[Question] What optimizations are enabled for GPU in O3 over O2 optimization levels?

[sanchitintel]

Hi,

Can you please point me to a list of compiler optimizations that are enabled for O3 but not O2 for GPUs?

Thanks!

[AlexeySachkov]

Hi @sanchitintel, at intel/llvm there are no GPU-specific optimizations and we use the same optimization pipeline for all targets.

We use mostly standard LLVM optimization pipeline, but there are some categories of passes which were disabled, like loop optimizations or vectorization. We discovered that they actually hurt performance more when applied for an IR that has no specific target in mind.

If you want to dig deeper into -O3 vs -O2 passes list, then you need to look into LLVM documentation, or start your journey here:

llvm/clang/lib/CodeGen/BackendUtil.cpp

Line 1211 in e22b8e4

MPM.addPass(PB.buildPerModuleDefaultPipeline(Level));

If you are curious about what device compilers do with different optimization levels, then it should be asked in different places. For example, for Intel GPUs the place to ask is https://github.com/intel/intel-graphics-compiler

[sanchitintel]

Thanks for the info! :)