Many of MIOpen kernels have parameters which affect their performance. Setting these parameters to optimal values allows reaching the best possible throughput. These optimal values depend on many things, including network configuration, GPU type, clock frequencies, ROCm version etc. Because of these dependencies and also due to enormous number of possible network configurations, it is virtually impossible to supply all values that users may need together with the library. Instead, MIOpen provides a set of pre-tuned values for the most applicable network configurations, and also means for expanding the set of optimized values. MIOpen's performance database contains these pre-tuned parameter values as well as optimized parameters tuned by users.
The performance database consists of two parts:
- System Performance Database, a system-wide storage which holds the pre-tuned values for the most applicable configurations,
- User Performance Database, a per-user storage which is intended to hold optimized values for arbitrary configurations.
User PerfDb always takes precedence over System PerfDb.
MIOpen also has auto-tuning functionality, which is able to find optimized kernel parameter values for a specific configuration. The auto-tune process may take a substantial amount of time, however, once the optimized values are found, they are stored in the User PerfDb. MIOpen then will automatically read and use these parameter values when needed again instead of running the expensive auto-tuning search.
By default, System PerfDb resides within MIOpen's install location, while User PerfDb resides in the user's home directory. See Setting up locations for more information.
The System PerfDb is not modified upon installation of MIOpen.
Auto-tuning the kernels.¶
MIOpen performs auto-tuning during the following MIOpen API calls:
During the call, auto-tuning is performed only for one problem configuration (implicitly defined by the tensor descriptors passed to API function).
The following conditions must be met for the auto-tune to begin:
- The applicable kernel(s) has tuning parameters.
- The passed value of
- Both System and User PerfDb do not yet contain values for the relevant problem configuration.
The latter two conditions may be overridden by enforcing the search by means of the following environment variables:
These variables may also be used for removing values from User PerfDb, see below.
Both symbolic (case-insensitive) and numeric values are supported.
Setting the value to "NONE", or "1" will have no change in the default behavior.
Auto-tune will not be skipped even if PerfDb already contains optimized values. If auto-tune is requested via API, then MIOpen will perform it and update PerfDb.
This mode can be used for fine-tuning the MIOpen installation on the user's system. When MIOpen is in this mode, the applications that use it may take quite long to finish.
MIOpen will perform auto-tune even if not requested via MIOpen API. In other words, the library will behave as if
exhaustiveSearch parameter set to
true even this is not really so. If optimized values already reside in PerfDb, then auto-tune will not be performed.
This mode allows for tuning the apps that do not anticipate means for getting the best performance from MIOpen. When MIOpen is in this mode, the first run of the user's app may take substantially longer time than expected.
A combination of SEARCH and DB_UPDATE. MIOpen performs auto-tune (and updates User PerfDb) on each
miopenFindConvolution*() call. It is not recommended to use this mode except for debugging purposes.
Use with care. MIOpen removes optimized values related to given problem configuration from the User PerfDb. Auto-tune is blocked, even if it is explicitly requested. System PerfDb left intact.
This variable allows for limiting the scope of
MIOPEN_FIND_ENFORCE, so that only forward, backward data or backward weights convolutions will be affected. Both symbolic (case-insensitive) and numeric values are supported, as shown below.
MIOPEN_FIND_ENFORCE affects all convolutions. This is the default.
MIOPEN_FIND_ENFORCE affects only Forward convolutions.
MIOPEN_FIND_ENFORCE affects only Backward Data convolutions.
MIOPEN_FIND_ENFORCE affects only Backward With Regard to Weights (a.k.a. WRW) convolutions.
Updating MIOpen and the User Db¶
It is important to note that if the user installs a new version of MIOpen, it is recommended that the user move, or delete their old user performance database file. This will prevent older database entries from poluting the configurations shipped with the newer system database. The user perf db is named
miopen.udb and is located at the user perf db path.