SimpleCGRA

SimpleCGRA: An Open-Source Platform for Temporal-Mapping CGRAs

We are still tidying the codes of SimpleCGRA. The codes of all modules are available. We will keep improving SimpleCGRA.

How to run SimpleCGRA?

• Most modules of SimpleCGRA are written in Python. We test SimpleCGRA on anaconda and python 3.7.

• Install the dependencies:

conda create -n scip python=3.7
conda activate scip
conda install matplotlib scikit-learn networkx xmltodict scip graphviz pyscipopt cvxpy -c conda-forge
pip install graphviz

• Run the benchmark

export SET=express
export BENCH=cosine2
python3 ./test/pnr.py ./benchmark/${SET}/${BENCH}/${BENCH}_DFG.txt ./benchmark/${SET}/${BENCH}/${BENCH}_compat.txt \
        -p ./benchmark/${SET}/${BENCH}/${BENCH}_param.txt 

• The default mapping parameters will be used if ./test/pnr.py is run without "-p xxx_param.txt ".

• Compilation is optional. It is needed if you want to use the front-end or use the C++ interface.

• You can compile SimpleCGRA by running "make". The front-end is tested in LLVM 11.

Description about the code

• arch/: code about CGRA modeling
• benchmark/: CGRA-ME/ExPRESS/standard benchmarks
• common/: provides basic functions
• dataflow/: code about dataflow modeling
• mapping/: code about CGRA mapping (scheduling, placement, routing)
• test/: code about experiments

• test/pnr.py: CGRA mapping
• test/dse.py: design space exploration
• test/dfgpart.py: dataflow graph partition

Experiments

Mapping

IIs on CGRA-ME benchmarks (ADRES-4x4)

Benchmark	ILP	RAMP	SA	HybridMap
accumulate	2	2	1	1
cap	3	3	3	3
conv2	2	2	1	1
conv3	2	4	1	1
mac	1	-	1	1
mac2	3	-	1	1
matrixmultiply	2	2	1	1
mults1	3	5	2	2
mults2	3	-	2	2
nomem1	1	-	1	1
simple	1	2	1	1
simple2	1	2	1	1
sum	1	2	1	1

Mapping times on CGRA-ME benchmarks

Benchmark	ILP	RAMP	SA	HybridMap
accumulate	5m50s	5.58s	2.49s	1.95s
cap	8m28s	1m32s	1m2s	18.49s
conv2	2m31s	50.21s	1.41s	1.34s
conv3	4m10s	4m22s	27.97s	2.52s
mac	2.95s	-	0.99s	1.27s
mac2	28m22s	-	4m19s	45.68s
matrixmultiply	14.43s	0.44s	3.73s	2.36s
mults1	20m17s	1m22s	12.32s	12.01s
mults2	23m19s	-	14.24s	40.56s
nomem1	7.83s	-	0.90s	1.13s
simple	12.75s	1.14s	0.96s	1.47s
simple2	34.33s	1.09s	0.93s	1.43s
sum	2.32s	0.37s	0.87s	1.18s

IIs on ExPRESS benchmarks

Benchmark	SA	Analytic
arf	2	2
cosine1	6	6
cosine2	10	10
ewf	3	3
feedback	4	4
fir1	6	6
fir2	6	5
horner	1	1
motion	3	3
matmul	7	7
matinv	20	20

Mapping times on ExPRESS benchmarks

Benchmark	SA	Analytic
arf	3m27s	1m36s
cosine1	1m49s	1m13s
cosine2	23.24s	32.34s
ewf	2m6s	57.48s
feedback	7m51s	2m31s
fir1	6.10s	11.23s
fir2	12.10s	15.91s
horner	27.02s	3.33s
motion	4m51s	1min21s
matmul	13m37s	3m54s
matinv	31m4s	10m13s

IIs on standard benchmarks

Benchmark	SA	Analytic
aes	6	6
bitcount	3	3
cap	1	1
fir	3	3
fourier	3	3
gsm	37	37
interpol	10	10
isqrt	3	3
mac	1	1
mac2	2	1
mults1	2	1
mults2	1	1
sha	4	4
susan	10	10

Mapping times on standard benchmarks

Benchmark	SA	Analytic
aes	39.91s	43s
bitcount	1m21s	39.77s
cap	4.63s	6.09s
fir	3.16s	4.11s
fourier	1m21s	1m16s
gsm	7m2s	5m37s
interpol	3m28s	1m15s
isqrt	1m29s	33.68s
mac	0.94s	1.41s
mac2	1.33s	1.84s
mults1	9.17s	2.78s
mults2	1.22s	2.08s
sha	26.55s	17.56s
susan	12.57s	18.65s

CGRA-ME benchmarks, ADRES vs. HyCUBE, II comparison

Benchmark	ADRES	HyCUBE
accumulate	1	1
cap	3	1
conv2	1	1
conv3	1	1
mac	1	1
mac2	1	1
matrixmultiply	1	1
mults1	2	1
mults2	2	1
nomem1	1	1
simple	1	1
simple2	1	1
sum	1	1

CGRA-ME benchmarks, ADRES vs. HyCUBE, mapping time comparison

Benchmark	HybridMap	HyCUBE
accumulate	1.95s	1.95s
cap	18.49s	2.71s
conv2	1.34s	1.40s
conv3	2.52s	2.35s
mac	1.27s	1.09s
mac2	45.68s	2.96s
matrixmultiply	2.36s	1.90s
mults1	12.01s	3.62s
mults2	40.56s	3.30s
nomem1	1.13s	0.47s
simple	1.47s	1.03s
simple2	1.43s	1.11s
sum	1.18s	0.55s

ExPRESS benchmarks, ADRES vs. HyCUBE, II comparison

Benchmark	Analytic	HyCUBE
arf	2	2
cosine1	6	6
cosine2	10	10
ewf	3	3
feedback	4	4
fir1	6	6
fir2	5	5
horner	1	1
motion	3	2
matmul	7	7
matinv	20	20

ExPRESS benchmarks, ADRES vs. HyCUBE, mapping time comparison

Benchmark	Analytic	HyCUBE
arf	1m36s	16.83s
cosine1	1m13s	23.33s
cosine2	32.34s	27.35s
ewf	57.48s	23.19s
feedback	2m31s	15.52s
fir1	11.23s	8.77s
fir2	15.91s	9.18s
horner	3.33s	2.72s
motion	1min21s	11.13s
matmul	3m54s	30.99s
matinv	10m13s	1m29s

standard benchmarks, ADRES vs. HyCUBE, II comparison

Benchmark	Analytic	HyCUBE
aes	6	4
bitcount	3	3
cap	1	1
fir	3	3
fourier	3	3
gsm	37	25
interpol	10	6
isqrt	3	3
mac	1	1
mac2	1	1
mults1	1	1
mults2	1	1
sha	4	4
susan	10	10

standard benchmarks, ADRES vs. HyCUBE, mapping time comparison

Benchmark	Analytic	HyCUBE
aes	43.00s	33.05s
bitcount	39.77s	6.77s
cap	6.09s	3.95s
fir	4.11s	3.26s
fourier	1m16s	16.09s
gsm	5m37s	2m7s
interpol	1m15s	54.24s
isqrt	33.68s	39.35s
mac	1.41s	0.48s
mac2	1.84s	1.58s
mults1	2.78s	2.50s
mults2	2.08s	2.95s
sha	17.56s	7.97s
susan	18.65s	17.11s

Partition

Benchmark	Ops	II(M.)	II(P.)	Sub-DFGs	Time
cosine1	66	8	9	2	31s
cosine2	82	10	10	2	35s
matmul	109	7	10	2	43s
matinv	333	20	36	8	527s