23
23
'this': 'bounce memchr memcpy memset strchr strcpy strlen',
24
'bionic-a9': 'memcmp memcpy memset strcmp strcpy strlen',
25
'bionic-a15': 'memcmp memcpy memset strcmp strcpy strlen',
24
'bionic': 'memcmp memcpy memset strcmp strcpy strlen',
27
26
'csl': 'memcpy memset',
28
27
'glibc': 'memcpy memset strlen',
33
32
'plain': 'memset memcpy strcmp strcpy',
36
BOUNCE_ALIGNMENTS = ['1']
37
SINGLE_BUFFER_ALIGNMENTS = ['1', '2', '4', '8', '16', '32']
38
DUAL_BUFFER_ALIGNMENTS = ['1:32', '2:32', '4:32', '8:32', '16:32', '32:32']
41
'bounce': BOUNCE_ALIGNMENTS,
42
'memchr': SINGLE_BUFFER_ALIGNMENTS,
43
'memset': SINGLE_BUFFER_ALIGNMENTS,
44
'strchr': SINGLE_BUFFER_ALIGNMENTS,
45
'strlen': SINGLE_BUFFER_ALIGNMENTS,
46
'memcmp': DUAL_BUFFER_ALIGNMENTS,
47
'memcpy': DUAL_BUFFER_ALIGNMENTS,
48
'strcmp': DUAL_BUFFER_ALIGNMENTS,
49
'strcpy': DUAL_BUFFER_ALIGNMENTS,
54
def run(cache, variant, function, bytes, loops, alignment, run_id, quiet=False):
35
def run(cache, variant, function, bytes, loops, alignment=8, quiet=False):
55
36
"""Perform a single run, exercising the cache as appropriate."""
56
key = ':'.join('%s' % x for x in (variant, function, bytes, loops, alignment, run_id))
37
key = ':'.join('%s' % x for x in (variant, function, bytes, loops, alignment))
62
cmd = '%(xbuild)s%(variant)s -t %(function)s -c %(bytes)s -l %(loops)s -a %(alignment)s -r %(run_id)s' % locals()
43
cmd = '%(xbuild)s%(variant)s -t %(function)s -c %(bytes)s -l %(loops)s -a %(alignment)s' % locals()
65
46
got = subprocess.check_output(cmd.split()).strip()
80
def run_many(cache, variants, bytes, all_functions):
61
def run_many(cache, variants, bytes, alignments):
81
62
# We want the data to come out in a useful order. So fix an
82
63
# alignment and function, and do all sizes for a variant first
83
64
bytes = sorted(bytes)
84
mid = bytes[int(len(bytes)/1.5)]
87
# Use the ordering in 'this' as the default
88
all_functions = HAS['this'].split()
90
# Find all other functions
91
for functions in HAS.values():
92
for function in functions.split():
93
if function not in all_functions:
94
all_functions.append(function)
96
for function in all_functions:
97
for alignment in ALIGNMENTS[function]:
65
mid = bytes[len(bytes)/2]
67
# Use the ordering in 'this' as the default
68
all_functions = HAS['this'].split()
70
# Find all other functions
71
for functions in HAS.values():
72
for function in functions.split():
73
if function not in all_functions:
74
all_functions.append(function)
76
for alignment in alignments:
77
for function in all_functions:
98
78
for variant in variants:
99
79
if function not in HAS[variant].split():
109
89
loops = int(f / math.sqrt(max(1, mid)))
110
took = run(cache, variant, function, mid, loops, alignment, 0,
90
took = run(cache, variant, function, mid, loops, alignment, quiet=True)
112
91
# Keep it reasonable for silly routines like bounce
113
92
factor = min(20, max(0.05, want/took))
120
99
for b in sorted(bytes):
121
100
# Figure out the number of loops to give a roughly consistent run
122
101
loops = int(f / math.sqrt(max(1, b)))
123
for run_id in range(0, NUM_RUNS):
124
run(cache, variant, function, b, loops, alignment,
102
run(cache, variant, function, b, loops, alignment)
127
104
def run_top(cache):
128
105
variants = sorted(HAS.keys())
129
functions = sys.argv[1:]
131
107
# Upper limit in bytes to test to
143
119
steps = int(round(math.log(top) / math.log(step)))
144
120
bytes.extend([int(step**x) for x in range(0, steps+1)])
146
run_many(cache, variants, bytes, functions)
122
alignments = [8, 16, 4, 1, 2, 32]
124
run_many(cache, variants, bytes, alignments)
149
127
cachename = 'cache.txt'