EHLS.Runtime.AtomVM (Elixir HLS v0.1.0)

Copy Markdown View Source

The AtomVM runtime backend (priv/backend/atomvm): the generated functions call the :hwstream NIFs, which drive the LiteX backend's CSRs and DMA.

generate/1 writes <out>/gen/hw_kernels.erl (kernel / pipeline / param / state indices). A target builds into AtomVM and the application:

  • src/hwstream_nif.c (the NIFs; needs hwstream_csr.h) and src/hwstream_pack.c (a list of elements to the packed format of a CPU end and back; no hardware),
  • gen_csr_header.py <csr.json> <out dir> [package root]... (that header from the SoC's csr.json, and the headers the block packages' csr_headers.json ask for),
  • lib/hwstream.erl (the NIF stubs) and lib/hls_check.ex (EHLS.Check: a kernel on the FPGA against the CPU, on the board).

With mix ehls.compile --target cpu (generate_nif/1) the defhls functions run on the CPU as C instead of Elixir: gen/ehls_nif_kernels.c and gen/hw_nif.erl with src/ehls_nif.c (hwnif:info/1, init/1, run/4; src/ehls_nif.h), lib/hwnif.erl and lib/nif_run.ex (EHLS.NIFRun, which cuts a long input into chunks so that the VM is not held).

The generated CPU code stays within what AtomVM runs (immediate terms, no Enum.map_reduce/3).

Summary

Functions

For mix ehls.compile --target cpu: <out>/gen/ehls_nif_kernels.c (the kernels' C and their table, ehls_nif.h) and <out>/gen/hw_nif.erl (their numbers and layouts). A target builds src/ehls_nif.c with them and lib/hwnif.erl and lib/nif_run.ex (EHLS.NIFRun) into AtomVM and the application.

Functions

generate_nif(out)

For mix ehls.compile --target cpu: <out>/gen/ehls_nif_kernels.c (the kernels' C and their table, ehls_nif.h) and <out>/gen/hw_nif.erl (their numbers and layouts). A target builds src/ehls_nif.c with them and lib/hwnif.erl and lib/nif_run.ex (EHLS.NIFRun) into AtomVM and the application.