33
33
'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):
36
def run(cache, variant, function, bytes, loops, alignment=8, quiet=False):
55
37
"""Perform a single run, exercising the cache as appropriate."""
56
key = ':'.join('%s' % x for x in (variant, function, bytes, loops, alignment, run_id))
38
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()
44
cmd = '%(xbuild)s%(variant)s -t %(function)s -c %(bytes)s -l %(loops)s -a %(alignment)s' % locals()
65
47
got = subprocess.check_output(cmd.split()).strip()
80
def run_many(cache, variants, bytes, all_functions):
62
def run_many(cache, variants, bytes, alignments, all_functions):
81
63
# We want the data to come out in a useful order. So fix an
82
64
# alignment and function, and do all sizes for a variant first
83
65
bytes = sorted(bytes)
84
mid = bytes[int(len(bytes)/1.5)]
66
mid = bytes[len(bytes)/2]
86
68
if not all_functions:
87
69
# Use the ordering in 'this' as the default
93
75
if function not in all_functions:
94
76
all_functions.append(function)
96
for function in all_functions:
97
for alignment in ALIGNMENTS[function]:
78
for alignment in alignments:
79
for function in all_functions:
98
80
for variant in variants:
99
81
if function not in HAS[variant].split():
109
91
loops = int(f / math.sqrt(max(1, mid)))
110
took = run(cache, variant, function, mid, loops, alignment, 0,
92
took = run(cache, variant, function, mid, loops, alignment, quiet=True)
112
93
# Keep it reasonable for silly routines like bounce
113
94
factor = min(20, max(0.05, want/took))
120
101
for b in sorted(bytes):
121
102
# Figure out the number of loops to give a roughly consistent run
122
103
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,
104
run(cache, variant, function, b, loops, alignment)
127
106
def run_top(cache):
128
107
variants = sorted(HAS.keys())
143
122
steps = int(round(math.log(top) / math.log(step)))
144
123
bytes.extend([int(step**x) for x in range(0, steps+1)])
146
run_many(cache, variants, bytes, functions)
125
alignments = [8, 16, 4, 1, 2, 32]
127
run_many(cache, variants, bytes, alignments, functions)
149
130
cachename = 'cache.txt'