1 /*
2 ** Copyright 2010 The Android Open Source Project
3 **
4 ** Licensed under the Apache License, Version 2.0 (the "License");
5 ** you may not use this file except in compliance with the License.
6 ** You may obtain a copy of the License at
7 **
8 **     http://www.apache.org/licenses/LICENSE-2.0
9 **
10 ** Unless required by applicable law or agreed to in writing, software
11 ** distributed under the License is distributed on an "AS IS" BASIS,
12 ** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 ** See the License for the specific language governing permissions and
14 ** limitations under the License.
15 */
16 
17 /*
18  * Micro-benchmarking of sleep/cpu speed/memcpy/memset/memory reads/strcmp.
19  */
20 
21 #include <stdio.h>
22 #include <stdlib.h>
23 #include <string.h>
24 #include <ctype.h>
25 #include <math.h>
26 #include <sched.h>
27 #include <sys/resource.h>
28 #include <time.h>
29 #include <unistd.h>
30 
31 // The default size of data that will be manipulated in each iteration of
32 // a memory benchmark. Can be modified with the --data_size option.
33 #define DEFAULT_DATA_SIZE       1000000000
34 
35 // The amount of memory allocated for the cold benchmarks to use.
36 #define DEFAULT_COLD_DATA_SIZE  128*1024*1024
37 
38 // The default size of the stride between each buffer for cold benchmarks.
39 #define DEFAULT_COLD_STRIDE_SIZE  4096
40 
41 // Number of nanoseconds in a second.
42 #define NS_PER_SEC              1000000000
43 
44 // The maximum number of arguments that a benchmark will accept.
45 #define MAX_ARGS    2
46 
47 // Default memory alignment of malloc.
48 #define DEFAULT_MALLOC_MEMORY_ALIGNMENT   8
49 
50 // Contains information about benchmark options.
51 typedef struct {
52     bool print_average;
53     bool print_each_iter;
54 
55     int dst_align;
56     int dst_or_mask;
57     int src_align;
58     int src_or_mask;
59 
60     int cpu_to_lock;
61 
62     int data_size;
63     int dst_str_size;
64     int cold_data_size;
65     int cold_stride_size;
66 
67     int args[MAX_ARGS];
68     int num_args;
69 } command_data_t;
70 
71 typedef void *(*void_func_t)();
72 typedef void *(*memcpy_func_t)(void *, const void *, size_t);
73 typedef void *(*memset_func_t)(void *, int, size_t);
74 typedef int (*strcmp_func_t)(const char *, const char *);
75 typedef char *(*str_func_t)(char *, const char *);
76 typedef size_t (*strlen_func_t)(const char *);
77 
78 // Struct that contains a mapping of benchmark name to benchmark function.
79 typedef struct {
80     const char *name;
81     int (*ptr)(const char *, const command_data_t &, void_func_t func);
82     void_func_t func;
83 } function_t;
84 
85 // Get the current time in nanoseconds.
nanoTime()86 uint64_t nanoTime() {
87   struct timespec t;
88 
89   t.tv_sec = t.tv_nsec = 0;
90   clock_gettime(CLOCK_MONOTONIC, &t);
91   return static_cast<uint64_t>(t.tv_sec) * NS_PER_SEC + t.tv_nsec;
92 }
93 
94 // Static analyzer warns about potential memory leak of orig_ptr
95 // in getAlignedMemory. That is true and the callers in this program
96 // do not free orig_ptr. But, we don't care about that in this
97 // going-obsolete test program. So, here is a hack to trick the
98 // static analyzer.
99 static void *saved_orig_ptr;
100 
101 // Allocate memory with a specific alignment and return that pointer.
102 // This function assumes an alignment value that is a power of 2.
103 // If the alignment is 0, then use the pointer returned by malloc.
getAlignedMemory(uint8_t * orig_ptr,int alignment,int or_mask)104 uint8_t *getAlignedMemory(uint8_t *orig_ptr, int alignment, int or_mask) {
105   uint64_t ptr = reinterpret_cast<uint64_t>(orig_ptr);
106   saved_orig_ptr = orig_ptr;
107   if (alignment > 0) {
108       // When setting the alignment, set it to exactly the alignment chosen.
109       // The pointer returned will be guaranteed not to be aligned to anything
110       // more than that.
111       ptr += alignment - (ptr & (alignment - 1));
112       ptr |= alignment | or_mask;
113   }
114 
115   return reinterpret_cast<uint8_t*>(ptr);
116 }
117 
118 // Allocate memory with a specific alignment and return that pointer.
119 // This function assumes an alignment value that is a power of 2.
120 // If the alignment is 0, then use the pointer returned by malloc.
allocateAlignedMemory(size_t size,int alignment,int or_mask)121 uint8_t *allocateAlignedMemory(size_t size, int alignment, int or_mask) {
122   uint64_t ptr = reinterpret_cast<uint64_t>(malloc(size + 3 * alignment));
123   if (!ptr)
124       return NULL;
125   return getAlignedMemory((uint8_t*)ptr, alignment, or_mask);
126 }
127 
initString(uint8_t * buf,size_t size)128 void initString(uint8_t *buf, size_t size) {
129     for (size_t i = 0; i < size - 1; i++) {
130         buf[i] = static_cast<char>(32 + (i % 96));
131     }
132     buf[size-1] = '\0';
133 }
134 
computeAverage(uint64_t time_ns,size_t size,size_t copies)135 static inline double computeAverage(uint64_t time_ns, size_t size, size_t copies) {
136     return ((size/1024.0) * copies) / ((double)time_ns/NS_PER_SEC);
137 }
138 
computeRunningAvg(double avg,double running_avg,size_t cur_idx)139 static inline double computeRunningAvg(double avg, double running_avg, size_t cur_idx) {
140     return (running_avg / (cur_idx + 1)) * cur_idx + (avg / (cur_idx + 1));
141 }
142 
computeRunningSquareAvg(double avg,double square_avg,size_t cur_idx)143 static inline double computeRunningSquareAvg(double avg, double square_avg, size_t cur_idx) {
144     return (square_avg / (cur_idx + 1)) * cur_idx + (avg / (cur_idx + 1)) * avg;
145 }
146 
computeStdDev(double square_avg,double running_avg)147 static inline double computeStdDev(double square_avg, double running_avg) {
148     return sqrt(square_avg - running_avg * running_avg);
149 }
150 
printIter(uint64_t time_ns,const char * name,size_t size,size_t copies,double avg)151 static inline void printIter(uint64_t time_ns, const char *name, size_t size, size_t copies, double avg) {
152     printf("%s %zux%zu bytes took %.06f seconds (%f MB/s)\n",
153            name, copies, size, (double)time_ns/NS_PER_SEC, avg/1024.0);
154 }
155 
printSummary(uint64_t,const char * name,size_t size,size_t copies,double running_avg,double std_dev,double min,double max)156 static inline void printSummary(uint64_t /*time_ns*/, const char *name, size_t size, size_t copies, double running_avg, double std_dev, double min, double max) {
157     printf("  %s %zux%zu bytes average %.2f MB/s std dev %.4f min %.2f MB/s max %.2f MB/s\n",
158            name, copies, size, running_avg/1024.0, std_dev/1024.0, min/1024.0,
159            max/1024.0);
160 }
161 
162 // For the cold benchmarks, a large buffer will be created which
163 // contains many "size" buffers. This function will figure out the increment
164 // needed between each buffer so that each one is aligned to "alignment".
getAlignmentIncrement(size_t size,int alignment)165 int getAlignmentIncrement(size_t size, int alignment) {
166     if (alignment == 0) {
167         alignment = DEFAULT_MALLOC_MEMORY_ALIGNMENT;
168     }
169     alignment *= 2;
170     return size + alignment - (size % alignment);
171 }
172 
getColdBuffer(int num_buffers,size_t incr,int alignment,int or_mask)173 uint8_t *getColdBuffer(int num_buffers, size_t incr, int alignment, int or_mask) {
174     uint8_t *buffers = reinterpret_cast<uint8_t*>(malloc(num_buffers * incr + 3 * alignment));
175     if (!buffers) {
176         return NULL;
177     }
178     return getAlignedMemory(buffers, alignment, or_mask);
179 }
180 
computeColdAverage(uint64_t time_ns,size_t size,size_t copies,size_t num_buffers)181 static inline double computeColdAverage(uint64_t time_ns, size_t size, size_t copies, size_t num_buffers) {
182     return ((size/1024.0) * copies * num_buffers) / ((double)time_ns/NS_PER_SEC);
183 }
184 
printColdIter(uint64_t time_ns,const char * name,size_t size,size_t copies,size_t num_buffers,double avg)185 static void inline printColdIter(uint64_t time_ns, const char *name, size_t size, size_t copies, size_t num_buffers, double avg) {
186     printf("%s %zux%zux%zu bytes took %.06f seconds (%f MB/s)\n",
187            name, copies, num_buffers, size, (double)time_ns/NS_PER_SEC, avg/1024.0);
188 }
189 
printColdSummary(uint64_t,const char * name,size_t size,size_t copies,size_t num_buffers,double running_avg,double square_avg,double min,double max)190 static void inline printColdSummary(
191         uint64_t /*time_ns*/, const char *name, size_t size, size_t copies, size_t num_buffers,
192         double running_avg, double square_avg, double min, double max) {
193     printf("  %s %zux%zux%zu bytes average %.2f MB/s std dev %.4f min %.2f MB/s max %.2f MB/s\n",
194            name, copies, num_buffers, size, running_avg/1024.0,
195            computeStdDev(running_avg, square_avg)/1024.0, min/1024.0, max/1024.0);
196 }
197 
198 #define MAINLOOP(cmd_data, BENCH, COMPUTE_AVG, PRINT_ITER, PRINT_AVG) \
199     uint64_t time_ns;                                                 \
200     int iters = cmd_data.args[1];                                     \
201     bool print_average = cmd_data.print_average;                      \
202     bool print_each_iter = cmd_data.print_each_iter;                  \
203     double min = 0.0, max = 0.0, running_avg = 0.0, square_avg = 0.0; \
204     double avg;                                                       \
205     for (int i = 0; iters == -1 || i < iters; i++) {                  \
206         time_ns = nanoTime();                                         \
207         BENCH;                                                        \
208         time_ns = nanoTime() - time_ns;                               \
209         avg = COMPUTE_AVG;                                            \
210         if (print_average) {                                          \
211             running_avg = computeRunningAvg(avg, running_avg, i);     \
212             square_avg = computeRunningSquareAvg(avg, square_avg, i); \
213             if (min == 0.0 || avg < min) {                            \
214                 min = avg;                                            \
215             }                                                         \
216             if (avg > max) {                                          \
217                 max = avg;                                            \
218             }                                                         \
219         }                                                             \
220         if (print_each_iter) {                                        \
221             PRINT_ITER;                                               \
222         }                                                             \
223     }                                                                 \
224     if (print_average) {                                              \
225         PRINT_AVG;                                                    \
226     }
227 
228 #define MAINLOOP_DATA(name, cmd_data, size, BENCH)                    \
229     size_t copies = cmd_data.data_size/size;                          \
230     size_t j;                                                         \
231     MAINLOOP(cmd_data,                                                \
232              for (j = 0; j < copies; j++) {                           \
233                  BENCH;                                               \
234              },                                                       \
235              computeAverage(time_ns, size, copies),                   \
236              printIter(time_ns, name, size, copies, avg),             \
237              double std_dev = computeStdDev(square_avg, running_avg); \
238              printSummary(time_ns, name, size, copies, running_avg,   \
239                           std_dev, min, max));
240 
241 #define MAINLOOP_COLD(name, cmd_data, size, num_incrs, BENCH)                 \
242     size_t num_strides = num_buffers / num_incrs;                             \
243     if ((num_buffers % num_incrs) != 0) {                                     \
244         num_strides--;                                                        \
245     }                                                                         \
246     size_t copies = 1;                                                        \
247     num_buffers = num_incrs * num_strides;                                    \
248     if (num_buffers * size < static_cast<size_t>(cmd_data.data_size)) {       \
249         copies = cmd_data.data_size / (num_buffers * size);                   \
250     }                                                                         \
251     if (num_strides == 0) {                                                   \
252         printf("%s: Chosen options lead to no copies, aborting.\n", name);    \
253         return -1;                                                            \
254     }                                                                         \
255     size_t j, k;                                                              \
256     MAINLOOP(cmd_data,                                                        \
257              for (j = 0; j < copies; j++) {                                   \
258                  for (k = 0; k < num_incrs; k++) {                            \
259                      BENCH;                                                   \
260                 }                                                             \
261             },                                                                \
262             computeColdAverage(time_ns, size, copies, num_buffers),           \
263             printColdIter(time_ns, name, size, copies, num_buffers, avg),     \
264             printColdSummary(time_ns, name, size, copies, num_buffers,        \
265                              running_avg, square_avg, min, max));
266 
267 // This version of the macro creates a single buffer of the given size and
268 // alignment. The variable "buf" will be a pointer to the buffer and should
269 // be used by the BENCH code.
270 // INIT - Any specialized code needed to initialize the data. This will only
271 //        be executed once.
272 // BENCH - The actual code to benchmark and is timed.
273 #define BENCH_ONE_BUF(name, cmd_data, INIT, BENCH)                            \
274     size_t size = cmd_data.args[0]; \
275     uint8_t *buf = allocateAlignedMemory(size, cmd_data.dst_align, cmd_data.dst_or_mask); \
276     if (!buf)                                                                 \
277         return -1;                                                            \
278     INIT;                                                                     \
279     MAINLOOP_DATA(name, cmd_data, size, BENCH);
280 
281 // This version of the macro creates two buffers of the given sizes and
282 // alignments. The variables "buf1" and "buf2" will be pointers to the
283 // buffers and should be used by the BENCH code.
284 // INIT - Any specialized code needed to initialize the data. This will only
285 //        be executed once.
286 // BENCH - The actual code to benchmark and is timed.
287 #define BENCH_TWO_BUFS(name, cmd_data, INIT, BENCH)                           \
288     size_t size = cmd_data.args[0];                                           \
289     uint8_t *buf1 = allocateAlignedMemory(size, cmd_data.src_align, cmd_data.src_or_mask); \
290     if (!buf1)                                                                \
291         return -1;                                                            \
292     size_t total_size = size;                                                 \
293     if (cmd_data.dst_str_size > 0)                                            \
294         total_size += cmd_data.dst_str_size;                                  \
295     uint8_t *buf2 = allocateAlignedMemory(total_size, cmd_data.dst_align, cmd_data.dst_or_mask); \
296     if (!buf2)                                                                \
297         return -1;                                                            \
298     INIT;                                                                     \
299     MAINLOOP_DATA(name, cmd_data, size, BENCH);
300 
301 // This version of the macro attempts to benchmark code when the data
302 // being manipulated is not in the cache, thus the cache is cold. It does
303 // this by creating a single large buffer that is designed to be larger than
304 // the largest cache in the system. The variable "buf" will be one slice
305 // of the buffer that the BENCH code should use that is of the correct size
306 // and alignment. In order to avoid any algorithms that prefetch past the end
307 // of their "buf" and into the next sequential buffer, the code strides
308 // through the buffer. Specifically, as "buf" values are iterated in BENCH
309 // code, the end of "buf" is guaranteed to be at least "stride_size" away
310 // from the next "buf".
311 // INIT - Any specialized code needed to initialize the data. This will only
312 //        be executed once.
313 // BENCH - The actual code to benchmark and is timed.
314 #define COLD_ONE_BUF(name, cmd_data, INIT, BENCH)                             \
315     size_t size = cmd_data.args[0];                                           \
316     size_t incr = getAlignmentIncrement(size, cmd_data.dst_align);            \
317     size_t num_buffers = cmd_data.cold_data_size / incr;                      \
318     size_t buffer_size = num_buffers * incr;                                  \
319     uint8_t *buffer = getColdBuffer(num_buffers, incr, cmd_data.dst_align, cmd_data.dst_or_mask); \
320     if (!buffer)                                                              \
321         return -1;                                                            \
322     size_t num_incrs = cmd_data.cold_stride_size / incr + 1;                  \
323     size_t stride_incr = incr * num_incrs;                                    \
324     uint8_t *buf;                                                             \
325     size_t l;                                                                 \
326     INIT;                                                                     \
327     MAINLOOP_COLD(name, cmd_data, size, num_incrs,                            \
328                   buf = buffer + k * incr;                                    \
329                   for (l = 0; l < num_strides; l++) {                         \
330                       BENCH;                                                  \
331                       buf += stride_incr;                                     \
332                   });
333 
334 // This version of the macro attempts to benchmark code when the data
335 // being manipulated is not in the cache, thus the cache is cold. It does
336 // this by creating two large buffers each of which is designed to be
337 // larger than the largest cache in the system. Two variables "buf1" and
338 // "buf2" will be the two buffers that BENCH code should use. In order
339 // to avoid any algorithms that prefetch past the end of either "buf1"
340 // or "buf2" and into the next sequential buffer, the code strides through
341 // both buffers. Specifically, as "buf1" and "buf2" values are iterated in
342 // BENCH code, the end of "buf1" and "buf2" is guaranteed to be at least
343 // "stride_size" away from the next "buf1" and "buf2".
344 // INIT - Any specialized code needed to initialize the data. This will only
345 //        be executed once.
346 // BENCH - The actual code to benchmark and is timed.
347 #define COLD_TWO_BUFS(name, cmd_data, INIT, BENCH)                            \
348     size_t size = cmd_data.args[0];                                           \
349     size_t buf1_incr = getAlignmentIncrement(size, cmd_data.src_align);       \
350     size_t total_size = size;                                                 \
351     if (cmd_data.dst_str_size > 0)                                            \
352         total_size += cmd_data.dst_str_size;                                  \
353     size_t buf2_incr = getAlignmentIncrement(total_size, cmd_data.dst_align); \
354     size_t max_incr = (buf1_incr > buf2_incr) ? buf1_incr : buf2_incr;        \
355     size_t num_buffers = cmd_data.cold_data_size / max_incr;                  \
356     size_t buffer1_size = num_buffers * buf1_incr;                            \
357     size_t buffer2_size = num_buffers * buf2_incr;                            \
358     uint8_t *buffer1 = getColdBuffer(num_buffers, buf1_incr, cmd_data.src_align, cmd_data.src_or_mask); \
359     if (!buffer1)                                                             \
360         return -1;                                                            \
361     uint8_t *buffer2 = getColdBuffer(num_buffers, buf2_incr, cmd_data.dst_align, cmd_data.dst_or_mask); \
362     if (!buffer2)                                                             \
363         return -1;                                                            \
364     size_t min_incr = (buf1_incr < buf2_incr) ? buf1_incr : buf2_incr;        \
365     size_t num_incrs = cmd_data.cold_stride_size / min_incr + 1;              \
366     size_t buf1_stride_incr = buf1_incr * num_incrs;                          \
367     size_t buf2_stride_incr = buf2_incr * num_incrs;                          \
368     size_t l;                                                                 \
369     uint8_t *buf1;                                                            \
370     uint8_t *buf2;                                                            \
371     INIT;                                                                     \
372     MAINLOOP_COLD(name, cmd_data, size, num_incrs,                            \
373                   buf1 = buffer1 + k * buf1_incr;                             \
374                   buf2 = buffer2 + k * buf2_incr;                             \
375                   for (l = 0; l < num_strides; l++) {                         \
376                       BENCH;                                                  \
377                       buf1 += buf1_stride_incr;                               \
378                       buf2 += buf2_stride_incr;                               \
379                   });
380 
benchmarkSleep(const char *,const command_data_t & cmd_data,void_func_t)381 int benchmarkSleep(const char* /*name*/, const command_data_t &cmd_data, void_func_t /*func*/) {
382     int delay = cmd_data.args[0];
383     MAINLOOP(cmd_data, sleep(delay),
384              (double)time_ns/NS_PER_SEC,
385              printf("sleep(%d) took %.06f seconds\n", delay, avg);,
386              printf("  sleep(%d) average %.06f seconds std dev %f min %.06f seconds max %0.6f seconds\n", \
387                     delay, running_avg, computeStdDev(square_avg, running_avg), \
388                     min, max));
389 
390     return 0;
391 }
392 
benchmarkMemset(const char * name,const command_data_t & cmd_data,void_func_t func)393 int benchmarkMemset(const char *name, const command_data_t &cmd_data, void_func_t func) {
394     memset_func_t memset_func = reinterpret_cast<memset_func_t>(func);
395     BENCH_ONE_BUF(name, cmd_data, ;, memset_func(buf, i, size));
396 
397     return 0;
398 }
399 
benchmarkMemsetCold(const char * name,const command_data_t & cmd_data,void_func_t func)400 int benchmarkMemsetCold(const char *name, const command_data_t &cmd_data, void_func_t func) {
401     memset_func_t memset_func = reinterpret_cast<memset_func_t>(func);
402     COLD_ONE_BUF(name, cmd_data, ;, memset_func(buf, l, size));
403 
404     return 0;
405 }
406 
benchmarkMemcpy(const char * name,const command_data_t & cmd_data,void_func_t func)407 int benchmarkMemcpy(const char *name, const command_data_t &cmd_data, void_func_t func) {
408     memcpy_func_t memcpy_func = reinterpret_cast<memcpy_func_t>(func);
409 
410     BENCH_TWO_BUFS(name, cmd_data,
411                    memset(buf1, 0xff, size); \
412                    memset(buf2, 0, size),
413                    memcpy_func(buf2, buf1, size));
414 
415     return 0;
416 }
417 
benchmarkMemcpyCold(const char * name,const command_data_t & cmd_data,void_func_t func)418 int benchmarkMemcpyCold(const char *name, const command_data_t &cmd_data, void_func_t func) {
419     memcpy_func_t memcpy_func = reinterpret_cast<memcpy_func_t>(func);
420 
421     COLD_TWO_BUFS(name, cmd_data,
422                   memset(buffer1, 0xff, buffer1_size); \
423                   memset(buffer2, 0x0, buffer2_size),
424                   memcpy_func(buf2, buf1, size));
425 
426     return 0;
427 }
428 
benchmarkMemmoveBackwards(const char * name,const command_data_t & cmd_data,void_func_t func)429 int benchmarkMemmoveBackwards(const char *name, const command_data_t &cmd_data, void_func_t func) {
430     memcpy_func_t memmove_func = reinterpret_cast<memcpy_func_t>(func);
431 
432     size_t size = cmd_data.args[0];
433     size_t alloc_size = size * 2 + 3 * cmd_data.dst_align;
434     uint8_t* src = allocateAlignedMemory(size, cmd_data.src_align, cmd_data.src_or_mask);
435     if (!src)
436         return -1;
437     // Force memmove to do a backwards copy by getting a pointer into the source buffer.
438     uint8_t* dst = getAlignedMemory(src+1, cmd_data.dst_align, cmd_data.dst_or_mask);
439     if (!dst)
440         return -1;
441     MAINLOOP_DATA(name, cmd_data, size, memmove_func(dst, src, size));
442     return 0;
443 }
444 
benchmarkMemread(const char * name,const command_data_t & cmd_data,void_func_t)445 int benchmarkMemread(const char *name, const command_data_t &cmd_data, void_func_t /*func*/) {
446     int size = cmd_data.args[0];
447 
448     uint32_t *src = reinterpret_cast<uint32_t*>(malloc(size));
449     if (!src)
450         return -1;
451     memset(src, 0xff, size);
452 
453     // Use volatile so the compiler does not optimize away the reads.
454     volatile int foo;
455     size_t k;
456     MAINLOOP_DATA(name, cmd_data, size,
457                   for (k = 0; k < size/sizeof(uint32_t); k++) foo = src[k]);
458     free(src);
459 
460     return 0;
461 }
462 
benchmarkStrcmp(const char * name,const command_data_t & cmd_data,void_func_t func)463 int benchmarkStrcmp(const char *name, const command_data_t &cmd_data, void_func_t func) {
464     strcmp_func_t strcmp_func = reinterpret_cast<strcmp_func_t>(func);
465 
466     int retval;
467     BENCH_TWO_BUFS(name, cmd_data,
468                    initString(buf1, size); \
469                    initString(buf2, size),
470                    retval = strcmp_func(reinterpret_cast<char*>(buf1), reinterpret_cast<char*>(buf2)); \
471                    if (retval != 0) printf("%s failed, return value %d\n", name, retval));
472 
473     return 0;
474 }
475 
benchmarkStrcmpCold(const char * name,const command_data_t & cmd_data,void_func_t func)476 int benchmarkStrcmpCold(const char *name, const command_data_t &cmd_data, void_func_t func) {
477     strcmp_func_t strcmp_func = reinterpret_cast<strcmp_func_t>(func);
478 
479     int retval;
480     COLD_TWO_BUFS(name, cmd_data,
481                   memset(buffer1, 'a', buffer1_size); \
482                   memset(buffer2, 'a', buffer2_size); \
483                   for (size_t i =0; i < num_buffers; i++) { \
484                       buffer1[size-1+buf1_incr*i] = '\0'; \
485                       buffer2[size-1+buf2_incr*i] = '\0'; \
486                   },
487                   retval = strcmp_func(reinterpret_cast<char*>(buf1), reinterpret_cast<char*>(buf2)); \
488                   if (retval != 0) printf("%s failed, return value %d\n", name, retval));
489 
490     return 0;
491 }
492 
benchmarkStrlen(const char * name,const command_data_t & cmd_data,void_func_t func)493 int benchmarkStrlen(const char *name, const command_data_t &cmd_data, void_func_t func) {
494     size_t real_size;
495     strlen_func_t strlen_func = reinterpret_cast<strlen_func_t>(func);
496     BENCH_ONE_BUF(name, cmd_data,
497                   initString(buf, size),
498                   real_size = strlen_func(reinterpret_cast<char*>(buf)); \
499                   if (real_size + 1 != size) { \
500                       printf("%s failed, expected %zu, got %zu\n", name, size, real_size); \
501                       return -1; \
502                   });
503 
504     return 0;
505 }
506 
benchmarkStrlenCold(const char * name,const command_data_t & cmd_data,void_func_t func)507 int benchmarkStrlenCold(const char *name, const command_data_t &cmd_data, void_func_t func) {
508     strlen_func_t strlen_func = reinterpret_cast<strlen_func_t>(func);
509     size_t real_size;
510     COLD_ONE_BUF(name, cmd_data,
511                  memset(buffer, 'a', buffer_size); \
512                  for (size_t i = 0; i < num_buffers; i++) { \
513                      buffer[size-1+incr*i] = '\0'; \
514                  },
515                  real_size = strlen_func(reinterpret_cast<char*>(buf)); \
516                  if (real_size + 1 != size) { \
517                      printf("%s failed, expected %zu, got %zu\n", name, size, real_size); \
518                      return -1; \
519                  });
520     return 0;
521 }
522 
benchmarkStrcat(const char * name,const command_data_t & cmd_data,void_func_t func)523 int benchmarkStrcat(const char *name, const command_data_t &cmd_data, void_func_t func) {
524     str_func_t str_func = reinterpret_cast<str_func_t>(func);
525 
526     int dst_str_size = cmd_data.dst_str_size;
527     if (dst_str_size <= 0) {
528         printf("%s requires --dst_str_size to be set to a non-zero value.\n",
529                name);
530         return -1;
531     }
532     BENCH_TWO_BUFS(name, cmd_data,
533                    initString(buf1, size); \
534                    initString(buf2, dst_str_size),
535                    str_func(reinterpret_cast<char*>(buf2), reinterpret_cast<char*>(buf1)); buf2[dst_str_size-1] = '\0');
536 
537     return 0;
538 }
539 
benchmarkStrcatCold(const char * name,const command_data_t & cmd_data,void_func_t func)540 int benchmarkStrcatCold(const char *name, const command_data_t &cmd_data, void_func_t func) {
541     str_func_t str_func = reinterpret_cast<str_func_t>(func);
542 
543     int dst_str_size = cmd_data.dst_str_size;
544     if (dst_str_size <= 0) {
545         printf("%s requires --dst_str_size to be set to a non-zero value.\n",
546                name);
547         return -1;
548     }
549     COLD_TWO_BUFS(name, cmd_data,
550                   memset(buffer1, 'a', buffer1_size); \
551                   memset(buffer2, 'b', buffer2_size); \
552                   for (size_t i = 0; i < num_buffers; i++) { \
553                       buffer1[size-1+buf1_incr*i] = '\0'; \
554                       buffer2[dst_str_size-1+buf2_incr*i] = '\0'; \
555                   },
556                   str_func(reinterpret_cast<char*>(buf2), reinterpret_cast<char*>(buf1)); buf2[dst_str_size-1] = '\0');
557 
558     return 0;
559 }
560 
561 
benchmarkStrcpy(const char * name,const command_data_t & cmd_data,void_func_t func)562 int benchmarkStrcpy(const char *name, const command_data_t &cmd_data, void_func_t func) {
563     str_func_t str_func = reinterpret_cast<str_func_t>(func);
564 
565     BENCH_TWO_BUFS(name, cmd_data,
566                    initString(buf1, size); \
567                    memset(buf2, 0, size),
568                    str_func(reinterpret_cast<char*>(buf2), reinterpret_cast<char*>(buf1)));
569 
570     return 0;
571 }
572 
benchmarkStrcpyCold(const char * name,const command_data_t & cmd_data,void_func_t func)573 int benchmarkStrcpyCold(const char *name, const command_data_t &cmd_data, void_func_t func) {
574     str_func_t str_func = reinterpret_cast<str_func_t>(func);
575 
576     COLD_TWO_BUFS(name, cmd_data,
577                   memset(buffer1, 'a', buffer1_size); \
578                   for (size_t i = 0; i < num_buffers; i++) { \
579                      buffer1[size-1+buf1_incr*i] = '\0'; \
580                   } \
581                   memset(buffer2, 0, buffer2_size),
582                   str_func(reinterpret_cast<char*>(buf2), reinterpret_cast<char*>(buf1)));
583 
584     return 0;
585 }
586 
587 // Create the mapping structure.
588 function_t function_table[] = {
589     { "memcpy", benchmarkMemcpy, reinterpret_cast<void_func_t>(memcpy) },
590     { "memcpy_cold", benchmarkMemcpyCold, reinterpret_cast<void_func_t>(memcpy) },
591     { "memmove_forward", benchmarkMemcpy, reinterpret_cast<void_func_t>(memmove) },
592     { "memmove_backward", benchmarkMemmoveBackwards, reinterpret_cast<void_func_t>(memmove) },
593     { "memread", benchmarkMemread, NULL },
594     { "memset", benchmarkMemset, reinterpret_cast<void_func_t>(memset) },
595     { "memset_cold", benchmarkMemsetCold, reinterpret_cast<void_func_t>(memset) },
596     { "sleep", benchmarkSleep, NULL },
597     { "strcat", benchmarkStrcat, reinterpret_cast<void_func_t>(strcat) },
598     { "strcat_cold", benchmarkStrcatCold, reinterpret_cast<void_func_t>(strcat) },
599     { "strcmp", benchmarkStrcmp, reinterpret_cast<void_func_t>(strcmp) },
600     { "strcmp_cold", benchmarkStrcmpCold, reinterpret_cast<void_func_t>(strcmp) },
601     { "strcpy", benchmarkStrcpy, reinterpret_cast<void_func_t>(strcpy) },
602     { "strcpy_cold", benchmarkStrcpyCold, reinterpret_cast<void_func_t>(strcpy) },
603     { "strlen", benchmarkStrlen, reinterpret_cast<void_func_t>(strlen) },
604     { "strlen_cold", benchmarkStrlenCold, reinterpret_cast<void_func_t>(strlen) },
605 };
606 
usage()607 void usage() {
608     printf("Usage:\n");
609     printf("  micro_bench [--data_size DATA_BYTES] [--print_average]\n");
610     printf("              [--no_print_each_iter] [--lock_to_cpu CORE]\n");
611     printf("              [--src_align ALIGN] [--src_or_mask OR_MASK]\n");
612     printf("              [--dst_align ALIGN] [--dst_or_mask OR_MASK]\n");
613     printf("              [--dst_str_size SIZE] [--cold_data_size DATA_BYTES]\n");
614     printf("              [--cold_stride_size SIZE]\n");
615     printf("    --data_size DATA_BYTES\n");
616     printf("      For the data benchmarks (memcpy/memset/memread) the approximate\n");
617     printf("      size of data, in bytes, that will be manipulated in each iteration.\n");
618     printf("    --print_average\n");
619     printf("      Print the average and standard deviation of all iterations.\n");
620     printf("    --no_print_each_iter\n");
621     printf("      Do not print any values in each iteration.\n");
622     printf("    --lock_to_cpu CORE\n");
623     printf("      Lock to the specified CORE. The default is to use the last core found.\n");
624     printf("    --dst_align ALIGN\n");
625     printf("      If the command supports it, align the destination pointer to ALIGN.\n");
626     printf("      The default is to use the value returned by malloc.\n");
627     printf("    --dst_or_mask OR_MASK\n");
628     printf("      If the command supports it, or in the OR_MASK on to the destination pointer.\n");
629     printf("      The OR_MASK must be smaller than the dst_align value.\n");
630     printf("      The default value is 0.\n");
631 
632     printf("    --src_align ALIGN\n");
633     printf("      If the command supports it, align the source pointer to ALIGN. The default is to use the\n");
634     printf("      value returned by malloc.\n");
635     printf("    --src_or_mask OR_MASK\n");
636     printf("      If the command supports it, or in the OR_MASK on to the source pointer.\n");
637     printf("      The OR_MASK must be smaller than the src_align value.\n");
638     printf("      The default value is 0.\n");
639     printf("    --dst_str_size SIZE\n");
640     printf("      If the command supports it, create a destination string of this length.\n");
641     printf("      The default is to not update the destination string.\n");
642     printf("    --cold_data_size DATA_SIZE\n");
643     printf("      For _cold benchmarks, use this as the total amount of memory to use.\n");
644     printf("      The default is 128MB, and the number should be larger than the cache on the chip.\n");
645     printf("      This value is specified in bytes.\n");
646     printf("    --cold_stride_size SIZE\n");
647     printf("      For _cold benchmarks, use this as the minimum stride between iterations.\n");
648     printf("      The default is 4096 bytes and the number should be larger than the amount of data\n");
649     printf("      pulled in to the cache by each run of the benchmark.\n");
650     printf("    ITERS\n");
651     printf("      The number of iterations to execute each benchmark. If not\n");
652     printf("      passed in then run forever.\n");
653     printf("  micro_bench cpu UNUSED [ITERS]\n");
654     printf("  micro_bench [--dst_align ALIGN] [--dst_or_mask OR_MASK] memcpy NUM_BYTES [ITERS]\n");
655     printf("  micro_bench memread NUM_BYTES [ITERS]\n");
656     printf("  micro_bench [--dst_align ALIGN] [--dst_or_mask OR_MASK] memset NUM_BYTES [ITERS]\n");
657     printf("  micro_bench sleep TIME_TO_SLEEP [ITERS]\n");
658     printf("    TIME_TO_SLEEP\n");
659     printf("      The time in seconds to sleep.\n");
660     printf("  micro_bench [--src_align ALIGN] [--src_or_mask OR_MASK] [--dst_align ALIGN] [--dst_or_mask] [--dst_str_size SIZE] strcat NUM_BYTES [ITERS]\n");
661     printf("  micro_bench [--src_align ALIGN] [--src_or_mask OR_MASK] [--dst_align ALIGN] [--dst_or_mask OR_MASK] strcmp NUM_BYTES [ITERS]\n");
662     printf("  micro_bench [--src_align ALIGN] [--src_or_mask OR_MASK] [--dst_align ALIGN] [--dst_or_mask] strcpy NUM_BYTES [ITERS]\n");
663     printf("  micro_bench [--dst_align ALIGN] [--dst_or_mask OR_MASK] strlen NUM_BYTES [ITERS]\n");
664     printf("\n");
665     printf("  In addition, memcpy/memcpy/memset/strcat/strcpy/strlen have _cold versions\n");
666     printf("  that will execute the function on a buffer not in the cache.\n");
667 }
668 
processOptions(int argc,char ** argv,command_data_t * cmd_data)669 function_t *processOptions(int argc, char **argv, command_data_t *cmd_data) {
670     function_t *command = NULL;
671 
672     // Initialize the command_flags.
673     cmd_data->print_average = false;
674     cmd_data->print_each_iter = true;
675     cmd_data->dst_align = 0;
676     cmd_data->src_align = 0;
677     cmd_data->src_or_mask = 0;
678     cmd_data->dst_or_mask = 0;
679     cmd_data->num_args = 0;
680     cmd_data->cpu_to_lock = -1;
681     cmd_data->data_size = DEFAULT_DATA_SIZE;
682     cmd_data->dst_str_size = -1;
683     cmd_data->cold_data_size = DEFAULT_COLD_DATA_SIZE;
684     cmd_data->cold_stride_size = DEFAULT_COLD_STRIDE_SIZE;
685     for (int i = 0; i < MAX_ARGS; i++) {
686         cmd_data->args[i] = -1;
687     }
688 
689     for (int i = 1; i < argc; i++) {
690         if (argv[i][0] == '-') {
691             int *save_value = NULL;
692             if (strcmp(argv[i], "--print_average") == 0) {
693                 cmd_data->print_average = true;
694             } else if (strcmp(argv[i], "--no_print_each_iter") == 0) {
695                 cmd_data->print_each_iter = false;
696             } else if (strcmp(argv[i], "--dst_align") == 0) {
697                 save_value = &cmd_data->dst_align;
698             } else if (strcmp(argv[i], "--src_align") == 0) {
699                 save_value = &cmd_data->src_align;
700             } else if (strcmp(argv[i], "--dst_or_mask") == 0) {
701                 save_value = &cmd_data->dst_or_mask;
702             } else if (strcmp(argv[i], "--src_or_mask") == 0) {
703                 save_value = &cmd_data->src_or_mask;
704             } else if (strcmp(argv[i], "--lock_to_cpu") == 0) {
705                 save_value = &cmd_data->cpu_to_lock;
706             } else if (strcmp(argv[i], "--data_size") == 0) {
707                 save_value = &cmd_data->data_size;
708             } else if (strcmp(argv[i], "--dst_str_size") == 0) {
709                 save_value = &cmd_data->dst_str_size;
710             } else if (strcmp(argv[i], "--cold_data_size") == 0) {
711                 save_value = &cmd_data->cold_data_size;
712             } else if (strcmp(argv[i], "--cold_stride_size") == 0) {
713                 save_value = &cmd_data->cold_stride_size;
714             } else {
715                 printf("Unknown option %s\n", argv[i]);
716                 return NULL;
717             }
718             if (save_value) {
719                 // Checking both characters without a strlen() call should be
720                 // safe since as long as the argument exists, one character will
721                 // be present (\0). And if the first character is '-', then
722                 // there will always be a second character (\0 again).
723                 if (i == argc - 1 || (argv[i + 1][0] == '-' && !isdigit(argv[i + 1][1]))) {
724                     printf("The option %s requires one argument.\n",
725                            argv[i]);
726                     return NULL;
727                 }
728                 *save_value = (int)strtol(argv[++i], NULL, 0);
729             }
730         } else if (!command) {
731             for (size_t j = 0; j < sizeof(function_table)/sizeof(function_t); j++) {
732                 if (strcmp(argv[i], function_table[j].name) == 0) {
733                     command = &function_table[j];
734                     break;
735                 }
736             }
737             if (!command) {
738                 printf("Uknown command %s\n", argv[i]);
739                 return NULL;
740             }
741         } else if (cmd_data->num_args > MAX_ARGS) {
742             printf("More than %d number arguments passed in.\n", MAX_ARGS);
743             return NULL;
744         } else {
745             cmd_data->args[cmd_data->num_args++] = atoi(argv[i]);
746         }
747     }
748 
749     // Check the arguments passed in make sense.
750     if (cmd_data->num_args != 1 && cmd_data->num_args != 2) {
751         printf("Not enough arguments passed in.\n");
752         return NULL;
753     } else if (cmd_data->dst_align < 0) {
754         printf("The --dst_align option must be greater than or equal to 0.\n");
755         return NULL;
756     } else if (cmd_data->src_align < 0) {
757         printf("The --src_align option must be greater than or equal to 0.\n");
758         return NULL;
759     } else if (cmd_data->data_size <= 0) {
760         printf("The --data_size option must be a positive number.\n");
761         return NULL;
762     } else if ((cmd_data->dst_align & (cmd_data->dst_align - 1))) {
763         printf("The --dst_align option must be a power of 2.\n");
764         return NULL;
765     } else if ((cmd_data->src_align & (cmd_data->src_align - 1))) {
766         printf("The --src_align option must be a power of 2.\n");
767         return NULL;
768     } else if (!cmd_data->src_align && cmd_data->src_or_mask) {
769         printf("The --src_or_mask option requires that --src_align be set.\n");
770         return NULL;
771     } else if (!cmd_data->dst_align && cmd_data->dst_or_mask) {
772         printf("The --dst_or_mask option requires that --dst_align be set.\n");
773         return NULL;
774     } else if (cmd_data->src_or_mask > cmd_data->src_align) {
775         printf("The value of --src_or_mask cannot be larger that --src_align.\n");
776         return NULL;
777     } else if (cmd_data->dst_or_mask > cmd_data->dst_align) {
778         printf("The value of --src_or_mask cannot be larger that --src_align.\n");
779         return NULL;
780     }
781 
782     return command;
783 }
784 
raisePriorityAndLock(int cpu_to_lock)785 bool raisePriorityAndLock(int cpu_to_lock) {
786     cpu_set_t cpuset;
787 
788     if (setpriority(PRIO_PROCESS, 0, -20)) {
789         perror("Unable to raise priority of process.\n");
790         return false;
791     }
792 
793     CPU_ZERO(&cpuset);
794     if (sched_getaffinity(0, sizeof(cpuset), &cpuset) != 0) {
795         perror("sched_getaffinity failed");
796         return false;
797     }
798 
799     if (cpu_to_lock < 0) {
800         // Lock to the last active core we find.
801         for (int i = 0; i < CPU_SETSIZE; i++) {
802             if (CPU_ISSET(i, &cpuset)) {
803                 cpu_to_lock = i;
804             }
805         }
806     } else if (!CPU_ISSET(cpu_to_lock, &cpuset)) {
807         printf("Cpu %d does not exist.\n", cpu_to_lock);
808         return false;
809     }
810 
811     if (cpu_to_lock < 0) {
812         printf("Cannot find any valid cpu to lock.\n");
813         return false;
814     }
815 
816     CPU_ZERO(&cpuset);
817     CPU_SET(cpu_to_lock, &cpuset);
818     if (sched_setaffinity(0, sizeof(cpuset), &cpuset) != 0) {
819         perror("sched_setaffinity failed");
820         return false;
821     }
822 
823     return true;
824 }
825 
main(int argc,char ** argv)826 int main(int argc, char **argv) {
827     command_data_t cmd_data;
828 
829     function_t *command = processOptions(argc, argv, &cmd_data);
830     if (!command) {
831       usage();
832       return -1;
833     }
834 
835     if (!raisePriorityAndLock(cmd_data.cpu_to_lock)) {
836       return -1;
837     }
838 
839     printf("%s\n", command->name);
840     return (*command->ptr)(command->name, cmd_data, command->func);
841 }
842