| |
| """ttaw: shared infrastructure of the Autoware ports to one Tenstorrent Blackhole p150 (ttnn / tt-metalium). |
| |
| The package is the source of truth under ``common/ttaw`` and is vendored into every bundle as the sub-package |
| ``code/<pkg>/ttaw`` by ``common/tools/vendor.py``. Rules that keep one tree valid in both places: |
| |
| - relative imports only (``from .device import open_device``), never ``import ttaw``; |
| - importing any module has no side effects: no ``import ttnn`` / ``torch`` at module level, no device, no network, |
| no environment reads (ttnn and torch are imported inside the functions that need them); |
| - optimization knobs are read once at model build (:mod:`.knobs`) and each has an env A/B switch; |
| - custom kernel ``.cpp`` files are package data, located with :func:`.ops.kernel_path`. |
| |
| Modules (see ``API.md`` for the full guide): |
| |
| ================== ====== ============================================================================= |
| module id what it provides |
| ================== ====== ============================================================================= |
| ``device`` C01 ``open_device`` (ETH dispatch default, ``auto`` patch detection), ``describe_device`` |
| ``trace`` C02 ``TraceRunner``: persistent I/O, warm-up, capture, variants, 1CQ / 2CQ, state |
| ``tensors`` C03 tile padding, layout / dtype helpers, fp32 islands, zero-copy host staging buffers |
| ``weights`` C04 ``OnnxWeights``, BN folding in fp64, safetensors / ``.pth`` loading, weight cache |
| ``precision`` C05 explicit ``compute_kernel_config`` factory and per-module policy tables |
| ``knobs`` -- env A/B knobs read once at build |
| ``metrics`` C06 PCC variants, argmax agreement, IoU, top-K overlap, detection matching, ADE/FDE |
| ``golden`` C06 golden ``.npz`` files, tap registry, gate registry that never loosens, reports |
| ``profiling`` C07 stage bench (p50/p99), Tracy signposts, ops-CSV summarizer, AICLK sampler |
| ``io`` C08 point cloud / image / calibration decoding, output encoders |
| ``outputs`` C08 ``Detections3D`` ``Detections2D`` ``Segmentation3D`` ``Mask2D`` ``Trajectory`` |
| ``api_base`` C08 ``ModelBase`` (from_pretrained, weights resolution, warm-up, lock, info, close) |
| ``server`` C08 FastAPI app factory (``server.app``); stdlib client, smoke and reference checks |
| (``server.client``, ``server.smoke``) |
| ``image`` C14 bit-exact Autoware image pre-processing from cached LUTs (YOLOX letterbox, BEVDet crop) |
| ``image_area`` C14 METEOR preset: OpenCV ``INTER_AREA`` down-scaling of uint8 frames, bit-exact |
| ``image_linear`` C14 SceneSeg preset: OpenCV ``INTER_LINEAR`` / ``INTER_NEAREST`` resizes of uint8 frames and the |
| VisionPilot normalisation (BGR / RGB), bit-exact |
| ``image_triangle`` C14 StreamPETR preset: Autoware's anti-aliased triangle resize + crop + normalise kernel, |
| bit-exact (normalisation presets autoware_main / autoware_0.52 / awml_training) |
| ``geometry`` C10 rigid transforms (float64), Autoware's float32 sweep transform, yaw conventions |
| ``pointcloud`` C11 Autoware multi-sweep densification state machine (per stream), hygiene masks |
| ``voxelize`` C12 deterministic pillars, 9/10/11-feature decoration, canvas scatter / gather index |
| ``sparse`` C13 sparse-conv rulebooks: spconv-faithful sub-manifold / strided neighbour maps (gather |
| form), tile-ordered im2col and to-dense gather indices, capacity buckets, oracles |
| ``nms`` C15 circle NMS, perception_utils IoU-BEV NMS, area class remapper |
| ``decode`` C16 CenterHead dense decode (and at given cells), Autoware DetectedObject mapping |
| ``ops`` -- kernel package-data helpers (custom ``generic_op`` kernels live in ``ops/kernels``) |
| ``ops.conv`` C17 conv2d builders (weights prepared once, fused activations, DRAM slicing), K-split, |
| ConvTranspose k == s, feature-map glue |
| ``ops.upsample`` C18 nearest up-sampling, ``F.interpolate`` matrices, separable device resize |
| ``ops.attention`` C20 SDPA wrapper (non-causal, explicit scale, chunk table, aligned masked keys), heads |
| ``ops.gather`` C19 gather-form scatter (zero sentinel row + TILE / PADDED ``ttnn.embedding``), row gathers |
| ``ops.topk`` C21 top-k proposal selection (class-major ``topk_large_indices`` row, exact decode, set metrics) |
| ``ops.heatmap`` C22 sigmoid + 3x3 local max of TF / BF / PT heatmaps (one exact masked chain) |
| ``ops.deform`` C23 grid_sample helpers: fp32 affine / resize grids (exact identity), P9 rules, kernel emulation |
| ``ops.segment`` K1 per-segment max over rows sorted by segment (``generic_op`` kernel, exact, no |
| per-core args), the log-step stock-op fallback, CSR / shift tables, oracles |
| ``models`` C24 C25 device sub-networks: ``models.second`` (SECOND + SECONDFPN), ``models.pillars`` |
| C27 C26 (PillarFeatureNet + its input staging), ``models.centerhead`` (CenterPoint dense |
| head: K-split shared conv, merged heads), ``models.resnet`` (ResNet |
| bottleneck / basic-block builders, cameras as batch), ``models.transfusion_head`` |
| (TransFusion query head: selection, query init, decoder layer, merged heads) |
| ================== ====== ============================================================================= |
| """ |
|
|
| __version__ = "0.23.2" |
|
|
| __all__ = [ |
| "__version__", |
| "api_base", |
| "decode", |
| "device", |
| "geometry", |
| "golden", |
| "image", |
| "image_area", |
| "image_linear", |
| "image_triangle", |
| "io", |
| "knobs", |
| "metrics", |
| "models", |
| "nms", |
| "ops", |
| "outputs", |
| "pointcloud", |
| "precision", |
| "profiling", |
| "server", |
| "sparse", |
| "tensors", |
| "trace", |
| "voxelize", |
| "weights", |
| ] |
|
|
|
|
| def __getattr__(name: str): |
| """Load submodules on first attribute access (``ttaw.trace``), keeping ``import ttaw`` free of side effects.""" |
| if name in __all__ and name != "__version__": |
| import importlib |
|
|
| return importlib.import_module(f".{name}", __name__) |
| raise AttributeError(f"module {__name__!r} has no attribute {name!r}") |
|
|
|
|
| def __dir__(): |
| return sorted(set(globals()) | set(__all__)) |
|
|
