POINTOPS / COMPUTE FIELD APPLE SILICON · PYTORCH / MPS / METAL JS 2D TEACHING MODEL · NO GPU TIMING

DETERMINISTIC INPUT / 2D

POINTER OR ARROWS → QUERY / FPS START

Input pointsQuerySelected

FRAME / SCAN—

—

SELECTED INDICES / ORDER—

SELECTION RULE——

INTERPRETATION

This scene explains selection. Measurements were recorded separately under named hardware and input conditions.

Inspect measurements ↗

These rules lead into PyTorch/MPS execution.Open execution →

PyTorch → MPS → Metal

EXECUTION PATHPYTORCH → APPLE GPU

3D WORKLOAD

×CUDA-only operatorBlocked on Apple Silicon

01PyTorchTensor · model call

02mps-pointops compatSignature · device bridge

03MPS dispatchInput contract · path selection

04Metal kernelGPU computation

05MPS result tensorPassed to the next layer

CURRENT STEP—

—

Scope

Implementation stages

COMPUTE STACKPHASE 1 → 5

CURRENT LAYER—

—

MEASURE / VERIFY

Measurements

SAFE MATH——

FAST MATH——

CI—Recorded checks

Safe and Fast were separate runs on M5 Pro research source, not a combined PyPI package result.

Spatial index computation

Morton key computation

Reduced conceptual model—

Selected point’s key rank—

X Y Z

Drag on the grid or focus it and use the arrow keys. Use the number fields to edit full coordinates.

Morton order · 16 coordinates

Drag horizontally or swipe through samples. Select a point to load its coordinates.

Calculated Morton key · decimal

—

Interleaved bits · low bits on the right

—

AABB pruning

Reduced conceptual model—

Point distribution

Radius—

Orange bounds: AABBs actually pruned

Move the pointer or use arrow keys on the canvas to move the query. Adjust the radius with the slider.

—

RECORDED / QUERY

Measured comparison within one fixture

Measured records

Install and run

INSTALL / RELEASED

python -m pip install mps-pointops==1.0.0

COMPATIBILITY SHIM / OPT-IN

import mps_pointops.compat
mps_pointops.compat.install()

Place these two lines before CUDA-oriented imports to enable the compatibility path. Registered namespaces and coverage vary by operation.

PROVENANCE / LIMITS

Provenance and remaining limits

Published package: —Version DOI ↗
Measured source: ——
Safe/Fast test source: —

—

Physical M1 Chamfer · pinned PyTorch3D CPU oracle · Safe/Fast each: 480 base + 252 normals/Pointclouds cases, 0 failures, max |Δ| 1.91e-6. Chamfer evidence ↗ Physical M1 sparse operators · Safe/Fast each: 92 passed, 0 skipped at 6959fd59. Sparse scope & raw logs ↗

—

Exact physical GPU memory peak remains unmeasured. v1.0.0 includes a bounded, opt-in BVH path.