* Updates:
- [API/CLI] rsmi_dev_*_partition_set &
rsmi_dev_*_partition_reset - exposed RSMI_STATUS_BUSY for
EBUSY writes + cleaned up accidental map insertions
(maplookup[] can insert values that are not in the map,
map.at(key) fixes this potential issue)
- [API] rsmi_dev_gpu_metrics_info_get() - returns
RSMI_STATUS_NOT_SUPPORTED for unsupported metric tables
outside of 1v1/1v2/1v3
- [API] writeDevInfoStr() - exposes RSMI_STATUS_BUSY for
EBUSY write errors; kept backward compatibility
for other writes which do not care about these states
- [API] rsmi_dev_od_volt_info_get()
& rsmi_dev_od_volt_curve_regions_get() have better logging
+ Expose more details on why they are erroring
- [Utils/logs/example] Expose AMD GPU gfx target version to aid in
system troubleshooting
- [Utils] Added test methods that look at od volt
freq & regions into here - for easier access across
several tests
- [Utils] Updated getRSMIStatusString(new argument - fullstatus;
default to true for backwards compatibility)
-> true shows shortened RSMI STATUS response
- [Utils] Added splitString to cut out noisy return responses
(used in getRSMIStatusString(), when fullstatus = true)
- [Utils] Added getFileCreationDate() to expose build date
of the library - helpful for local builds or experimental builds
- [Utils] Macro cleanup
- [Example] Added a few gpu_metric checks - helpful for upcoming
updates
- [Device] SYSFS/DebugFS - now have better r/w displayed in logs
- [LOGS] Expose library build date - see above for details
- [Tests] Add more warnings/errors to test builds
- [Tests] Moved up Partition tests for ordered test runs - helped
identify issues with GPU BUSY writes
- [Tests] compute_partition_read_write - handles RSMI_STATUS_BUSY
with waits for busy status found & cleaned up how we checked
for partition changes - with RSMI responses exposed more clearly
- [Tests] perf_determinism - multi gpu now properly runs through
with full resets as needed
- [Tests] volt_freq_curv_read - better error handling with more
verbose output
Change-Id: Ie94c6abb6a9aab95c345996d3ad3843cf6734977
Signed-off-by: Charis Poag <Charis.Poag@amd.com>
[ROCm/amdsmi commit: 57b6135e54]
ROCm System Management Interface (ROCm SMI) Library
The ROCm System Management Interface Library, or ROCm SMI library, is part of the Radeon Open Compute ROCm software stack . It is a C library for Linux that provides a user space interface for applications to monitor and control GPU applications.
For additional information refer to ROCm Documentation
DISCLAIMER
The information contained herein is for informational purposes only, and is subject to change without notice. In addition, any stated support is planned and is also subject to change. While every precaution has been taken in the preparation of this document, it may contain technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to update or otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties with respect to the accuracy or completeness of the contents of this document, and assumes no liability of any kind, including the implied warranties of noninfringement, merchantability or fitness for particular purposes, with respect to the operation or use of AMD hardware, software or other products described herein.
© 2022-2023 Advanced Micro Devices, Inc. All Rights Reserved.
Building ROCm SMI
Additional Required software for building
In order to build the ROCm SMI library, the following components are required. Note that the software versions listed are what was used in development. Earlier versions are not guaranteed to work:
- CMake (v3.5.0)
- g++ (5.4.0)
In order to build the latest documentation, the following are required:
- Python 3.8+
- NPM (sass)
The source code for ROCm SMI is available on Github.
After the ROCm SMI library git repository has been cloned to a local Linux machine, building the library is achieved by following the typical CMake build sequence. Specifically,
mkdir -p build
cd build
cmake ..
make -j $(nproc)
# Install library file and header; default location is /opt/rocm
make install
The built library will appear in the build folder.
To build the rpm and deb packages follow the above steps with:
make package
Documentation
The following is an example of how to build the docs:
python3 -m venv .venv
.venv/bin/python3 -m pip install -r docs/.sphinx/requirements.txt
.venv/bin/python3 -m sphinx -T -E -b html -d docs/_build/doctrees -D language=en docs docs/_build/html
Building the Tests
In order to verify the build and capability of ROCm SMI on your system and to see an example of how ROCm SMI can be used, you may build and run the tests that are available in the repo. To build the tests, follow these steps:
mkdir build
cd build
cmake -DBUILD_TESTS=ON ..
make -j $(nproc)
To run the test, execute the program rsmitst that is built from the steps above.
Usage Basics
Device Indices
Many of the functions in the library take a "device index". The device index is a number greater than or equal to 0, and less than the number of devices detected, as determined by rsmi_num_monitor_devices(). The index is used to distinguish the detected devices from one another. It is important to note that a device may end up with a different index after a reboot, so an index should not be relied upon to be constant over reboots.
Hello ROCm SMI
The only required ROCm-SMI call for any program that wants to use ROCm-SMI is the rsmi_init() call. This call initializes some internal data structures that will be used by subsequent ROCm-SMI calls.
When ROCm-SMI is no longer being used, rsmi_shut_down() should be called. This provides a way to do any releasing of resources that ROCm-SMI may have held. In many cases, this may have no effect, but may be necessary in future versions of the library.
A simple "Hello World" type program that displays the device ID of detected devices would look like this:
#include <stdint.h>
#include "rocm_smi/rocm_smi.h"
int main() {
rsmi_status_t ret;
uint32_t num_devices;
uint16_t dev_id;
// We will skip return code checks for this example, but it
// is recommended to always check this as some calls may not
// apply for some devices or ROCm releases
ret = rsmi_init(0);
ret = rsmi_num_monitor_devices(&num_devices);
for (int i=0; i < num_devices; ++i) {
ret = rsmi_dev_id_get(i, &dev_id);
// dev_id holds the device ID of device i, upon a
// successful call
}
ret = rsmi_shut_down();
return 0;
}