xiaorui-yin / spatz

Spatz is a compact RISC-V-based vector processor meant for high-performance, small computing clusters.

Spatz

Spatz is a compact vector processor based on RISC-V's Vector Extension (RVV) v1.0. Spatz acts as a coprocessor of Snitch, a tiny 64-bit scalar core. It is developed as part of the PULP project, a joint effort between ETH Zurich and the University of Bologna.

Getting started

Make sure you download all necessary dependencies:

make all

The Makefile target will automatically download and compile tested versions of LLVM, GCC, Spike, and Verilator. It might take a while. If you have issues cloning the GitHub modules, you might need to remove the folders in sw/toolchain.

ETH users can source the toolchains and initialize the environment by doing:

source util/iis-env.sh

make init

The Spatz cluster system (hw/system/spatz_cluster) is a fundamental system around a Snitch core and a Spatz coprocessor. The cluster can be configured using a config file. The configuration parameters are documented using JSON schema, and documentation is generated for the schema. The cluster testbench simulates an infinite memory. The RISC-V ELF file is preloaded using RISC-V's Front-end Server (fesvr).

Simulating the system

In hw/system/spatz_cluster:

• Compile the software and the binaries:
  • Verilator:

    make sw.vlt

• QuestaSim:

    make sw.vsim

• VCS:

    make sw.vcs

• Run a binary on the simulator:
  • Verilator:

bin/spatz_cluster.vlt path/to/riscv/binary

• QuestaSim:

# Headless
bin/spatz_cluster.vsim path/to/riscv/binary
# GUI
bin/spatz_cluster.vsim.gui path/to/riscv/binary

• VCS

bin/spatz_cluster.vcs path/to/riscv/binary

• Build the traces in .logs/trace_hart_X.txt with the help of spike-dasm:

make traces

• Annotate the traces in .logs/trace_hart_X.s with the source code related to the retired instructions:

make annotate

• Get an overview of all Makefile targets:

make help

Configure the Cluster

To configure the cluster with a different configuration, either edit the configuration files in the cfg folder or create a new configuration file and pass it to the Makefile:

make bin/spatz_cluster.vlt CFG=cfg/spatz_cluster.default.hjson

The default config is in cfg/spatz_cluster.default.hjson. Alternatively, you can also set your CFG environment variable, the Makefile will pick it up and override the standard config.

Architecture

Spatz cluster

Spatz was not designed for full compliance with RVV. Check Ara for an open-source vector processor fully compliant with RVV (and by the same authors!). Instead, Spatz implements some instructions of the vector extension, enough to build a compact and highly efficient embedded vector processor. Thanks to its small size, Spatz is highly scalable, and we rely on multi-core vector processing to scale up the system.

The default Spatz cluster has two Snitch-Spatz core complexes (CCs), each with 2 KiB of latch-based VRF. Each CC has four trans-precision FPUs with support for Spatz-specific SDOTP extensions for low-precision computing. The two Spatz-based CCs share access to 128 KiB of L1 scratchpad memory, divided into 16 SRAM banks.

Spatz core

Each Spatz has three functional units:

• The Vector Arithmetic Unit (VAU), hosting F trans-precision FPUs and an integer computation unit. Each FPU supports fp8, fp16, fp32, and fp64 computation. Each IPU supports 8, 16, 32, and 64-bit computation. All units maintain a throughput of 64 bit/cycle regardless of the current Selected Element Width. The VAU also supports integer and floating-point reductions.
• The Vector Load/Store Unit (VLSU), with support for unit-strided, constant-strided, and indexed memory accesses. The VLSU supports a parametric number of 64-bit-wide memory interfaces. Thanks to the multiple narrow interfaces, Spatz can accelerate memory operations. By default, the number of 64-bit memory interfaces matches the number of FPUs in the design. Important, Spatz' VLSU cannot access the cluster's L2 memory. Ensure that all vector memory requests go to the local L1 memory (we provide the snrt_l1alloc and snrt_dma_start_1d functions for L1 initialization).
• The Vector Slide Unit (VSLDU) executes vector permutation instructions. As of now, we support vector slide up/down and vector moves.

Supported instructions

The most up-to-date list of supported vector instructions can be found in sw/riscvTests/CMakeLists.txt. Spatz does not yet understand vector masking (although this is a work in progress), or fixed-point computation. It also does not understand many of the shuffling and permutation instructions of RVV (e.g., vrgather), and users are asked to shuffle data in memory through indexed memory operations. We very much welcome contributions that expand Spatz' capabilities as a vector coprocessor!

License

Spatz is being made available under permissive open-source licenses.

The following files are released under Apache License 2.0 (Apache-2.0) see LICENSE:

• sw/
• util/
• docs/schema

The following files are released under Solderpad v0.51 (SHL-0.51) see hw/LICENSE:

• hw/

The following files are released under Creative Commons BY 4.0 (CC-BY-4.0) see docs/fig/LICENSE:

• docs/fig

The following directories contains third-party sources that come with their licenses. See the respective folder for the licenses used.

• sw/snRuntime/vendor
• sw/toolchain/
• util/vendor

Publications

If you want to use Spatz, you can cite us:

@Article{Spatz2023,
  title         = {Spatz: Clustering Compact RISC-V-Based Vector Units to Maximize Computing Efficiency},
  author        = {Matheus Cavalcante and Matteo Perotti and Samuel Riedel and Luca Benini},
  year          = {2023},
  month         = sep,
  eprint        = {2309.10137},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AR}
}