Repository navigation
A lockstep batch observes every member's field at every plane, even when the scan keeps none #299
Description
What happens
At each saved plane _save calls runner.observe_members(), which drains each member's full observation. The lone-run protocol can do better: observe(field=False) is in the backend protocol, and a batch has no such route. So Batch with fields="none" still moves every member's field, which by reading means a copy to the host on a GPU build. SEP 22 left this out when it gave single runs the field retention.
Batch runs modal propagation only, so the fields are small next to a spatial run's, and the gain has not been measured. Read, not run.
To do
- Time
observe_members()against the step on a native GPUBatch(8), to see whether there is anything to gain, and stop here if it is under 5 % of a plane interval. - If there is: add a field-free member observation to the batch protocol, on
numpy_refandnative, and use it where the retention keeps no field of that plane. The statistics that need a field still get one.
Must not change: a member's stored output against the same point run alone, to rounding.
Done when
The measurement is recorded here with machine and shape, and either the change lands with a test that a fields="none" batch takes no field copy, or this issue closes on the number.
Checked against commit 8a9f764. Takes over the second half of the roadmap's entry "Runner reuse and field-free members in scans."
Activity
Leave a batch member's field where it is at a plane no one reads
jtravs commentedon Nov 5, 2026
Measured and fixed in #520. On an NVIDIA H200 (CUDA 13.1),
Batch(8)of a fibre on native, 8 steps a plane, one process, at8cd170e2:n_omegaThe share was also 13 to 16 % on an AMD Instinct MI325X, and 1 to 4 % on CPU, which is why the gain is a device's. The saving falls as planes get further apart, and the statistics at each plane, 0.8 to 2.1 ms, are unchanged.
BatchBackend.observe_membersnow takesfield=, asobservedoes, and aBatchscan asks for a field only at a plane the scan reads, by the rule a lone point of that scan keeps. Native's own byte count says no member's field crossed, and a scan's stored points match the same points run alone, intests/driver/test_scan_batch.pyandtests/backends/test_batch_lockstep.py. #520 will close this.