TransDot — Transprecision Dot-Product FPU

TransDot extends FPnew with transprecision dot-product (DP) support across FP32, FP16, FP8, and FP4 formats, plus SIMD FMA operations — all in a single fused datapath with area parity to the FPnew baseline.

Quickstart

source sourceme.sh

Full TransDot Regression (FP32/FP16/FP8 scalar + SIMD + DP)

cd tb/sv_tb_new
export FPNEW_HOME=$(git rev-parse --show-toplevel)
vcs -sverilog -full64 -f filelist.f -top tb_fpnew -o simv -timescale=1ns/1ps
./simv +TESTMODE=simd    # runs all formats: scalar, SIMD, DP

Expected: 2048/2048 vectors pass (256 per format × 8 formats).

No-DP Variant Regression (FP32/FP16/FP8 scalar + SIMD only)

cd tb/sv_tb_new
export FPNEW_HOME=$(git rev-parse --show-toplevel)
vcs -sverilog -full64 \
  -f $FPNEW_HOME/src/transdot_no_dp/filelist_best.f \
  +define+SIMULATION \
  $FPNEW_HOME/tb/sv_tb_new/tb_fpnew.sv \
  -top tb_fpnew -o simv_nodp -timescale=1ns/1ps
./simv_nodp +TESTMODE=simd

Expected: 1280/1280 vectors pass (256 per format × 5 formats). DP tests show 0/256 (expected — DP disabled).

Gate-Level Simulation

After synthesis, verify the netlist:

cd tb/sv_tb_new
vcs -sverilog -full64 \
  -f filelist_syn.f \
  -top tb_fpnew_syn -o simv_syn -timescale=1ns/1ps
./simv_syn +TESTMODE=scalar
./simv_syn +TESTMODE=simd

Repository Structure

src/                          # RTL source
  transdot_fp4_fp8_fp16_fp32_fma_opt.sv  # Full TransDot FMA (DP + SIMD)
  transdot_decomp_*.sv                    # Decomposed datapath modules
  fpnew_*.sv                              # FPnew base modules
  transdot_no_dp/                         # No-DP single-module variant
    transdot_fp16_fp32_fma_simd_base.sv   # Best no-DP design (5-12% smaller than FPnew)
    filelist_best.f                        # Filelist for synthesis/simulation
tb/                           # Testbenches and test data
syn/                          # Synthesis scripts (gitignored outputs)
instances/                    # FPU wrapper instances
docs/                         # Documentation

Synthesis

For the no-DP variant, comment out auto_ungroup none in the synthesis script to enable flattening (yields 5-12% area savings vs FPnew baseline at all timing points).

---

FPnew - New Floating-Point Unit with Transprecision Capabilities

Parametric floating-point unit with support for standard RISC-V formats and operations as well as transprecision formats, written in SystemVerilog.

Maintainers: Pasquale Davide Schiavone davide@openhwgroup.org, Pascal Gouedo pascal.gouedo@dolphin.fr
Authors: Stefan Mach smach@iis.ee.ethz.ch, Luca Bertaccini lbertaccini@iis.ee.ethz.ch

Features

The FPU is a parametric design that allows generating FP hardware units for various use cases. Even though mainly designed for use in RISC-V processors, the FPU or its sub-blocks can easily be utilized in other environments. Our design aims to be compliant with IEEE 754-2008 and provides the following features:

Formats

Any IEEE 754-2008 style binary floating-point format can be supported, including single-, double-, quad- and half-precision (binary32, binary64, binary128, binary16). Formats can be defined with arbitrary number of exponent and mantissa bits through parameters and are always symmetrically biased. Multiple FP formats can be supported concurrently, and the number of formats supported is not limited.

Multiple integer formats with arbitrary number of bits (as source or destionation of conversions) can also be defined.

Operations

• Addition/Subtraction
• Multiplication
• Fused multiply-add in four flavours (fmadd, fmsub, fnmadd, fnmsub)
• Division^1,2
• Square root^1,2
• Minimum/Maximum³
• Comparisons
• Sign-Injections (copy, abs, negate, copySign etc.)
• Conversions among all supported FP formats
• Conversions between FP formats and integers (signed & unsigned) and vice versa
• Classification

Multi-format FMA operations (i.e. multiplication in one format, accumulation in another) are optionally supported.

Optionally, packed-SIMD versions of all the above operations can be generated for formats narrower than the FPU datapath width. E.g.: Support for double-precision (64bit) operations and two simultaneous single-precision (32bit) operations.

It is also possible to generate only a subset of operations if e.g. divisions are not needed.

¹Some compliance issues with IEEE 754-2008 are currently known to exist for the PULP DivSqrt unit (Rounding mismatches have been reported in GitHub issues. This can lead to results being off by 1ulp, and the inexact flag not being properly raised in these cases as well)
²Two DivSqrt units are supported: the multi-format PULP DivSqrt unit and a 32-bit unit integrated from the T-Head OpenE906. The PulpDivsqrt parameter can be set to 1 or 0 to select the former or the latter unit, respectively.
³Implementing IEEE 754-201x minimumNumber and maximumNumber, respectively

Rounding modes

All IEEE 754-2008 rounding modes are supported, namely

• roundTiesToEven
• roundTiesToAway
• roundTowardPositive
• roundTowardNegative
• roundTowardZero

Status Flags

All IEEE 754-2008 status flags are supported, namely

• Invalid operation (NV)
• Division by zero (DZ)
• Overflow (OF)
• Underflow (UF)
• Inexact (NX)

Getting Started

Dependencies

FPnew currently depends on the following:

• lzc and rr_arb_tree from the common_cells repository (https://github.com/pulp-platform/common_cells.git)
• optional: Divider and square-root unit from the fpu-div-sqrt-mvp repository (https://github.com/pulp-platform/fpu_div_sqrt_mvp.git)

These two repositories are included in the source code directory as git submodules, use

git submodule update --init --recursive

if you want to load these dependencies there.

Consider using Bender for managing dependencies in your projects. FPnew comes with Bender support!

Usage

The top-level module of the FPU is called fpnew_top and can be directly instantiated in your design. Make sure you compile the package fpnew_pkg ahead of any files making references to types, parameters or functions defined there.

It is discouraged to import all of fpnew_pkg into your source files. Instead, explicitly scope references into the package like so: fpnew_pkg::foo.

Example Instantiation

// FPU instance
fpnew_top #(
  .Features       ( fpnew_pkg::RV64D          ),
  .Implementation ( fpnew_pkg::DEFAULT_NOREGS ),
  .TagType        ( logic                     )
) i_fpnew_top (
  .clk_i,
  .rst_ni,
  .operands_i,
  .rnd_mode_i,
  .op_i,
  .op_mod_i,
  .src_fmt_i,
  .dst_fmt_i,
  .int_fmt_i,
  .vectorial_op_i,
  .simd_mask_i,
  .tag_i,
  .in_valid_i,
  .in_ready_o,
  .flush_i,
  .result_o,
  .status_o,
  .tag_o,
  .out_valid_o,
  .out_ready_i,
  .busy_o
);

TransDot Mode Encoding (Breaking API)

TransDot mode control is opcode-driven in fpnew_pkg::operation_e. The legacy public sideband controls were removed from top-level/wrapper interfaces:

• dp_enable_i
• simd_enable_i
• fp4_enable_i

Use explicit operation IDs instead:

• TDOT_SIMD_FMADD (16): merged SIMD FMA mode
• TDOT_DP_FMADD (17): dot-product accumulation mode
• TDOT_FP4_DP_FMADD (18): FP4 (E2M1) dot-product accumulation mode

FP4 is now a first-class format (FP4) selected through src_fmt_i/dst_fmt_i.

Documentation

More in-depth documentation on the FPnew configuration, interfaces and architecture is provided in docs/README.md.

Issues and Contributing

In case you find any issues with FPnew that have not been reported yet, don't hesitate to open a new issue here on Github. Please, don't use the issue tracker for support questions. Instead, consider contacting the maintainers or consulting the PULP forums.

In case you would like to contribute to the project, please refer to the contributing guidelines in docs/CONTRIBUTING.md before opening a pull request.

Repository Structure

HDL source code can be found in the src directory while documentation is located in docs. A changelog is kept at docs/CHANGELOG.md.

This repository loosely follows the GitFlow branching model. This means that the master branch is considered stable and used to publish releases of the FPU while the develop branch contains features and bugfixes that have not yet been properly released.

Furthermore, this repository tries to adhere to SemVer, as outlined in the changelog.

Licensing

FPnew is released under the SolderPad Hardware License, which is a permissive license based on Apache 2.0. Please refer to the SolderPad license file for further information.

The T-Head E906 DivSqrt unit, integrated into FPnew in vendor/opene906, is reseased under the Apache License, Version 2.0. Please refer to the Apache 2.0 license file for further information.

Publication

If you use FPnew in your work, you can cite us:

FPnew Publication

@article{mach2020fpnew,
  title={Fpnew: An open-source multiformat floating-point unit architecture for energy-proportional transprecision computing},
  author={Mach, Stefan and Schuiki, Fabian and Zaruba, Florian and Benini, Luca},
  journal={IEEE Transactions on Very Large Scale Integration (VLSI) Systems},
  volume={29},
  number={4},
  pages={774--787},
  year={2020},
  publisher={IEEE}
}

Acknowledgement

This project has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 732631.

For further information, visit oprecomp.eu.