The RISC-V Floating-point Coprocessor Model (rvfpm) is a customizable model for verification, design exploration and performance modeling. It provides an easy to configure model, supporting the 32- and 64-bit RISC-V "F" and "Zfinx" extensions, a variety of pipeline structures and optimizations using the eXtension Interface for ease of integration.
Compiled, tested and verified on x86_64 using GCC 12.0.0. Ibex verification done in Ibex fork.
| Extension | Status | Comment |
|---|---|---|
| "F" | Implemented. 100% compliance using Berkeley TestFloat | Default extension. |
| "Zfinx" | Implemented. | Verified functionally using Ibex. |
| "D" | Implemented. 100% compliance using Berkeley TestFloat. Not tested functionally. | Enable by setting EXT_D in config. |
| "Zdinx" | Implemented. Not verified funtionally. | Enable by setting EXT_D and Zfinx in config. |
| "RV64F/D" | Implemented. 100% compliance using Berkeley TestFloat. Not tested functionally. | Enable by setting RV64 in config. |
| "Q" | In progress. | Branch "precisions" |
| "Zfh", "Zfhmin" | In progress. | Branch "precisions" |
| "Zfa" | Not started. | |
| "bfloat" | Not started. | Dependant on the Zfh extension |
The rvfpm core is structured as above. All files relevant to it, and its verification can be found in work/. Headerfiles are located in work/include, and need to be included during compilation. Core cpp files are located in work/src.
The RISC-V Floating-point Coprocessor Model implements all but the compressed interface of the eXtension Interface. Using rvfpm.sv located in work/src, relevant functions in the C++ core are called using DPI-C.
Prerequisites:
- Python
- Pyyaml:
pip3 install pyyaml - Simulator tool - depending on use-case of rvfpm.
The core is configured using Yaml configurations in work/run. If no config is specified using CONFIG="path/to/config/" the default config will be used.
Depending on the target, adapt compiler, flags and simulation tools in the Makefile in work/.
By running make setup CONFIG=<optional_path>, the C++ core is compiled into a shared library which can be loaded into a variety of simulators (verified for QuestaSim 2020) along with config.svh, in_xif.sv and rvfpm.sv .
Testing for compliance to the IEEE 754-2008 standard done using Berkeley TestFloat. Binaries are located in work/bin. The C++ interface for the test-suite available in work/src, note that conversion functions are partially generated by ChatGPT.
To test the system using Berkeley Testfloat, run
make TestFloat LEVEL=1/2
Results are written to work/tests/<target>/rm-<rounding mode>/<test>.txt
A summary of tests performed, and errors encountered can be seen below. Refer to the summary in work/tests for additional details.
| Extension | Tests preformed | Errors |
|---|---|---|
| F | 58,260,633,824 | 0 |
| D | 295,640,668,058 | 0 |
| RV64F | 580,448 | 0 |
For instructions not available in Berkeley TestFloat, functional verification were done by integrating rvfpm into lowRISCs Ibex in a seperate Ibex fork.