Celerity Simulation Quick Start V2
Bespoke Silicon Group @ University of Washington (http://bjump.org)
1. Checkout Repos
1.1. git clone https://bitbucket.org/taylor-bsg/bsg_celerity_benchmarks.git
1.2. cd bsg_celerity_benchmarks
1.3. Checkout baseline:
baseline is a set of coherent versions of dependent repos. Upto now following
baselines are defined:
BASELINE=CELERITY_V1_ACTUAL : celertity v1 tapeout
=CELERITY_V1_INTENDED: default, intended v1 tapeout.
=CELERITY_V2 : v2 tapeout, backend fix.
=CELERITY_V3 : v3, with remote-load and atomics.
=LATEST : Warning, latest repo. Not fully test.
1.3.1. Have bsg_cadenv access:
make checkout-repos BASELINE=<baseline-name> BSG_CADENV=1
1.3.2. Do not have bsg_cadenv access:
make checkout-repos BASELINE=<baseline-name>
1.3.3. If you do not have access to bsg_cadenv repository:
a) Setup the VCS license.
b) Setup the VCS_HOME
c) After checkout all of the repos (see Section 2),
in vcs-build/bsg_rocket/rockets/coyote/testing/rtl/Makefile
replace include ../../../../../cad/common/mk/cadenv.mk”
with “export VCS_BIN=<Path To VCS Bin Directory>”
2. Build toolchains and Test
Only need to do once.
2.1. Make sure GCC>=4.8 available. Otherwise:
2.1.1. Checkout comments in vcs-build/bsg_riscv/Makefile for how to install
gcc-4.8 if not available.
2.1.2. Check the CC and CXX in following Makefile if gcc is not default.
2.1.2.1. vcs-build/bsg_riscv/Makefile
2.1.2.2. vcs-build/bsg_rocket/rockets/coyote/testing/rtl_five/Makefile
2.1.2.3. vcs-build/bsg_manycore/software/riscv-tools/Makefile
2.2. make build-manycore-tools ## EST time: 12 mins
2.3. make build-rocket-tools ## EST time: 16 mins
2.4. make loopback-test ## EST time: <1 mins
3. Run Benchmarks
3.1. cd vcs-build/bsg_rocket/rockets/coyote/testing/rtl_five
3.2. make run BENCHMARK_0=bsg_rocket_loopback # without waveform
3.3. make run_debug BENCHMARK_0=bsg_rocket_loopback # with waveform
3.4. make soft_run BENCHMARK_0=bsg_rocket_loopback
# run without recompiling RTL. Faster.
4. Run different benchmarks
4.1. Benchmark directory: All benchmark hex files should be placed in this directory
bsg_rocket/common/benchmark
4.2. Run different benchmark:
make run BENCHMARK_0=<benchmark_name>
## without .riscv.hex suffix
4.3. Run on different Rocket:
make run BENCHMARK_<rocket_num>=<benchmark_name>
## rocket_num can be 0, 1, 2, 3, 4
## But some benchmark can only run on specific
## rocket
4.4. Recompile benchmarks:
cd bsg_rocket/rockets/coyote/testing/rtl_five
./compile_bmark <benchmark_name>
Celerity Simulation Quick Start V1
Bespoke Silicon Group @ University of Washington (http://bjump.org)
1. Get code
1.1. mkdir celerity
1.2. cd celerity
1.3. Manycore repo: git clone https://bitbucket.org/taylor-bsg/bsg_manycore
1.4. IP repo: git clone https://bitbucket.org/taylor-bsg/bsg_ip_cores
1.5. Top Design repo:git clone https://bitbucket.org/taylor-bsg/bsg_designs
git checkout Synopsys_Benchmark
1.6. Package repo: git clone https://bitbucket.org/taylor-bsg/bsg_packaging
1.7. RISC-V tools: git clone https://bitbucket.org/taylor-bsg/bsg_riscv
1.8. Sim Framework: git clone https://bitbucket.org/taylor-bsg/bsg_rocket
1.9. CAD setup:
1.9.1. For Others:
a) Setup the VCS license.
b) Setup the VCS_HOME
c) Replace the first line:
“include ../../../../../cad/common/mk/cadenv.mk”
in bsg_rocket/rockets/coyote/testing/rtl/Makefile
with “export VCS_BIN=<Path To VCS Bin Directory>”
1.9.2. d) Alternatively, you can modify the rules to use Verilator or Vivado
command line simulator.
1.9.3. For Bespoke Silicon Group Members:
git clone https://bitbucket.org/taylor-bsg/cad
git checkout bsg_tsmc180
2. Build RISC-V toolchain
2.1. Prerequisite
2.1.1. Make sure host machine has gcc >= 4.8
2.1.2. Checkout comment in bsg_rocket/Makefile for how to install gcc-4.8 if not
available.
2.2. Toolchain for Berkeley Rocket:
2.2.1. cd bsg_riscv
2.2.2. git checkout bsg_celerity
[ check CC and CXX in Makefile if gcc is not default. ]
2.2.3. make checkout-all ## EST time: 1 mins
2.2.4. make build-riscv-tools-newlib ## EST time: 4 mins
2.2.5. make test-spike-hello ## Verify the installation
2.3. Toolchain for Manycore:
2.3.1. cd bsg_manycore/software/riscv-tools
[ check CC and CXX Makefile if gcc is not default.]
2.3.2. make checkout-all ## EST time: 9 mins
2.3.3. make build-riscv-tools ## EST time: 41 mins
3. Run simulation
3.1. cd expobsg_rocket/rockets/coyote/testing/rtl_five
[ check CC and CXX in Makefile if gcc is not default. ]
3.2. make run BENCHMARK_0=bsg_rocket_loopback # without waveform
3.3. make run_debug BENCHMARK_0=bsg_rocket_loopback # with waveform
4. Run different benchmarks
4.1. Benchmark directory:
bsg_rocket/common/benchmark
4.2. Run different benchmark:
make run BENCHMARK_0=<benchmark_name>
## without .riscv.hex suffix
4.3. Run on different Rocket:
make run BENCHMARK_<rocket_num>=<benchmark_name>
## rocket_num can be 0, 1, 2, 3, 4
## But some benchmark can only run on specific
## rocket
4.4. Recompile benchmarks:
cd bsg_rocket/rockets/coyote/testing/rtl_five
./compile_bmark <benchmark_name>
5. Some running time metrix:
Following table shows some benchmarks we have ran on our server.
● Benchmarks with ‘bnn’ will activate the Binary Neural Network Accelerator.
● Benchmarks with ‘streambuf’ will stream weights that stored in manycore to the BNN
● Benchmarks with ‘layers’ means only run for single layer of the BNN
● Benchmarks without ‘layers’ means to run the full BNN.
Type following commands will run the benchmarks and generate the running result at a directory
named by the date & time:
>> cd bsg_rocket/rockets/coyote/testing/rtl_five
>> make run_all
Server Configuration:
CPU: Intel(R) Xeon(R) CPU E3-1241 v3 @ 3.50GHz, 4 cores
Mem: 16GB
------------------------------------------------------------------------------------------------------------------------------
BENCHMARK_0 BENCHMARK_4/21 TIME2
------------------------------------------------------------------------------------------------------------------------------
bsg_rocket_loopback: NULL 1 mins
bsg_rocket_manycore_loopback: NULL 1 mins
NULL: bnn_loopback 1 mins
bsg_rocket_manycore_token_queue: NULL 3 mins
NULL bnn_layer_1 8 mins
manycore_streambuf_layer_1 bnn_layer_1_sneakpath 14/908 mins
manycore_streambuf_layer_2 bnn_layer_2_sneakpath 16 mins
manycore_streambuf_layer_3 bnn_layer_3_sneakpath 16 mins
manycore_streambuf_layer_4 bnn_layer_4_sneakpath 23 mins
manycore_streambuf_layer_5 bnn_layer_5_sneakpath 32 mins
manycore_streambuf_layer_6 bnn_layer_6_sneakpath 57 mins
manycore_streambuf_layer_7 bnn_layer_7_sneakpath 202/1771 mins
manycore_streambuf_layer_8 bnn_layer_8_sneakpath 22 mins
manycore_streambuf_single_image bnn_sneakpath 1141/9893 mins
------------------------------------------------------------------------------------------------------------------------------
1
For Versions after CELERITY_V1, use ‘BENCHMARK_2’; for Versions “CELERITY_V1_*”, use
“BENCHMARK_4”.
2
Time for 16x31/64x64 Configurations.
