The AccelSeeker© framework is a tool for automatically identifying and selecting HW accelerators directly from the application source files. It is built within LLVM8 compiler infrastructure and consists of Analysis Passes that estimate Software (SW) latency, Hardware (HW) latency, Area and I/O requirements. Subsequently, an exact selection algorithm selects the subset of HW accelerators that maximizes performance (speedup) under a user defined area (HW resources) budget.
If you use AccelSeeker in your research, we would appreciate a citation to:
Compiler-Assisted Selection of Hardware Acceleration Candidates from Application Source Code.
Georgios Zacharopoulos, Lorenzo Ferretti, Giovanni Ansaloni, Giuseppe Di Guglielmo, Luca Carloni, Laura Pozzi.
https://ieeexplore.ieee.org/abstract/document/8988767
First we need to install all necessary tools. (LLVM8 and AccelSeeker Analysis passes)
If you already have LLVM8 installed you may skip A. and proceed to B. Otherwise please proceed to A. (suggested).
A.
./bootstrap.8.0.sh
The bootstrap.8.0.sh script downloads and builds LLVM8, which is needed to compile and load the AccelSeeker passes.
If, for any reason, you move/rename LLVM8 source tree, then you have to modify the "LLVM_BUILD" Paths in the Makefile and the run_sys_aw.sh script, that invokes the AccelSeeker passes, inside the directory for each benchmark accordingly.
-- If LLVM8 is already installed: --
B.
All necessary files containing the analysis passes need to be copied to the LLVM8 source tree. This line requires to be edited according to the path of your LLVM8 installation:
export LLVM_SRC_TREE="path/to/llvm/source/root"
In order to provide the correct path to your LLVM source tree.
./bootstrap_AS_passes.sh
LLVM8 can then be recompiled using make and a new Shared Object (SO) should be created in order to load the AccelSeeker passes.
cd "path/to/llvm/build" && make
For testing the H.264 decoder https://github.com/adsc-hls/synthesizable_h264 synthesizable version by Xinheng Liu et al of University of Illinois at Urbana-Champaign is used.
cd h264_decoder
1) Collect dynamic profiling information and generate the annotated Intermediate Representation (IR) files.
cd sim
We make sure that the LLVM lines in "Makefile_AccelSeeker" point to the path of the LLVM8 build and lib directory:
BIN_DIR_LLVM=path/to/llvm/build/bin
LIB_DIR_LLVM=path/to/llvm/build/lib
Then we run the instrumented binary with the appropriate input parameters and generate the annotated IR files using the profiling information.
make profile
2) Identification of candidates for acceleration and estimation of Latency, Area and I/O requirements.
We make sure that the LLVM_BUILD line in "run_sys_aw.sh" points to the path of the LLVM8 build directory:
LLVM_BUILD=path/to/llvm/build
The following script invokes the AccelSeeker Analysis passes and generates the files needed to construct the final Merit/Cost estimation. The files generated are: FCI.txt IO.txt LA.txt
./run_sys_aw.sh
The following script generates the Merit/Cost (MC) file along with the implementation of the Overlapping rule in the final Merit/Cost/Indexes (MCI) file. The files generated are: MCI.txt MC.txt
./generate_accelcands_list.sh
The MCI.txt format is as follows:
BENCHMARK-NAME ACCELERATOR-NAME MERIT(CYCLES SAVED) COST(LUTS) FUNCTION_INDEXES
The MCI.txt file can be used subsequently by the exact selection algorithm in order to select the subsets of the AccelSeeker candidates list that maximize Merit (Speedup) under various Costs (Area budgets or HW resources).
../accel_selection_algo_src/accel-find MCI.txt AREA_BUDGET
e.g. ../accel_selection_algo_src/accel-find MCI.txt 1000
The unit used for area bucget by default is Look-Up Tables (LUTs).
To delete all data files use:
../scripts/delete_all_data_files.sh
AccelSeeker performs identification of valid candidates for acceleration (AccelCands) and estimates their performance in terms of speedup gains (cycles saved or hereafter called Merit) and hardware resources required (area or hereafter called Cost).
The process is as follows:
a) Dynamic Profiling of the runtime.
The existing tools of LLVM are used to generate an instrumented version of the binary of a provided benchmark.
http://clang.llvm.org/docs/UsersManual.html#profiling-with-instrumentation
e.g.
clang -O3 -fprofile-instr-generate bench.c -o bench_instrumented
The bench.profdata file generated and used to generate the respective *.ir files of the benchmark.
bench_instrumented $(BENCH_COMMAND_LINE_PARAMETERS)
llvm-profdata merge -output=$(BENCH).profdata default.profraw
clang -S -emit-llvm -O3 -fprofile-instr-use=$(BENCH).profdata -o bench.ir bench.c
b) AccelCands Identification and performance/area estimation (Merit/Cost estimation).
The Analysis Passes are invoked. We are providing as input the *.ir files that were generated in the previous step.
The output from loading this pass provides us with a full analysis of the AccelCands regarding their Merit, Cost estimation and the application of the overlapping rule, so that candidates whose computation completely
overlaps with the computation of other candidates are mutually exclusive.
c) AccelCands Selection (HW accelerators selection)
The Selection phase takes place where a subset of the initial set of potential candidates for acceleration
are selected, so that their cumulative speedup is maximized and their cumulative area does not exceed a user-defined area budget.
** Modifications are needed to comply for every benchmark. **
Georgios Zacharopoulos georgios@seas.harvard.edu Date: May, 2020