tigris compile

Compile an ONNX model into a binary .tgrs execution plan for deployment on embedded devices.

Usage#

tigris compile MODEL [OPTIONS]

Options#

Compilation Pipeline#

The compiler runs a 7-stage pipeline:

1. Load. Import the ONNX model, fold constants, canonicalize the graph.
2. Normalize. Fold QDQ patterns, extract quantization parameters, lower ops to TiGrIS op types.
3. Lifetimes. Compute tensor lifetimes from the execution order.
4. Memory Timeline. Build the activation memory timeline, compute peak memory.
5. Temporal Partition. Split the graph into stages that each fit within the SRAM budget, inserting spill/reload ops at stage boundaries.
6. Spatial Partition. For stages that still exceed the budget, compute spatial tiling plans (tile height, halo, receptive field) and detect chain-tileable stage sequences.
7. Binary Emit. Serialize to the .tgrs binary format.

The output is a single .tgrs file designed for zero-copy, zero-alloc loading on the target device.

XIP (Execute In Place)#

When --xip is enabled, the runtime reads weights directly from flash via memory-mapped I/O instead of copying them to RAM. The plan binary format is designed for memory-mapped access, so the C loader returns pointers directly into the mapped buffer. This is the normal mode for embedded deployment.

Weight Compression#

LZ4 compression reduces plan size on flash at the cost of a small SRAM overhead for decompression at runtime.

tigris compile model.onnx -m 256K -c lz4 -o model.tgrs

When compression is enabled:

• Weights are compressed per stage into individual blocks
• At runtime, the executor decompresses one stage’s weights at a time into a reserved prefix of the SRAM arena
• The decompression overhead is tigris_weight_decompression_overhead(&plan) bytes, which must be added to your fast buffer allocation
• The plan reports both compressed and uncompressed sizes

Examples#

Compile with a 256K SRAM budget:

tigris compile ds_cnn.onnx -m 256K -o ds_cnn.tgrs

Output:

Binary plan written to ds_cnn.tgrs
  28 ops, 3 stages @ 256.00 KiB budget
  plan size: 87.42 KiB

Compile with LZ4 compression and flash budget check:

tigris compile mobilenetv2.onnx -m 256K -c lz4 -f 4M -o mobilenetv2.tgrs

Output:

Binary plan written to mobilenetv2.tgrs (LZ4 compressed)
  53 ops, 12 stages @ 256.00 KiB budget
  plan size: 2.85 MiB (uncompressed: 3.41 MiB, ratio: 0.84x)
  flash 4.00 MiB: fits

Two-pool memory (SRAM + PSRAM):

tigris compile yolov5n.onnx -m 232K -m 6M -o yolov5n.tgrs