iDMA

What iDMA Is

iDMA moves data between memories and peripherals across different bus protocols (AXI, OBI, TileLink) and supports multi-dimensional transfers. It splits the problem into three layers so you can mix and match integration styles without changing the core engine.

• Frontend: software-visible interface for requests (selectable register interface, Snitch ISA, descriptor rings).
• Midend: optional ND/RT decomposition and multicore split.
• Backend: protocol execution on the bus, with optional on-the-fly compute.

Figure placeholder

Diagram: end-to-end request lifecycle (frontend request, optional midend split, backend legalizer + transport, response). Show where idma_req_t and idma_rsp_t are created and consumed.

Start Here

If you are integrating iDMA into an SoC, read these in order:

• Quickstart
• Programming Model
• System Integration

If you are deep-diving the design:

• Frontend
• Midend
• Backend
• Compute
• Interfaces and Types

Supported Protocols

The backend supports these protocols via protocol_e in idma_pkg.sv:

Enum Value	Protocol	Description	Typical Use
0	AXI	Full AXI4	Main memory, high bandwidth
1	OBI	OBI	Simple peripherals, low area
2	AXILITE	AXI4-Lite	Register access
3	TILELINK	TileLink-UH	TL-based SoCs (via TLToAXI4)
4	INIT	Init protocol	Efficient zeroing (Occamy)
5	AXI_STREAM	AXI Stream	Streaming endpoints

On-the-Fly Compute

The backend can optionally transform data as it moves, in the transport datapath, with no extra memory traffic. One operation is applied per transfer:

• Transpose: tiled matrix transpose (idma_otf_transpose).
• MX quantize / dequantize: OCP microscaling between FP32/FP16 and MXFP8 (E5M2 elements with an E8M0-labelled block scale, 32-element / 33 B blocks). These are size-changing transfers, legalized in the legalizer.

Enabled via EnableCompute plus a per-op ComputeOps mask. See Compute.

Code Generation (High Level)

iDMA uses the MARIO generator to produce protocol-specific RTL from templates and YAML capability databases. This is how a single codebase produces many backend variants.

Key locations:

• src/db/*.yml - protocol capability databases
• src/backend/tpl/ - backend templates
• src/frontend/reg/tpl/ - register frontend templates
• util/gen_idma.py - generator entry point
• util/mario/ - generator modules

Where to Find Details

• Full data type definitions: Interfaces and Types
• Performance constraints and tradeoffs: Performance and Limitations
• Testbench and job files: Verification
• Documentation QA checklist: Docs Verification Plan