| 1 |
/* ****************************************************************** |
| 2 |
* hist : Histogram functions |
| 3 |
* part of Finite State Entropy project |
| 4 |
* Copyright (c) 2013-2020, Yann Collet, Facebook, Inc. |
| 5 |
* |
| 6 |
* You can contact the author at : |
| 7 |
* - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy |
| 8 |
* - Public forum : https://groups.google.com/forum/#!forum/lz4c |
| 9 |
* |
| 10 |
* This source code is licensed under both the BSD-style license (found in the |
| 11 |
* LICENSE file in the root directory of this source tree) and the GPLv2 (found |
| 12 |
* in the COPYING file in the root directory of this source tree). |
| 13 |
* You may select, at your option, one of the above-listed licenses. |
| 14 |
****************************************************************** */ |
| 15 |
|
| 16 |
/* --- dependencies --- */ |
| 17 |
#include "../common/mem.h" /* U32, BYTE, etc. */ |
| 18 |
#include "../common/debug.h" /* assert, DEBUGLOG */ |
| 19 |
#include "../common/error_private.h" /* ERROR */ |
| 20 |
#include "hist.h" |
| 21 |
|
| 22 |
|
| 23 |
/* --- Error management --- */ |
| 24 |
unsigned HIST_isError(size_t code) { return ERR_isError(code); } |
| 25 |
|
| 26 |
/*-************************************************************** |
| 27 |
* Histogram functions |
| 28 |
****************************************************************/ |
| 29 |
unsigned HIST_count_simple(unsigned* count, unsigned* maxSymbolValuePtr, |
| 30 |
const void* src, size_t srcSize) |
| 31 |
{ |
| 32 |
const BYTE* ip = (const BYTE*)src; |
| 33 |
const BYTE* const end = ip + srcSize; |
| 34 |
unsigned maxSymbolValue = *maxSymbolValuePtr; |
| 35 |
unsigned largestCount=0; |
| 36 |
|
| 37 |
memset(count, 0, (maxSymbolValue+1) * sizeof(*count)); |
| 38 |
if (srcSize==0) { *maxSymbolValuePtr = 0; return 0; } |
| 39 |
|
| 40 |
while (ip<end) { |
| 41 |
assert(*ip <= maxSymbolValue); |
| 42 |
count[*ip++]++; |
| 43 |
} |
| 44 |
|
| 45 |
while (!count[maxSymbolValue]) maxSymbolValue--; |
| 46 |
*maxSymbolValuePtr = maxSymbolValue; |
| 47 |
|
| 48 |
{ U32 s; |
| 49 |
for (s=0; s<=maxSymbolValue; s++) |
| 50 |
if (count[s] > largestCount) largestCount = count[s]; |
| 51 |
} |
| 52 |
|
| 53 |
return largestCount; |
| 54 |
} |
| 55 |
|
| 56 |
typedef enum { trustInput, checkMaxSymbolValue } HIST_checkInput_e; |
| 57 |
|
| 58 |
/* HIST_count_parallel_wksp() : |
| 59 |
* store histogram into 4 intermediate tables, recombined at the end. |
| 60 |
* this design makes better use of OoO cpus, |
| 61 |
* and is noticeably faster when some values are heavily repeated. |
| 62 |
* But it needs some additional workspace for intermediate tables. |
| 63 |
* `workSpace` size must be a table of size >= HIST_WKSP_SIZE_U32. |
| 64 |
* @return : largest histogram frequency, |
| 65 |
* or an error code (notably when histogram would be larger than *maxSymbolValuePtr). */ |
| 66 |
static size_t HIST_count_parallel_wksp( |
| 67 |
unsigned* count, unsigned* maxSymbolValuePtr, |
| 68 |
const void* source, size_t sourceSize, |
| 69 |
HIST_checkInput_e check, |
| 70 |
U32* const workSpace) |
| 71 |
{ |
| 72 |
const BYTE* ip = (const BYTE*)source; |
| 73 |
const BYTE* const iend = ip+sourceSize; |
| 74 |
unsigned maxSymbolValue = *maxSymbolValuePtr; |
| 75 |
unsigned max=0; |
| 76 |
U32* const Counting1 = workSpace; |
| 77 |
U32* const Counting2 = Counting1 + 256; |
| 78 |
U32* const Counting3 = Counting2 + 256; |
| 79 |
U32* const Counting4 = Counting3 + 256; |
| 80 |
|
| 81 |
memset(workSpace, 0, 4*256*sizeof(unsigned)); |
| 82 |
|
| 83 |
/* safety checks */ |
| 84 |
if (!sourceSize) { |
| 85 |
memset(count, 0, maxSymbolValue + 1); |
| 86 |
*maxSymbolValuePtr = 0; |
| 87 |
return 0; |
| 88 |
} |
| 89 |
if (!maxSymbolValue) maxSymbolValue = 255; /* 0 == default */ |
| 90 |
|
| 91 |
/* by stripes of 16 bytes */ |
| 92 |
{ U32 cached = MEM_read32(ip); ip += 4; |
| 93 |
while (ip < iend-15) { |
| 94 |
U32 c = cached; cached = MEM_read32(ip); ip += 4; |
| 95 |
Counting1[(BYTE) c ]++; |
| 96 |
Counting2[(BYTE)(c>>8) ]++; |
| 97 |
Counting3[(BYTE)(c>>16)]++; |
| 98 |
Counting4[ c>>24 ]++; |
| 99 |
c = cached; cached = MEM_read32(ip); ip += 4; |
| 100 |
Counting1[(BYTE) c ]++; |
| 101 |
Counting2[(BYTE)(c>>8) ]++; |
| 102 |
Counting3[(BYTE)(c>>16)]++; |
| 103 |
Counting4[ c>>24 ]++; |
| 104 |
c = cached; cached = MEM_read32(ip); ip += 4; |
| 105 |
Counting1[(BYTE) c ]++; |
| 106 |
Counting2[(BYTE)(c>>8) ]++; |
| 107 |
Counting3[(BYTE)(c>>16)]++; |
| 108 |
Counting4[ c>>24 ]++; |
| 109 |
c = cached; cached = MEM_read32(ip); ip += 4; |
| 110 |
Counting1[(BYTE) c ]++; |
| 111 |
Counting2[(BYTE)(c>>8) ]++; |
| 112 |
Counting3[(BYTE)(c>>16)]++; |
| 113 |
Counting4[ c>>24 ]++; |
| 114 |
} |
| 115 |
ip-=4; |
| 116 |
} |
| 117 |
|
| 118 |
/* finish last symbols */ |
| 119 |
while (ip<iend) Counting1[*ip++]++; |
| 120 |
|
| 121 |
if (check) { /* verify stats will fit into destination table */ |
| 122 |
U32 s; for (s=255; s>maxSymbolValue; s--) { |
| 123 |
Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s]; |
| 124 |
if (Counting1[s]) return ERROR(maxSymbolValue_tooSmall); |
| 125 |
} } |
| 126 |
|
| 127 |
{ U32 s; |
| 128 |
if (maxSymbolValue > 255) maxSymbolValue = 255; |
| 129 |
for (s=0; s<=maxSymbolValue; s++) { |
| 130 |
count[s] = Counting1[s] + Counting2[s] + Counting3[s] + Counting4[s]; |
| 131 |
if (count[s] > max) max = count[s]; |
| 132 |
} } |
| 133 |
|
| 134 |
while (!count[maxSymbolValue]) maxSymbolValue--; |
| 135 |
*maxSymbolValuePtr = maxSymbolValue; |
| 136 |
return (size_t)max; |
| 137 |
} |
| 138 |
|
| 139 |
/* HIST_countFast_wksp() : |
| 140 |
* Same as HIST_countFast(), but using an externally provided scratch buffer. |
| 141 |
* `workSpace` is a writable buffer which must be 4-bytes aligned, |
| 142 |
* `workSpaceSize` must be >= HIST_WKSP_SIZE |
| 143 |
*/ |
| 144 |
size_t HIST_countFast_wksp(unsigned* count, unsigned* maxSymbolValuePtr, |
| 145 |
const void* source, size_t sourceSize, |
| 146 |
void* workSpace, size_t workSpaceSize) |
| 147 |
{ |
| 148 |
if (sourceSize < 1500) /* heuristic threshold */ |
| 149 |
return HIST_count_simple(count, maxSymbolValuePtr, source, sourceSize); |
| 150 |
if ((size_t)workSpace & 3) return ERROR(GENERIC); /* must be aligned on 4-bytes boundaries */ |
| 151 |
if (workSpaceSize < HIST_WKSP_SIZE) return ERROR(workSpace_tooSmall); |
| 152 |
return HIST_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, trustInput, (U32*)workSpace); |
| 153 |
} |
| 154 |
|
| 155 |
/* fast variant (unsafe : won't check if src contains values beyond count[] limit) */ |
| 156 |
size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr, |
| 157 |
const void* source, size_t sourceSize) |
| 158 |
{ |
| 159 |
unsigned tmpCounters[HIST_WKSP_SIZE_U32]; |
| 160 |
return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters, sizeof(tmpCounters)); |
| 161 |
} |
| 162 |
|
| 163 |
/* HIST_count_wksp() : |
| 164 |
* Same as HIST_count(), but using an externally provided scratch buffer. |
| 165 |
* `workSpace` size must be table of >= HIST_WKSP_SIZE_U32 unsigned */ |
| 166 |
size_t HIST_count_wksp(unsigned* count, unsigned* maxSymbolValuePtr, |
| 167 |
const void* source, size_t sourceSize, |
| 168 |
void* workSpace, size_t workSpaceSize) |
| 169 |
{ |
| 170 |
if ((size_t)workSpace & 3) return ERROR(GENERIC); /* must be aligned on 4-bytes boundaries */ |
| 171 |
if (workSpaceSize < HIST_WKSP_SIZE) return ERROR(workSpace_tooSmall); |
| 172 |
if (*maxSymbolValuePtr < 255) |
| 173 |
return HIST_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, checkMaxSymbolValue, (U32*)workSpace); |
| 174 |
*maxSymbolValuePtr = 255; |
| 175 |
return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, workSpace, workSpaceSize); |
| 176 |
} |
| 177 |
|
| 178 |
size_t HIST_count(unsigned* count, unsigned* maxSymbolValuePtr, |
| 179 |
const void* src, size_t srcSize) |
| 180 |
{ |
| 181 |
unsigned tmpCounters[HIST_WKSP_SIZE_U32]; |
| 182 |
return HIST_count_wksp(count, maxSymbolValuePtr, src, srcSize, tmpCounters, sizeof(tmpCounters)); |
| 183 |
} |