hipbaslt-patch-series-perf.md

Load Logic Perf

Subject	Time	Peak Memory	vs Baseline
[hipblaslt] Use multiprocessing.Pool for TensileCreateLibrary	18.531s	879.0 MB	baseline
[hipblaslt] CustomKernels: lru_cache for 20x speedup of some logic files	18.275s	878.3 MB	0% faster, 0% less memory
[hipblaslt] reduce memory usage during logic load	15.500s	873.7 MB	15% faster, 0% less memory
[hipblaslt] Remove unused key arg for getPrimitiveParameterValueAbbreviation	15.535s	873.4 MB	15% faster, 0% less memory
[hipblaslt] intern strings to reduce duplicate memory for solution keys	15.001s	721.9 MB	17.5% faster, 17.5% less memory
[hipblaslt] tensilelite: gc in parallel worker	15.167s	722.1 MB	17.5% faster, 17.5% less memory
[hipblaslt] tensilelite: teach state_key_ordering slots	14.827s	722.5 MB	20% faster, 17.5% less memory
[hipblaslt] tensilelite: intern FreeIndex, BatchIndex, BoundIndex and SizeMapping	14.970s	704.1 MB	17.5% faster, 20% less memory
[hipblaslt] tensilelite: remove unused targetObjFilename	14.998s	702.3 MB	17.5% faster, 20% less memory
[hipblaslt] tensilelite: record code object file index without mutations	14.448s	584.1 MB	22.5% faster, 32.5% less memory

---

Full ISA TensileCreateLibrary

Time before

Total time (s): 4630.94
Total kernels processed: 196454
Kernels processed per second: 42.42
KernelHelperObjs: 328
# Peak memory ~34GB not logged

Time After

Total time (s): 3418.73
Total kernels processed: 196454
Kernels processed per second: 57.46
KernelHelperObjs: 328
Peak memory usage (MB): 24,266.2
Current memory usage (MB): 21,930.3