Merge Vector instruction COT project #22
Conversation
…th an example on both rl and c language, with their IR outputs and yaml pass remarks
…ne print from rlc to llvm
…e name and line number
… and specification of a Target Machine for the Pass Manager Builder. This makes rlc -O2 vectorization equals to opt -O2 for rlc/tool/rlc/test/examples/vectorization/dict.rl
No more get target machine procedure on runOptimizer() but passed as a func arg
…nt bench, add base case to grow()
…ench script and dict test Soa_dictionary is a structure of array version of the dictionary implementation, in this way it possible to achieve vectorization. Now contains() get vectorized
The new merge_bench separate every different measurement in different files. In this way we can compare the same benchmark from different commits
Fix _grow() bug in size 0 case, contains() improved with early exit if first element of stride empty
contains() and _insert() were bugged, the accumulator of condition was in or instead of and, _insert() had a totally wrong implementation of distance check. get() and remove() vectorized, remove() to be finished. To check if the 4 above get vectorized.
…() vectorized Update merge_bench.sh description
Soa_dictionary seems to work fine if using _index_binary_search(), but it does not vectorize get() and remove(). Add print_dict() for testing and many print for debugging purposes
Add different approaches to index_of, none works
There was a problem hiding this comment.
i guess this is the file with the optimized dictionary implementation? if it is better than the previous one it should just entirelly replace that
| mlir::StringAttr noaliasAttr = rewriter.getStringAttr("llvm.noalias"); | ||
|
|
||
| for (size_t i = 1; i < newF.getFunctionBody().getNumArguments(); i++){ | ||
| newF.setArgAttr( |
There was a problem hiding this comment.
this we cannot add to the final commit because it is not always true that they never alias
There was a problem hiding this comment.
Ok, I will delete them. If these attributes will lead to better benchmarks, I will just write it into the report
| passManager.addPass( | ||
| MPM.addPass( | ||
| createModuleToFunctionPassAdaptor(std::move(functionPassManager))); | ||
| if (emitSanitizerInstrumentation and not targetIsWindows) |
There was a problem hiding this comment.
there is something going on here that happen probably due to the rebasing. it seem that ode that was there before has been accidentally removed.
There was a problem hiding this comment.
These instructions were repeated in both branches' end, I just moved them down
| }; | ||
| }; | ||
|
|
||
| struct mlir::rlc::TargetInfoImpl |
There was a problem hiding this comment.
i know that this is probably due to the rebase, but does this piece of code really need to be moved away from its original location?
There was a problem hiding this comment.
I need it to be before runOptimizer() because one of its arguments has this structure type
| @@ -0,0 +1,126 @@ | |||
| #!/bin/bash | |||
There was a problem hiding this comment.
is this file only needed when measuring performance from two different patches, for the purpose of comparing them?
There was a problem hiding this comment.
Yes it is, if you do not need it in the repo, I will remove it
2 participants
The following is a brief per file description of the changes:
BackEnd.cpp: the changes enable the SLP vectorizer and to run optimization passes leveraging the target machine capabilities. These changes match the vectorized code produced by
optConversion.cpp: I still need to figure out if these changes are beneficial, I will benchmark them. If they are detrimental, I will remove them from the merge
DictIntBenchBenchmark.cpp: just an increase to the dictionary benchmark range
merge_benchs.sh: this script is meant to group benchmarks from different files in order to have a single plot of the same benchmark in different runs.
It is a work in progress, ignore it for now.SoA_dictionary.rl: This is a structure of array version of the dictionary container. It allows vectorization of get(), remove() and contains(). Still to be benchmarked, I expect a performance gain for large dictionaries
dictionary.rl: edit a wrong function comment