NVDLA FireMarshal Workload

This repository uses FireMarshal to build a Linux image/binary with the NVDLA driver.

Updating the Linux kernel to support NVDLA

You must have the CONFIG_DRM_GEM_CMA_HELPER string in your kernel .config file to build the driver. However, by default in FireMarshal's Linux port, this is not set from linux-config. To set this you need to change the riscv-linux port in the following way:

• In drivers/gpu/drm/Kconfig change the bool under config DRM_GEM_CMA_HELPER to bool "GEM CMA Helper Functions"

Additionally, with the bump to Linux 5.7, the use of dma_<declare/release>_coherent_memory was prohibited in driver code. Until this is updated in the NVDLA driver code (when it is eventually bumped to a newer version of Linux), these functions had to be readded.

To get all kernel modifications for NVDLA (for Linux 5.7), please checkout the nvdla-linux-v57 branch in FireMarshal's riscv-linux

To build the test workloads

1. Build the Linux kernel, UMD, KMD, and images: ./marshal -v build $NVDLA_WORKLOAD/marshal-configs/<workload>.json.
2. Then you can continue with the rest of the FireMarshal commands.