21b6655c10
Enables a fairer comparison between core-coupled tensor core to Hopper tensor core, where the latter benefits from coalesced full-throughput moveout to GMEM because it does not use the 1x2 interleaved register mapping. This means the result matrix will be stored swizzled in the GMEM, without breaking correctness.
Vortex GPGPU
Vortex is a full-stack open-source RISC-V GPGPU.
Specifications
- Support RISC-V RV32IMAF and RV64IMAFD
- Microarchitecture:
- configurable number of cores, warps, and threads.
- configurable number of ALU, FPU, LSU, and SFU units per core.
- configurable pipeline issue width.
- optional shared memory, L1, L2, and L3 caches.
- Software:
- OpenCL 1.2 Support.
- Supported FPGAs:
- Altera Arria 10
- Altera Stratix 10
- Xilinx Alveo U50, U250, U280
- Xilinx Versal VCK5000
Directory structure
doc: Documentation.hw: Hardware sources.driver: Host drivers repository.runtime: Kernel Runtime software.sim: Simulators repository.tests: Tests repository.ci: Continuous integration scripts.miscs: Miscellaneous resources.
Build Instructions
More detailed build instructions can be found here.
Supported OS Platforms
- Ubuntu 18.04, 20.04
- Centos 7
Toolchain Dependencies
Install development tools
$ sudo apt-get install build-essential
$ sudo apt-get install git
Install Vortex codebase
$ git clone --recursive https://github.com/vortexgpgpu/vortex.git
$ cd Vortex
Install prebuilt toolchain
By default, the toolchain will install to /opt folder which requires sudo access.
You can install the toolchain to a different location of your choice by setting TOOLDIR (e.g. export TOOLDIR=$HOME/tools).
$ export TOOLDIR=/opt
$ ./ci/toolchain_install.sh --all
$ source ./ci/toolchain_env.sh
Build Vortex sources
$ make -s
Quick demo running vecadd OpenCL kernel on 2 cores
$ ./ci/blackbox.sh --cores=2 --app=vecadd