Difference between
modified post c21e by Anonymous on Sat 26th Nov 2022 18:53
original post 0bbd6628 by Anonymous on Sat 26th Nov 2022 15:49
Download diff
Show old version | new version | both versions

rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ make tests

1		-	rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ make tests
	1	+	rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ banner "test 104"
2		-	gcc -std=gnu17 -m32 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_32
	2	+	# ### #
3		-	gcc -std=gnu17 -m64 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_64
	3	+	##### ###### #### ##### ## # # # #
4		-	ksh93 -c 'set -o xtrace ; \
	4	+	# # # # # # # # # # #
5		-	cat x86_rdtscp2_perf.c ; \
	5	+	# ##### #### # # # # # #######
6		-	cat Makefile ; \
	6	+	# # # # # # # # #
7		-	time ./x86_rdtscp2_perf_32 ; \
	7	+	# # # # # # # # #
8		-	time ./x86_rdtscp2_perf_64'
	8	+	# ###### #### # ##### ### #
9		-	+ cat x86_rdtscp2_perf.c
	9	+
10		-	/*
	10	+	rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ make clean
11		-	* x86_rdtscp2_perf.c - compare execution time x86 RDTSCP instruction
	11	+	rm -f \
12		-	* vs. \|sched_getcpu()\|
	12	+	x86_rdtscp2_perf_32 \
13		-	*
	13	+	x86_rdtscp2_perf_64
14		-	* Compile with:
	14	+	rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ make tests
15		-	* $ gcc -std=gnu17 -m32 -Wall x86_rdtscp2_perf.c -lpthread
	15	+	gcc -std=gnu17 -m32 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_32
16		-	*
	16	+	gcc -std=gnu17 -m64 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_64
17		-	* Written by Roland Mainz <roland.mainz@nrubsig.org>
	17	+	ksh93 -c 'set -o xtrace ; \
	18	+	cat x86_rdtscp2_perf.c ; \
18	26		*
	27	+	* Compile with:
19	32		*/
20	33
21	34		#define _XOPEN_SOURCE 700
22	35		#define _GNU_SOURCE 1
23	36
24	37		#include <stdio.h>
25	38		#include <stdlib.h>
26	39		#include <errno.h>
27	40
28	41		#include <pthread.h>
29	42		#include <sched.h>
30	43
31	44		#include <x86intrin.h>
32	45
33	46
34	47		int main(int ac, char *av[])
35	48		{
36	49		(void)puts("#start.");
37	50
38		-	unsigned int A = 0;
	51	+	unsigned int A = 0;
	52	+	long long li; /* loop iterator */
39	53
40	54		#define USE_FIXED_CPU_NUMBER 1
41	55		#if USE_FIXED_CPU_NUMBER
42	56		int cpu;
43	57		cpu_set_t cpuset;
44	58		pthread_t thread;
45	59
46	60		thread = pthread_self();
47	61		cpu = sched_getcpu();
48	62
49	63		CPU_ZERO(&cpuset);
50	64		CPU_SET(cpu, &cpuset);
51	65
52	66		if (pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset) == 0) {
53	67		(void)printf("# pthread_setaffinity_np(), "
54	68		"thread fixed to cpu %d\n",
55	69		cpu);
56	70		}
57	71		else {
58	72		/*
59	73		* If you specify a CPU number beyond the maximum number
60	74		* of CPUs, then \|pthread_setaffinity_np()\| can have a
61	75		* \|errno == 0\|
62	76		*/
63	77		perror("pthread_setaffinity_np");
64	78		}
65	79		#endif /* USE_FIXED_CPU_NUMBER */
66	80
67	81		struct timespec start, stop;
68	82		long long accum;
69	83		long long dummy = 0LL; /* use dummy variable to defeat optimiser */
70	84
71	85		if (clock_gettime(CLOCK_REALTIME, &start) == -1) {
72	86		perror("clock gettime");
73	87		exit(EXIT_FAILURE);
74	88		}
	89	+	#define NUMTESTITERATIONS (1000000000LL)
77	91		dummy += _rdtscp(&A);
78	92		dummy += A;
79	93		}
80	94
81	95		if (clock_gettime(CLOCK_REALTIME, &stop) == -1 ) {
82	96		perror("clock gettime");
83	97		exit(EXIT_FAILURE);
84	98		}
85	99
86	100		#define NSECPERSEC (1000000000LL)
87	101		accum = (stop.tv_sec - start.tv_sec)*NSECPERSEC +
88	102		(stop.tv_nsec - start.tv_nsec);
89		-	(void)printf("# time needed for x86 rdtscp instruction:\t%20.20f\n",
	103	+	(void)printf("# x86 rdtscp instruction:\t"
	104	+	"iterations=%lld\t"
90	107		(double)(((long double)accum)/((long double)NSECPERSEC)));
91	108
92	109
93	110
94	111		if (clock_gettime(CLOCK_REALTIME, &start) == -1) {
95	112		perror( "clock gettime" );
96	113		exit(EXIT_FAILURE);
97	114		}
98	115
99		-	for (long l = 0 ; l < 100000000 ; l++) {
	116	+	for (li = 0 ; li < NUMTESTITERATIONS ; li++) {
100	117		dummy += sched_getcpu();
101	118		}
102	119
103	120
104	121		if (clock_gettime(CLOCK_REALTIME, &stop) == -1) {
105	122		perror("clock gettime");
106	123		exit(EXIT_FAILURE);
107	124		}
108	125
109	126		accum = (stop.tv_sec - start.tv_sec)*NSECPERSEC +
110	127		(stop.tv_nsec - start.tv_nsec);
111		-	(void)printf("# time needed for \|sched_getcpu()\|: \t%20.20f\n",
	128	+	(void)printf("# \|sched_getcpu()\|: \t"
	129	+	"iterations=%lld\t"
112	132		(double)(((long double)accum)/((long double)NSECPERSEC)));
113	133
114	134		(void)dummy;
115	135
116	136		(void)puts("#done.");
117	137
118	138		return EXIT_SUCCESS;
119	139		}
120	140		+ cat Makefile
121	141		#
122	142		# Makefile for x86_rdtscp2_perf.c - compare execution time x86 RDTSCP
123	143		# instruction vs. \|sched_getcpu()\|
124	144		#
125	145
126	146		all: \
127	147		x86_rdtscp2_perf_32 \
128	148		x86_rdtscp2_perf_64
129	149
130	150		# explicitly use -O0 to disable peephole optimisations
131	151		x86_rdtscp2_perf_32: x86_rdtscp2_perf.c
132	152		gcc -std=gnu17 -m32 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_32
133	153
134	154		x86_rdtscp2_perf_64: x86_rdtscp2_perf.c
135	155		gcc -std=gnu17 -m64 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_64
136	156
137	157
138	158		tests: \
139	159		x86_rdtscp2_perf_32 \
140	160		x86_rdtscp2_perf_64
141	161		ksh93 -c 'set -o xtrace ; \
142	162		cat x86_rdtscp2_perf.c ; \
143	163		cat Makefile ; \
144	164		time ./x86_rdtscp2_perf_32 ; \
145	165		time ./x86_rdtscp2_perf_64'
146	166
147	167		clean:
148	168		rm -f \
149	169		x86_rdtscp2_perf_32 \
150	170		x86_rdtscp2_perf_64
151	171		# EOF.
152	172		+ ./x86_rdtscp2_perf_32
153	173		#start.
154		-	# pthread_setaffinity_np(), thread fixed to cpu 1
	174	+	# pthread_setaffinity_np(), thread fixed to cpu 5
155		-	# time needed for x86 rdtscp instruction: 1.32610883200000007065
	175	+	# x86 rdtscp instruction: iterations=1000000000 time=14.56760232699999946249 secs
156		-	# time needed for \|sched_getcpu()\|: 33.35152474899999930358
	176	+	# \|sched_getcpu()\|: iterations=1000000000 time=333.57820389099998692473 secs
157	177		#done.
158	178
159		-	real 0m34.68s
	179	+	real 5m48.15s
160		-	user 0m22.13s
	180	+	user 3m42.13s
161		-	sys 0m12.53s
	181	+	sys 2m5.94s
162	182		+ ./x86_rdtscp2_perf_64
163	183		#start.
164		-	# pthread_setaffinity_np(), thread fixed to cpu 0
	184	+	# pthread_setaffinity_np(), thread fixed to cpu 3
165		-	# time needed for x86 rdtscp instruction: 1.23842487100000009370
	185	+	# x86 rdtscp instruction: iterations=1000000000 time=12.15419475899999923740 secs
166		-	# time needed for \|sched_getcpu()\|: 1.95299446999999992691
	186	+	# \|sched_getcpu()\|: iterations=1000000000 time=13.76809202599999970573 secs
167	187		#done.
168	188
169		-	real 0m3.19s
	189	+	real 0m25.94s
170		-	user 0m3.19s
	190	+	user 0m25.92s
171	191		sys 0m0.00s
172	192