tt-wavelet

tt-wavelet implements one-level FP32 lifting wavelet transforms on Tenstorrent Wormhole and Blackhole.

Setup

git clone --recurse-submodules https://github.com/ke1rro/tt-wavelet.git
cd tt-wavelet
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Build

build.sh uses the TT-Metal revision pinned in the tt-metal submodule and builds the complete local stack:

• TT-Metal and the TTNN Python bindings;
• TTNN-Wavelet, linked into TT-Metal from this repository's single ttnn-wavelet source tree; and
• the standalone lwt, ilwt, lwt_2d, ilwt_2d, and benchmark binaries.

On a new machine, install TT-Metal's system and Python dependencies first:

./build.sh --bootstrap

For a normal incremental build:

./build.sh

To rebuild one target after a focused source change, pass its name:

./build.sh lwt
./build.sh --target ttnn --jobs $(nproc)

Supported targets are:

• ttnn – TTNN runtime, Python bindings, and the linked TTNN-Wavelet operation.
• lwt – standalone forward 1D lifting wavelet transform.
• ilwt – standalone inverse 1D lifting wavelet transform.
• lwt_2d – standalone forward 2D lifting wavelet transform.
• ilwt_2d – standalone inverse 2D lifting wavelet transform.
• tt_wavelet_benchmark_runner – standalone benchmark runner used by the benchmark scripts.

./build.sh --jobs 16
./build.sh --type Debug

After a build, enable the local runtime before running a binary or importing TTNN:

source ./scripts/set_env.sh