some torch.profiles come with truncated event names which leads to not finding corresponding kernel categories or ideal cycles (if requested and passed via log file). The problem is about to get fixed at the source. In the meantime, this patch can help:
diff --git a/src/aiu_trace_analyzer/pipeline/rcu_utilization.py b/src/aiu_trace_analyzer/pipeline/rcu_utilization.py
index 3013a85..4eb0b25 100644
--- a/src/aiu_trace_analyzer/pipeline/rcu_utilization.py
+++ b/src/aiu_trace_analyzer/pipeline/rcu_utilization.py
@@ -23,6 +23,7 @@ RCU_pt_util_counter_unit = "Percent"
_default_fprint_len = 25
_kernel_db_feature_implemented = False
+_kernel_name_trunc_limit = 63
class RCUTableFingerprint():
def __init__(self, datalimit: int = -1, initial_data: str = ''):
@@ -184,7 +185,7 @@ class RCUUtilizationContext(AbstractContext, PipelineContextTool):
category = "NotAvailable"
else:
category = "Total"
- kernel = kernel_and_cat[0]+" Cmpt Exec"
+ kernel = kernel_and_cat[0][:_kernel_name_trunc_limit]+" Cmpt Exec"
fprint.add(kernel)
if kernel not in current_table:
some torch.profiles come with truncated event names which leads to not finding corresponding kernel categories or ideal cycles (if requested and passed via log file). The problem is about to get fixed at the source. In the meantime, this patch can help: