RLE_DICTIONARY), |
| 1047 | dictionary_(AllocateBuffer(pool, 0)), |
| 1048 | dictionary_length_(0), |
| 1049 | byte_array_data_(AllocateBuffer(pool, 0)), |
| 1050 | byte_array_offsets_(AllocateBuffer(pool, 0)), |
| 1051 | indices_scratch_space_(AllocateBuffer(pool, 0)) {} |
| 1052 | |
| 1053 | // Perform type-specific initiatialization |
| 1054 | void SetDict(TypedDecoder<Type>* dictionary) override; |
| 1055 | |
| 1056 | void SetData(int num_values, const uint8_t* data, int len) override { |
| 1057 | num_values_ = num_values; |
| 1058 | if (len == 0) return; |
| 1059 | uint8_t bit_width = *data; |
| 1060 | ++data; |
| 1061 | --len; |
| 1062 | idx_decoder_ = arrow::util::RleDecoder(data, len, bit_width); |
| 1063 | } |
| 1064 | |
| 1065 | int Decode(T* buffer, int num_values) override { |
| 1066 | num_values = std::min(num_values, num_values_); |
| 1067 | int decoded_values = idx_decoder_.GetBatchWithDict( |
| 1068 | reinterpret_cast<const T*>(dictionary_->data()), buffer, num_values); |
| 1069 | if (decoded_values != num_values) { |
| 1070 | ParquetException::EofException(); |
| 1071 | } |
| 1072 | num_values_ -= num_values; |
| 1073 | return num_values; |
| 1074 | } |
| 1075 | |
| 1076 | int DecodeSpaced(T* buffer, int num_values, int null_count, const uint8_t* valid_bits, |
| 1077 | int64_t valid_bits_offset) override { |
| 1078 | num_values = std::min(num_values, num_values_); |
| 1079 | if (num_values != idx_decoder_.GetBatchWithDictSpaced( |
| 1080 | reinterpret_cast<const T*>(dictionary_->data()), buffer, |
| 1081 | num_values, null_count, valid_bits, valid_bits_offset)) { |
| 1082 | ParquetException::EofException(); |
| 1083 | } |
| 1084 | num_values_ -= num_values; |
| 1085 | return num_values; |
| 1086 | } |
| 1087 | |
| 1088 | void InsertDictionary(arrow::ArrayBuilder* builder) override; |
| 1089 | |
| 1090 | int DecodeIndicesSpaced(int num_values, int null_count, const uint8_t* valid_bits, |
| 1091 | int64_t valid_bits_offset, |
| 1092 | arrow::ArrayBuilder* builder) override { |
| 1093 | if (num_values > 0) { |
| 1094 | // TODO(wesm): Refactor to batch reads for improved memory use. It is not |
| 1095 | // trivial because the null_count is relative to the entire bitmap |
| 1096 | PARQUET_THROW_NOT_OK(indices_scratch_space_->TypedResize<int32_t>( |
| 1097 | num_values, /*shrink_to_fit=*/false)); |
| 1098 | } |
| 1099 | |
| 1100 | auto indices_buffer = |
| 1101 | reinterpret_cast<int32_t*>(indices_scratch_space_->mutable_data()); |
| 1102 | |
| 1103 | if (num_values != idx_decoder_.GetBatchSpaced(num_values, null_count, valid_bits, |
| 1104 | valid_bits_offset, indices_buffer)) { |
| 1105 | ParquetException::EofException(); |
| 1106 | } |
| 1107 | |
| 1108 | /// XXX(wesm): Cannot append "valid bits" directly to the builder |
| 1109 | std::vector<uint8_t> valid_bytes(num_values); |
| 1110 | arrow::internal::BitmapReader bit_reader(valid_bits, valid_bits_offset, num_values); |
| 1111 | for (int64_t i = 0; i < num_values; ++i) { |
| 1112 | valid_bytes[i] = static_cast<uint8_t>(bit_reader.IsSet()); |
| 1113 | bit_reader.Next(); |
| 1114 | } |
| 1115 | |
| 1116 | auto binary_builder = checked_cast<arrow::BinaryDictionary32Builder*>(builder); |
| 1117 | PARQUET_THROW_NOT_OK( |
| 1118 | binary_builder->AppendIndices(indices_buffer, num_values, valid_bytes.data())); |
| 1119 | num_values_ -= num_values - null_count; |
| 1120 | return num_values - null_count; |
| 1121 | } |
| 1122 | |
| 1123 | int DecodeIndices(int num_values, arrow::ArrayBuilder* builder) override { |
| 1124 | num_values = std::min(num_values, num_values_); |
| 1125 | num_values = std::min(num_values, num_values_); |
| 1126 | if (num_values > 0) { |
| 1127 | // TODO(wesm): Refactor to batch reads for improved memory use. This is |
| 1128 | // relatively simple here because we don't have to do any bookkeeping of |
| 1129 | // nulls |
| 1130 | PARQUET_THROW_NOT_OK(indices_scratch_space_->TypedResize<int32_t>( |
| 1131 | num_values, /*shrink_to_fit=*/false)); |
| 1132 | } |
| 1133 | auto indices_buffer = |
| 1134 | reinterpret_cast<int32_t*>(indices_scratch_space_->mutable_data()); |
| 1135 | if (num_values != idx_decoder_.GetBatch(indices_buffer, num_values)) { |
| 1136 | ParquetException::EofException(); |
| 1137 | } |
| 1138 | auto binary_builder = checked_cast<arrow::BinaryDictionary32Builder*>(builder); |
| 1139 | PARQUET_THROW_NOT_OK(binary_builder->AppendIndices(indices_buffer, num_values)); |
| 1140 | num_values_ -= num_values; |
| 1141 | return num_values; |
| 1142 | } |
| 1143 | |
| 1144 | protected: |
| 1145 | inline void DecodeDict(TypedDecoder<Type>* dictionary) { |
| 1146 | dictionary_length_ = static_cast<int32_t>(dictionary->values_left()); |
| 1147 | PARQUET_THROW_NOT_OK(dictionary_->Resize(dictionary_length_ * sizeof(T), |
| 1148 | /*shrink_to_fit=*/false)); |
| 1149 | dictionary->Decode(reinterpret_cast<T*>(dictionary_->mutable_data()), |
| 1150 | dictionary_length_); |
| 1151 | } |
| 1152 | |
| 1153 | // Only one is set. |
| 1154 | std::shared_ptr<ResizableBuffer> dictionary_; |
| 1155 | |
| 1156 | int32_t dictionary_length_; |
| 1157 | |
| 1158 | // Data that contains the byte array data (byte_array_dictionary_ just has the |
| 1159 | // pointers). |
| 1160 | std::shared_ptr<ResizableBuffer> byte_array_data_; |
| 1161 | |
| 1162 | // Arrow-style byte offsets for each dictionary value. We maintain two |
| 1163 | // representations of the dictionary, one as ByteArray* for non-Arrow |
| 1164 | // consumers and this one for Arrow conumers. Since dictionaries are |
| 1165 | // generally pretty small to begin with this doesn't mean too much extra |
| 1166 | // memory use in most cases |
| 1167 | std::shared_ptr<ResizableBuffer> byte_array_offsets_; |
| 1168 | |
| 1169 | // Reusable buffer for decoding dictionary indices to be appended to a |
| 1170 | // BinaryDictionary32Builder |
| 1171 | std::shared_ptr<ResizableBuffer> indices_scratch_space_; |
| 1172 | |
| 1173 | arrow::util::RleDecoder idx_decoder_; |
| 1174 | }; |
| 1175 | |
| 1176 | template <typename Type> |
| 1177 | void DictDecoderImpl<Type>::SetDict(TypedDecoder<Type>* dictionary) { |
| 1178 | DecodeDict(dictionary); |
| 1179 | } |
| 1180 | |
| 1181 | template <> |
| 1182 | void DictDecoderImpl<BooleanType>::SetDict(TypedDecoder<BooleanType>* dictionary) { |
| 1183 | ParquetException::NYI("Dictionary encoding is not implemented for boolean values" ); |
| 1184 | } |
| 1185 | |
| 1186 | template <> |
| 1187 | void DictDecoderImpl<ByteArrayType>::SetDict(TypedDecoder<ByteArrayType>* dictionary) { |
| 1188 | DecodeDict(dictionary); |
| 1189 | |
| 1190 | auto dict_values = reinterpret_cast<ByteArray*>(dictionary_->mutable_data()); |
| 1191 | |
| 1192 | int total_size = 0; |
| 1193 | for (int i = 0; i < dictionary_length_; ++i) { |
| 1194 | total_size += dict_values[i].len; |
| 1195 | } |
| 1196 | if (total_size > 0) { |
| 1197 | PARQUET_THROW_NOT_OK(byte_array_data_->Resize(total_size, |
| 1198 | /*shrink_to_fit=*/false)); |
| 1199 | PARQUET_THROW_NOT_OK( |
| 1200 | byte_array_offsets_->Resize((dictionary_length_ + 1) * sizeof(int32_t), |
| 1201 | /*shrink_to_fit=*/false)); |
| 1202 | } |
| 1203 | |
| 1204 | int32_t offset = 0; |
| 1205 | uint8_t* bytes_data = byte_array_data_->mutable_data(); |
| 1206 | int32_t* bytes_offsets = |
| 1207 | reinterpret_cast<int32_t*>(byte_array_offsets_->mutable_data()); |
| 1208 | for (int i = 0; i < dictionary_length_; ++i) { |
| 1209 | memcpy(bytes_data + offset, dict_values[i].ptr, dict_values[i].len); |
| 1210 | bytes_offsets[i] = offset; |
| 1211 | dict_values[i].ptr = bytes_data + offset; |
| 1212 | offset += dict_values[i].len; |
| 1213 | } |
| 1214 | bytes_offsets[dictionary_length_] = offset; |
| 1215 | } |
| 1216 | |
| 1217 | template <> |
| 1218 | inline void DictDecoderImpl<FLBAType>::SetDict(TypedDecoder<FLBAType>* dictionary) { |
| 1219 | DecodeDict(dictionary); |
| 1220 | |
| 1221 | auto dict_values = reinterpret_cast<FLBA*>(dictionary_->mutable_data()); |
| 1222 | |
| 1223 | int fixed_len = descr_->type_length(); |
| 1224 | int total_size = dictionary_length_ * fixed_len; |
| 1225 | |
| 1226 | PARQUET_THROW_NOT_OK(byte_array_data_->Resize(total_size, |
| 1227 | /*shrink_to_fit=*/false)); |
| 1228 | uint8_t* bytes_data = byte_array_data_->mutable_data(); |
| 1229 | for (int32_t i = 0, offset = 0; i < dictionary_length_; ++i, offset += fixed_len) { |
| 1230 | memcpy(bytes_data + offset, dict_values[i].ptr, fixed_len); |
| 1231 | dict_values[i].ptr = bytes_data + offset; |
| 1232 | } |
| 1233 | } |
| 1234 | |
| 1235 | template <typename Type> |
| 1236 | void DictDecoderImpl<Type>::InsertDictionary(arrow::ArrayBuilder* builder) { |
| 1237 | ParquetException::NYI("InsertDictionary only implemented for BYTE_ARRAY types" ); |
| 1238 | } |
| 1239 | |
| 1240 | template <> |
| 1241 | void DictDecoderImpl<ByteArrayType>::InsertDictionary(arrow::ArrayBuilder* builder) { |
| 1242 | auto binary_builder = checked_cast<arrow::BinaryDictionary32Builder*>(builder); |
| 1243 | |
| 1244 | // Make an BinaryArray referencing the internal dictionary data |
| 1245 | auto arr = std::make_shared<arrow::BinaryArray>(dictionary_length_, byte_array_offsets_, |
| 1246 | byte_array_data_); |
| 1247 | PARQUET_THROW_NOT_OK(binary_builder->InsertMemoValues(*arr)); |
| 1248 | } |
| 1249 | |
| 1250 | class DictByteArrayDecoderImpl : public DictDecoderImpl<ByteArrayType>, |
| 1251 | virtual public ByteArrayDecoder { |
| 1252 | public: |
| 1253 | using BASE = DictDecoderImpl<ByteArrayType>; |
| 1254 | using BASE::DictDecoderImpl; |
| 1255 | |
| 1256 | int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits, |
| 1257 | int64_t valid_bits_offset, |
| 1258 | arrow::BinaryDictionary32Builder* builder) override { |
| 1259 | int result = 0; |
| 1260 | PARQUET_THROW_NOT_OK(DecodeArrow(num_values, null_count, valid_bits, |
| 1261 | valid_bits_offset, builder, &result)); |
| 1262 | return result; |
| 1263 | } |
| 1264 | |
| 1265 | int DecodeArrowNonNull(int num_values, |
| 1266 | arrow::BinaryDictionary32Builder* builder) override { |
| 1267 | int result = 0; |
| 1268 | PARQUET_THROW_NOT_OK(DecodeArrowNonNull(num_values, builder, &result)); |
| 1269 | return result; |
| 1270 | } |
| 1271 | |
| 1272 | int DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits, |
| 1273 | int64_t valid_bits_offset, ArrowBinaryAccumulator* out) override { |
| 1274 | int result = 0; |
| 1275 | PARQUET_THROW_NOT_OK(DecodeArrowDense(num_values, null_count, valid_bits, |
| 1276 | valid_bits_offset, out, &result)); |
| 1277 | return result; |
| 1278 | } |
| 1279 | |
| 1280 | int DecodeArrowNonNull(int num_values, ArrowBinaryAccumulator* out) override { |
| 1281 | int result = 0; |
| 1282 | PARQUET_THROW_NOT_OK(DecodeArrowDenseNonNull(num_values, out, &result)); |
| 1283 | return result; |
| 1284 | } |
| 1285 | |
| 1286 | private: |
| 1287 | Status DecodeArrowDense(int num_values, int null_count, const uint8_t* valid_bits, |
| 1288 | int64_t valid_bits_offset, ArrowBinaryAccumulator* out, |
| 1289 | int* out_num_values) { |
| 1290 | constexpr int32_t buffer_size = 1024; |
| 1291 | int32_t indices_buffer[buffer_size]; |
| 1292 | |
| 1293 | ArrowBinaryHelper helper(out); |
| 1294 | |
| 1295 | arrow::internal::BitmapReader bit_reader(valid_bits, valid_bits_offset, num_values); |
| 1296 | |
| 1297 | auto dict_values = reinterpret_cast<const ByteArray*>(dictionary_->data()); |
| 1298 | int values_decoded = 0; |
| 1299 | int num_appended = 0; |
| 1300 | while (num_appended < num_values) { |
| 1301 | bool is_valid = bit_reader.IsSet(); |
| 1302 | bit_reader.Next(); |
| 1303 | |
| 1304 | if (is_valid) { |
| 1305 | int32_t batch_size = |
| 1306 | std::min<int32_t>(buffer_size, num_values - num_appended - null_count); |
| 1307 | int num_indices = idx_decoder_.GetBatch(indices_buffer, batch_size); |
| 1308 | |
| 1309 | int i = 0; |
| 1310 | while (true) { |
| 1311 | // Consume all indices |
| 1312 | if (is_valid) { |
| 1313 | const auto& val = dict_values[indices_buffer[i]]; |
| 1314 | if (ARROW_PREDICT_FALSE(!helper.CanFit(val.len))) { |
| 1315 | RETURN_NOT_OK(helper.PushChunk()); |
| 1316 | } |
| 1317 | RETURN_NOT_OK(helper.Append(val.ptr, static_cast<int32_t>(val.len))); |
| 1318 | ++i; |
| 1319 | ++values_decoded; |
| 1320 | } else { |
| 1321 | RETURN_NOT_OK(helper.AppendNull()); |
| 1322 | --null_count; |
| 1323 | } |
| 1324 | ++num_appended; |
| 1325 | if (i == num_indices) { |
| 1326 | // Do not advance the bit_reader if we have fulfilled the decode |
| 1327 | // request |
| 1328 | break; |
| 1329 | } |
| 1330 | is_valid = bit_reader.IsSet(); |
| 1331 | bit_reader.Next(); |
| 1332 | } |
| 1333 | } else { |
| 1334 | RETURN_NOT_OK(helper.AppendNull()); |
| 1335 | --null_count; |
| 1336 | ++num_appended; |
| 1337 | } |
| 1338 | } |
| 1339 | *out_num_values = values_decoded; |
| 1340 | return Status::OK(); |
| 1341 | } |
| 1342 | |
| 1343 | Status DecodeArrowDenseNonNull(int num_values, ArrowBinaryAccumulator* out, |
| 1344 | int* out_num_values) { |
| 1345 | constexpr int32_t buffer_size = 2048; |
| 1346 | int32_t indices_buffer[buffer_size]; |
| 1347 | int values_decoded = 0; |
| 1348 | |
| 1349 | ArrowBinaryHelper helper(out); |
| 1350 | auto dict_values = reinterpret_cast<const ByteArray*>(dictionary_->data()); |
| 1351 | |
| 1352 | while (values_decoded < num_values) { |
| 1353 | int32_t batch_size = std::min<int32_t>(buffer_size, num_values - values_decoded); |
| 1354 | int num_indices = idx_decoder_.GetBatch(indices_buffer, batch_size); |
| 1355 | if (num_indices == 0) ParquetException::EofException(); |
| 1356 | for (int i = 0; i < num_indices; ++i) { |
| 1357 | const auto& val = dict_values[indices_buffer[i]]; |
| 1358 | if (ARROW_PREDICT_FALSE(!helper.CanFit(val.len))) { |
| 1359 | RETURN_NOT_OK(helper.PushChunk()); |
| 1360 | } |
| 1361 | RETURN_NOT_OK(helper.Append(val.ptr, static_cast<int32_t>(val.len))); |
| 1362 | } |
| 1363 | values_decoded += num_indices; |
| 1364 | } |
| 1365 | *out_num_values = values_decoded; |
| 1366 | return Status::OK(); |
| 1367 | } |
| 1368 | |
| 1369 | template <typename BuilderType> |
| 1370 | Status DecodeArrow(int num_values, int null_count, const uint8_t* valid_bits, |
| 1371 | int64_t valid_bits_offset, BuilderType* builder, |
| 1372 | int* out_num_values) { |
| 1373 | constexpr int32_t buffer_size = 1024; |
| 1374 | int32_t indices_buffer[buffer_size]; |
| 1375 | |
| 1376 | RETURN_NOT_OK(builder->Reserve(num_values)); |
| 1377 | arrow::internal::BitmapReader bit_reader(valid_bits, valid_bits_offset, num_values); |
| 1378 | |
| 1379 | auto dict_values = reinterpret_cast<const ByteArray*>(dictionary_->data()); |
| 1380 | |
| 1381 | int values_decoded = 0; |
| 1382 | int num_appended = 0; |
| 1383 | while (num_appended < num_values) { |
| 1384 | bool is_valid = bit_reader.IsSet(); |
| 1385 | bit_reader.Next(); |
| 1386 | |
| 1387 | if (is_valid) { |
| 1388 | int32_t batch_size = |
| 1389 | std::min<int32_t>(buffer_size, num_values - num_appended - null_count); |
| 1390 | int num_indices = idx_decoder_.GetBatch(indices_buffer, batch_size); |
| 1391 | |
| 1392 | int i = 0; |
| 1393 | while (true) { |
| 1394 | // Consume all indices |
| 1395 | if (is_valid) { |
| 1396 | const auto& val = dict_values[indices_buffer[i]]; |
| 1397 | RETURN_NOT_OK(builder->Append(val.ptr, val.len)); |
| 1398 | ++i; |
| 1399 | ++values_decoded; |
| 1400 | } else { |
| 1401 | RETURN_NOT_OK(builder->AppendNull()); |
| 1402 | --null_count; |
| 1403 | } |
| 1404 | ++num_appended; |
| 1405 | if (i == num_indices) { |
| 1406 | // Do not advance the bit_reader if we have fulfilled the decode |
| 1407 | // request |
| 1408 | break; |
| 1409 | } |
| 1410 | is_valid = bit_reader.IsSet(); |
| 1411 | bit_reader.Next(); |
| 1412 | } |
| 1413 | } else { |
| 1414 | RETURN_NOT_OK(builder->AppendNull()); |
| 1415 | --null_count; |
| 1416 | ++num_appended; |
| 1417 | } |
| 1418 | } |
| 1419 | *out_num_values = values_decoded; |
| 1420 | return Status::OK(); |
| 1421 | } |
| 1422 | |
| 1423 | template <typename BuilderType> |
| 1424 | Status DecodeArrowNonNull(int num_values, BuilderType* builder, int* out_num_values) { |
| 1425 | constexpr int32_t buffer_size = 2048; |
| 1426 | int32_t indices_buffer[buffer_size]; |
| 1427 | int values_decoded = 0; |
| 1428 | RETURN_NOT_OK(builder->Reserve(num_values)); |
| 1429 | |
| 1430 | auto dict_values = reinterpret_cast<const ByteArray*>(dictionary_->data()); |
| 1431 | |
| 1432 | while (values_decoded < num_values) { |
| 1433 | int32_t batch_size = std::min<int32_t>(buffer_size, num_values - values_decoded); |
| 1434 | int num_indices = idx_decoder_.GetBatch(indices_buffer, batch_size); |
| 1435 | if (num_indices == 0) ParquetException::EofException(); |
| 1436 | for (int i = 0; i < num_indices; ++i) { |
| 1437 | const auto& val = dict_values[indices_buffer[i]]; |
| 1438 | RETURN_NOT_OK(builder->Append(val.ptr, val.len)); |
| 1439 | } |
| 1440 | values_decoded += num_indices; |
| 1441 | } |
| 1442 | *out_num_values = values_decoded; |
| 1443 | return Status::OK(); |
| 1444 | } |
| 1445 | }; |
| 1446 | |
| 1447 | class DictFLBADecoder : public DictDecoderImpl<FLBAType>, virtual public FLBADecoder { |
| 1448 | public: |
| 1449 | using BASE = DictDecoderImpl<FLBAType>; |
| 1450 | using BASE::DictDecoderImpl; |
| 1451 | }; |
| 1452 | |
| 1453 | // ---------------------------------------------------------------------- |
| 1454 | // DeltaBitPackDecoder |
| 1455 | |
| 1456 | template <typename DType> |
| 1457 | class DeltaBitPackDecoder : public DecoderImpl, virtual public TypedDecoder<DType> { |
| 1458 | public: |
| 1459 | typedef typename DType::c_type T; |
| 1460 | |
| 1461 | explicit DeltaBitPackDecoder(const ColumnDescriptor* descr, |
| 1462 | MemoryPool* pool = arrow::default_memory_pool()) |
| 1463 | : DecoderImpl(descr, Encoding::DELTA_BINARY_PACKED), pool_(pool) { |
| 1464 | if (DType::type_num != Type::INT32 && DType::type_num != Type::INT64) { |
| 1465 | throw ParquetException("Delta bit pack encoding should only be for integer data." ); |
| 1466 | } |
| 1467 | } |
| 1468 | |
| 1469 | virtual void SetData(int num_values, const uint8_t* data, int len) { |
| 1470 | this->num_values_ = num_values; |
| 1471 | decoder_ = arrow::BitUtil::BitReader(data, len); |
| 1472 | values_current_block_ = 0; |
| 1473 | values_current_mini_block_ = 0; |
| 1474 | } |
| 1475 | |
| 1476 | virtual int Decode(T* buffer, int max_values) { |
| 1477 | return GetInternal(buffer, max_values); |
| 1478 | } |
| 1479 | |
| 1480 | private: |
| 1481 | void InitBlock() { |
| 1482 | int32_t block_size; |
| 1483 | if (!decoder_.GetVlqInt(&block_size)) ParquetException::EofException(); |
| 1484 | if (!decoder_.GetVlqInt(&num_mini_blocks_)) ParquetException::EofException(); |
| 1485 | if (!decoder_.GetVlqInt(&values_current_block_)) { |
| 1486 | ParquetException::EofException(); |
| 1487 | } |
| 1488 | if (!decoder_.GetZigZagVlqInt(&last_value_)) ParquetException::EofException(); |
| 1489 | |
| 1490 | delta_bit_widths_ = AllocateBuffer(pool_, num_mini_blocks_); |
| 1491 | uint8_t* bit_width_data = delta_bit_widths_->mutable_data(); |
| 1492 | |
| 1493 | if (!decoder_.GetZigZagVlqInt(&min_delta_)) ParquetException::EofException(); |
| 1494 | for (int i = 0; i < num_mini_blocks_; ++i) { |
| 1495 | if (!decoder_.GetAligned<uint8_t>(1, bit_width_data + i)) { |
| 1496 | ParquetException::EofException(); |
| 1497 | } |
| 1498 | } |
| 1499 | values_per_mini_block_ = block_size / num_mini_blocks_; |
| 1500 | mini_block_idx_ = 0; |
| 1501 | delta_bit_width_ = bit_width_data[0]; |
| 1502 | values_current_mini_block_ = values_per_mini_block_; |
| 1503 | } |
| 1504 | |
| 1505 | template <typename T> |
| 1506 | int GetInternal(T* buffer, int max_values) { |
| 1507 | max_values = std::min(max_values, this->num_values_); |
| 1508 | const uint8_t* bit_width_data = delta_bit_widths_->data(); |
| 1509 | for (int i = 0; i < max_values; ++i) { |
| 1510 | if (ARROW_PREDICT_FALSE(values_current_mini_block_ == 0)) { |
| 1511 | ++mini_block_idx_; |
| 1512 | if (mini_block_idx_ < static_cast<size_t>(delta_bit_widths_->size())) { |
| 1513 | delta_bit_width_ = bit_width_data[mini_block_idx_]; |
| 1514 | values_current_mini_block_ = values_per_mini_block_; |
| 1515 | } else { |
| 1516 | InitBlock(); |
| 1517 | buffer[i] = last_value_; |
| 1518 | continue; |
| 1519 | } |
| 1520 | } |
| 1521 | |
| 1522 | // TODO: the key to this algorithm is to decode the entire miniblock at once. |
| 1523 | int64_t delta; |
| 1524 | if (!decoder_.GetValue(delta_bit_width_, &delta)) ParquetException::EofException(); |
| 1525 | delta += min_delta_; |
| 1526 | last_value_ += static_cast<int32_t>(delta); |
| 1527 | buffer[i] = last_value_; |
| 1528 | --values_current_mini_block_; |
| 1529 | } |
| 1530 | this->num_values_ -= max_values; |
| 1531 | return max_values; |
| 1532 | } |
| 1533 | |
| 1534 | MemoryPool* pool_; |
| 1535 | arrow::BitUtil::BitReader decoder_; |
| 1536 | int32_t values_current_block_; |
| 1537 | int32_t num_mini_blocks_; |
| 1538 | uint64_t values_per_mini_block_; |
| 1539 | uint64_t values_current_mini_block_; |
| 1540 | |
| 1541 | int32_t min_delta_; |
| 1542 | size_t mini_block_idx_; |
| 1543 | std::shared_ptr<ResizableBuffer> delta_bit_widths_; |
| 1544 | int delta_bit_width_; |
| 1545 | |
| 1546 | int32_t last_value_; |
| 1547 | }; |
| 1548 | |
| 1549 | // ---------------------------------------------------------------------- |
| 1550 | // DELTA_LENGTH_BYTE_ARRAY |
| 1551 | |
| 1552 | class DeltaLengthByteArrayDecoder : public DecoderImpl, |
| 1553 | virtual public TypedDecoder<ByteArrayType> { |
| 1554 | public: |
| 1555 | explicit DeltaLengthByteArrayDecoder(const ColumnDescriptor* descr, |
| 1556 | MemoryPool* pool = arrow::default_memory_pool()) |
| 1557 | : DecoderImpl(descr, Encoding::DELTA_LENGTH_BYTE_ARRAY), |
| 1558 | len_decoder_(nullptr, pool) {} |
| 1559 | |
| 1560 | virtual void SetData(int num_values, const uint8_t* data, int len) { |
| 1561 | num_values_ = num_values; |
| 1562 | if (len == 0) return; |
| 1563 | int total_lengths_len = arrow::util::SafeLoadAs<int32_t>(data); |
| 1564 | data += 4; |
| 1565 | this->len_decoder_.SetData(num_values, data, total_lengths_len); |
| 1566 | data_ = data + total_lengths_len; |
| 1567 | this->len_ = len - 4 - total_lengths_len; |
| 1568 | } |
| 1569 | |
| 1570 | virtual int Decode(ByteArray* buffer, int max_values) { |
| 1571 | max_values = std::min(max_values, num_values_); |
| 1572 | std::vector<int> lengths(max_values); |
| 1573 | len_decoder_.Decode(lengths.data(), max_values); |
| 1574 | for (int i = 0; i < max_values; ++i) { |
| 1575 | buffer[i].len = lengths[i]; |
| 1576 | buffer[i].ptr = data_; |
| 1577 | this->data_ += lengths[i]; |
| 1578 | this->len_ -= lengths[i]; |
| 1579 | } |
| 1580 | this->num_values_ -= max_values; |
| 1581 | return max_values; |
| 1582 | } |
| 1583 | |
| 1584 | private: |
| 1585 | DeltaBitPackDecoder<Int32Type> len_decoder_; |
| 1586 | }; |
| 1587 | |
| 1588 | // ---------------------------------------------------------------------- |
| 1589 | // DELTA_BYTE_ARRAY |
| 1590 | |
| 1591 | class DeltaByteArrayDecoder : public DecoderImpl, |
| 1592 | virtual public TypedDecoder<ByteArrayType> { |
| 1593 | public: |
| 1594 | explicit DeltaByteArrayDecoder(const ColumnDescriptor* descr, |
| 1595 | MemoryPool* pool = arrow::default_memory_pool()) |
| 1596 | : DecoderImpl(descr, Encoding::DELTA_BYTE_ARRAY), |
| 1597 | prefix_len_decoder_(nullptr, pool), |
| 1598 | suffix_decoder_(nullptr, pool), |
| 1599 | last_value_(0, nullptr) {} |
| 1600 | |
| 1601 | virtual void SetData(int num_values, const uint8_t* data, int len) { |
| 1602 | num_values_ = num_values; |
| 1603 | if (len == 0) return; |
| 1604 | int prefix_len_length = arrow::util::SafeLoadAs<int32_t>(data); |
| 1605 | data += 4; |
| 1606 | len -= 4; |
| 1607 | prefix_len_decoder_.SetData(num_values, data, prefix_len_length); |
| 1608 | data += prefix_len_length; |
| 1609 | len -= prefix_len_length; |
| 1610 | suffix_decoder_.SetData(num_values, data, len); |
| 1611 | } |
| 1612 | |
| 1613 | // TODO: this doesn't work and requires memory management. We need to allocate |
| 1614 | // new strings to store the results. |
| 1615 | virtual int Decode(ByteArray* buffer, int max_values) { |
| 1616 | max_values = std::min(max_values, this->num_values_); |
| 1617 | for (int i = 0; i < max_values; ++i) { |
| 1618 | int prefix_len = 0; |
| 1619 | prefix_len_decoder_.Decode(&prefix_len, 1); |
| 1620 | ByteArray suffix = {0, nullptr}; |
| 1621 | suffix_decoder_.Decode(&suffix, 1); |
| 1622 | buffer[i].len = prefix_len + suffix.len; |
| 1623 | |
| 1624 | uint8_t* result = reinterpret_cast<uint8_t*>(malloc(buffer[i].len)); |
| 1625 | memcpy(result, last_value_.ptr, prefix_len); |
| 1626 | memcpy(result + prefix_len, suffix.ptr, suffix.len); |
| 1627 | |
| 1628 | buffer[i].ptr = result; |
| 1629 | last_value_ = buffer[i]; |
| 1630 | } |
| 1631 | this->num_values_ -= max_values; |
| 1632 | return max_values; |
| 1633 | } |
| 1634 | |
| 1635 | private: |
| 1636 | DeltaBitPackDecoder<Int32Type> prefix_len_decoder_; |
| 1637 | DeltaLengthByteArrayDecoder suffix_decoder_; |
| 1638 | ByteArray last_value_; |
| 1639 | }; |
| 1640 | |
| 1641 | // ---------------------------------------------------------------------- |
| 1642 | |
| 1643 | std::unique_ptr<Decoder> MakeDecoder(Type::type type_num, Encoding::type encoding, |
| 1644 | const ColumnDescriptor* descr) { |
| 1645 | if (encoding == Encoding::PLAIN) { |
| 1646 | switch (type_num) { |
| 1647 | case Type::BOOLEAN: |
| 1648 | return std::unique_ptr<Decoder>(new PlainBooleanDecoder(descr)); |
| 1649 | case Type::INT32: |
| 1650 | return std::unique_ptr<Decoder>(new PlainDecoder<Int32Type>(descr)); |
| 1651 | case Type::INT64: |
| 1652 | return std::unique_ptr<Decoder>(new PlainDecoder<Int64Type>(descr)); |
| 1653 | case Type::INT96: |
| 1654 | return std::unique_ptr<Decoder>(new PlainDecoder<Int96Type>(descr)); |
| 1655 | case Type::FLOAT: |
| 1656 | return std::unique_ptr<Decoder>(new PlainDecoder<FloatType>(descr)); |
| 1657 | case Type::DOUBLE: |
| 1658 | return std::unique_ptr<Decoder>(new PlainDecoder<DoubleType>(descr)); |
| 1659 | case Type::BYTE_ARRAY: |
| 1660 | return std::unique_ptr<Decoder>(new PlainByteArrayDecoder(descr)); |
| 1661 | case Type::FIXED_LEN_BYTE_ARRAY: |
| 1662 | return std::unique_ptr<Decoder>(new PlainFLBADecoder(descr)); |
| 1663 | default: |
| 1664 | break; |
| 1665 | } |
| 1666 | } else { |
| 1667 | ParquetException::NYI("Selected encoding is not supported" ); |
| 1668 | } |
| 1669 | DCHECK(false) << "Should not be able to reach this code" ; |
| 1670 | return nullptr; |
| 1671 | } |
| 1672 | |
| 1673 | namespace detail { |
| 1674 | |
| 1675 | std::unique_ptr<Decoder> MakeDictDecoder(Type::type type_num, |
| 1676 | const ColumnDescriptor* descr, |
| 1677 | MemoryPool* pool) { |
| 1678 | switch (type_num) { |
| 1679 | case Type::BOOLEAN: |
| 1680 | ParquetException::NYI("Dictionary encoding not implemented for boolean type" ); |
| 1681 | case Type::INT32: |
| 1682 | return std::unique_ptr<Decoder>(new DictDecoderImpl<Int32Type>(descr, pool)); |
| 1683 | case Type::INT64: |
| 1684 | return std::unique_ptr<Decoder>(new DictDecoderImpl<Int64Type>(descr, pool)); |
| 1685 | case Type::INT96: |
| 1686 | return std::unique_ptr<Decoder>(new DictDecoderImpl<Int96Type>(descr, pool)); |
| 1687 | case Type::FLOAT: |
| 1688 | return std::unique_ptr<Decoder>(new DictDecoderImpl<FloatType>(descr, pool)); |
| 1689 | case Type::DOUBLE: |
| 1690 | return std::unique_ptr<Decoder>(new DictDecoderImpl<DoubleType>(descr, pool)); |
| 1691 | case Type::BYTE_ARRAY: |
| 1692 | return std::unique_ptr<Decoder>(new DictByteArrayDecoderImpl(descr, pool)); |
| 1693 | case Type::FIXED_LEN_BYTE_ARRAY: |
| 1694 | return std::unique_ptr<Decoder>(new DictFLBADecoder(descr, pool)); |
| 1695 | default: |
| 1696 | break; |
| 1697 | } |
| 1698 | DCHECK(false) << "Should not be able to reach this code" ; |
| 1699 | return nullptr; |
| 1700 | } |
| 1701 | |
| 1702 | } // namespace detail |
| 1703 | } // namespace parquet |
| 1704 | |