Benchmarks¶

The benchmark suite mirrors the MLX Lattice benchmark vocabulary so CUDA and Metal results can be compared by case, layout, feature width, and mode.

Basic usage¶

uv run bench --preset smoke --group conv --device cuda
uv run bench --list

A fuller run can sweep sizes, layouts, channels, and modes:

uv run bench \
  --preset standard \
  --group tensor \
  --group conv \
  --group nn \
  --size 8192 \
  --size 32768 \
  --channels 32 \
  --channels 64 \
  --layout grid \
  --layout block4 \
  --mode cold_op \
  --mode hot_op \
  --output cuda-standard.json

Case groups¶

Group	Coverage
`tensor`	Sparse tensor construction, device/dtype paths, branch combination, global pooling, crop, activation, and normalization modules.
`hash`	Coordinate hashing, offset hashing, self-query, and count kernels.
`dense`	Dense materialization, voxelization, devoxelization, and interpolation weights.
`kmap`	Downsample/upsample helpers and kernel-map construction across CUDA dataflows.
`conv`	Pointwise, spatial, strided, Fetch-on-Demand, Gather-Scatter, target, and submanifold convolution paths.
`nn`	Composed sparse modules including residual, cat, classifier-style, and activation-chain graphs.
`train`	Forward/backward convolution paths.

Density layouts¶

Synthetic coordinate layouts include isolated, line, plane, grid, and dense local blocks from block2 through block8. They intentionally stress different kernel-map and memory-locality patterns.

Output¶

Reports contain environment metadata, per-case latency samples, median/min/p90 /p95 latency, throughput units, workload metrics, and skip/error notes. Relative --output paths are written under benchmarks/results with a text summary next to the JSON file.