Profiling MinkowskiEngine

MinkowskiEngine wraps its performance-critical backend calls in named torch.profiler.record_function ranges so a profiled model shows where time is spent instead of one opaque block. All ranges are prefixed ME::.

Usage

import torch
from torch.profiler import profile, ProfilerActivity

model = model.cuda()
inp = inp  # a MinkowskiEngine.SparseTensor on CUDA

with profile(activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA]) as prof:
    out = model(inp)
    out.F.sum().backward()

# Sort by CUDA time and filter to Minkowski ranges
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=20))

What the ranges mean

• ME::<Op>.forward / ME::<Op>.backward — the backend compute call for an autograd op (Convolution, ConvolutionTranspose, LocalPooling, LocalPoolingTranspose, GlobalPooling, DirectMaxPooling, Broadcast, Interpolation, SPMM, SPMMAverage).

• ME::CoordinateManager.<method> — coordinate-map / kernel-map construction (insert_and_map, insert_field, field_to_sparse_insert_and_map, stride, kernel_map, interpolation_map_weight). Time here is the sparse bookkeeping that precedes the actual op compute; on the first iteration of a new input shape it can dominate, then largely disappears once maps are cached.

Comparing the two groups tells you whether a step is bound by op math or by coordinate-map building. record_function is a no-op when no profiler is active, so these ranges add no measurable overhead to normal training or inference.