chaining_channels

Chaining Channels

This reference design demonstrates low-level DMA control and channel chaining on a Ryzen™ AI NPU (npu2_1col).

Overview

This design showcases advanced DMA programming techniques:

1. MemTile Buffer Initialization: A 1kB buffer in MemTile (row 1) is initialized with values 1-256 using the initial_value parameter.

2. Explicit DMA Control: The MemTile DMA is controlled via explicit locks rather than automatic ObjectFIFO management. The DMA waits for a lock release triggered by npu_write32 in the runtime sequence.

3. Low-Level Runtime Sequence: Uses explicit DMA buffer descriptor operations:

   • npu_writebd - Configure buffer descriptors with lock parameters
   • npu_address_patch - Patch buffer addresses into BDs
   • npu_push_queue - Push BDs to DMA queues
   • npu_sync - Wait for DMA completion

4. Lock-Based Sequencing: DMA operations are sequenced using locks in the buffer descriptors rather than tokens, providing explicit control over data flow timing.

5. Data Flow:

   • Write Path: MemTile (1KB initialized data) → DDR buffer A
   • Read Path: DDR buffer B (4KB) → ComputeTile (data discarded, locks toggled)

Architecture

The design uses:

• ShimTile (row 0) - Interfaces with external DDR memory
• MemTile (row 1) - Stores initialized data, controlled by lock at address 0xC0000
• ComputeTile (row 2) - Receives data and toggles locks in an infinite loop
• Explicit Flows - Connect MemTile→ShimTile and ShimTile→ComputeTile DMAs

Building and Running

To compile and run the design:

make        # Build without tracing
make run    # Run test

With Tracing

To enable ShimTile tracing for debugging DMA channels and data traffic:

make TRACE=1            # Build with tracing enabled
make run TRACE=1        # Run with tracing
make parse_trace        # Parse and analyze trace data

The trace will capture DMA channel activity, lock operations, and data traffic on the ShimTile. Trace data is written to trace.txt and can be analyzed with the parse_trace target, which generates trace_chaining_channels.json and displays a summary. You can then use perfetto to view the trace as shown below:

Cleaning

make clean              # Clean all build artifacts and trace files
make clean_trace        # Clean only trace files

Verification

The test verifies that buffer A contains the initialized pattern (values 1-256) that was written from the MemTile.