Previously, during the rsmi_init discovery process, the existence
of an hwmon# directory was used to distinguish between gpus nodes
and non-gpu nodes. This isn't reliable in some scenarios. Instead,
the existence of the vbios_version file is used as an
indicator that the node is indeed a gpu.
Change-Id: Icfbe5c42ed0970077b05f25c3d209308a31bec85
The purpose of this patch is to implement GPU reset functionality
in the LIB, and to call it from the rocm_smi python CLI.
Signed-off-by: Ori Messinger <Ori.Messinger@amd.com>
Change-Id: Iaf525f7016f8354a7fd93af0209ca2e97ef4fd56
A special mode of operation to achieve minimal performance variation by letting
the user have the ability to provide the desired frequency to be set as the soft limit.
The user can control the entry and exit to the mode via rocm-smi a mechanism to
enter / exit performance determinism mode as below.
Enter performance determinism mode:
- hold a lock
- write performance_determinism to power_dpm_force_performance_level
- write input clk_freq to pp_dpm_sclk
- release lock
Exit performance determinism_mode:
- hold a lock
- write auto to power_dpm_force_performance_level
- release lock
Signed-off-by: Divya Shikre <DivyaUday.Shikre@amd.com>
Change-Id: Ia24e27954cdf1c4337ffc83d8948fbdfaf4552d2
WKS and server cards have an FRU with product information, so try to use
that for product name and product SKU if it exists.
Signed-off-by: Kent Russell <kent.russell@amd.com>
Change-Id: I40bbd3bf62f4cb02e96015ed1630112691cacbc3
* Don't make different shared memory mutexes for different users
* Don't delete (unlink) the shared mutex file if the mutex
initialization fails. This may mess up other processes that
are using it. Instead, print a message on how to resolve the
situation, and then throw an error.
Note, this situation comes up when debug builds (usually)
either assert() or otherwise end execution without a proper
clean up.
* Remove cpplint from shared_mutex code
Change-Id: I5f8ca6150cac5c2405fb97007516da345093f966
Given a device index, return the corresponding NUMA node for the
device.
Also, add NUMA node tests to Sys Info Read test.
Change-Id: I0df4937470e6362e6737ccea568d4b3e5890c91a
This solution takes into account that some hwmons use
label files to map sensor types. The previous solution
did not take this into account.
Change-Id: I1d6204573cefa8197b2cfe0ffb412b545df3d80a
Given a process ID, give the device indices that process is
currently using.
Also:
* made corrections to how RSMI, amdgpu (ie, "card#") and
KFD indicies translate from one another
* add a few missing error codes to rsmi_status_string()
* fix some formatting
Change-Id: Icd2cae66bb4fec768da96af7cf9cf8b8b66ec7f9
Improvements include
* adding additional build flags that warn about stack-smashing
and type conversion errors
* run-time checks for valid function input values and adquate
space for the result of arithmetic operations.
* make sure default case for switch statements do something
besides just assert
* disable using env. var. debugging in release mode
Change-Id: I5f048310c5c56e05d9ec31bcc273404d6a0dd646
For device-getter functions, allow users to specify a nullptr
for the provided buffer. In those cases, the function will return
RSMI_STATUS_NOT_SUPPORTED if the hardware or system software does
not support the function. If the function is supported, then
RSMI_STATUS_INVALID_ARGS will be returned, unless a different
error is encountered.
Additionally, tests and documentation were updated to reflect
this change.
Change-Id: Ie7db3a4c8c66af97ebd7ee1e3b95cd331ace9d9c
Add support and testing for reading the vram vendor associated with
the GPU. The vram vendor can be found as a separate sysfs file at:
/sys/class/drm/card[X]/device/mem_info_vram_vendor
The vram vendor is displayed as a string value.
Change-Id: I12c8e56e57f45aa08d7d6c25338c4e468ed1c7fc
Signed-off-by: Ori Messinger <Ori.Messinger@amd.com>
The new functions added in this commit allow a caller to tell up
front what functions, function variants and monitors are
supported.
Also,
* fixed a few documentation/formatting issues
* fixed a process_info test issue
Change-Id: I2184ab1a4a6898f847e791f273e2185d556e78e9
If the 32-bit domain is found in the kfd node properties for
a device, then it will be used when constructing the bdfid.
If it's not present, it will continue to use the 16 bit version.
Also, whether or not 32b or 16b are used for the domain, the
domain will now be placed in the upper 32b of the 64b bdfid.
* Fixed some unrelated doxygen issues
Change-Id: Icb5116daa1ab45ee305bdbe6cd5df5736dd3ffa3
Also, don't return an error for empty sysfs files. The reserved memory
page file will often have no lines. We don't want it to appear that
this function is not supported if the file is empty.
Change-Id: I1d28bb184ea587bb578fe71dd75adc2a812d09a8
Added implementation of and tests for
rsmi_dev_compute_process_info_by_pid_get() and
rsmi_dev_compute_process_info_get()
Change-Id: I4c4f5f39fe6701da37916c9ad41449b5d35ac7af
Add support and testing for reading the Unique ID associated with a
specific GPU. This ID will persist across reboots, even if the GPU is
moved to a different machine. Note that this is per-GPU, not per-card,
as some cards have multiple GPUs, and each GPU will get a unique
identifier
Change-Id: Idce50c6febc2ceb1a4c1200d2489ec8b9d8fe174
This commit uses a pthread mutex in shared memory to prevent
almost all cases of multiple processes simultaneously
reading/writing to device sysfs files. The main existing race
condition is when 2 processes are starting at the same time,
setting up their shared memory and mutexes. Since this is meant
to prevent collisions among thread and processes, the small
shared memory segments (big enough for a pthread_mutex) will
persist until reboot.
Chris Freehill
Ori Messinger
Divya Shikre
Kent Russell
Mukul Joshi