NEC Deutschland
The Region Vectorizer (RV) is a general-purpose vectorization framework for LLVM. RV provides a unified interface to vectorize code regions, such as inner and outer loops, up to whole functions.
- Support for tail-predicated OLV through LLVM-VP.
- Support for OpenMP 4.5
#pragma omp simd
and#pragma omp declare simd
(pass-fopenmp -fplugin=libRV.so -mllvm -rv
to Clang and you are set). - Automatic outer-loop vectorization (preview feature) (pass
-mllvm -rv-autovec
to enable). - Support for inter-procedural/recursive vectorization.
- Implements Partial Control-Flow Linearization, S. Moll and S. Hack (PLDI '18).
- Automatically uses SLEEF vector math functions.
- Whole-Function vectorizer (
min -> min_avx2
). - Outer-loop vectorizer.
RV is an LLVM project and integrates into the LLVM build system.
Clone this repository into llvm-project/rv where llvm-project is your LLVM source directory.
To build RV along with LLVM, you need to tell cmake where to find RV.
This can be done by specifying -DLLVM_EXTERNAL_PROJECTS="rv" -DLLVM_EXTERNAL_RV_SOURCE_DIR=llvm-project/rv
to cmake.
Run git submodule update --init
to pull the SLEEF submodule.
To (optionally) enable vectorized complex arithmetic through compiler-rt checkout compiler-rt in llvm/runtimes and configure cmake with -DRV_ENABLE_CRT=on
.
- LLVM trunk (as of latest commit on this branch)
- Clang (for the vector math libraries)
- compiler-rt [optional] (for complex arithmetic functions)
Install LLVM+RV, go to rv/test/ and run ./test_rv.py.
RV ships with frontend passes for Outer-Loop and Whole-Function Vectorization.
The passes pick up on SIMD pragmas in your code to vectorize the region (loop or function) in question.
RV is designed to deal with any control flow inside those regions. However, in case of loop vectorization the annotated loops themselves need to be parallel counting loops.
RV supports a range of value reductions and recurrences, including conditional ones (e.g. if (i % 3 == 0) a += A[i];
).
Be aware that RV will exactly do as you annotated. Specifically, RV does not perform exhaustive legality checks nor is there cost modelling of any kind.
You'll get what you ordered.
- Annotate vectorizable loops with
#pragma clang loop vectorize(assume_safety) vectorize_width(W)
where W is the desired vectorization width. - Invoke clang with
-fplugin=libRV.so -mllvm -rv-loopvec
. We recommend to also disable loop unrolling-fno-unroll-loops
.
Users of RV should include its main header file include/rv/rv.h and supporting headers in include/rv. The command line tester (tool/rvTool.cpp) is a good starting point to learn how to use RVs API.
- include/ - header files
- src/ - source files
- vecmath/ - SIMD library sources
- test/ - tests
- tool/ - sources of rvTool
RV's diagnostic output can be configured through a couple of environment variables. These will be read by the Outer-Loop Vectorizer and rvTool.
To get a short diagnostic report from every transformation in RV, set the environment variable RV_REPORT
to any value but 0
.
To also get a report from RV's Outer-Loop Vectorizer, set the environment variable LV_DIAG
to a non-0
value.
RV_ENABLE_CRT:BOOL
Whether RV should inline and vectorize complex math functions. This makes use of the complex arithmetic implementations in compiler-rt. Requires compiler-rt to live in llvm/projects. Defaults to OFF.RV_TARGETS_TO_BUILD:ListOfTargets
List of LLVM targets, for which the SLEEF vector math library should be built. Same format asLLVM_TARGETS_TO_BUILD
. RV uses SLEEF to vectorize math functions. Clang has to be able to (cross-)compile for all of these targets or the build will fail. Defaults to "Native", the host target.RV_DEBUG:BOOL
If enabled, RV will produce (very) verbose debug output and run additional consistency checks. Make sure you compile with assertions. Recommended for debugging only. Defaults to OFF.LLVM_RVPLUG_LINK_INTO_TOOLS:BOOL
Enables the LLVM pass plugin mechanism to link RV into all LLVM tools (opt, clang, ..). Obviates the need to load libRV manually as a plugin on the command line.
The Region Vectorizer is distributed under the University of Illinois Open Source License. See LICENSE.TXT for details.