2. Simple profiling: logging

Logging is one of the simplest types of profiling. It is set up with logging environment variables which control the information that is logged as the program compiles and executes. Each framework has specific environment variables that control logging.

Table 2.1 Logging environment variables

Framework

Logging environment variables

Reference in framework user guide

PyTorch and PyTorch Geometric for the IPU

POPTORCH_LOG_LEVEL

Logging level

TensorFlow 2

TF_CPP_MIN_VLOG_LEVEL, TF_CPP_VMODULE, XLA_FLAGS

Retrieving information about compilation and execution

Poplar

POPLAR_PROFILER_LOG_LEVEL

Logging level

GCL

GCL_LOG_LEVEL, GCL_API_LOG_LEVEL, GCL_LOG_DEST

Logging

PopART

POPART_LOG_LEVEL, POPART_LOG_DEST, POPART_LOG_CONFIG

Turning on execution tracing and Logging

PopXL

POPART_LOG_LEVEL, POPART_LOG_DEST, POPART_LOG_CONFIG

Turning on execution tracing and Logging

PopEF

POPEF_LOG_LEVEL, POPEF_LOG_DEST

Environment variables

Model Runtime

MODEL_RUNTIME_LOG_LEVEL, MODEL_RUNTIME_LOG_DEST

Logging