Evaluate Profile-Guided Optimization (PGO) about plaidml HOT 3 CLOSED

PlaidML is a compiler infrastructure using LLVM as a framework. Our research moved on to LLVM/MLIR orphaning PlaidML or not coming up with new releases (no matter which branch we talk about).

In any case, PGO is well-known and I would like to understand if your suggestion is just based on "LLVM is a dependency". PlaidML is unlikely to benefit from PGO (as it uses runtime code generation aka JIT compilation).

We have bad experience with automatically derived conclusions (correct me if this is not the case here), and we prefer not to be involved into unspecific learnings or "make-work" projects.

from plaidml.

In any case, PGO is well-known and I would like to understand if your suggestion is just based on "LLVM is a dependency".

My suggestion is based on multiple factors:

• According to multiple benchmarks PGO helps a lot with optimizing compilers and interpreters' performance itself (not the performance of the generated by compilers/interpreters code).
• According to my benchmark on applying PGO to tfcompile project - PGO helps with optimizing the performance of the deep-learning compiler as well.

PlaidML is unlikely to benefit from PGO (as it uses runtime code generation aka JIT compilation).

You are probably mixing two different things: "performance of the generated code" and "performance of the compiler itself". With the former PGO does not help (at least in the proposed by me way). I am suggesting optimizing the performance of the PlaidML compiler/interpreter itself. Here you can check how PGO helps with optimizing JIT-like compilers performance (CPython, PHP, Perl, Ruby).

from plaidml.

You are probably mixing two different things: "performance of the generated code" and "performance of the compiler itself".

Make sense. We do not care about "performance of the compiler itself" and use something else if this is a constraint (not using LLVM or taking care in the first place). For "tfcompile", I can only see improved prototyping turnaround time; otherwise only two cases: (1) deep learning amortizes JIT-overhead, (2) inference should always work on optimized/frozen/precompiled/etc topologies and never run into JIT overhead.

from plaidml.