1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
|
import NvRules
def get_identifier():
return "TemplateRuleSpeedup"
def get_name():
return "Speedup Estimation Template"
def get_description():
return "A rule template containing a speedup estimation and focus metrics."
def apply(handle):
ctx = NvRules.get_context(handle)
action = ctx.range_by_idx(0).action_by_idx(0)
frontend = ctx.frontend()
# Get metrics for the speedup estimation
compute_throughput_name = "sm__throughput.avg.pct_of_peak_sustained_elapsed"
memory_throughput_name = (
"gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed"
)
compute_throughput = action[compute_throughput_name].value()
memory_throughput = action[memory_throughput_name].value()
# Calculate the potential speedup; we will later need a number in percent
if compute_throughput > memory_throughput:
dominated_by = "compute"
improvement_percent = 100 - compute_throughput
else:
dominated_by = "memory"
improvement_percent = 100 - memory_throughput
# Post a message to the frontend summarizing the outcome of the rule.
# Messages of type OPTIMIZATION with a speedup estimate will also be displayed
# on the summary page.
message_id = frontend.message(
NvRules.IFrontend.MsgType_MSG_OPTIMIZATION,
"This kernel is currently dominated by {}.".format(dominated_by),
"Compute vs Memory",
)
# Attach a speedup estimate to the last message.
# Since we did not make sure that the required metrics are collected by a
# parent-scope section, we ought to make sure that the metrics are collected manually
# or by another section. We express this here with SpeedupType_GLOBAL.
frontend.speedup(
message_id, NvRules.IFrontend.SpeedupType_GLOBAL, improvement_percent
)
# Attach the two metrics which entered the speedup estimation to the last message.
# These metrics serve as focus metrics or key performance indicators and should
# be tracked when optimizing the kernel according to this rule.
frontend.focus_metric(
message_id,
compute_throughput_name,
compute_throughput,
NvRules.IFrontend.Severity_SEVERITY_DEFAULT,
"Increase the compute throughput",
)
frontend.focus_metric(
message_id,
memory_throughput_name,
memory_throughput,
NvRules.IFrontend.Severity_SEVERITY_DEFAULT,
"Increase the memory throughput",
)
|
