v0.2.1 - centovalli

Coyote v0.2.1 includes the following fixes and features:

• Support for FPGA - GPU DMA without host involvement
• A fully RoCEv2-compatible, 100G RDMA stack
• An extensive simulation infrastructure, which allows the verification of vFPGAs with the same code, both on the host and in hardware
• Integration with Python unit tests
• 10 getting-started examples, including video tutorials, and improved documentation
• Improved run-time scheduler and a more robust background service
• Various bug fixes for:
  - User interrupts
  - Memory virtualisation with card memory
  - Partial reconfiguration build flow dependencies
  - Multiple RDMA connections & support for large message size
  - Clock domain crossing between user, networking and static regions
  - HBM splitting for non-U55C platforms

v0.2.0

Larger Changes

• Nested dynamic reconfiguration build flow
• RoCEv2 RDMA fully compliant with Mellanox NICs
• GPU direct memory access (only AMD gpus)
• Heterogeneous Memory Management (HMM API)
• HACC ETHZ multi-device support (both FPGA and GPU)
• Hypervisor updates, can run VMs on top of vFPGAs
• Credits in vFPGAs (both local and remote)
• Build flow supports static, dynamic and application layer builds
• SW API updates -> simplified interaction, interface built around the notion of a Coyote thread
• Multithreading example
• New examples
• Coyote can run as a background daemon for both local and remote clients
• Network submodule
• Hugepages and Transparent Hugepages support

To Do

• Update all examples
• Update documentation
• Retest hypervisor
• Extend CI coverage
• Create example build project repo
• Examine dmabuf API for local FPGA-side memory (alternative to SVM)