Updates in Nsight Python 0.9.6#

Enhancements#

Github Issue #11: Added support for multiple derived metrics in nsight.analyze.kernel() using derive_metric parameter. The derive_metric function can now return either a single value or a dictionary of multiple metrics.
Added metric parameter to nsight.analyze.plot() decorator to specify which metric to visualize when multiple metrics are collected.
Normalization improvements in nsight.analyze.kernel():
- Changed normalize_against to use standard normalization (current/baseline) instead of appending normalization info to metric names.
- Added Normalized column to the output dataframe to indicate which annotation is used for normalization.
Adaptive Thermovision in nsight.analyze.kernel():
- Replaced thermal_control boolean parameter with thermal_mode parameter that accepts "auto", "manual", or "off" values for more flexible thermal throttling control.
- Added thermal_wait parameter to specify the thermal headroom threshold (T.Limit in °C) that triggers cooling pause.
- Added thermal_cont parameter to specify the thermal headroom threshold (T.Limit in °C) to resume profiling after cooling.
- Added thermal_timeout parameter to specify the maximum wait time in seconds for GPU to cool down.

Github Issue #13: Fixed incorrect profiling results when making multiple function calls of the same decorated function.
Github Issue #17: Fixed ZeroDivisionError when handling zero-valued metrics (e.g., sm__idc_divergent_instructions.avg). Added test coverage for zero-valued metrics to ensure proper handling.