[tools] add unified runtime tracing tool #444

pbalcer · 2023-04-17T10:33:52Z

urtrace is a command-line tool for tracing and profiling Unified Runtime library calls in a process. It can be used to inspect the function arguments of each call, and provides options for filtering, output formatting, and profiling.

pbalcer · 2023-04-17T10:36:10Z

I also have a few other things in the pipeline for argument printing verbosity, but I didn't want to do them here due to potential conflicts with #411.

Example output (it will improve once #411 is merged):

unified-runtime/build/bin ‹urtrace*› » ./urtrace --null --profiling  ./hello_world 
urInit(.device_flags = 0) -> UR_RESULT_SUCCESS; (4.156us)
Platform initialized.
urPlatformGet(.NumEntries = 1, .phPlatforms = [], .pNumPlatforms = 0x7f1befb01820 (1)) -> UR_RESULT_SUCCESS; (3.148us)
urPlatformGet(.NumEntries = 1, .phPlatforms = [0x80800001], .pNumPlatforms = nullptr) -> UR_RESULT_SUCCESS; (1.033us)
urPlatformGetApiVersion(.hDriver = 0x80800001, .pVersion = 0x7f1befb018b0 (0.6)) -> UR_RESULT_SUCCESS; (1.119us)
API version: 0.6
urDeviceGet(.hPlatform = 0x80800001, .DeviceType = UR_DEVICE_TYPE_GPU, .NumEntries = 0, .phDevices = [], .pNumDevices = 0x7f1befb018c0 (1)) -> UR_RESULT_SUCCESS; (1.445us)
urDeviceGet(.hPlatform = 0x80800001, .DeviceType = UR_DEVICE_TYPE_GPU, .NumEntries = 1, .phDevices = [0x80800002], .pNumDevices = nullptr) -> UR_RESULT_SUCCESS; (1.053us)
urDeviceGetInfo(.hDevice = 0x80800002, .propName = UR_DEVICE_INFO_TYPE, .propSize = 4, .pPropValue = 0x7f1befb01960, .pPropSizeRet = nullptr) -> UR_RESULT_SUCCESS; (1.672us)
urDeviceGetInfo(.hDevice = 0x80800002, .propName = UR_DEVICE_INFO_NAME, .propSize = 1023, .pPropValue = 0x7f1befb01970, .pPropSizeRet = nullptr) -> UR_RESULT_SUCCESS; (2.3us)
Found a Null Device gpu.
urTearDown(.pParams = nullptr) -> UR_RESULT_SUCCESS; (1.044us)

igchor

LGTM

tools/urtrace/collector.cpp

PatKamin · 2023-04-18T13:58:32Z

cmake/match.cmake

 # with a known good match file.
 #

+find_package(Python3 COMPONENTS Interpreter)


This is done in the main CMakeLists.txt - do you have to run this again in this script to have python3 path defined?

Yes, because this is a separate instance of the cmake process.

add_test(NAME ${TEST_NAME} COMMAND ${CMAKE_COMMAND} -P ${PROJECT_SOURCE_DIR}/cmake/match.cmake ... )

This starts a new cmake process to run the provided script. It's unfortunate, but that's how it works. I could replace this entire match.cmake script with a python one, which would simplify things (but that's a separate PR).

Thanks, that's ok then.

cmake/match.py

PatKamin · 2023-04-19T06:54:23Z

cmake/match.cmake

 # with a known good match file.
 #

+find_package(Python3 COMPONENTS Interpreter)


Thanks, that's ok then.

tools/urtrace/urtrace

PatKamin · 2023-04-19T07:15:33Z

tools/urtrace/urtrace

@@ -0,0 +1,136 @@
+#!/usr/bin/env python
+


Could you give this file a .py extension?

No, because I want it to look like a normal binary if installed in the system.

test/CMakeLists.txt

tools/urtrace/collector.cpp

PatKamin · 2023-04-19T07:44:48Z

tools/urtrace/collector.cpp

+        ostr << dur.count() << "ns";
+    } break;
+    case TIME_UNIT_US: {
+        std::chrono::duration<double, std::micro> d = dur;


nit: Might be more readable:
std::chrono::microseconds d = dur

but it would only print whole integer microseconds, which is not useful...

Oh, I missed that, sure.

tools/urtrace/collector.cpp

PatKamin · 2023-04-19T09:31:11Z

tools/urtrace/collector.cpp

+    MAX_TIME_UNIT,
+};
+
+const char *time_unit_str[MAX_TIME_UNIT] = {"auto", "ns", "us", "ms", "s"};


Would making a time_unit and a const char *time_unit_str a map entry cause a performance hit here? It seems cleaner to use a map instead, however you won't be able to switch through all of the keys.

I disagree that using a map is cleaner. This allocates no memory and is very straightforward. It's a very common pattern for mapping enum values to strings...

tools/urtrace/urtrace

PatKamin · 2023-04-19T10:36:42Z

tools/urtrace/urtrace

+if args.debug:
+    print(result)
+
+exit(result.returncode)


No empty line at EOF.

That's not true? Git shows something like 'no newline at end of file' and on github in PRs there's a red bar if that's the case.

That's what I meant, to add an empty newline, sorry for misconfusion.

urtrace is a command-line tool for tracing and profiling Unified Runtime library calls in a process. It can be used to inspect the function arguments of each call, and provides options for filtering, output formatting, and profiling.

PatKamin · 2023-04-19T13:32:26Z

tools/urtrace/collector.cpp

+        ostr << dur.count() << "ns";
+    } break;
+    case TIME_UNIT_US: {
+        std::chrono::duration<double, std::micro> d = dur;


Oh, I missed that, sure.

PatKamin · 2023-04-19T13:34:02Z

tools/urtrace/urtrace

+if args.debug:
+    print(result)
+
+exit(result.returncode)


That's what I meant, to add an empty newline, sorry for misconfusion.

pbalcer mentioned this pull request Apr 17, 2023

[ur] Improve logger env var handling #438

Merged

igchor reviewed Apr 17, 2023

View reviewed changes

tools/urtrace/collector.cpp Outdated Show resolved Hide resolved

tools/urtrace/collector.cpp Outdated Show resolved Hide resolved

pbalcer force-pushed the urtrace branch 2 times, most recently from 084b45e to e0c026c Compare April 18, 2023 08:56

veselypeta approved these changes Apr 18, 2023

View reviewed changes

PatKamin reviewed Apr 18, 2023

View reviewed changes

PatKamin reviewed Apr 19, 2023

View reviewed changes

pbalcer force-pushed the urtrace branch 2 times, most recently from f83e33d to 549f701 Compare April 19, 2023 10:29

PatKamin reviewed Apr 19, 2023

View reviewed changes

pbalcer force-pushed the urtrace branch 2 times, most recently from 1b3cb44 to 9ce3535 Compare April 19, 2023 12:54

[tools] add unified runtime tracing tool

ef2ce32

urtrace is a command-line tool for tracing and profiling Unified Runtime library calls in a process. It can be used to inspect the function arguments of each call, and provides options for filtering, output formatting, and profiling.

pbalcer force-pushed the urtrace branch from 9ce3535 to ef2ce32 Compare April 19, 2023 13:08

PatKamin approved these changes Apr 19, 2023

View reviewed changes

pbalcer merged commit ed77900 into oneapi-src:main Apr 19, 2023

[tools] add unified runtime tracing tool #444

[tools] add unified runtime tracing tool #444

Uh oh!

Conversation

pbalcer commented Apr 17, 2023

Uh oh!

pbalcer commented Apr 17, 2023

Uh oh!

igchor left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pbalcer Apr 19, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

pbalcer Apr 19, 2023 •

edited

Loading