Add back memory-mapped hwpe-ctrl interface exploiting new SystemRDL register spec

[FrancescoConti]

This pull request reintroduces a memory-mapped register interface for RedMulE, based on a SystemRDL specification. The changes include a SystemRDL file that describes the register map, a script to generate SystemVerilog and C header outputs using PeakRDL, and integration of the generated interface into the hardware design.
The register interface follows the same structure of the XIF interface currently available in the ab/xif-1.0 branch.

To-do list before review:

• integration within RedMulE controller
• in-platform tests (Snitch cluster / Picobello)

Other actions to-do list (necessary, but can also be pushed to future PRs):

• in-platform tests (MAGIA)
• reintroduce out-of-context testbench and CI

[FrancescoConti]

@belanoa should this be rebased on top of something? I believe on my side the PR is now mature.

[Copilot]

Pull request overview

This pull request reintroduces a memory-mapped register interface for the RedMulE hardware accelerator, replacing the previous XIF (eXtension Interface) custom instruction approach with a more traditional memory-mapped control interface. The implementation uses SystemRDL specification to automatically generate register interface code via PeakRDL tools, following the same architectural patterns established in the existing XIF interface.

Changes:

• Added new CtrlIntfConfig parameter to select between XIF and HWPE_TARGET (memory-mapped) control interfaces at elaboration time
• Introduced SystemRDL-based register specification and auto-generated register interface modules
• Added MisalignedAccessSupport parameter for TCDM access configuration, propagated through the streamer hierarchy

Reviewed changes

Copilot reviewed 12 out of 13 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
rtl/redmule_top.sv	Adds conditional generation logic to instantiate either XIF decoder or memory-mapped target decoder based on CtrlIntfConfig parameter
rtl/redmule_pkg.sv	Defines control interface enum type and misaligned access support default parameter
rtl/redmule_streamer.sv	Propagates MisalignedAccessSupport parameter and updates HCI interface instantiations with explicit size parameters
rtl/redmule_tiler.sv	Adds assertion to validate DataW parameter relationship
rtl/redmule_ctrl.sv	Integrates target_clear_i signal into clear logic
rtl/redmule_inst_decoder.sv	Adds extensive inline documentation comments explaining XIF decoder logic (no functional changes)
rtl/ctrl/redmule_target_decoder.sv	New module implementing memory-mapped control interface decoder with register-to-config mapping
rtl/ctrl/redmule_regif.rdl	SystemRDL specification defining register map for RedMulE control interface
rtl/ctrl/regif/redmule_regif_pkg.sv	Auto-generated package containing register interface types
rtl/ctrl/regif/redmule_regif.sv	Auto-generated OBI-compatible register interface module
rtl/ctrl/gen_regif.sh	Script to generate SystemVerilog and C header files from SystemRDL specification
Bender.yml	Updates hwpe-ctrl to version 3.0.0 and hci to version 2.3.0
Bender.lock	Dependency lock file (requires regeneration)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[belanoa]

Hey Francesco, the PR looks great! My only comment is about the HWPE control interface. Should we use explicit signals instead of the interface? This way, we can avoid instantiating a dummy HWPE control interface every time we use the accelerator with the XIF, which is not very elegant and impractical IMO.

[FrancescoConti]

Hey Francesco, the PR looks great! My only comment is about the HWPE control interface. Should we use explicit signals instead of the interface? This way, we can avoid instantiating a dummy HWPE control interface every time we use the accelerator with the XIF, which is not very elegant and impractical IMO.

Personally I find the cleanest way to deal with this is to create a redmule_xif_top and a redmule_mm_top (or whatever names we prefer), "hiding" the unused interfaces (another example are the hwpe_stream connections to other RedMules). Whether this should be done at platform level (like in magia_redmule_wrap in pulp-platform/MAGIA#46) or inside the redmule repository, I don't really have a best solution - maybe a slight preference for the latter option.