Difference between
modified post 0bbd6628 by Anonymous on Sat 26th Nov 2022 15:49
original post 8504236b4 by Anonymous on Sat 26th Nov 2022 15:28
Download diff
Show old version | new version | both versions

rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ cat x86_rdtscp2_perf.c

1		-	rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ cat x86_rdtscp2_perf.c
	1	+	rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ make tests
	2	+	gcc -std=gnu17 -m32 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_32
2	10		/*
3	11		* x86_rdtscp2_perf.c - compare execution time x86 RDTSCP instruction
4	12		* vs. \|sched_getcpu()\|
5	13		*
6	14		* Compile with:
7	15		* $ gcc -std=gnu17 -m32 -Wall x86_rdtscp2_perf.c -lpthread
8	16		*
9	17		* Written by Roland Mainz <roland.mainz@nrubsig.org>
10	18		*
11	19		*/
12	20
13	21		#define _XOPEN_SOURCE 700
14	22		#define _GNU_SOURCE 1
15	23
16	24		#include <stdio.h>
17	25		#include <stdlib.h>
18	26		#include <errno.h>
19	27
20	28		#include <pthread.h>
21	29		#include <sched.h>
22	30
23	31		#include <x86intrin.h>
24	32
25	33
26	34		int main(int ac, char *av[])
27	35		{
28	36		(void)puts("#start.");
29	37
30	38		unsigned int A = 0;
31	39
32	40		#define USE_FIXED_CPU_NUMBER 1
33	41		#if USE_FIXED_CPU_NUMBER
34	42		int cpu;
35	43		cpu_set_t cpuset;
36	44		pthread_t thread;
37	45
38	46		thread = pthread_self();
39	47		cpu = sched_getcpu();
40	48
41	49		CPU_ZERO(&cpuset);
42	50		CPU_SET(cpu, &cpuset);
43	51
44	52		if (pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset) == 0) {
45	53		(void)printf("# pthread_setaffinity_np(), "
46	54		"thread fixed to cpu %d\n",
47	55		cpu);
48	56		}
49	57		else {
50	58		/*
51	59		* If you specify a CPU number beyond the maximum number
52	60		* of CPUs, then \|pthread_setaffinity_np()\| can have a
53	61		* \|errno == 0\|
54	62		*/
55	63		perror("pthread_setaffinity_np");
56	64		}
57	65		#endif /* USE_FIXED_CPU_NUMBER */
58	66
59	67		struct timespec start, stop;
60	68		long long accum;
61	69		long long dummy = 0LL; /* use dummy variable to defeat optimiser */
62	70
63	71		if (clock_gettime(CLOCK_REALTIME, &start) == -1) {
64	72		perror("clock gettime");
65	73		exit(EXIT_FAILURE);
66	74		}
67	75
68	76		for (long l = 0 ; l < 100000000 ; l++) {
69	77		dummy += _rdtscp(&A);
70	78		dummy += A;
71	79		}
72	80
73	81		if (clock_gettime(CLOCK_REALTIME, &stop) == -1 ) {
74	82		perror("clock gettime");
75	83		exit(EXIT_FAILURE);
76	84		}
77	85
78		-	#define NSECPERSEC (1000000000UL)
	86	+	#define NSECPERSEC (1000000000LL)
79	87		accum = (stop.tv_sec - start.tv_sec)*NSECPERSEC +
80	88		(stop.tv_nsec - start.tv_nsec);
81	89		(void)printf("# time needed for x86 rdtscp instruction:\t%20.20f\n",
82		-	(double)((long double)accum/(long double)NSECPERSEC));
	90	+	(double)(((long double)accum)/((long double)NSECPERSEC)));
83	91
84	92
85	93
86	94		if (clock_gettime(CLOCK_REALTIME, &start) == -1) {
87	95		perror( "clock gettime" );
88	96		exit(EXIT_FAILURE);
89	97		}
90	98
91	99		for (long l = 0 ; l < 100000000 ; l++) {
92	100		dummy += sched_getcpu();
93	101		}
94	102
95	103
96	104		if (clock_gettime(CLOCK_REALTIME, &stop) == -1) {
97	105		perror("clock gettime");
98	106		exit(EXIT_FAILURE);
99	107		}
100	108
101	109		accum = (stop.tv_sec - start.tv_sec)*NSECPERSEC +
102	110		(stop.tv_nsec - start.tv_nsec);
103	111		(void)printf("# time needed for \|sched_getcpu()\|: \t%20.20f\n",
104		-	(double)((long double)accum/(long double)NSECPERSEC));
	112	+	(double)(((long double)accum)/((long double)NSECPERSEC)));
105	113
106	114		(void)dummy;
107	115
108	116		(void)puts("#done.");
109	117
110	118		return EXIT_SUCCESS;
111	119		}
112		-	rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ cat Makefile
	120	+	+ cat Makefile
113	121		#
114	122		# Makefile for x86_rdtscp2_perf.c - compare execution time x86 RDTSCP
115	123		# instruction vs. \|sched_getcpu()\|
116	124		#
117	125
118	126		all: \
119	127		x86_rdtscp2_perf_32 \
120	128		x86_rdtscp2_perf_64
121	129
122	130		# explicitly use -O0 to disable peephole optimisations
123	131		x86_rdtscp2_perf_32: x86_rdtscp2_perf.c
124	132		gcc -std=gnu17 -m32 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_32
125	133
126	134		x86_rdtscp2_perf_64: x86_rdtscp2_perf.c
127	135		gcc -std=gnu17 -m64 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_64
128	136
129	137
130	138		tests: \
131	139		x86_rdtscp2_perf_32 \
132	140		x86_rdtscp2_perf_64
133		-	ksh93 -c 'set -o xtrace ; time ./x86_rdtscp2_perf_32 ; time ./x86_rdtscp2_perf_64'
	141	+	ksh93 -c 'set -o xtrace ; \
	142	+	cat x86_rdtscp2_perf.c ; \
134	146
135	147		clean:
136	148		rm -f \
137	149		x86_rdtscp2_perf_32 \
138	150		x86_rdtscp2_perf_64
139	151		# EOF.
142	152		+ ./x86_rdtscp2_perf_32
143	153		#start.
144		-	# pthread_setaffinity_np(), thread fixed to cpu 0
	154	+	# pthread_setaffinity_np(), thread fixed to cpu 1
145		-	# time needed for x86 rdtscp instruction: 1.29958247999999998434
	155	+	# time needed for x86 rdtscp instruction: 1.32610883200000007065
146		-	# time needed for \|sched_getcpu()\|: 3.39503941599999992107
	156	+	# time needed for \|sched_getcpu()\|: 33.35152474899999930358
147	157		#done.
148	158
149		-	real 0m34.77s
	159	+	real 0m34.68s
150		-	user 0m21.45s
	160	+	user 0m22.13s
151		-	sys 0m13.29s
	161	+	sys 0m12.53s
152	162		+ ./x86_rdtscp2_perf_64
153	163		#start.
154		-	# pthread_setaffinity_np(), thread fixed to cpu 6
	164	+	# pthread_setaffinity_np(), thread fixed to cpu 0
155		-	# time needed for x86 rdtscp instruction: 1.20903293999999994490
	165	+	# time needed for x86 rdtscp instruction: 1.23842487100000009370
156		-	# time needed for \|sched_getcpu()\|: 1.38745017499999989674
	166	+	# time needed for \|sched_getcpu()\|: 1.95299446999999992691
157	167		#done.
158	168
159		-	real 0m2.60s
	169	+	real 0m3.19s
160		-	user 0m2.60s
	170	+	user 0m3.19s
161	171		sys 0m0.00s
162	172