extlzham 0.0.1.PROTOTYPE
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +7 -0
- data/LICENSE.md +27 -0
- data/README.md +21 -0
- data/Rakefile +143 -0
- data/contrib/lzham/LICENSE +22 -0
- data/contrib/lzham/README.md +209 -0
- data/contrib/lzham/include/lzham.h +781 -0
- data/contrib/lzham/lzhamcomp/lzham_comp.h +38 -0
- data/contrib/lzham/lzhamcomp/lzham_lzbase.cpp +244 -0
- data/contrib/lzham/lzhamcomp/lzham_lzbase.h +45 -0
- data/contrib/lzham/lzhamcomp/lzham_lzcomp.cpp +608 -0
- data/contrib/lzham/lzhamcomp/lzham_lzcomp_internal.cpp +1966 -0
- data/contrib/lzham/lzhamcomp/lzham_lzcomp_internal.h +472 -0
- data/contrib/lzham/lzhamcomp/lzham_lzcomp_state.cpp +1413 -0
- data/contrib/lzham/lzhamcomp/lzham_match_accel.cpp +562 -0
- data/contrib/lzham/lzhamcomp/lzham_match_accel.h +146 -0
- data/contrib/lzham/lzhamcomp/lzham_null_threading.h +97 -0
- data/contrib/lzham/lzhamcomp/lzham_pthreads_threading.cpp +229 -0
- data/contrib/lzham/lzhamcomp/lzham_pthreads_threading.h +520 -0
- data/contrib/lzham/lzhamcomp/lzham_threading.h +12 -0
- data/contrib/lzham/lzhamcomp/lzham_win32_threading.cpp +220 -0
- data/contrib/lzham/lzhamcomp/lzham_win32_threading.h +368 -0
- data/contrib/lzham/lzhamdecomp/lzham_assert.cpp +66 -0
- data/contrib/lzham/lzhamdecomp/lzham_assert.h +40 -0
- data/contrib/lzham/lzhamdecomp/lzham_checksum.cpp +73 -0
- data/contrib/lzham/lzhamdecomp/lzham_checksum.h +13 -0
- data/contrib/lzham/lzhamdecomp/lzham_config.h +23 -0
- data/contrib/lzham/lzhamdecomp/lzham_core.h +264 -0
- data/contrib/lzham/lzhamdecomp/lzham_decomp.h +37 -0
- data/contrib/lzham/lzhamdecomp/lzham_helpers.h +54 -0
- data/contrib/lzham/lzhamdecomp/lzham_huffman_codes.cpp +262 -0
- data/contrib/lzham/lzhamdecomp/lzham_huffman_codes.h +14 -0
- data/contrib/lzham/lzhamdecomp/lzham_lzdecomp.cpp +1527 -0
- data/contrib/lzham/lzhamdecomp/lzham_lzdecompbase.cpp +131 -0
- data/contrib/lzham/lzhamdecomp/lzham_lzdecompbase.h +89 -0
- data/contrib/lzham/lzhamdecomp/lzham_math.h +142 -0
- data/contrib/lzham/lzhamdecomp/lzham_mem.cpp +284 -0
- data/contrib/lzham/lzhamdecomp/lzham_mem.h +112 -0
- data/contrib/lzham/lzhamdecomp/lzham_platform.cpp +157 -0
- data/contrib/lzham/lzhamdecomp/lzham_platform.h +284 -0
- data/contrib/lzham/lzhamdecomp/lzham_prefix_coding.cpp +351 -0
- data/contrib/lzham/lzhamdecomp/lzham_prefix_coding.h +146 -0
- data/contrib/lzham/lzhamdecomp/lzham_symbol_codec.cpp +1484 -0
- data/contrib/lzham/lzhamdecomp/lzham_symbol_codec.h +556 -0
- data/contrib/lzham/lzhamdecomp/lzham_timer.cpp +147 -0
- data/contrib/lzham/lzhamdecomp/lzham_timer.h +99 -0
- data/contrib/lzham/lzhamdecomp/lzham_traits.h +141 -0
- data/contrib/lzham/lzhamdecomp/lzham_types.h +97 -0
- data/contrib/lzham/lzhamdecomp/lzham_utils.h +58 -0
- data/contrib/lzham/lzhamdecomp/lzham_vector.cpp +75 -0
- data/contrib/lzham/lzhamdecomp/lzham_vector.h +588 -0
- data/contrib/lzham/lzhamlib/lzham_lib.cpp +179 -0
- data/examples/basic.rb +48 -0
- data/ext/extconf.rb +26 -0
- data/ext/extlzham.c +741 -0
- data/gemstub.rb +22 -0
- data/lib/extlzham/version.rb +5 -0
- data/lib/extlzham.rb +153 -0
- metadata +135 -0
@@ -0,0 +1,562 @@
|
|
1
|
+
// File: lzham_match_accel.cpp
|
2
|
+
// See Copyright Notice and license at the end of include/lzham.h
|
3
|
+
#include "lzham_core.h"
|
4
|
+
#include "lzham_match_accel.h"
|
5
|
+
#include "lzham_timer.h"
|
6
|
+
|
7
|
+
namespace lzham
|
8
|
+
{
|
9
|
+
static inline uint32 hash2_to_12(uint c0, uint c1)
|
10
|
+
{
|
11
|
+
return c0 ^ (c1 << 4);
|
12
|
+
}
|
13
|
+
|
14
|
+
static inline uint32 hash3_to_16(uint c0, uint c1, uint c2)
|
15
|
+
{
|
16
|
+
return (c0 | (c1 << 8)) ^ (c2 << 4);
|
17
|
+
}
|
18
|
+
|
19
|
+
search_accelerator::search_accelerator() :
|
20
|
+
m_pLZBase(NULL),
|
21
|
+
m_pTask_pool(NULL),
|
22
|
+
m_max_helper_threads(0),
|
23
|
+
m_max_dict_size(0),
|
24
|
+
m_max_dict_size_mask(0),
|
25
|
+
m_lookahead_pos(0),
|
26
|
+
m_lookahead_size(0),
|
27
|
+
m_cur_dict_size(0),
|
28
|
+
m_fill_lookahead_pos(0),
|
29
|
+
m_fill_lookahead_size(0),
|
30
|
+
m_fill_dict_size(0),
|
31
|
+
m_max_probes(0),
|
32
|
+
m_max_matches(0),
|
33
|
+
m_all_matches(false),
|
34
|
+
m_next_match_ref(0),
|
35
|
+
m_num_completed_helper_threads(0)
|
36
|
+
{
|
37
|
+
}
|
38
|
+
|
39
|
+
bool search_accelerator::init(CLZBase* pLZBase, task_pool* pPool, uint max_helper_threads, uint max_dict_size, uint max_matches, bool all_matches, uint max_probes)
|
40
|
+
{
|
41
|
+
LZHAM_ASSERT(pLZBase);
|
42
|
+
LZHAM_ASSERT(max_dict_size && math::is_power_of_2(max_dict_size));
|
43
|
+
LZHAM_ASSERT(max_probes);
|
44
|
+
|
45
|
+
m_max_probes = LZHAM_MIN(cMatchAccelMaxSupportedProbes, max_probes);
|
46
|
+
|
47
|
+
m_pLZBase = pLZBase;
|
48
|
+
m_pTask_pool = max_helper_threads ? pPool : NULL;
|
49
|
+
m_max_helper_threads = m_pTask_pool ? max_helper_threads : 0;
|
50
|
+
m_max_matches = LZHAM_MIN(m_max_probes, max_matches);
|
51
|
+
m_all_matches = all_matches;
|
52
|
+
|
53
|
+
m_max_dict_size = max_dict_size;
|
54
|
+
m_max_dict_size_mask = m_max_dict_size - 1;
|
55
|
+
m_cur_dict_size = 0;
|
56
|
+
m_lookahead_size = 0;
|
57
|
+
m_lookahead_pos = 0;
|
58
|
+
m_fill_lookahead_pos = 0;
|
59
|
+
m_fill_lookahead_size = 0;
|
60
|
+
m_fill_dict_size = 0;
|
61
|
+
m_num_completed_helper_threads = 0;
|
62
|
+
|
63
|
+
if (!m_dict.try_resize_no_construct(max_dict_size + LZHAM_MIN(m_max_dict_size, static_cast<uint>(CLZBase::cMaxHugeMatchLen))))
|
64
|
+
return false;
|
65
|
+
|
66
|
+
if (!m_hash.try_resize_no_construct(cHashSize))
|
67
|
+
return false;
|
68
|
+
|
69
|
+
if (!m_nodes.try_resize_no_construct(max_dict_size))
|
70
|
+
return false;
|
71
|
+
|
72
|
+
memset(m_hash.get_ptr(), 0, m_hash.size_in_bytes());
|
73
|
+
|
74
|
+
return true;
|
75
|
+
}
|
76
|
+
|
77
|
+
void search_accelerator::reset()
|
78
|
+
{
|
79
|
+
m_cur_dict_size = 0;
|
80
|
+
m_lookahead_size = 0;
|
81
|
+
m_lookahead_pos = 0;
|
82
|
+
m_fill_lookahead_pos = 0;
|
83
|
+
m_fill_lookahead_size = 0;
|
84
|
+
m_fill_dict_size = 0;
|
85
|
+
m_num_completed_helper_threads = 0;
|
86
|
+
|
87
|
+
// Clearing the hash tables is only necessary for determinism (otherwise, it's possible the matches returned after a reset will depend on the data processes before the reset).
|
88
|
+
if (m_hash.size())
|
89
|
+
memset(m_hash.get_ptr(), 0, m_hash.size_in_bytes());
|
90
|
+
if (m_digram_hash.size())
|
91
|
+
memset(m_digram_hash.get_ptr(), 0, m_digram_hash.size_in_bytes());
|
92
|
+
}
|
93
|
+
|
94
|
+
void search_accelerator::flush()
|
95
|
+
{
|
96
|
+
m_cur_dict_size = 0;
|
97
|
+
}
|
98
|
+
|
99
|
+
uint search_accelerator::get_max_add_bytes() const
|
100
|
+
{
|
101
|
+
uint add_pos = static_cast<uint>(m_lookahead_pos & (m_max_dict_size - 1));
|
102
|
+
return m_max_dict_size - add_pos;
|
103
|
+
}
|
104
|
+
|
105
|
+
static uint8 g_hamming_dist[256] =
|
106
|
+
{
|
107
|
+
0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4,
|
108
|
+
1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
|
109
|
+
1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
|
110
|
+
2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
|
111
|
+
1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
|
112
|
+
2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
|
113
|
+
2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
|
114
|
+
3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
|
115
|
+
1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
|
116
|
+
2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
|
117
|
+
2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
|
118
|
+
3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
|
119
|
+
2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
|
120
|
+
3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
|
121
|
+
3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
|
122
|
+
4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8
|
123
|
+
};
|
124
|
+
|
125
|
+
void search_accelerator::find_all_matches_callback(uint64 data, void* pData_ptr)
|
126
|
+
{
|
127
|
+
scoped_perf_section find_all_matches_timer("find_all_matches_callback");
|
128
|
+
|
129
|
+
LZHAM_NOTE_UNUSED(pData_ptr);
|
130
|
+
const uint thread_index = (uint)data;
|
131
|
+
|
132
|
+
dict_match temp_matches[cMatchAccelMaxSupportedProbes * 2];
|
133
|
+
|
134
|
+
uint fill_lookahead_pos = m_fill_lookahead_pos;
|
135
|
+
uint fill_dict_size = m_fill_dict_size;
|
136
|
+
uint fill_lookahead_size = m_fill_lookahead_size;
|
137
|
+
|
138
|
+
uint c0 = 0, c1 = 0;
|
139
|
+
if (fill_lookahead_size >= 2)
|
140
|
+
{
|
141
|
+
c0 = m_dict[fill_lookahead_pos & m_max_dict_size_mask];
|
142
|
+
c1 = m_dict[(fill_lookahead_pos & m_max_dict_size_mask) + 1];
|
143
|
+
}
|
144
|
+
|
145
|
+
const uint8* pDict = m_dict.get_ptr();
|
146
|
+
|
147
|
+
while (fill_lookahead_size >= 3)
|
148
|
+
{
|
149
|
+
uint insert_pos = fill_lookahead_pos & m_max_dict_size_mask;
|
150
|
+
|
151
|
+
uint c2 = pDict[insert_pos + 2];
|
152
|
+
uint h = hash3_to_16(c0, c1, c2);
|
153
|
+
c0 = c1;
|
154
|
+
c1 = c2;
|
155
|
+
|
156
|
+
LZHAM_ASSERT(!m_hash_thread_index.size() || (m_hash_thread_index[h] != UINT8_MAX));
|
157
|
+
|
158
|
+
// Only process those strings that this worker thread was assigned to - this allows us to manipulate multiple trees in parallel with no worries about synchronization.
|
159
|
+
if (m_hash_thread_index.size() && (m_hash_thread_index[h] != thread_index))
|
160
|
+
{
|
161
|
+
fill_lookahead_pos++;
|
162
|
+
fill_lookahead_size--;
|
163
|
+
fill_dict_size++;
|
164
|
+
continue;
|
165
|
+
}
|
166
|
+
|
167
|
+
dict_match* pDstMatch = temp_matches;
|
168
|
+
|
169
|
+
uint cur_pos = m_hash[h];
|
170
|
+
m_hash[h] = static_cast<uint>(fill_lookahead_pos);
|
171
|
+
|
172
|
+
uint *pLeft = &m_nodes[insert_pos].m_left;
|
173
|
+
uint *pRight = &m_nodes[insert_pos].m_right;
|
174
|
+
|
175
|
+
const uint max_match_len = LZHAM_MIN(static_cast<uint>(CLZBase::cMaxMatchLen), fill_lookahead_size);
|
176
|
+
uint best_match_len = 2;
|
177
|
+
|
178
|
+
const uint8* pIns = &pDict[insert_pos];
|
179
|
+
|
180
|
+
uint n = m_max_probes;
|
181
|
+
for ( ; ; )
|
182
|
+
{
|
183
|
+
uint delta_pos = fill_lookahead_pos - cur_pos;
|
184
|
+
if ((n-- == 0) || (!delta_pos) || (delta_pos >= fill_dict_size))
|
185
|
+
{
|
186
|
+
*pLeft = 0;
|
187
|
+
*pRight = 0;
|
188
|
+
break;
|
189
|
+
}
|
190
|
+
|
191
|
+
uint pos = cur_pos & m_max_dict_size_mask;
|
192
|
+
node *pNode = &m_nodes[pos];
|
193
|
+
|
194
|
+
// Unfortunately, the initial compare match_len must be 0 because of the way we hash and truncate matches at the end of each block.
|
195
|
+
uint match_len = 0;
|
196
|
+
const uint8* pComp = &pDict[pos];
|
197
|
+
|
198
|
+
#if LZHAM_PLATFORM_X360 || (LZHAM_USE_UNALIGNED_INT_LOADS == 0)
|
199
|
+
for ( ; match_len < max_match_len; match_len++)
|
200
|
+
if (pComp[match_len] != pIns[match_len])
|
201
|
+
break;
|
202
|
+
#else
|
203
|
+
// Compare a qword at a time for a bit more efficiency.
|
204
|
+
const uint64* pComp_end = reinterpret_cast<const uint64*>(pComp + max_match_len - 7);
|
205
|
+
const uint64* pComp_cur = reinterpret_cast<const uint64*>(pComp);
|
206
|
+
const uint64* pIns_cur = reinterpret_cast<const uint64*>(pIns);
|
207
|
+
while (pComp_cur < pComp_end)
|
208
|
+
{
|
209
|
+
if (*pComp_cur != *pIns_cur)
|
210
|
+
break;
|
211
|
+
pComp_cur++;
|
212
|
+
pIns_cur++;
|
213
|
+
}
|
214
|
+
uint alt_match_len = static_cast<uint>(reinterpret_cast<const uint8*>(pComp_cur) - reinterpret_cast<const uint8*>(pComp));
|
215
|
+
for ( ; alt_match_len < max_match_len; alt_match_len++)
|
216
|
+
if (pComp[alt_match_len] != pIns[alt_match_len])
|
217
|
+
break;
|
218
|
+
#ifdef LZVERIFY
|
219
|
+
for ( ; match_len < max_match_len; match_len++)
|
220
|
+
if (pComp[match_len] != pIns[match_len])
|
221
|
+
break;
|
222
|
+
LZHAM_VERIFY(alt_match_len == match_len);
|
223
|
+
#endif
|
224
|
+
match_len = alt_match_len;
|
225
|
+
#endif
|
226
|
+
|
227
|
+
if (match_len > best_match_len)
|
228
|
+
{
|
229
|
+
pDstMatch->m_len = static_cast<uint16>(match_len - CLZBase::cMinMatchLen);
|
230
|
+
pDstMatch->m_dist = delta_pos;
|
231
|
+
pDstMatch++;
|
232
|
+
|
233
|
+
best_match_len = match_len;
|
234
|
+
|
235
|
+
if (match_len == max_match_len)
|
236
|
+
{
|
237
|
+
*pLeft = pNode->m_left;
|
238
|
+
*pRight = pNode->m_right;
|
239
|
+
break;
|
240
|
+
}
|
241
|
+
}
|
242
|
+
else if (m_all_matches)
|
243
|
+
{
|
244
|
+
pDstMatch->m_len = static_cast<uint16>(match_len - CLZBase::cMinMatchLen);
|
245
|
+
pDstMatch->m_dist = delta_pos;
|
246
|
+
pDstMatch++;
|
247
|
+
}
|
248
|
+
else if ((best_match_len > 2) && (best_match_len == match_len))
|
249
|
+
{
|
250
|
+
uint bestMatchDist = pDstMatch[-1].m_dist;
|
251
|
+
uint compMatchDist = delta_pos;
|
252
|
+
|
253
|
+
uint bestMatchSlot, bestMatchSlotOfs;
|
254
|
+
m_pLZBase->compute_lzx_position_slot(bestMatchDist, bestMatchSlot, bestMatchSlotOfs);
|
255
|
+
|
256
|
+
uint compMatchSlot, compMatchOfs;
|
257
|
+
m_pLZBase->compute_lzx_position_slot(compMatchDist, compMatchSlot, compMatchOfs);
|
258
|
+
|
259
|
+
// If both matches uses the same match slot, choose the one with the offset containing the lowest nibble as these bits separately entropy coded.
|
260
|
+
// This could choose a match which is further away in the absolute sense, but closer in a coding sense.
|
261
|
+
if ( (compMatchSlot < bestMatchSlot) ||
|
262
|
+
((compMatchSlot >= 8) && (compMatchSlot == bestMatchSlot) && ((compMatchOfs & 15) < (bestMatchSlotOfs & 15))) )
|
263
|
+
{
|
264
|
+
LZHAM_ASSERT((pDstMatch[-1].m_len + (uint)CLZBase::cMinMatchLen) == best_match_len);
|
265
|
+
pDstMatch[-1].m_dist = delta_pos;
|
266
|
+
}
|
267
|
+
else if ((match_len < max_match_len) && (compMatchSlot <= bestMatchSlot))
|
268
|
+
{
|
269
|
+
// Choose the match which has lowest hamming distance in the mismatch byte for a tiny win on binary files.
|
270
|
+
// TODO: This competes against the prev. optimization.
|
271
|
+
uint desired_mismatch_byte = pIns[match_len];
|
272
|
+
|
273
|
+
uint cur_mismatch_byte = pDict[(insert_pos - bestMatchDist + match_len) & m_max_dict_size_mask];
|
274
|
+
uint cur_mismatch_dist = g_hamming_dist[cur_mismatch_byte ^ desired_mismatch_byte];
|
275
|
+
|
276
|
+
uint new_mismatch_byte = pComp[match_len];
|
277
|
+
uint new_mismatch_dist = g_hamming_dist[new_mismatch_byte ^ desired_mismatch_byte];
|
278
|
+
if (new_mismatch_dist < cur_mismatch_dist)
|
279
|
+
{
|
280
|
+
LZHAM_ASSERT((pDstMatch[-1].m_len + (uint)CLZBase::cMinMatchLen) == best_match_len);
|
281
|
+
pDstMatch[-1].m_dist = delta_pos;
|
282
|
+
}
|
283
|
+
}
|
284
|
+
}
|
285
|
+
|
286
|
+
uint new_pos;
|
287
|
+
if (pComp[match_len] < pIns[match_len])
|
288
|
+
{
|
289
|
+
*pLeft = cur_pos;
|
290
|
+
pLeft = &pNode->m_right;
|
291
|
+
new_pos = pNode->m_right;
|
292
|
+
}
|
293
|
+
else
|
294
|
+
{
|
295
|
+
*pRight = cur_pos;
|
296
|
+
pRight = &pNode->m_left;
|
297
|
+
new_pos = pNode->m_left;
|
298
|
+
}
|
299
|
+
if (new_pos == cur_pos)
|
300
|
+
break;
|
301
|
+
cur_pos = new_pos;
|
302
|
+
}
|
303
|
+
|
304
|
+
const uint num_matches = (uint)(pDstMatch - temp_matches);
|
305
|
+
|
306
|
+
if (num_matches)
|
307
|
+
{
|
308
|
+
pDstMatch[-1].m_dist |= 0x80000000;
|
309
|
+
|
310
|
+
const uint num_matches_to_write = LZHAM_MIN(num_matches, m_max_matches);
|
311
|
+
|
312
|
+
const uint match_ref_ofs = static_cast<uint>(atomic_exchange_add(&m_next_match_ref, num_matches_to_write));
|
313
|
+
|
314
|
+
memcpy(&m_matches[match_ref_ofs],
|
315
|
+
temp_matches + (num_matches - num_matches_to_write),
|
316
|
+
sizeof(temp_matches[0]) * num_matches_to_write);
|
317
|
+
|
318
|
+
// FIXME: This is going to really hurt on platforms requiring export barriers.
|
319
|
+
LZHAM_MEMORY_EXPORT_BARRIER
|
320
|
+
|
321
|
+
atomic_exchange32((atomic32_t*)&m_match_refs[static_cast<uint>(fill_lookahead_pos - m_fill_lookahead_pos)], match_ref_ofs);
|
322
|
+
}
|
323
|
+
else
|
324
|
+
{
|
325
|
+
atomic_exchange32((atomic32_t*)&m_match_refs[static_cast<uint>(fill_lookahead_pos - m_fill_lookahead_pos)], -2);
|
326
|
+
}
|
327
|
+
|
328
|
+
fill_lookahead_pos++;
|
329
|
+
fill_lookahead_size--;
|
330
|
+
fill_dict_size++;
|
331
|
+
}
|
332
|
+
|
333
|
+
while (fill_lookahead_size)
|
334
|
+
{
|
335
|
+
uint insert_pos = fill_lookahead_pos & m_max_dict_size_mask;
|
336
|
+
m_nodes[insert_pos].m_left = 0;
|
337
|
+
m_nodes[insert_pos].m_right = 0;
|
338
|
+
|
339
|
+
atomic_exchange32((atomic32_t*)&m_match_refs[static_cast<uint>(fill_lookahead_pos - m_fill_lookahead_pos)], -2);
|
340
|
+
|
341
|
+
fill_lookahead_pos++;
|
342
|
+
fill_lookahead_size--;
|
343
|
+
fill_dict_size++;
|
344
|
+
}
|
345
|
+
|
346
|
+
atomic_increment32(&m_num_completed_helper_threads);
|
347
|
+
}
|
348
|
+
|
349
|
+
bool search_accelerator::find_len2_matches()
|
350
|
+
{
|
351
|
+
if (!m_digram_hash.size())
|
352
|
+
{
|
353
|
+
if (!m_digram_hash.try_resize(cDigramHashSize))
|
354
|
+
return false;
|
355
|
+
}
|
356
|
+
|
357
|
+
if (m_digram_next.size() < m_lookahead_size)
|
358
|
+
{
|
359
|
+
if (!m_digram_next.try_resize(m_lookahead_size))
|
360
|
+
return false;
|
361
|
+
}
|
362
|
+
|
363
|
+
uint lookahead_dict_pos = m_lookahead_pos & m_max_dict_size_mask;
|
364
|
+
|
365
|
+
for (int lookahead_ofs = 0; lookahead_ofs < ((int)m_lookahead_size - 1); ++lookahead_ofs, ++lookahead_dict_pos)
|
366
|
+
{
|
367
|
+
uint c0 = m_dict[lookahead_dict_pos];
|
368
|
+
uint c1 = m_dict[lookahead_dict_pos + 1];
|
369
|
+
|
370
|
+
uint h = hash2_to_12(c0, c1) & (cDigramHashSize - 1);
|
371
|
+
|
372
|
+
m_digram_next[lookahead_ofs] = m_digram_hash[h];
|
373
|
+
m_digram_hash[h] = m_lookahead_pos + lookahead_ofs;
|
374
|
+
}
|
375
|
+
|
376
|
+
m_digram_next[m_lookahead_size - 1] = 0;
|
377
|
+
|
378
|
+
return true;
|
379
|
+
}
|
380
|
+
|
381
|
+
uint search_accelerator::get_len2_match(uint lookahead_ofs)
|
382
|
+
{
|
383
|
+
if ((m_fill_lookahead_size - lookahead_ofs) < 2)
|
384
|
+
return 0;
|
385
|
+
|
386
|
+
uint cur_pos = m_lookahead_pos + lookahead_ofs;
|
387
|
+
|
388
|
+
uint next_match_pos = m_digram_next[cur_pos - m_fill_lookahead_pos];
|
389
|
+
|
390
|
+
uint match_dist = cur_pos - next_match_pos;
|
391
|
+
|
392
|
+
if ((!match_dist) || (match_dist > CLZBase::cMaxLen2MatchDist) || (match_dist > (m_cur_dict_size + lookahead_ofs)))
|
393
|
+
return 0;
|
394
|
+
|
395
|
+
const uint8* pCur = &m_dict[cur_pos & m_max_dict_size_mask];
|
396
|
+
const uint8* pMatch = &m_dict[next_match_pos & m_max_dict_size_mask];
|
397
|
+
|
398
|
+
if ((pCur[0] == pMatch[0]) && (pCur[1] == pMatch[1]))
|
399
|
+
return match_dist;
|
400
|
+
|
401
|
+
return 0;
|
402
|
+
}
|
403
|
+
|
404
|
+
bool search_accelerator::find_all_matches(uint num_bytes)
|
405
|
+
{
|
406
|
+
if (!m_matches.try_resize_no_construct(m_max_probes * num_bytes))
|
407
|
+
return false;
|
408
|
+
|
409
|
+
if (!m_match_refs.try_resize_no_construct(num_bytes))
|
410
|
+
return false;
|
411
|
+
|
412
|
+
memset(m_match_refs.get_ptr(), 0xFF, m_match_refs.size_in_bytes());
|
413
|
+
|
414
|
+
m_fill_lookahead_pos = m_lookahead_pos;
|
415
|
+
m_fill_lookahead_size = num_bytes;
|
416
|
+
m_fill_dict_size = m_cur_dict_size;
|
417
|
+
|
418
|
+
m_next_match_ref = 0;
|
419
|
+
|
420
|
+
if (!m_pTask_pool)
|
421
|
+
{
|
422
|
+
find_all_matches_callback(0, NULL);
|
423
|
+
|
424
|
+
m_num_completed_helper_threads = 0;
|
425
|
+
}
|
426
|
+
else
|
427
|
+
{
|
428
|
+
if (!m_hash_thread_index.try_resize_no_construct(0x10000))
|
429
|
+
return false;
|
430
|
+
|
431
|
+
memset(m_hash_thread_index.get_ptr(), 0xFF, m_hash_thread_index.size_in_bytes());
|
432
|
+
|
433
|
+
uint next_thread_index = 0;
|
434
|
+
const uint8* pDict = &m_dict[m_lookahead_pos & m_max_dict_size_mask];
|
435
|
+
uint num_unique_trigrams = 0;
|
436
|
+
|
437
|
+
if (num_bytes >= 3)
|
438
|
+
{
|
439
|
+
uint c0 = pDict[0];
|
440
|
+
uint c1 = pDict[1];
|
441
|
+
|
442
|
+
const int limit = ((int)num_bytes - 2);
|
443
|
+
for (int i = 0; i < limit; i++)
|
444
|
+
{
|
445
|
+
uint c2 = pDict[2];
|
446
|
+
uint t = hash3_to_16(c0, c1, c2);
|
447
|
+
c0 = c1;
|
448
|
+
c1 = c2;
|
449
|
+
|
450
|
+
pDict++;
|
451
|
+
|
452
|
+
if (m_hash_thread_index[t] == UINT8_MAX)
|
453
|
+
{
|
454
|
+
num_unique_trigrams++;
|
455
|
+
|
456
|
+
m_hash_thread_index[t] = static_cast<uint8>(next_thread_index);
|
457
|
+
if (++next_thread_index == m_max_helper_threads)
|
458
|
+
next_thread_index = 0;
|
459
|
+
}
|
460
|
+
}
|
461
|
+
}
|
462
|
+
|
463
|
+
m_num_completed_helper_threads = 0;
|
464
|
+
|
465
|
+
if (!m_pTask_pool->queue_multiple_object_tasks(this, &search_accelerator::find_all_matches_callback, 0, m_max_helper_threads))
|
466
|
+
return false;
|
467
|
+
}
|
468
|
+
|
469
|
+
return find_len2_matches();
|
470
|
+
}
|
471
|
+
|
472
|
+
bool search_accelerator::add_bytes_begin(uint num_bytes, const uint8* pBytes)
|
473
|
+
{
|
474
|
+
LZHAM_ASSERT(num_bytes <= m_max_dict_size);
|
475
|
+
LZHAM_ASSERT(!m_lookahead_size);
|
476
|
+
|
477
|
+
uint add_pos = m_lookahead_pos & m_max_dict_size_mask;
|
478
|
+
LZHAM_ASSERT((add_pos + num_bytes) <= m_max_dict_size);
|
479
|
+
|
480
|
+
memcpy(&m_dict[add_pos], pBytes, num_bytes);
|
481
|
+
|
482
|
+
uint dict_bytes_to_mirror = LZHAM_MIN(static_cast<uint>(CLZBase::cMaxHugeMatchLen), m_max_dict_size);
|
483
|
+
if (add_pos < dict_bytes_to_mirror)
|
484
|
+
memcpy(&m_dict[m_max_dict_size], &m_dict[0], dict_bytes_to_mirror);
|
485
|
+
|
486
|
+
m_lookahead_size = num_bytes;
|
487
|
+
|
488
|
+
uint max_possible_dict_size = m_max_dict_size - num_bytes;
|
489
|
+
m_cur_dict_size = LZHAM_MIN(m_cur_dict_size, max_possible_dict_size);
|
490
|
+
|
491
|
+
m_next_match_ref = 0;
|
492
|
+
|
493
|
+
return find_all_matches(num_bytes);
|
494
|
+
}
|
495
|
+
|
496
|
+
void search_accelerator::add_bytes_end()
|
497
|
+
{
|
498
|
+
if (m_pTask_pool)
|
499
|
+
{
|
500
|
+
m_pTask_pool->join();
|
501
|
+
}
|
502
|
+
|
503
|
+
LZHAM_ASSERT((uint)m_next_match_ref <= m_matches.size());
|
504
|
+
}
|
505
|
+
|
506
|
+
dict_match* search_accelerator::find_matches(uint lookahead_ofs, bool spin)
|
507
|
+
{
|
508
|
+
LZHAM_ASSERT(lookahead_ofs < m_lookahead_size);
|
509
|
+
|
510
|
+
const uint match_ref_ofs = static_cast<uint>(m_lookahead_pos - m_fill_lookahead_pos + lookahead_ofs);
|
511
|
+
|
512
|
+
int match_ref;
|
513
|
+
uint spin_count = 0;
|
514
|
+
|
515
|
+
// This may spin until the match finder job(s) catch up to the caller's lookahead position.
|
516
|
+
for ( ; ; )
|
517
|
+
{
|
518
|
+
match_ref = static_cast<int>(m_match_refs[match_ref_ofs]);
|
519
|
+
if (match_ref == -2)
|
520
|
+
return NULL;
|
521
|
+
else if (match_ref != -1)
|
522
|
+
break;
|
523
|
+
|
524
|
+
spin_count++;
|
525
|
+
const uint cMaxSpinCount = 1000;
|
526
|
+
if ((spin) && (spin_count < cMaxSpinCount))
|
527
|
+
{
|
528
|
+
lzham_yield_processor();
|
529
|
+
lzham_yield_processor();
|
530
|
+
lzham_yield_processor();
|
531
|
+
lzham_yield_processor();
|
532
|
+
lzham_yield_processor();
|
533
|
+
lzham_yield_processor();
|
534
|
+
lzham_yield_processor();
|
535
|
+
lzham_yield_processor();
|
536
|
+
|
537
|
+
LZHAM_MEMORY_IMPORT_BARRIER
|
538
|
+
}
|
539
|
+
else
|
540
|
+
{
|
541
|
+
spin_count = cMaxSpinCount;
|
542
|
+
|
543
|
+
lzham_sleep(1);
|
544
|
+
}
|
545
|
+
}
|
546
|
+
|
547
|
+
LZHAM_MEMORY_IMPORT_BARRIER
|
548
|
+
|
549
|
+
return &m_matches[match_ref];
|
550
|
+
}
|
551
|
+
|
552
|
+
void search_accelerator::advance_bytes(uint num_bytes)
|
553
|
+
{
|
554
|
+
LZHAM_ASSERT(num_bytes <= m_lookahead_size);
|
555
|
+
|
556
|
+
m_lookahead_pos += num_bytes;
|
557
|
+
m_lookahead_size -= num_bytes;
|
558
|
+
|
559
|
+
m_cur_dict_size += num_bytes;
|
560
|
+
LZHAM_ASSERT(m_cur_dict_size <= m_max_dict_size);
|
561
|
+
}
|
562
|
+
}
|
@@ -0,0 +1,146 @@
|
|
1
|
+
// File: lzham_match_accel.h
|
2
|
+
// See Copyright Notice and license at the end of include/lzham.h
|
3
|
+
#pragma once
|
4
|
+
#include "lzham_lzbase.h"
|
5
|
+
#include "lzham_threading.h"
|
6
|
+
|
7
|
+
namespace lzham
|
8
|
+
{
|
9
|
+
const uint cMatchAccelMaxSupportedProbes = 128;
|
10
|
+
|
11
|
+
struct node
|
12
|
+
{
|
13
|
+
uint m_left;
|
14
|
+
uint m_right;
|
15
|
+
};
|
16
|
+
|
17
|
+
LZHAM_DEFINE_BITWISE_MOVABLE(node);
|
18
|
+
|
19
|
+
#pragma pack(push, 1)
|
20
|
+
struct dict_match
|
21
|
+
{
|
22
|
+
uint m_dist;
|
23
|
+
uint16 m_len;
|
24
|
+
|
25
|
+
inline uint get_dist() const { return m_dist & 0x7FFFFFFF; }
|
26
|
+
inline uint get_len() const { return m_len + 2; }
|
27
|
+
inline bool is_last() const { return (int)m_dist < 0; }
|
28
|
+
};
|
29
|
+
#pragma pack(pop)
|
30
|
+
|
31
|
+
LZHAM_DEFINE_BITWISE_MOVABLE(dict_match);
|
32
|
+
|
33
|
+
class search_accelerator
|
34
|
+
{
|
35
|
+
public:
|
36
|
+
search_accelerator();
|
37
|
+
|
38
|
+
// If all_matches is true, the match finder returns all found matches with no filtering.
|
39
|
+
// Otherwise, the finder will tend to return lists of matches with mostly unique lengths.
|
40
|
+
// For each length, it will discard matches with worse distances (in the coding sense).
|
41
|
+
bool init(CLZBase* pLZBase, task_pool* pPool, uint max_helper_threads, uint max_dict_size, uint max_matches, bool all_matches, uint max_probes);
|
42
|
+
|
43
|
+
void reset();
|
44
|
+
void flush();
|
45
|
+
|
46
|
+
inline uint get_max_dict_size() const { return m_max_dict_size; }
|
47
|
+
inline uint get_max_dict_size_mask() const { return m_max_dict_size_mask; }
|
48
|
+
inline uint get_cur_dict_size() const { return m_cur_dict_size; }
|
49
|
+
|
50
|
+
inline uint get_lookahead_pos() const { return m_lookahead_pos; }
|
51
|
+
inline uint get_lookahead_size() const { return m_lookahead_size; }
|
52
|
+
|
53
|
+
inline uint get_char(int delta_pos) const { return m_dict[(m_lookahead_pos + delta_pos) & m_max_dict_size_mask]; }
|
54
|
+
inline uint get_char(uint cur_dict_pos, int delta_pos) const { return m_dict[(cur_dict_pos + delta_pos) & m_max_dict_size_mask]; }
|
55
|
+
inline const uint8* get_ptr(uint pos) const { return &m_dict[pos]; }
|
56
|
+
|
57
|
+
uint get_max_helper_threads() const { return m_max_helper_threads; }
|
58
|
+
|
59
|
+
inline uint operator[](uint pos) const { return m_dict[pos]; }
|
60
|
+
|
61
|
+
uint get_max_add_bytes() const;
|
62
|
+
bool add_bytes_begin(uint num_bytes, const uint8* pBytes);
|
63
|
+
inline atomic32_t get_num_completed_helper_threads() const { return m_num_completed_helper_threads; }
|
64
|
+
void add_bytes_end();
|
65
|
+
|
66
|
+
// Returns the lookahead's raw position/size/dict_size at the time add_bytes_begin() is called.
|
67
|
+
inline uint get_fill_lookahead_pos() const { return m_fill_lookahead_pos; }
|
68
|
+
inline uint get_fill_lookahead_size() const { return m_fill_lookahead_size; }
|
69
|
+
inline uint get_fill_dict_size() const { return m_fill_dict_size; }
|
70
|
+
|
71
|
+
uint get_len2_match(uint lookahead_ofs);
|
72
|
+
dict_match* find_matches(uint lookahead_ofs, bool spin = true);
|
73
|
+
|
74
|
+
void advance_bytes(uint num_bytes);
|
75
|
+
|
76
|
+
LZHAM_FORCE_INLINE uint get_match_len(uint lookahead_ofs, int dist, uint max_match_len, uint start_match_len = 0) const
|
77
|
+
{
|
78
|
+
LZHAM_ASSERT(lookahead_ofs < m_lookahead_size);
|
79
|
+
LZHAM_ASSERT(start_match_len <= max_match_len);
|
80
|
+
LZHAM_ASSERT(max_match_len <= (get_lookahead_size() - lookahead_ofs));
|
81
|
+
|
82
|
+
const int find_dict_size = m_cur_dict_size + lookahead_ofs;
|
83
|
+
if (dist > find_dict_size)
|
84
|
+
return 0;
|
85
|
+
|
86
|
+
const uint comp_pos = static_cast<uint>((m_lookahead_pos + lookahead_ofs - dist) & m_max_dict_size_mask);
|
87
|
+
const uint lookahead_pos = (m_lookahead_pos + lookahead_ofs) & m_max_dict_size_mask;
|
88
|
+
|
89
|
+
const uint8* pComp = &m_dict[comp_pos];
|
90
|
+
const uint8* pLookahead = &m_dict[lookahead_pos];
|
91
|
+
|
92
|
+
uint match_len;
|
93
|
+
for (match_len = start_match_len; match_len < max_match_len; match_len++)
|
94
|
+
if (pComp[match_len] != pLookahead[match_len])
|
95
|
+
break;
|
96
|
+
|
97
|
+
return match_len;
|
98
|
+
}
|
99
|
+
|
100
|
+
public:
|
101
|
+
CLZBase* m_pLZBase;
|
102
|
+
task_pool* m_pTask_pool;
|
103
|
+
uint m_max_helper_threads;
|
104
|
+
|
105
|
+
uint m_max_dict_size;
|
106
|
+
uint m_max_dict_size_mask;
|
107
|
+
|
108
|
+
uint m_lookahead_pos;
|
109
|
+
uint m_lookahead_size;
|
110
|
+
|
111
|
+
uint m_cur_dict_size;
|
112
|
+
|
113
|
+
lzham::vector<uint8> m_dict;
|
114
|
+
|
115
|
+
enum { cHashSize = 65536 };
|
116
|
+
lzham::vector<uint> m_hash;
|
117
|
+
lzham::vector<node> m_nodes;
|
118
|
+
|
119
|
+
lzham::vector<dict_match> m_matches;
|
120
|
+
lzham::vector<atomic32_t> m_match_refs;
|
121
|
+
|
122
|
+
lzham::vector<uint8> m_hash_thread_index;
|
123
|
+
|
124
|
+
enum { cDigramHashSize = 4096 };
|
125
|
+
lzham::vector<uint> m_digram_hash;
|
126
|
+
lzham::vector<uint> m_digram_next;
|
127
|
+
|
128
|
+
uint m_fill_lookahead_pos;
|
129
|
+
uint m_fill_lookahead_size;
|
130
|
+
uint m_fill_dict_size;
|
131
|
+
|
132
|
+
uint m_max_probes;
|
133
|
+
uint m_max_matches;
|
134
|
+
|
135
|
+
bool m_all_matches;
|
136
|
+
|
137
|
+
volatile atomic32_t m_next_match_ref;
|
138
|
+
|
139
|
+
volatile atomic32_t m_num_completed_helper_threads;
|
140
|
+
|
141
|
+
void find_all_matches_callback(uint64 data, void* pData_ptr);
|
142
|
+
bool find_all_matches(uint num_bytes);
|
143
|
+
bool find_len2_matches();
|
144
|
+
};
|
145
|
+
|
146
|
+
} // namespace lzham
|