OLD | NEW |
1 /* Copyright 2013 Google Inc. All Rights Reserved. | 1 /* Copyright 2013 Google Inc. All Rights Reserved. |
2 | 2 |
3 Distributed under MIT license. | 3 Distributed under MIT license. |
4 See file LICENSE for detail or copy at https://opensource.org/licenses/MIT | 4 See file LICENSE for detail or copy at https://opensource.org/licenses/MIT |
5 */ | 5 */ |
6 | 6 |
7 // Sliding window over the input data. | 7 /* Sliding window over the input data. */ |
8 | 8 |
9 #ifndef BROTLI_ENC_RINGBUFFER_H_ | 9 #ifndef BROTLI_ENC_RINGBUFFER_H_ |
10 #define BROTLI_ENC_RINGBUFFER_H_ | 10 #define BROTLI_ENC_RINGBUFFER_H_ |
11 | 11 |
12 #include <cstdlib> /* free, realloc */ | 12 #include <string.h> /* memcpy */ |
13 | 13 |
| 14 #include <brotli/types.h> |
| 15 #include "./memory.h" |
14 #include "./port.h" | 16 #include "./port.h" |
15 #include "./types.h" | 17 #include "./quality.h" |
16 | 18 |
17 namespace brotli { | 19 #if defined(__cplusplus) || defined(c_plusplus) |
| 20 extern "C" { |
| 21 #endif |
18 | 22 |
19 // A RingBuffer(window_bits, tail_bits) contains `1 << window_bits' bytes of | 23 /* A RingBuffer(window_bits, tail_bits) contains `1 << window_bits' bytes of |
20 // data in a circular manner: writing a byte writes it to: | 24 data in a circular manner: writing a byte writes it to: |
21 // `position() % (1 << window_bits)'. | 25 `position() % (1 << window_bits)'. |
22 // For convenience, the RingBuffer array contains another copy of the | 26 For convenience, the RingBuffer array contains another copy of the |
23 // first `1 << tail_bits' bytes: | 27 first `1 << tail_bits' bytes: |
24 // buffer_[i] == buffer_[i + (1 << window_bits)], if i < (1 << tail_bits), | 28 buffer_[i] == buffer_[i + (1 << window_bits)], if i < (1 << tail_bits), |
25 // and another copy of the last two bytes: | 29 and another copy of the last two bytes: |
26 // buffer_[-1] == buffer_[(1 << window_bits) - 1] and | 30 buffer_[-1] == buffer_[(1 << window_bits) - 1] and |
27 // buffer_[-2] == buffer_[(1 << window_bits) - 2]. | 31 buffer_[-2] == buffer_[(1 << window_bits) - 2]. */ |
28 class RingBuffer { | 32 typedef struct RingBuffer { |
29 public: | 33 /* Size of the ring-buffer is (1 << window_bits) + tail_size_. */ |
30 RingBuffer(int window_bits, int tail_bits) | |
31 : size_(1u << window_bits), | |
32 mask_((1u << window_bits) - 1), | |
33 tail_size_(1u << tail_bits), | |
34 total_size_(size_ + tail_size_), | |
35 cur_size_(0), | |
36 pos_(0), | |
37 data_(0), | |
38 buffer_(0) {} | |
39 | |
40 ~RingBuffer(void) { | |
41 free(data_); | |
42 } | |
43 | |
44 // Allocates or re-allocates data_ to the given length + plus some slack | |
45 // region before and after. Fills the slack regions with zeros. | |
46 inline void InitBuffer(const uint32_t buflen) { | |
47 static const size_t kSlackForEightByteHashingEverywhere = 7; | |
48 cur_size_ = buflen; | |
49 data_ = static_cast<uint8_t*>(realloc( | |
50 data_, 2 + buflen + kSlackForEightByteHashingEverywhere)); | |
51 buffer_ = data_ + 2; | |
52 buffer_[-2] = buffer_[-1] = 0; | |
53 for (size_t i = 0; i < kSlackForEightByteHashingEverywhere; ++i) { | |
54 buffer_[cur_size_ + i] = 0; | |
55 } | |
56 } | |
57 | |
58 // Push bytes into the ring buffer. | |
59 void Write(const uint8_t *bytes, size_t n) { | |
60 if (pos_ == 0 && n < tail_size_) { | |
61 // Special case for the first write: to process the first block, we don't | |
62 // need to allocate the whole ringbuffer and we don't need the tail | |
63 // either. However, we do this memory usage optimization only if the | |
64 // first write is less than the tail size, which is also the input block | |
65 // size, otherwise it is likely that other blocks will follow and we | |
66 // will need to reallocate to the full size anyway. | |
67 pos_ = static_cast<uint32_t>(n); | |
68 InitBuffer(pos_); | |
69 memcpy(buffer_, bytes, n); | |
70 return; | |
71 } | |
72 if (cur_size_ < total_size_) { | |
73 // Lazily allocate the full buffer. | |
74 InitBuffer(total_size_); | |
75 // Initialize the last two bytes to zero, so that we don't have to worry | |
76 // later when we copy the last two bytes to the first two positions. | |
77 buffer_[size_ - 2] = 0; | |
78 buffer_[size_ - 1] = 0; | |
79 } | |
80 const size_t masked_pos = pos_ & mask_; | |
81 // The length of the writes is limited so that we do not need to worry | |
82 // about a write | |
83 WriteTail(bytes, n); | |
84 if (PREDICT_TRUE(masked_pos + n <= size_)) { | |
85 // A single write fits. | |
86 memcpy(&buffer_[masked_pos], bytes, n); | |
87 } else { | |
88 // Split into two writes. | |
89 // Copy into the end of the buffer, including the tail buffer. | |
90 memcpy(&buffer_[masked_pos], bytes, | |
91 std::min(n, total_size_ - masked_pos)); | |
92 // Copy into the beginning of the buffer | |
93 memcpy(&buffer_[0], bytes + (size_ - masked_pos), | |
94 n - (size_ - masked_pos)); | |
95 } | |
96 buffer_[-2] = buffer_[size_ - 2]; | |
97 buffer_[-1] = buffer_[size_ - 1]; | |
98 pos_ += static_cast<uint32_t>(n); | |
99 if (pos_ > (1u << 30)) { /* Wrap, but preserve not-a-first-lap feature. */ | |
100 pos_ = (pos_ & ((1u << 30) - 1)) | (1u << 30); | |
101 } | |
102 } | |
103 | |
104 void Reset(void) { | |
105 pos_ = 0; | |
106 } | |
107 | |
108 // Logical cursor position in the ring buffer. | |
109 uint32_t position(void) const { return pos_; } | |
110 | |
111 // Bit mask for getting the physical position for a logical position. | |
112 uint32_t mask(void) const { return mask_; } | |
113 | |
114 uint8_t *start(void) { return &buffer_[0]; } | |
115 const uint8_t *start(void) const { return &buffer_[0]; } | |
116 | |
117 private: | |
118 void WriteTail(const uint8_t *bytes, size_t n) { | |
119 const size_t masked_pos = pos_ & mask_; | |
120 if (PREDICT_FALSE(masked_pos < tail_size_)) { | |
121 // Just fill the tail buffer with the beginning data. | |
122 const size_t p = size_ + masked_pos; | |
123 memcpy(&buffer_[p], bytes, std::min(n, tail_size_ - masked_pos)); | |
124 } | |
125 } | |
126 | |
127 // Size of the ringbuffer is (1 << window_bits) + tail_size_. | |
128 const uint32_t size_; | 34 const uint32_t size_; |
129 const uint32_t mask_; | 35 const uint32_t mask_; |
130 const uint32_t tail_size_; | 36 const uint32_t tail_size_; |
131 const uint32_t total_size_; | 37 const uint32_t total_size_; |
132 | 38 |
133 uint32_t cur_size_; | 39 uint32_t cur_size_; |
134 // Position to write in the ring buffer. | 40 /* Position to write in the ring buffer. */ |
135 uint32_t pos_; | 41 uint32_t pos_; |
136 // The actual ring buffer containing the copy of the last two bytes, the data, | 42 /* The actual ring buffer containing the copy of the last two bytes, the data, |
137 // and the copy of the beginning as a tail. | 43 and the copy of the beginning as a tail. */ |
138 uint8_t *data_; | 44 uint8_t *data_; |
139 // The start of the ringbuffer. | 45 /* The start of the ring-buffer. */ |
140 uint8_t *buffer_; | 46 uint8_t *buffer_; |
141 }; | 47 } RingBuffer; |
142 | 48 |
143 } // namespace brotli | 49 static BROTLI_INLINE void RingBufferInit(RingBuffer* rb) { |
| 50 rb->cur_size_ = 0; |
| 51 rb->pos_ = 0; |
| 52 rb->data_ = 0; |
| 53 rb->buffer_ = 0; |
| 54 } |
144 | 55 |
145 #endif // BROTLI_ENC_RINGBUFFER_H_ | 56 static BROTLI_INLINE void RingBufferSetup( |
| 57 const BrotliEncoderParams* params, RingBuffer* rb) { |
| 58 int window_bits = ComputeRbBits(params); |
| 59 int tail_bits = params->lgblock; |
| 60 *(uint32_t*)&rb->size_ = 1u << window_bits; |
| 61 *(uint32_t*)&rb->mask_ = (1u << window_bits) - 1; |
| 62 *(uint32_t*)&rb->tail_size_ = 1u << tail_bits; |
| 63 *(uint32_t*)&rb->total_size_ = rb->size_ + rb->tail_size_; |
| 64 } |
| 65 |
| 66 static BROTLI_INLINE void RingBufferFree(MemoryManager* m, RingBuffer* rb) { |
| 67 BROTLI_FREE(m, rb->data_); |
| 68 } |
| 69 |
| 70 /* Allocates or re-allocates data_ to the given length + plus some slack |
| 71 region before and after. Fills the slack regions with zeros. */ |
| 72 static BROTLI_INLINE void RingBufferInitBuffer( |
| 73 MemoryManager* m, const uint32_t buflen, RingBuffer* rb) { |
| 74 static const size_t kSlackForEightByteHashingEverywhere = 7; |
| 75 uint8_t* new_data = BROTLI_ALLOC( |
| 76 m, uint8_t, 2 + buflen + kSlackForEightByteHashingEverywhere); |
| 77 size_t i; |
| 78 if (BROTLI_IS_OOM(m)) return; |
| 79 if (rb->data_) { |
| 80 memcpy(new_data, rb->data_, |
| 81 2 + rb->cur_size_ + kSlackForEightByteHashingEverywhere); |
| 82 BROTLI_FREE(m, rb->data_); |
| 83 } |
| 84 rb->data_ = new_data; |
| 85 rb->cur_size_ = buflen; |
| 86 rb->buffer_ = rb->data_ + 2; |
| 87 rb->buffer_[-2] = rb->buffer_[-1] = 0; |
| 88 for (i = 0; i < kSlackForEightByteHashingEverywhere; ++i) { |
| 89 rb->buffer_[rb->cur_size_ + i] = 0; |
| 90 } |
| 91 } |
| 92 |
| 93 static BROTLI_INLINE void RingBufferWriteTail( |
| 94 const uint8_t *bytes, size_t n, RingBuffer* rb) { |
| 95 const size_t masked_pos = rb->pos_ & rb->mask_; |
| 96 if (BROTLI_PREDICT_FALSE(masked_pos < rb->tail_size_)) { |
| 97 /* Just fill the tail buffer with the beginning data. */ |
| 98 const size_t p = rb->size_ + masked_pos; |
| 99 memcpy(&rb->buffer_[p], bytes, |
| 100 BROTLI_MIN(size_t, n, rb->tail_size_ - masked_pos)); |
| 101 } |
| 102 } |
| 103 |
| 104 /* Push bytes into the ring buffer. */ |
| 105 static BROTLI_INLINE void RingBufferWrite( |
| 106 MemoryManager* m, const uint8_t *bytes, size_t n, RingBuffer* rb) { |
| 107 if (rb->pos_ == 0 && n < rb->tail_size_) { |
| 108 /* Special case for the first write: to process the first block, we don't |
| 109 need to allocate the whole ring-buffer and we don't need the tail |
| 110 either. However, we do this memory usage optimization only if the |
| 111 first write is less than the tail size, which is also the input block |
| 112 size, otherwise it is likely that other blocks will follow and we |
| 113 will need to reallocate to the full size anyway. */ |
| 114 rb->pos_ = (uint32_t)n; |
| 115 RingBufferInitBuffer(m, rb->pos_, rb); |
| 116 if (BROTLI_IS_OOM(m)) return; |
| 117 memcpy(rb->buffer_, bytes, n); |
| 118 return; |
| 119 } |
| 120 if (rb->cur_size_ < rb->total_size_) { |
| 121 /* Lazily allocate the full buffer. */ |
| 122 RingBufferInitBuffer(m, rb->total_size_, rb); |
| 123 if (BROTLI_IS_OOM(m)) return; |
| 124 /* Initialize the last two bytes to zero, so that we don't have to worry |
| 125 later when we copy the last two bytes to the first two positions. */ |
| 126 rb->buffer_[rb->size_ - 2] = 0; |
| 127 rb->buffer_[rb->size_ - 1] = 0; |
| 128 } |
| 129 { |
| 130 const size_t masked_pos = rb->pos_ & rb->mask_; |
| 131 /* The length of the writes is limited so that we do not need to worry |
| 132 about a write */ |
| 133 RingBufferWriteTail(bytes, n, rb); |
| 134 if (BROTLI_PREDICT_TRUE(masked_pos + n <= rb->size_)) { |
| 135 /* A single write fits. */ |
| 136 memcpy(&rb->buffer_[masked_pos], bytes, n); |
| 137 } else { |
| 138 /* Split into two writes. |
| 139 Copy into the end of the buffer, including the tail buffer. */ |
| 140 memcpy(&rb->buffer_[masked_pos], bytes, |
| 141 BROTLI_MIN(size_t, n, rb->total_size_ - masked_pos)); |
| 142 /* Copy into the beginning of the buffer */ |
| 143 memcpy(&rb->buffer_[0], bytes + (rb->size_ - masked_pos), |
| 144 n - (rb->size_ - masked_pos)); |
| 145 } |
| 146 } |
| 147 rb->buffer_[-2] = rb->buffer_[rb->size_ - 2]; |
| 148 rb->buffer_[-1] = rb->buffer_[rb->size_ - 1]; |
| 149 rb->pos_ += (uint32_t)n; |
| 150 if (rb->pos_ > (1u << 30)) { |
| 151 /* Wrap, but preserve not-a-first-lap feature. */ |
| 152 rb->pos_ = (rb->pos_ & ((1u << 30) - 1)) | (1u << 30); |
| 153 } |
| 154 } |
| 155 |
| 156 #if defined(__cplusplus) || defined(c_plusplus) |
| 157 } /* extern "C" */ |
| 158 #endif |
| 159 |
| 160 #endif /* BROTLI_ENC_RINGBUFFER_H_ */ |
OLD | NEW |