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
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
|
# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
from collections import defaultdict
import NvRules
from RequestedMetrics import MetricRequest, RequestedMetricsParser, Importance
requested_metrics = [
MetricRequest("memory_l1_wavefronts_shared", "shared_wavefronts"),
MetricRequest("memory_l1_wavefronts_shared_ideal", "shared_wavefronts_ideal"),
MetricRequest("derived__memory_l1_wavefronts_shared_excessive", "excessive_wavefronts"),
MetricRequest("l1tex__cycles_active.sum", "l1tex_cycles_active", Importance.OPTIONAL, 0),
MetricRequest("l1tex__cycles_elapsed.sum", "l1tex_cycles_elapsed", Importance.OPTIONAL, 0),
]
def get_identifier():
return "UncoalescedSharedAccess"
def get_name():
return "Uncoalesced Shared Accesses"
def get_description():
return "Uncoalesced Shared Accesses"
def get_section_identifier():
return "SourceCounters"
def get_parent_rules_identifiers():
return ["Memory"]
def get_estimated_speedup(metrics):
"""Estimate potential speedup from reducing uncoalesced shared memory accesses.
The performance improvement is approximated as relative proportion of excessive
wavefronts weighted by time spent in the L1TEX unit.
"""
active_cycles = metrics["l1tex_cycles_active"].value()
elapsed_cycles = metrics["l1tex_cycles_elapsed"].value()
excessive_wavefronts = metrics["excessive_wavefronts"].value()
total_wavefronts = metrics["shared_wavefronts"].value()
if (elapsed_cycles > 0) and (total_wavefronts > 0):
improvement_percent = (
(active_cycles / elapsed_cycles)
* (excessive_wavefronts / total_wavefronts)
* 100
)
return NvRules.IFrontend.SpeedupType_GLOBAL, improvement_percent
else:
return NvRules.IFrontend.SpeedupType_LOCAL, 0
def apply(handle):
ctx = NvRules.get_context(handle)
action = ctx.range_by_idx(0).action_by_idx(0)
fe = ctx.frontend()
metrics = RequestedMetricsParser(handle, action).parse(requested_metrics)
shared_wavefronts_metric = metrics["shared_wavefronts"]
shared_wavefronts_correlation_ids = shared_wavefronts_metric.correlation_ids()
ideal_shared_wavefronts_metric = metrics["shared_wavefronts_ideal"]
total_shared_wavefronts = shared_wavefronts_metric.value()
total_ideal_shared_wavefronts = ideal_shared_wavefronts_metric.value()
# No need to check further if total shared wavefronts match with the ideal value
if total_shared_wavefronts <= total_ideal_shared_wavefronts:
return
num_shared_wavefronts_instances = shared_wavefronts_metric.num_instances()
num_ideal_shared_wavefronts_instances = ideal_shared_wavefronts_metric.num_instances()
# We cannot execute the rule if we don't get the same instance count for both metrics
if num_shared_wavefronts_instances != num_ideal_shared_wavefronts_instances:
return
total_diff = 0
excess_by_line = defaultdict(int)
total_by_line = defaultdict(int)
for i in range(num_shared_wavefronts_instances):
per_instance_shared_wavefronts = shared_wavefronts_metric.as_uint64(i)
per_instance_ideal_shared_wavefronts = ideal_shared_wavefronts_metric.as_uint64(i)
if (per_instance_shared_wavefronts != per_instance_ideal_shared_wavefronts):
total_diff += abs(per_instance_ideal_shared_wavefronts - per_instance_shared_wavefronts)
# If there are excessive wavefronts, create source markers in the appropriate places
if (per_instance_shared_wavefronts > per_instance_ideal_shared_wavefronts):
address = shared_wavefronts_correlation_ids.as_uint64(i)
source_info = action.source_info(address)
excess = abs(per_instance_ideal_shared_wavefronts - per_instance_shared_wavefronts)
# Create source marker in the SASS file
fe.source_marker("{:.2f}% of this line's shared wavefronts are excessive.".format(excess / per_instance_shared_wavefronts * 100), address, NvRules.IFrontend.MarkerKind_SASS, NvRules.IFrontend.MsgType_MSG_WARNING)
# Aggregate diffs per line for the Source file marker
if source_info != None:
line = source_info.line()
file_name = source_info.file_name()
excess_by_line[line] += excess
total_by_line[line] += per_instance_shared_wavefronts
for line_number, local_diff in excess_by_line.items():
# Create source marker in the Source file per affected line
fe.source_marker("{:.2f}% of this line's shared wavefronts are excessive.".format(local_diff / total_by_line[line_number] * 100), line_number, NvRules.IFrontend.MarkerKind_SOURCE, file_name, NvRules.IFrontend.MsgType_MSG_WARNING)
if total_diff > 0:
message = "This kernel has uncoalesced shared accesses resulting in a total of {} excessive wavefronts ({:.0f}% of the total {} wavefronts)." \
" Check the L1 Wavefronts Shared Excessive table for the primary source locations." \
" The @url:CUDA Best Practices Guide:https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#shared-memory-in-matrix-multiplication-c-ab@ has an example on optimizing shared memory accesses." \
.format(total_diff, 100. * total_diff / total_shared_wavefronts, total_shared_wavefronts)
msg_id = fe.message(NvRules.IFrontend.MsgType_MSG_OPTIMIZATION, message)
speedup_type, speedup_value = get_estimated_speedup(metrics)
fe.speedup(msg_id, speedup_type, speedup_value)
fe.focus_metric(msg_id, metrics["excessive_wavefronts"].name(), total_diff, NvRules.IFrontend.Severity_SEVERITY_DEFAULT, "Reduce the number of excessive wavefronts in L1TEX")
fe.load_chart_from_file("UncoalescedSharedAccess.chart")
|
