Floating Point Library

[ADTSemi]

Hi iam designing an Accelerator for which i need floating point package but where can i find detailed usage guide for it

[quark17]

There doesn't appear to be any documentation. It would be good if someone can write some, for the libraries guide. The packages are open source, so you can see what the data types and functions are. And examples of usage can be found in Bluespec Inc's CPUs, such as Flute and Toooba.

The source code is in src/Libraries/Base3-Math/. The package FloatingPoint.bsv defines the data type and provides some basic conversion functions, typeclass instances, and combinational operators There are also the packages Divide.bsv and SquareRoot.bsv that define modules for computing floating-point divide and square root, in both pipelined and non-pipelined versions. (They also contain modules for fixed point divide and square root.) The FloatingPoint.bsv package has some modules which are wrapper around a hardware implementation, that add the logic for handling NaN etc; these modules have an argument for passing in the hardware for the core numerical operation, so you can pass in the pipelined or non-pipelined versions, or you own implementation.

An example of using the FloatingPoint library can be seen in Bluespec Inc's Flute CPU. The floating point operations are all handled in a single file, FPU.bsv, which is short and readable. The FPU is then imported into FBox_Core.bsv, which decodes the floating point instructions and sets up the requests to the FPU.

The BSC testsuite also includes tests for FloatingPoint in the testsuite/bsc.libs/FloatingPoint directory. The file FloatTest.bsv demonstrates using a few of the types, functions, and modules.

[quark17]

However, be aware that there are known issues with the floating point library, such as #489 and #497, and limitations such as #534. As mentioned in this comment on #489, IBM provides a set of tests for IEEE 754R and we are aware of failures on that testsuite. (Despite what the comment said, no one is currently working on fixing that. We should probably open a separate GitHub issue for it.)

An alternative to the BSC library is to use someone else's implementation. There are probably Verilog implementations that can be imported in BSV. For example, this comment on #534 mentions a LibBF library that is available under the MIT license. (If someone does import that library, we would welcome the contribution, either in this repo or the bsc-contrib repo.)