| 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 | // From Apache Impala (incubating) as of 2016-01-29 |
| 19 | |
| 20 | #ifndef ARROW_UTIL_BIT_STREAM_UTILS_H |
| 21 | #define ARROW_UTIL_BIT_STREAM_UTILS_H |
| 22 | |
| 23 | #include <string.h> |
| 24 | #include <algorithm> |
| 25 | #include <cstdint> |
| 26 | |
| 27 | #include "arrow/util/bit_util.h" |
| 28 | #include "arrow/util/bpacking.h" |
| 29 | #include "arrow/util/logging.h" |
| 30 | #include "arrow/util/macros.h" |
| 31 | |
| 32 | namespace arrow { |
| 33 | namespace BitUtil { |
| 34 | |
| 35 | /// Utility class to write bit/byte streams. This class can write data to either be |
| 36 | /// bit packed or byte aligned (and a single stream that has a mix of both). |
| 37 | /// This class does not allocate memory. |
| 38 | class BitWriter { |
| 39 | public: |
| 40 | /// buffer: buffer to write bits to. Buffer should be preallocated with |
| 41 | /// 'buffer_len' bytes. |
| 42 | BitWriter(uint8_t* buffer, int buffer_len) : buffer_(buffer), max_bytes_(buffer_len) { |
| 43 | Clear(); |
| 44 | } |
| 45 | |
| 46 | void Clear() { |
| 47 | buffered_values_ = 0; |
| 48 | byte_offset_ = 0; |
| 49 | bit_offset_ = 0; |
| 50 | } |
| 51 | |
| 52 | /// The number of current bytes written, including the current byte (i.e. may include a |
| 53 | /// fraction of a byte). Includes buffered values. |
| 54 | int bytes_written() const { |
| 55 | return byte_offset_ + static_cast<int>(BitUtil::BytesForBits(bit_offset_)); |
| 56 | } |
| 57 | uint8_t* buffer() const { return buffer_; } |
| 58 | int buffer_len() const { return max_bytes_; } |
| 59 | |
| 60 | /// Writes a value to buffered_values_, flushing to buffer_ if necessary. This is bit |
| 61 | /// packed. Returns false if there was not enough space. num_bits must be <= 32. |
| 62 | bool PutValue(uint64_t v, int num_bits); |
| 63 | |
| 64 | /// Writes v to the next aligned byte using num_bytes. If T is larger than |
| 65 | /// num_bytes, the extra high-order bytes will be ignored. Returns false if |
| 66 | /// there was not enough space. |
| 67 | template <typename T> |
| 68 | bool PutAligned(T v, int num_bytes); |
| 69 | |
| 70 | /// Write a Vlq encoded int to the buffer. Returns false if there was not enough |
| 71 | /// room. The value is written byte aligned. |
| 72 | /// For more details on vlq: |
| 73 | /// en.wikipedia.org/wiki/Variable-length_quantity |
| 74 | bool PutVlqInt(uint32_t v); |
| 75 | |
| 76 | // Writes an int zigzag encoded. |
| 77 | bool PutZigZagVlqInt(int32_t v); |
| 78 | |
| 79 | /// Get a pointer to the next aligned byte and advance the underlying buffer |
| 80 | /// by num_bytes. |
| 81 | /// Returns NULL if there was not enough space. |
| 82 | uint8_t* GetNextBytePtr(int num_bytes = 1); |
| 83 | |
| 84 | /// Flushes all buffered values to the buffer. Call this when done writing to |
| 85 | /// the buffer. If 'align' is true, buffered_values_ is reset and any future |
| 86 | /// writes will be written to the next byte boundary. |
| 87 | void Flush(bool align = false); |
| 88 | |
| 89 | private: |
| 90 | uint8_t* buffer_; |
| 91 | int max_bytes_; |
| 92 | |
| 93 | /// Bit-packed values are initially written to this variable before being memcpy'd to |
| 94 | /// buffer_. This is faster than writing values byte by byte directly to buffer_. |
| 95 | uint64_t buffered_values_; |
| 96 | |
| 97 | int byte_offset_; // Offset in buffer_ |
| 98 | int bit_offset_; // Offset in buffered_values_ |
| 99 | }; |
| 100 | |
| 101 | /// Utility class to read bit/byte stream. This class can read bits or bytes |
| 102 | /// that are either byte aligned or not. It also has utilities to read multiple |
| 103 | /// bytes in one read (e.g. encoded int). |
| 104 | class BitReader { |
| 105 | public: |
| 106 | /// 'buffer' is the buffer to read from. The buffer's length is 'buffer_len'. |
| 107 | BitReader(const uint8_t* buffer, int buffer_len) |
| 108 | : buffer_(buffer), max_bytes_(buffer_len), byte_offset_(0), bit_offset_(0) { |
| 109 | int num_bytes = std::min(8, max_bytes_ - byte_offset_); |
| 110 | memcpy(&buffered_values_, buffer_ + byte_offset_, num_bytes); |
| 111 | } |
| 112 | |
| 113 | BitReader() |
| 114 | : buffer_(NULL), |
| 115 | max_bytes_(0), |
| 116 | buffered_values_(0), |
| 117 | byte_offset_(0), |
| 118 | bit_offset_(0) {} |
| 119 | |
| 120 | void Reset(const uint8_t* buffer, int buffer_len) { |
| 121 | buffer_ = buffer; |
| 122 | max_bytes_ = buffer_len; |
| 123 | byte_offset_ = 0; |
| 124 | bit_offset_ = 0; |
| 125 | int num_bytes = std::min(8, max_bytes_ - byte_offset_); |
| 126 | memcpy(&buffered_values_, buffer_ + byte_offset_, num_bytes); |
| 127 | } |
| 128 | |
| 129 | /// Gets the next value from the buffer. Returns true if 'v' could be read or false if |
| 130 | /// there are not enough bytes left. num_bits must be <= 32. |
| 131 | template <typename T> |
| 132 | bool GetValue(int num_bits, T* v); |
| 133 | |
| 134 | /// Get a number of values from the buffer. Return the number of values actually read. |
| 135 | template <typename T> |
| 136 | int GetBatch(int num_bits, T* v, int batch_size); |
| 137 | |
| 138 | /// Reads a 'num_bytes'-sized value from the buffer and stores it in 'v'. T |
| 139 | /// needs to be a little-endian native type and big enough to store |
| 140 | /// 'num_bytes'. The value is assumed to be byte-aligned so the stream will |
| 141 | /// be advanced to the start of the next byte before 'v' is read. Returns |
| 142 | /// false if there are not enough bytes left. |
| 143 | template <typename T> |
| 144 | bool GetAligned(int num_bytes, T* v); |
| 145 | |
| 146 | /// Reads a vlq encoded int from the stream. The encoded int must start at |
| 147 | /// the beginning of a byte. Return false if there were not enough bytes in |
| 148 | /// the buffer. |
| 149 | bool GetVlqInt(int32_t* v); |
| 150 | |
| 151 | // Reads a zigzag encoded int `into` v. |
| 152 | bool GetZigZagVlqInt(int32_t* v); |
| 153 | |
| 154 | /// Returns the number of bytes left in the stream, not including the current |
| 155 | /// byte (i.e., there may be an additional fraction of a byte). |
| 156 | int bytes_left() { |
| 157 | return max_bytes_ - |
| 158 | (byte_offset_ + static_cast<int>(BitUtil::BytesForBits(bit_offset_))); |
| 159 | } |
| 160 | |
| 161 | /// Maximum byte length of a vlq encoded int |
| 162 | static const int MAX_VLQ_BYTE_LEN = 5; |
| 163 | |
| 164 | private: |
| 165 | const uint8_t* buffer_; |
| 166 | int max_bytes_; |
| 167 | |
| 168 | /// Bytes are memcpy'd from buffer_ and values are read from this variable. This is |
| 169 | /// faster than reading values byte by byte directly from buffer_. |
| 170 | uint64_t buffered_values_; |
| 171 | |
| 172 | int byte_offset_; // Offset in buffer_ |
| 173 | int bit_offset_; // Offset in buffered_values_ |
| 174 | }; |
| 175 | |
| 176 | inline bool BitWriter::PutValue(uint64_t v, int num_bits) { |
| 177 | // TODO: revisit this limit if necessary (can be raised to 64 by fixing some edge cases) |
| 178 | DCHECK_LE(num_bits, 32); |
| 179 | DCHECK_EQ(v >> num_bits, 0) << "v = " << v << ", num_bits = " << num_bits; |
| 180 | |
| 181 | if (ARROW_PREDICT_FALSE(byte_offset_ * 8 + bit_offset_ + num_bits > max_bytes_ * 8)) |
| 182 | return false; |
| 183 | |
| 184 | buffered_values_ |= v << bit_offset_; |
| 185 | bit_offset_ += num_bits; |
| 186 | |
| 187 | if (ARROW_PREDICT_FALSE(bit_offset_ >= 64)) { |
| 188 | // Flush buffered_values_ and write out bits of v that did not fit |
| 189 | memcpy(buffer_ + byte_offset_, &buffered_values_, 8); |
| 190 | buffered_values_ = 0; |
| 191 | byte_offset_ += 8; |
| 192 | bit_offset_ -= 64; |
| 193 | buffered_values_ = v >> (num_bits - bit_offset_); |
| 194 | } |
| 195 | DCHECK_LT(bit_offset_, 64); |
| 196 | return true; |
| 197 | } |
| 198 | |
| 199 | inline void BitWriter::Flush(bool align) { |
| 200 | int num_bytes = static_cast<int>(BitUtil::BytesForBits(bit_offset_)); |
| 201 | DCHECK_LE(byte_offset_ + num_bytes, max_bytes_); |
| 202 | memcpy(buffer_ + byte_offset_, &buffered_values_, num_bytes); |
| 203 | |
| 204 | if (align) { |
| 205 | buffered_values_ = 0; |
| 206 | byte_offset_ += num_bytes; |
| 207 | bit_offset_ = 0; |
| 208 | } |
| 209 | } |
| 210 | |
| 211 | inline uint8_t* BitWriter::GetNextBytePtr(int num_bytes) { |
| 212 | Flush(/* align */ true); |
| 213 | DCHECK_LE(byte_offset_, max_bytes_); |
| 214 | if (byte_offset_ + num_bytes > max_bytes_) return NULL; |
| 215 | uint8_t* ptr = buffer_ + byte_offset_; |
| 216 | byte_offset_ += num_bytes; |
| 217 | return ptr; |
| 218 | } |
| 219 | |
| 220 | template <typename T> |
| 221 | inline bool BitWriter::PutAligned(T val, int num_bytes) { |
| 222 | uint8_t* ptr = GetNextBytePtr(num_bytes); |
| 223 | if (ptr == NULL) return false; |
| 224 | memcpy(ptr, &val, num_bytes); |
| 225 | return true; |
| 226 | } |
| 227 | |
| 228 | inline bool BitWriter::PutVlqInt(uint32_t v) { |
| 229 | bool result = true; |
| 230 | while ((v & 0xFFFFFF80) != 0L) { |
| 231 | result &= PutAligned<uint8_t>(static_cast<uint8_t>((v & 0x7F) | 0x80), 1); |
| 232 | v >>= 7; |
| 233 | } |
| 234 | result &= PutAligned<uint8_t>(static_cast<uint8_t>(v & 0x7F), 1); |
| 235 | return result; |
| 236 | } |
| 237 | |
| 238 | namespace detail { |
| 239 | |
| 240 | template <typename T> |
| 241 | inline void GetValue_(int num_bits, T* v, int max_bytes, const uint8_t* buffer, |
| 242 | int* bit_offset, int* byte_offset, uint64_t* buffered_values) { |
| 243 | #ifdef _MSC_VER |
| 244 | #pragma warning(push) |
| 245 | #pragma warning(disable : 4800) |
| 246 | #endif |
| 247 | *v = static_cast<T>(BitUtil::TrailingBits(*buffered_values, *bit_offset + num_bits) >> |
| 248 | *bit_offset); |
| 249 | #ifdef _MSC_VER |
| 250 | #pragma warning(pop) |
| 251 | #endif |
| 252 | *bit_offset += num_bits; |
| 253 | if (*bit_offset >= 64) { |
| 254 | *byte_offset += 8; |
| 255 | *bit_offset -= 64; |
| 256 | |
| 257 | int bytes_remaining = max_bytes - *byte_offset; |
| 258 | if (ARROW_PREDICT_TRUE(bytes_remaining >= 8)) { |
| 259 | memcpy(buffered_values, buffer + *byte_offset, 8); |
| 260 | } else { |
| 261 | memcpy(buffered_values, buffer + *byte_offset, bytes_remaining); |
| 262 | } |
| 263 | #ifdef _MSC_VER |
| 264 | #pragma warning(push) |
| 265 | #pragma warning(disable : 4800 4805) |
| 266 | #endif |
| 267 | // Read bits of v that crossed into new buffered_values_ |
| 268 | *v = *v | static_cast<T>(BitUtil::TrailingBits(*buffered_values, *bit_offset) |
| 269 | << (num_bits - *bit_offset)); |
| 270 | #ifdef _MSC_VER |
| 271 | #pragma warning(pop) |
| 272 | #endif |
| 273 | DCHECK_LE(*bit_offset, 64); |
| 274 | } |
| 275 | } |
| 276 | |
| 277 | } // namespace detail |
| 278 | |
| 279 | template <typename T> |
| 280 | inline bool BitReader::GetValue(int num_bits, T* v) { |
| 281 | return GetBatch(num_bits, v, 1) == 1; |
| 282 | } |
| 283 | |
| 284 | template <typename T> |
| 285 | inline int BitReader::GetBatch(int num_bits, T* v, int batch_size) { |
| 286 | DCHECK(buffer_ != NULL); |
| 287 | // TODO: revisit this limit if necessary |
| 288 | DCHECK_LE(num_bits, 32); |
| 289 | DCHECK_LE(num_bits, static_cast<int>(sizeof(T) * 8)); |
| 290 | |
| 291 | int bit_offset = bit_offset_; |
| 292 | int byte_offset = byte_offset_; |
| 293 | uint64_t buffered_values = buffered_values_; |
| 294 | int max_bytes = max_bytes_; |
| 295 | const uint8_t* buffer = buffer_; |
| 296 | |
| 297 | uint64_t needed_bits = num_bits * batch_size; |
| 298 | uint64_t remaining_bits = (max_bytes - byte_offset) * 8 - bit_offset; |
| 299 | if (remaining_bits < needed_bits) { |
| 300 | batch_size = static_cast<int>(remaining_bits) / num_bits; |
| 301 | } |
| 302 | |
| 303 | int i = 0; |
| 304 | if (ARROW_PREDICT_FALSE(bit_offset != 0)) { |
| 305 | for (; i < batch_size && bit_offset != 0; ++i) { |
| 306 | detail::GetValue_(num_bits, &v[i], max_bytes, buffer, &bit_offset, &byte_offset, |
| 307 | &buffered_values); |
| 308 | } |
| 309 | } |
| 310 | |
| 311 | if (sizeof(T) == 4) { |
| 312 | int num_unpacked = |
| 313 | internal::unpack32(reinterpret_cast<const uint32_t*>(buffer + byte_offset), |
| 314 | reinterpret_cast<uint32_t*>(v + i), batch_size - i, num_bits); |
| 315 | i += num_unpacked; |
| 316 | byte_offset += num_unpacked * num_bits / 8; |
| 317 | } else { |
| 318 | const int buffer_size = 1024; |
| 319 | uint32_t unpack_buffer[buffer_size]; |
| 320 | while (i < batch_size) { |
| 321 | int unpack_size = std::min(buffer_size, batch_size - i); |
| 322 | int num_unpacked = |
| 323 | internal::unpack32(reinterpret_cast<const uint32_t*>(buffer + byte_offset), |
| 324 | unpack_buffer, unpack_size, num_bits); |
| 325 | if (num_unpacked == 0) { |
| 326 | break; |
| 327 | } |
| 328 | for (int k = 0; k < num_unpacked; ++k) { |
| 329 | #ifdef _MSC_VER |
| 330 | #pragma warning(push) |
| 331 | #pragma warning(disable : 4800) |
| 332 | #endif |
| 333 | v[i + k] = static_cast<T>(unpack_buffer[k]); |
| 334 | #ifdef _MSC_VER |
| 335 | #pragma warning(pop) |
| 336 | #endif |
| 337 | } |
| 338 | i += num_unpacked; |
| 339 | byte_offset += num_unpacked * num_bits / 8; |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | int bytes_remaining = max_bytes - byte_offset; |
| 344 | if (bytes_remaining >= 8) { |
| 345 | memcpy(&buffered_values, buffer + byte_offset, 8); |
| 346 | } else { |
| 347 | memcpy(&buffered_values, buffer + byte_offset, bytes_remaining); |
| 348 | } |
| 349 | |
| 350 | for (; i < batch_size; ++i) { |
| 351 | detail::GetValue_(num_bits, &v[i], max_bytes, buffer, &bit_offset, &byte_offset, |
| 352 | &buffered_values); |
| 353 | } |
| 354 | |
| 355 | bit_offset_ = bit_offset; |
| 356 | byte_offset_ = byte_offset; |
| 357 | buffered_values_ = buffered_values; |
| 358 | |
| 359 | return batch_size; |
| 360 | } |
| 361 | |
| 362 | template <typename T> |
| 363 | inline bool BitReader::GetAligned(int num_bytes, T* v) { |
| 364 | DCHECK_LE(num_bytes, static_cast<int>(sizeof(T))); |
| 365 | int bytes_read = static_cast<int>(BitUtil::BytesForBits(bit_offset_)); |
| 366 | if (ARROW_PREDICT_FALSE(byte_offset_ + bytes_read + num_bytes > max_bytes_)) |
| 367 | return false; |
| 368 | |
| 369 | // Advance byte_offset to next unread byte and read num_bytes |
| 370 | byte_offset_ += bytes_read; |
| 371 | memcpy(v, buffer_ + byte_offset_, num_bytes); |
| 372 | byte_offset_ += num_bytes; |
| 373 | |
| 374 | // Reset buffered_values_ |
| 375 | bit_offset_ = 0; |
| 376 | int bytes_remaining = max_bytes_ - byte_offset_; |
| 377 | if (ARROW_PREDICT_TRUE(bytes_remaining >= 8)) { |
| 378 | memcpy(&buffered_values_, buffer_ + byte_offset_, 8); |
| 379 | } else { |
| 380 | memcpy(&buffered_values_, buffer_ + byte_offset_, bytes_remaining); |
| 381 | } |
| 382 | return true; |
| 383 | } |
| 384 | |
| 385 | inline bool BitReader::GetVlqInt(int32_t* v) { |
| 386 | *v = 0; |
| 387 | int shift = 0; |
| 388 | int num_bytes = 0; |
| 389 | uint8_t byte = 0; |
| 390 | do { |
| 391 | if (!GetAligned<uint8_t>(1, &byte)) return false; |
| 392 | *v |= (byte & 0x7F) << shift; |
| 393 | shift += 7; |
| 394 | DCHECK_LE(++num_bytes, MAX_VLQ_BYTE_LEN); |
| 395 | } while ((byte & 0x80) != 0); |
| 396 | return true; |
| 397 | } |
| 398 | |
| 399 | inline bool BitWriter::PutZigZagVlqInt(int32_t v) { |
| 400 | // Note negative left shift is undefined |
| 401 | uint32_t u = (static_cast<uint32_t>(v) << 1) ^ (v >> 31); |
| 402 | return PutVlqInt(u); |
| 403 | } |
| 404 | |
| 405 | inline bool BitReader::GetZigZagVlqInt(int32_t* v) { |
| 406 | int32_t u_signed; |
| 407 | if (!GetVlqInt(&u_signed)) return false; |
| 408 | uint32_t u = static_cast<uint32_t>(u_signed); |
| 409 | *reinterpret_cast<uint32_t*>(v) = (u >> 1) ^ -(static_cast<int32_t>(u & 1)); |
| 410 | return true; |
| 411 | } |
| 412 | |
| 413 | } // namespace BitUtil |
| 414 | } // namespace arrow |
| 415 | |
| 416 | #endif // ARROW_UTIL_BIT_STREAM_UTILS_H |
| 417 | |