Implementig FPU on CVE32E20

[Jovann2424]

Hello,

I am working with the OpenHW CORE-V CV32E20 (CVE2) core and I need to evaluate floating-point performance after integrating an external FP/FPU unit.

I am specifically looking for a suitable benchmark or microbenchmark for measuring the latency and inverse throughput of FP instructions, especially FADD.S.

I am interested in tests that could distinguish between:

dependent FADD.S instructions, for measuring latency;
independent FADD.S instructions, for measuring throughput.

Does the OpenHW/CORE-V project already provide or recommend any suitable FP benchmarks or microbenchmarks for this purpose?

Thank you!

[MikeOpenHWGroup]

Hi @Jovann2424, and thanks for your interest in CORE-V.

The short answer to your question is 'no', we do not provide any performance benchmarks. I know that several members have integrated benchmarks into one or more of our testbenches:

• coremark is widely used and doesn't take too long to run. You can find these at cv32e20-dv/tests/programs/custom/coremark and it should "just run" in the core testbench.
• Also in cv32e40s-dv, and several other testbenches, you will find the embench test suite. The developers of embench claim that it is superior to coremark because it does a better job of "real world" modeling. (I am in no position to judge.)

I am unaware if either of these performance suites specifically target floating point.

BTW, what are you using for an FPU? I am unaware of any open source FPU co-processors that support the CV-eXtension Interface.