xref: /aosp_15_r20/external/zstd/lib/compress/hist.c (revision 01826a4963a0d8a59bc3812d29bdf0fb76416722)
1*01826a49SYabin Cui /* ******************************************************************
2*01826a49SYabin Cui  * hist : Histogram functions
3*01826a49SYabin Cui  * part of Finite State Entropy project
4*01826a49SYabin Cui  * Copyright (c) Meta Platforms, Inc. and affiliates.
5*01826a49SYabin Cui  *
6*01826a49SYabin Cui  *  You can contact the author at :
7*01826a49SYabin Cui  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
8*01826a49SYabin Cui  *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
9*01826a49SYabin Cui  *
10*01826a49SYabin Cui  * This source code is licensed under both the BSD-style license (found in the
11*01826a49SYabin Cui  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
12*01826a49SYabin Cui  * in the COPYING file in the root directory of this source tree).
13*01826a49SYabin Cui  * You may select, at your option, one of the above-listed licenses.
14*01826a49SYabin Cui ****************************************************************** */
15*01826a49SYabin Cui 
16*01826a49SYabin Cui /* --- dependencies --- */
17*01826a49SYabin Cui #include "../common/mem.h"             /* U32, BYTE, etc. */
18*01826a49SYabin Cui #include "../common/debug.h"           /* assert, DEBUGLOG */
19*01826a49SYabin Cui #include "../common/error_private.h"   /* ERROR */
20*01826a49SYabin Cui #include "hist.h"
21*01826a49SYabin Cui 
22*01826a49SYabin Cui 
23*01826a49SYabin Cui /* --- Error management --- */
HIST_isError(size_t code)24*01826a49SYabin Cui unsigned HIST_isError(size_t code) { return ERR_isError(code); }
25*01826a49SYabin Cui 
26*01826a49SYabin Cui /*-**************************************************************
27*01826a49SYabin Cui  *  Histogram functions
28*01826a49SYabin Cui  ****************************************************************/
HIST_count_simple(unsigned * count,unsigned * maxSymbolValuePtr,const void * src,size_t srcSize)29*01826a49SYabin Cui unsigned HIST_count_simple(unsigned* count, unsigned* maxSymbolValuePtr,
30*01826a49SYabin Cui                            const void* src, size_t srcSize)
31*01826a49SYabin Cui {
32*01826a49SYabin Cui     const BYTE* ip = (const BYTE*)src;
33*01826a49SYabin Cui     const BYTE* const end = ip + srcSize;
34*01826a49SYabin Cui     unsigned maxSymbolValue = *maxSymbolValuePtr;
35*01826a49SYabin Cui     unsigned largestCount=0;
36*01826a49SYabin Cui 
37*01826a49SYabin Cui     ZSTD_memset(count, 0, (maxSymbolValue+1) * sizeof(*count));
38*01826a49SYabin Cui     if (srcSize==0) { *maxSymbolValuePtr = 0; return 0; }
39*01826a49SYabin Cui 
40*01826a49SYabin Cui     while (ip<end) {
41*01826a49SYabin Cui         assert(*ip <= maxSymbolValue);
42*01826a49SYabin Cui         count[*ip++]++;
43*01826a49SYabin Cui     }
44*01826a49SYabin Cui 
45*01826a49SYabin Cui     while (!count[maxSymbolValue]) maxSymbolValue--;
46*01826a49SYabin Cui     *maxSymbolValuePtr = maxSymbolValue;
47*01826a49SYabin Cui 
48*01826a49SYabin Cui     {   U32 s;
49*01826a49SYabin Cui         for (s=0; s<=maxSymbolValue; s++)
50*01826a49SYabin Cui             if (count[s] > largestCount) largestCount = count[s];
51*01826a49SYabin Cui     }
52*01826a49SYabin Cui 
53*01826a49SYabin Cui     return largestCount;
54*01826a49SYabin Cui }
55*01826a49SYabin Cui 
56*01826a49SYabin Cui typedef enum { trustInput, checkMaxSymbolValue } HIST_checkInput_e;
57*01826a49SYabin Cui 
58*01826a49SYabin Cui /* HIST_count_parallel_wksp() :
59*01826a49SYabin Cui  * store histogram into 4 intermediate tables, recombined at the end.
60*01826a49SYabin Cui  * this design makes better use of OoO cpus,
61*01826a49SYabin Cui  * and is noticeably faster when some values are heavily repeated.
62*01826a49SYabin Cui  * But it needs some additional workspace for intermediate tables.
63*01826a49SYabin Cui  * `workSpace` must be a U32 table of size >= HIST_WKSP_SIZE_U32.
64*01826a49SYabin Cui  * @return : largest histogram frequency,
65*01826a49SYabin Cui  *           or an error code (notably when histogram's alphabet is larger than *maxSymbolValuePtr) */
HIST_count_parallel_wksp(unsigned * count,unsigned * maxSymbolValuePtr,const void * source,size_t sourceSize,HIST_checkInput_e check,U32 * const workSpace)66*01826a49SYabin Cui static size_t HIST_count_parallel_wksp(
67*01826a49SYabin Cui                                 unsigned* count, unsigned* maxSymbolValuePtr,
68*01826a49SYabin Cui                                 const void* source, size_t sourceSize,
69*01826a49SYabin Cui                                 HIST_checkInput_e check,
70*01826a49SYabin Cui                                 U32* const workSpace)
71*01826a49SYabin Cui {
72*01826a49SYabin Cui     const BYTE* ip = (const BYTE*)source;
73*01826a49SYabin Cui     const BYTE* const iend = ip+sourceSize;
74*01826a49SYabin Cui     size_t const countSize = (*maxSymbolValuePtr + 1) * sizeof(*count);
75*01826a49SYabin Cui     unsigned max=0;
76*01826a49SYabin Cui     U32* const Counting1 = workSpace;
77*01826a49SYabin Cui     U32* const Counting2 = Counting1 + 256;
78*01826a49SYabin Cui     U32* const Counting3 = Counting2 + 256;
79*01826a49SYabin Cui     U32* const Counting4 = Counting3 + 256;
80*01826a49SYabin Cui 
81*01826a49SYabin Cui     /* safety checks */
82*01826a49SYabin Cui     assert(*maxSymbolValuePtr <= 255);
83*01826a49SYabin Cui     if (!sourceSize) {
84*01826a49SYabin Cui         ZSTD_memset(count, 0, countSize);
85*01826a49SYabin Cui         *maxSymbolValuePtr = 0;
86*01826a49SYabin Cui         return 0;
87*01826a49SYabin Cui     }
88*01826a49SYabin Cui     ZSTD_memset(workSpace, 0, 4*256*sizeof(unsigned));
89*01826a49SYabin Cui 
90*01826a49SYabin Cui     /* by stripes of 16 bytes */
91*01826a49SYabin Cui     {   U32 cached = MEM_read32(ip); ip += 4;
92*01826a49SYabin Cui         while (ip < iend-15) {
93*01826a49SYabin Cui             U32 c = cached; cached = MEM_read32(ip); ip += 4;
94*01826a49SYabin Cui             Counting1[(BYTE) c     ]++;
95*01826a49SYabin Cui             Counting2[(BYTE)(c>>8) ]++;
96*01826a49SYabin Cui             Counting3[(BYTE)(c>>16)]++;
97*01826a49SYabin Cui             Counting4[       c>>24 ]++;
98*01826a49SYabin Cui             c = cached; cached = MEM_read32(ip); ip += 4;
99*01826a49SYabin Cui             Counting1[(BYTE) c     ]++;
100*01826a49SYabin Cui             Counting2[(BYTE)(c>>8) ]++;
101*01826a49SYabin Cui             Counting3[(BYTE)(c>>16)]++;
102*01826a49SYabin Cui             Counting4[       c>>24 ]++;
103*01826a49SYabin Cui             c = cached; cached = MEM_read32(ip); ip += 4;
104*01826a49SYabin Cui             Counting1[(BYTE) c     ]++;
105*01826a49SYabin Cui             Counting2[(BYTE)(c>>8) ]++;
106*01826a49SYabin Cui             Counting3[(BYTE)(c>>16)]++;
107*01826a49SYabin Cui             Counting4[       c>>24 ]++;
108*01826a49SYabin Cui             c = cached; cached = MEM_read32(ip); ip += 4;
109*01826a49SYabin Cui             Counting1[(BYTE) c     ]++;
110*01826a49SYabin Cui             Counting2[(BYTE)(c>>8) ]++;
111*01826a49SYabin Cui             Counting3[(BYTE)(c>>16)]++;
112*01826a49SYabin Cui             Counting4[       c>>24 ]++;
113*01826a49SYabin Cui         }
114*01826a49SYabin Cui         ip-=4;
115*01826a49SYabin Cui     }
116*01826a49SYabin Cui 
117*01826a49SYabin Cui     /* finish last symbols */
118*01826a49SYabin Cui     while (ip<iend) Counting1[*ip++]++;
119*01826a49SYabin Cui 
120*01826a49SYabin Cui     {   U32 s;
121*01826a49SYabin Cui         for (s=0; s<256; s++) {
122*01826a49SYabin Cui             Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s];
123*01826a49SYabin Cui             if (Counting1[s] > max) max = Counting1[s];
124*01826a49SYabin Cui     }   }
125*01826a49SYabin Cui 
126*01826a49SYabin Cui     {   unsigned maxSymbolValue = 255;
127*01826a49SYabin Cui         while (!Counting1[maxSymbolValue]) maxSymbolValue--;
128*01826a49SYabin Cui         if (check && maxSymbolValue > *maxSymbolValuePtr) return ERROR(maxSymbolValue_tooSmall);
129*01826a49SYabin Cui         *maxSymbolValuePtr = maxSymbolValue;
130*01826a49SYabin Cui         ZSTD_memmove(count, Counting1, countSize);   /* in case count & Counting1 are overlapping */
131*01826a49SYabin Cui     }
132*01826a49SYabin Cui     return (size_t)max;
133*01826a49SYabin Cui }
134*01826a49SYabin Cui 
135*01826a49SYabin Cui /* HIST_countFast_wksp() :
136*01826a49SYabin Cui  * Same as HIST_countFast(), but using an externally provided scratch buffer.
137*01826a49SYabin Cui  * `workSpace` is a writable buffer which must be 4-bytes aligned,
138*01826a49SYabin Cui  * `workSpaceSize` must be >= HIST_WKSP_SIZE
139*01826a49SYabin Cui  */
HIST_countFast_wksp(unsigned * count,unsigned * maxSymbolValuePtr,const void * source,size_t sourceSize,void * workSpace,size_t workSpaceSize)140*01826a49SYabin Cui size_t HIST_countFast_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
141*01826a49SYabin Cui                           const void* source, size_t sourceSize,
142*01826a49SYabin Cui                           void* workSpace, size_t workSpaceSize)
143*01826a49SYabin Cui {
144*01826a49SYabin Cui     if (sourceSize < 1500) /* heuristic threshold */
145*01826a49SYabin Cui         return HIST_count_simple(count, maxSymbolValuePtr, source, sourceSize);
146*01826a49SYabin Cui     if ((size_t)workSpace & 3) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
147*01826a49SYabin Cui     if (workSpaceSize < HIST_WKSP_SIZE) return ERROR(workSpace_tooSmall);
148*01826a49SYabin Cui     return HIST_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, trustInput, (U32*)workSpace);
149*01826a49SYabin Cui }
150*01826a49SYabin Cui 
151*01826a49SYabin Cui /* HIST_count_wksp() :
152*01826a49SYabin Cui  * Same as HIST_count(), but using an externally provided scratch buffer.
153*01826a49SYabin Cui  * `workSpace` size must be table of >= HIST_WKSP_SIZE_U32 unsigned */
HIST_count_wksp(unsigned * count,unsigned * maxSymbolValuePtr,const void * source,size_t sourceSize,void * workSpace,size_t workSpaceSize)154*01826a49SYabin Cui size_t HIST_count_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
155*01826a49SYabin Cui                        const void* source, size_t sourceSize,
156*01826a49SYabin Cui                        void* workSpace, size_t workSpaceSize)
157*01826a49SYabin Cui {
158*01826a49SYabin Cui     if ((size_t)workSpace & 3) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
159*01826a49SYabin Cui     if (workSpaceSize < HIST_WKSP_SIZE) return ERROR(workSpace_tooSmall);
160*01826a49SYabin Cui     if (*maxSymbolValuePtr < 255)
161*01826a49SYabin Cui         return HIST_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, checkMaxSymbolValue, (U32*)workSpace);
162*01826a49SYabin Cui     *maxSymbolValuePtr = 255;
163*01826a49SYabin Cui     return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, workSpace, workSpaceSize);
164*01826a49SYabin Cui }
165*01826a49SYabin Cui 
166*01826a49SYabin Cui #ifndef ZSTD_NO_UNUSED_FUNCTIONS
167*01826a49SYabin Cui /* fast variant (unsafe : won't check if src contains values beyond count[] limit) */
HIST_countFast(unsigned * count,unsigned * maxSymbolValuePtr,const void * source,size_t sourceSize)168*01826a49SYabin Cui size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr,
169*01826a49SYabin Cui                      const void* source, size_t sourceSize)
170*01826a49SYabin Cui {
171*01826a49SYabin Cui     unsigned tmpCounters[HIST_WKSP_SIZE_U32];
172*01826a49SYabin Cui     return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters, sizeof(tmpCounters));
173*01826a49SYabin Cui }
174*01826a49SYabin Cui 
HIST_count(unsigned * count,unsigned * maxSymbolValuePtr,const void * src,size_t srcSize)175*01826a49SYabin Cui size_t HIST_count(unsigned* count, unsigned* maxSymbolValuePtr,
176*01826a49SYabin Cui                  const void* src, size_t srcSize)
177*01826a49SYabin Cui {
178*01826a49SYabin Cui     unsigned tmpCounters[HIST_WKSP_SIZE_U32];
179*01826a49SYabin Cui     return HIST_count_wksp(count, maxSymbolValuePtr, src, srcSize, tmpCounters, sizeof(tmpCounters));
180*01826a49SYabin Cui }
181*01826a49SYabin Cui #endif
182