Significant Performance Regression from 2017(est) vs latest.
Zelex opened this issue · 3 comments
We tried integrating the latest tiny_exr, but we saw a 2x performance degradation vs an older version of the lib.
Old version is roughly.... https://github.com/syoyo/tinyexr/blob/c2f7635a1d7653915b026a1f046bfc08a246cb1a/tinyexr.h
Some of the differences I notice vs newer is are
- latest has unaligned memory access protection via reading/writing individual bytes
- latest has more std::vector usage vs just flat C arrays
I'm sure there's more possible things to look into, but those are what I remember while I am typing this.
latest has unaligned memory access protection via reading/writing individual bytes
This is required to support loading data in unaligned address.
latest has more std::vector usage vs just flat C arrays
This is required to avoid stack overflow. Optimizing std::vector may be possible by writing custom lock free STL allocator
I think first figuring out why the 2019 build slowed down so much would be worthwhile.
This is required to support loading data in unaligned address.
I would do some profiling to see if this is where the slowdown is, but a 2x speed hit for unaligned might be better on an optional TINYEXR_ALLOW_UNALIGNED switch.
Nothing to do without reproducible build/run procedure and benchmarks