source: tests/zombies/string-perf/prog-allocn.cfa@ e0332dd3

Last change on this file since e0332dd3 was 08ed947, checked in by Michael Brooks <mlbrooks@…>, 4 years ago

Roll up of string changes for performance testing/improvement, and a couple API features supporting them.

String API changes:
Defining a tuning knob to control the heap growth policy (relapaces former 10% hardcode, downgraded to a default)
Implementing findFrom (allowing find-non-first); leaving find as find-first.

String implementation perf improvements:
Calling C-malloc directly instead of via CFA-alloc.
Replacings loops that copy with memmove calls.
Replacings loops that search for a value with memchr calls.

String perf testing realized:
Makefile supporting several prog-*.cfa, chosen by OPERATION value (implies prog.cfa changes to support the adjusted protocol)
Adjusting the starter/accumulater declarations in PEQ and PTA to behave consistently in cfa v cpp.
Adding tests: allocation, find, normalize, pass-by-val, pass-by-x.
Adding helper shell scripts for: generating flame graphs, collecting/crunching allocation stats using Mubeen's malloc wrappers

  • Property mode set to 100644
File size: 5.8 KB
RevLine 
[08ed947]1#if defined IMPL_STL_NA_NA
2 #define IMPL_STL
3#endif
4
5#if defined IMPL_BUHR94_NA_NA
6 #define IMPL_BUHR94
7#endif
8
9#if defined IMPL_STL
10 #include <string>
11 #include <iostream>
12 #include <cstdio>
13 using namespace std;
14 #define IMPL_CXX
15
16#elif defined IMPL_CFA_HL_SHARE
17 #define IMPL_CFA_HL
18 #define IMPL_CFA
19
20#elif defined IMPL_CFA_LL_SHARE
21 #define IMPL_CFA_LL
22 #define IMPL_CFA
23
24#elif defined IMPL_CFA_HL_NOSHARE
25 #define IMPL_CFA_HL
26 #define CFA_NOSHARE
27 #define IMPL_CFA
28
29#elif defined IMPL_CFA_LL_NOSHARE
30 #define IMPL_CFA_LL
31 #define CFA_NOSHARE
32 #define IMPL_CFA
33
34#elif defined IMPL_BUHR94
35 #include <iostream>
36 #include <cstdio>
37 #include "/u0/mlbrooks/usys1/sm/string/StringSharing/src/string.h"
38 #define IMPL_CXX
39
40#else
41 #error Bad IMPL_
42#endif
43
44
45#if defined IMPL_CFA_HL
46 #include <string.hfa>
47 extern void TUNING_set_string_heap_liveness_threshold(double); // in string_res.hfa
48#elif defined IMPL_CFA_LL
49 #include <string_res.hfa>
50#endif
51
52#if defined CFA_NOSHARE
53 #include <string_sharectx.hfa>
54 #define STRING_SHARING_CONTROL \
55 string_sharectx c = { NO_SHARING };
56#else
57 #define STRING_SHARING_CONTROL
58#endif
59
60#if defined IMPL_CFA
61 #include <math.hfa>
62 extern "C" {
63 void malloc_stats( void );
64 }
65#elif defined IMPL_CXX
66 #include <algorithm>
67 using std::min;
68 #include <malloc.h>
69#endif
70
71#include <time.h>
72#include <stdlib.h> // atoi
73#include <string.h> // strlen, only during setup
74
75#if defined IMPL_STL || defined IMPL_BUHR94
76 #define PRINT(s) std::cout << s << std::endl
77#elif defined IMPL_CFA_HL || defined IMPL_CFA_LL
78 #define PRINT(s) sout | s;
79#else
80 #error Unhandled print case
81#endif
82
83#if defined IMPL_CFA_LL
84 #define STRING_T string_res
85 #define ASSIGN_CHAR(str, idx, val) assignAt(str, idx, val)
86#else
87 #define STRING_T string
88 #define ASSIGN_CHAR(str, idx, val) str[idx] = val
89#endif
90
91double meanLen(int N, char ** strings) {
92 int totalLen = 0;
93 for (int i = 0 ; i < N; i ++) {
94 totalLen += strlen(strings[i]);
95 }
96 return (double)totalLen / (double)N;
97}
98
99volatile int checkthis = 0;
100#define MAYBE( op ) if (checkthis) { op; }
101
102int corpuslen = 0;
103char ** corpus = (char**) 0;
104size_t corpus_next_pos = 0;
105
106double repsPerLevel;
107double repBalance = 0.0000001;
108
109clock_t start, endTarget, end_actual;
110size_t allocationCountTarget = 0;
111
112size_t allocationCountActual = 0;
113//size_t allocationBytesActual = 0;
114
115void helper( int depth ) {
116
117 if (depth == 0) return;
118
119 corpus_next_pos += 1;
120 corpus_next_pos %= corpuslen;
121
122 STRING_T q = corpus[corpus_next_pos];
123// ASSIGN_CHAR(q, 0, '@'); // Turns out my implementation is slow at this step. A separate test could work it. It's inessential to the allocation test, given the assumption that both string reps allocate eagerly in their heaps from a constant. In the STL, that assumption is upheld by my observation that commenting out this line didn't speed it up. In CFA-share, I know it to be true of the implementation.
124
125 allocationCountActual += 1;
126// allocationBytesActual += q`len;
127
128// if (depth > 0) {
129
130 repBalance += repsPerLevel;
131 int curRepLimit = repBalance;
132 repBalance -= curRepLimit;
133
134 for ( int i = 0 ; i < curRepLimit; i++ ) {
135 if ((allocationCountActual+1) % allocationCountTarget == 0 && clock() > endTarget) return;
136 helper(depth-1);
137 }
138// }
139
140 MAYBE(PRINT(q));
141}
142
143int main( int argc, char ** argv ) {
144
145 STRING_SHARING_CONTROL
146
147
148 const char * usage_args[] = {"Depth RepsPerLevel ExpansionThreshold ExecTimeSecs Corpus...",
149 "Depth RepsPerLevel ExpansionThreshold -w WorkAllocCount Corpus..."};
150 const int static_arg_posns = 5;
151 int used_arg_posns = static_arg_posns;
152
153 int launchDepth = -1;
154 double expansionThreshold = -1.0;
155 int execTimeSecs = -1;
156
157 switch (min(argc, static_arg_posns)) {
158 case 5: if ( strcmp(argv[4], "-w") == 0 ) {
159 used_arg_posns ++;
160 allocationCountTarget = atoi(argv[5]);
161 } else {
162 execTimeSecs = atoi(argv[4]);
163 }
164 case 4: expansionThreshold = atof(argv[3]);
165 case 3: repsPerLevel = atof(argv[2]);
166 case 2: launchDepth = atoi(argv[1]);
167 }
168
169 corpuslen = argc - used_arg_posns;
170 corpus = argv + used_arg_posns;
171
172 if (launchDepth < 1 || repsPerLevel < 1.0 || (execTimeSecs < 1 && allocationCountTarget < 1) || corpuslen < 1 ||
173 (expansionThreshold != -1.0 && (expansionThreshold <= 0.0 || expansionThreshold >= 1.0))) {
174 for (int u = 0; u < sizeof(usage_args) / sizeof(*usage_args); u++) {
175 printf("usage: %s %s\n", argv[0], usage_args[u]);
176 }
177 printf("output:\nxxx,corpusItemCount,corpusMeanLenChars,allocationCountActual,execTimeActualSec,topIters\n");
178 exit(1);
179 }
180
181 if (expansionThreshold != -1.0 ) {
182 #if defined IMPL_CFA
183 TUNING_set_string_heap_liveness_threshold(expansionThreshold);
184 #else
185 printf("cannot set expansion threshold on non-CFA implementation");
186 exit(1);
187 #endif
188 }
189
190 double meanCorpusLen = meanLen(corpuslen, corpus);
191
192 // time driven experiment: re-check time every 10000 allocations
193 if (execTimeSecs > 0) allocationCountTarget = 10000;
194
195 start = clock();
196 endTarget = start + CLOCKS_PER_SEC * max(0, execTimeSecs);
197
198 size_t top_iters = 0;
199
200 for(;;) {
201 #if defined OP_PALL
202 helper( launchDepth );
203 #else
204 #error Bad OP_
205 #endif
206
207 top_iters++;
208
209 if ((allocationCountActual+1) % allocationCountTarget == 0 && clock() > endTarget) break;
210 }
211 end_actual = clock();
212 double elapsed = ((double) (end_actual - start)) / CLOCKS_PER_SEC;
213 printf("xxx,%d,%f,%ld,%f,%ld\n", corpuslen, meanCorpusLen, allocationCountActual, elapsed, top_iters);
214
215 // malloc_stats();
216
217 return 0;
218}
Note: See TracBrowser for help on using the repository browser.