Performance enhancements #139
Conversation
|
Could you please rebase onto experimental? |
Odd, I thought this was based on current experimental, it looks like that to me? |
|
Any updates here? |
|
Sorry - nothing much has happened this month - other distractions kicked in! I'll try and get back to this next week. |
CMakeLists.txt
Outdated
| app/calcmask.cc | ||
| app/utils.cc | ||
| ) | ||
| # special compile options for blend.cc to use SIMD optimiser | ||
| set_source_files_properties(app/blend.cc PROPERTIES COMPILE_OPTIONS "-march=native;-O9") |
There was a problem hiding this comment.
I'm not sure.. I guess that if one module is targetted / optimised for the current CPU, then it may as well all be. Initially I only wanted to mess with / break one specific part of the code. I was also thinking about compiling run-time swappable implementations for different targets (this gets talked about a bit in the GCC manuals when using intrinsics) - that is probably over-engineering.
There was a problem hiding this comment.
I think as we are compiling all 3rd party deps ourselves we may as well compile everything for -march=native. Let's split this into another PR, so we could merge this one for now.
Having run-time swappable implementations is probably overkill for us right now. Let's handle this separately.
|
Please try to always rebase on the target branch instead of merging things … |
Specifically to
alpha_blendandconvert_rgb_to_yuyvas per #138This first draft addresses
alpha_blend, reducing run time by a factor of 3-4x simply by separating the function and turning on full compiler optimisation for the machine that we are compiling on (-march=native). This may be a problem if compiling for a different target that doesn't have the same CPU features...