Skip to content

Latest commit

 

History

History
117 lines (88 loc) · 3.89 KB

README.md

File metadata and controls

117 lines (88 loc) · 3.89 KB
Raw

Benchmarks

This directory contains benchmarks that can be run on ibex simple system. Benchmarks may rely on code external to this directory (e.g. it may be found in vendor/) see the specific benchmark information below for details on how to build and run each benchmark and where benchmark code is located.

Building Simulation

All of these benchmarks run on Simple System. A verilator simulation suitable for running them can be built with:

fusesoc --cores-root=. run --target=sim --setup --build lowrisc:ibex:ibex_simple_system --RV32E=0 --RV32M=ibex_pkg::RV32MFast

See examples/simple_system/README.md for full details.

CoreMark

CoreMark (https://www.eembc.org/coremark/ https://github.com/eembc/coremark) is an industry standard benchmark with results available for a wide variety of systems.

The CoreMark source is vendored into the Ibex repository at vendor/eembc_coremark. Support structure and a makefile to build CoreMark for running on simple system is found in examples/sw/benchmarks/coremark.

To build CoreMark:

make -C ./examples/sw/benchmarks/coremark/

To run CoreMark (after building a suitable simulator binary, see above):

build/lowrisc_ibex_ibex_simple_system_0/sim-verilator/Vibex_simple_system --meminit=ram,examples/sw/benchmarks/coremark/coremark.elf

The simulator outputs the performance counter values observed for the benchmark (the counts do not include anything from pre or post benchmark loops).

CoreMark should output (to ibex_simple_system.log) something like the following:

2K performance run parameters for coremark.
CoreMark Size    : 666
Total ticks      : 4244465
Total time (secs): 8
Iterations/Sec   : 1
Iterations       : 10
Compiler version : GCC
Compiler flags   :
Memory location  :
seedcrc          : 0xe9f5
[0]crclist       : 0xe714
[0]crcmatrix     : 0x1fd7
[0]crcstate      : 0x8e3a
[0]crcfinal      : 0xfcaf
Correct operation validated. See README.md for run and reporting rules.

Choice of ISA string

Different ISAs (to choose different RISC-V ISA extensions) can be selected by passing the desired ISA string into RV_ISA when invoking make.

make -C ./examples/sw/benchmarks/coremark clean
make -C ./examples/sw/benchmarks/coremark RV_ISA=rv32imc

This will build CoreMark using the 'C' extension (compressed instructions).

When changing RV_ISA, you must clean out any old build with make clean and rebuild.

The following ISA strings give the best performance for the Ibex configurations listed in the README:

Config Best ISA
"small" rv32im
"maxperf" rv32im
"maxperf-pmp-bmfull" rv32imcb

CoreMark score

A CoreMark score is given as the number of iterations executed per second. The CoreMark binary is hard-coded to execute 10 iterations (see examples/sw/benchmarks/coremark/Makefile if you wish to alter this). To obtain a useful CoreMark score from the simulation you need to choose a clock speed the Ibex implementation you are interested in would run at, e.g. 100 MHz, taking the above example:

  • 10 iterations take 4244465 clock cycles
  • So at 100 MHz Ibex would execute (100 * 10^6) / (4244465 / 10) = 235.6 Iterations in 1 second.
  • CoreMark (at 100 MHz) is 235.6

CoreMark/MHz is often used instead of a raw CoreMark score. The example above gives a CoreMark/MHz of 2.36 (235.6 / 100 rounded to 2 decimal places).

To directly produce CoreMark/MHz from the number of iterations (I) and total ticks (T) use the follow formula:

CoreMark/MHz = (10 ^ 6) * I / T

Note that core_main.c from CoreMark has had a minor modification to prevent it from reporting an error if it executes for less than 10 seconds. This violates the run reporting rules (though does not effect benchmark execution). It is trivial to restore core_main.c to the version supplied by EEMBC in the CoreMark repository if an official result is desired.