[rocprofv3-benchmark] SDK and rocprofv3 Benchmarking Suite (#157)
* Adding Benchmarking Stg1
* config fix
* reset
* add jpeg and decode traces in iteration
* address comments benchmark config files.
* address comments.
* address comments.
* address comments: revert cntrl ctx.
* address comments: revert csv output.
* resolve merge conflits.
* format.
* build fix.
* fix hip runtime api traces.
* loop cb services.
* format.
* bug fix.
* Fix operator>
- public C++ comparison operator
* Update configuration options
- support selected regions (--selected-regions)
- support writing output config json (--output-config)
- update serialization data
* rocprofv3 tool library misc updates
- lambda for starting context
- support for writing config json
* Tool library updates
- Finished support for all benchmarking modes
- Added build spec support to config json
* Fix ROCPROFILER_SOVERSION
- this value should not be multiplied by 10,000
* Minor tweak to rocprofv3
* Benchmarking scripts
* formatting
* Fix duplicate include
* Add reproducible-dispatch-count test app
- used in benchmarking
* registration logging
- report number of registered contexts and active contexts after client initialization
* Serialize environment in rocprofv3 output config
* ROCPROFILER_BUILD_BENCHMARK CMake option
* Update benchmark SQL schema
- hash_id is text
- add md5sum to benchmarked_app
- remove app_id from benchmarked_sdk
- add sdk_id to benchmark_config
- separate hip_trace into hip_runtime_trace and hip_compiler_trace
- use INT instead of INTEGER for MySQL compatibility
- add count column in benchmark_statistics
- allow std_dev to be NULL in benchmark_statistics
* Update rocprofv3-benchmark.py
- use md5 instead of python hash (which includes random seed)
- use args.mysql_database
- compute md5sum of executable
- fix insert_benchmark_config
- marker trace fixes
- memory allocation fixes
- split hip_trace into hip_{runtime,compiler}_trace
- remove app_id from benchmarked_sdk
- support warmup runs
- count field in benchmark_statistics
* Support launcher and environment in YAML
* Update reproducible-dispatch-count.cpp
- support mode which doesn't use hip event timing
* Misc rocprofv3-benchmark.py updates
- fix some MySQL support
- remove some unnecessary logging
* support mysql db.
* Format.
* Updated SQL input files
- moved benchmark_schema.sql to benchmark_table.sql
- added benchmark_views.sql
- uses {{metric}} syntax for variable substitution
* cmake formatting
* update rocprofv3-benchmark.py
- benchmark config labels
- overhead views
* Encode rocprofv3-benchmark PID in rocprofv3 and timem output files
* Minor tweak to benchmark_views.sql
- include count
- reorder fields for readability
* split statements and use IS if values is NONE.
* use backtick instead of double quotes and add IS before NOT NULL.:
* Adding Mandelbrot Benchmark App
* Adding Dockerfile example
* Update dockerfile
* Update dockerfile
* [SDK] rocprofiler_query_external_correlation_id_request_kind_name
* Execution-profile benchmark mode
* Execution profile SQL support
* Rename mandlebrot folder + misc clang-tidy
* [rocprofv3-benchmark] Execution profile support
* Update installation
* add work dir when setting git revision, useful when building outside src.
* Set FULL_VERSION_STRING and ROCPROFILER_SDK_GIT_REVISION
- when benchmark folder is top-level
* Remove unused python packages from requirements.txt
* Use ldd/pyelftools to include linked libs for md5sum
- also add --filter-benchmark and --filter-rocprofv3 options
- support labeling the rocprofv3 options
- use more argparse groups
- more generic application of filters
- support variable substitution in environment, e.g. PATH=/some/path:$PATH
* Environment improvements
- improve reproducibility when env set via input file vs. shell
- support "environment-ignore" to remove environment variables
* Misc formatting
* Misc. fix
* use backticks for defining new columns name
* Support shuffling the order of benchmark modes/rocprofv3 args
* Address review comments
* Update Dockerfile
- rename to Dockerfile
- reduce to one layer
* Support docker build arg BRANCH
---------
Co-authored-by: Ammar ELWazir <aelwazir@amd.com>
Co-authored-by: Kandula, Venkateshwar reddy <Venkateshwarreddy.Kandula@amd.com>
Co-authored-by: Venkateshwar Reddy Kandula <vkandula@amd.com>
Co-authored-by: Madsen, Jonathan <Jonathan.Madsen@amd.com>
Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
This commit is contained in:
Elwazir, Ammar
GitHub
@@ -248,6 +248,11 @@ For MPI applications (or other job launchers such as SLURM), place rocprofv3 ins
|
||||
choices=("csv", "json", "pftrace", "otf2"),
|
||||
type=str.lower,
|
||||
)
|
||||
add_parser_bool_argument(
|
||||
io_options,
|
||||
"--output-config",
|
||||
help="Generate a output file of the rocprofv3 configuration, e.g. out_config.json",
|
||||
)
|
||||
io_options.add_argument(
|
||||
"--log-level",
|
||||
help="Set the desired log level",
|
||||
@@ -518,6 +523,11 @@ For MPI applications (or other job launchers such as SLURM), place rocprofv3 ins
|
||||
type=str,
|
||||
choices=("hour", "min", "sec", "msec", "usec", "nsec"),
|
||||
)
|
||||
add_parser_bool_argument(
|
||||
filter_options,
|
||||
"--selected-regions",
|
||||
help="If set, rocprofv3 will only profile regions of code surrounded by roctxProfilerResume(0) and roctxProfilerPause(0)",
|
||||
)
|
||||
|
||||
perfetto_options = parser.add_argument_group("Perfetto-specific options")
|
||||
|
||||
@@ -592,6 +602,19 @@ For MPI applications (or other job launchers such as SLURM), place rocprofv3 ins
|
||||
"--realpath",
|
||||
help=argparse.SUPPRESS,
|
||||
)
|
||||
advanced_options.add_argument(
|
||||
"--benchmark-mode",
|
||||
choices=(
|
||||
"disabled-sdk-contexts",
|
||||
"sdk-buffer-overhead",
|
||||
"sdk-callback-overhead",
|
||||
"tool-runtime-overhead",
|
||||
"execution-profile",
|
||||
),
|
||||
help=argparse.SUPPRESS,
|
||||
default=None,
|
||||
type=str.lower,
|
||||
)
|
||||
advanced_options.add_argument(
|
||||
"-A",
|
||||
"--agent-index",
|
||||
@@ -1034,6 +1057,7 @@ def run(app_args, args, **kwargs):
|
||||
|
||||
update_env("ROCPROF_OUTPUT_FILE_NAME", _output_file)
|
||||
update_env("ROCPROF_OUTPUT_PATH", _output_path)
|
||||
update_env("ROCPROF_OUTPUT_CONFIG_FILE", args.output_config, overwrite_if_true=True)
|
||||
if app_pass is not None and args.sub_directory is not None:
|
||||
app_env["ROCPROF_OUTPUT_PATH"] = os.path.join(
|
||||
f"{_output_path}", f"{args.sub_directory}{app_pass}"
|
||||
@@ -1198,6 +1222,11 @@ def run(app_args, args, **kwargs):
|
||||
args.truncate_kernels,
|
||||
overwrite_if_true=True,
|
||||
)
|
||||
update_env(
|
||||
"ROCPROF_SELECTED_REGIONS",
|
||||
args.selected_regions,
|
||||
overwrite_if_true=True,
|
||||
)
|
||||
|
||||
if args.list_avail:
|
||||
update_env(
|
||||
@@ -1241,6 +1270,24 @@ def run(app_args, args, **kwargs):
|
||||
args.log_level,
|
||||
)
|
||||
|
||||
if args.benchmark_mode:
|
||||
if args.benchmark_mode == "execution-profile":
|
||||
if args.group_by_queue is None:
|
||||
update_env(
|
||||
"ROCPROF_GROUP_BY_QUEUE",
|
||||
False,
|
||||
overwrite=True,
|
||||
)
|
||||
elif args.group_by_queue:
|
||||
fatal_error(
|
||||
"rocprofv3 requires --group-by-queue=false for --benchmark-mode=execution-profile"
|
||||
)
|
||||
|
||||
update_env(
|
||||
"ROCPROF_BENCHMARK_MODE",
|
||||
args.benchmark_mode,
|
||||
)
|
||||
|
||||
for opt, env_val in dict(
|
||||
[
|
||||
["kernel_rename", "KERNEL_RENAME"],
|
||||
|
||||
مرجع در شماره جدید
Block a user