1.5. Multi-level Logging¶
Similar to rocBLAS logging, rocSOLVER provides logging facilities that can be used to output information on rocSOLVER function calls. Three modes of logging are supported: trace logging, bench logging, and profile logging.
Note that performance will degrade when logging is enabled.
Table of contents
Trace logging outputs a line each time an internal rocSOLVER or rocBLAS routine is called, outputting the function name and the values of its arguments (excluding stride arguments). The maximum depth of nested function calls that can appear in the log is specified by the user.
Bench logging outputs a line each time a public rocSOLVER routine is called (excluding
auxiliary library functions), outputting a line that can be used with the executable
rocsolver-bench to call the function with the same size arguments.
Profile logging, upon calling
or terminating the logging session using
rocsolver_log_end, will output statistics on each
called internal rocSOLVER and rocBLAS routine. These include the number of times each function
was called, the total program runtime occupied by the function, and the total program runtime
occupied by its nested function calls. As with trace logging, the maximum depth of nested output
is specified by the user. Note that, when profile logging is enabled, the stream will be synchronized
after every internal function call.
In order to use rocSOLVER’s logging facilities, the user must first call
in order to allocate the internal data structures used for logging and begin the logging session.
The user may then specify a layer mode and max level depth, either programmatically using
rocsolver_log_set_max_levels, or by setting the corresponding
The layer mode specifies which logging type(s) are activated, and can be
or a bitwise combination of these. The max level depth specifies the default maximum depth of nested
function calls that may appear in the trace and profile logging.
Both the default layer mode and max level depth can be specified using environment variables.
If these variables are not set, the layer mode will default to
rocblas_layer_mode_none and the
max level depth will default to 1. These defaults can be restored by calling the function
ROCSOLVER_LAYER is a bitwise OR of zero or more bit masks as follows:
ROCSOLVER_LAYERis not set, then there is no logging
(ROCSOLVER_LAYER & 1) != 0, then there is trace logging
(ROCSOLVER_LAYER & 2) != 0, then there is bench logging
(ROCSOLVER_LAYER & 4) != 0, then there is profile logging
Three environment variables can set the full path name for a log file:
ROCSOLVER_LOG_TRACE_PATHsets the full path name for trace logging
ROCSOLVER_LOG_BENCH_PATHsets the full path name for bench logging
ROCSOLVER_LOG_PROFILE_PATHsets the full path name for profile logging
If one of these environment variables is not set, then
ROCSOLVER_LOG_PATH sets the full path
for the corresponding logging, if it is set. If neither the above nor
set, then the corresponding logging output is streamed to standard error.
The results of profile logging, if enabled, can be printed using
rocsolver_log_flush_profile. Once logging facilities are no longer required (e.g. at
program termination), the user must call
rocsolver_log_end to free the data structures used
for logging. If the profile log has not been flushed beforehand, then
will also output the results of profile logging.
For more details on the mentioned logging functions, see the Logging functions section on the rocSOLVER API document.
Code examples that illustrate the use of rocSOLVER’s multi-level logging facilities can be found
in this section or in the
example_logging.cpp file in the
The following example shows some basic use: enabling trace and profile logging, and setting the max depth for their output.
// initialization rocblas_handle handle; rocblas_create_handle(&handle); rocsolver_log_begin(); // begin trace logging and profile logging (max depth = 5) rocsolver_log_set_layer_mode(rocblas_layer_mode_log_trace | rocblas_layer_mode_log_profile); rocsolver_log_set_max_levels(5); // call rocSOLVER functions... // terminate logging and print profile results rocsolver_log_flush_profile(); rocsolver_log_end(); rocblas_destroy_handle(handle);
Alternatively, users may control which logging modes are enabled by using environment variables.
The benefit of this approach is that the program does not need to be recompiled if a different
logging environment is desired. This requires that
rocsolver_log_set_max_levels are not called in the code, e.g.
// initialization rocblas_handle handle; rocblas_create_handle(&handle); rocsolver_log_begin(); // call rocSOLVER functions... // termination rocsolver_log_end(); rocblas_destroy_handle(handle);
The user may then set the desired logging modes and max depth on the command line as follows:
export ROCSOLVER_LAYER=5 export ROCSOLVER_LEVELS=5
Kernel launches from within rocSOLVER can be added to the trace and profile logs using an
additional layer mode flag. The flag
rocblas_layer_mode_ex_log_kernel can be combined with
rocblas_layer_mode flags and passed to
rocsolver_log_set_layer_mode in order to enable
kernel logging. Alternatively, the environment variable
ROCSOLVER_LAYER can be set such that
(ROCSOLVER_LAYER & 16) != 0:
(ROCSOLVER_LAYER & 17) != 0, then kernel calls will be added to the trace log
(ROCSOLVER_LAYER & 20) != 0, then kernel calls will be added to the profile log
The logging facilities for rocSOLVER assume that each
rocblas_handle is associated with at
most one host thread. When using rocSOLVER’s multi-level logging setup, it is recommended to
create a separate
rocblas_handle for each host thread.
The rocsolver_log_* functions are not thread-safe. Calling a log function while any rocSOLVER routine is executing on another host thread will result in undefined behaviour. Once enabled, logging data collection is thread-safe. However, note that trace logging will likely result in garbled trace trees if rocSOLVER routines are called from multiple host threads.