Uh oh!
There was an error while loading. Please reload this page.
This repository was archived by the owner on Aug 20, 2026. It is now read-only.
pushpull: export pull/push/count/grad_backward (CUDA side) - #32
Merged
Conversation
Mirrors the cpu-lib export one-for-one: same public signatures, same DISPATCH_PP matrix, same accumulate/overwrite contract, `abs` exposed only on `grad_backward`. Calls the new CUHOST launchers in cuda-impl and forwards the CUDA stream. Like cpu-lib, the adjoints get their own module (`pushpull_backward.cpp`, sharing the new private `pushpull_dispatch.h` with `pushpull.cpp`) so the second copy of the instantiation matrix is a separate, parallelisable translation unit. That matters more here than on the CPU side: `pushpull` is already this build's heaviest module and ptxas *memory*, not just wall time, is the binding constraint -- exactly why reg_field/reg_field_rls were split (fastfields-cuda-lib#28/#30). Folding the adjoints into pushpull.cpp would have put both copies in one ptxas job. Compile+link only -- there is no GPU in CI, so the math is validated on the CPU side (fastfields-cpu-lib/tests/test_pushpull_backward.cpp, 6381 finite-difference checks against the *shared* kernels) and this file is validated by nvcc accepting it. Measured with nvcc 12.0: all-Dynamic policy 25s default BOUNDFLAGS/SPLINEFLAGS (the CI legs) 3m40s Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z
This was referenced Jul 31, 2026
…flict Both main (via cuda-impl#21's merge, 1e7a38d) and this branch (via cuda-impl#22's merge, 81079a8) independently re-pinned impl since their common ancestor. 81079a8 is a descendant of 1e7a38d (#22 was built on top of #21's merged state), so it is the correct resolution. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z
Uh oh!
There was an error while loading. Please reload this page.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for freeto subscribe to this conversation on GitHub.
Already have an account?
Sign in.
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part of the stack for fastfields-torch#16 (backpropagate through the pushpull sampling
grid).What
Mirrors the cpu-lib export one-for-one: same public signatures, same
DISPATCH_PPmatrix, same accumulate/overwrite contract,absexposed only ongrad_backward. Calls the newCUHOSTlaunchers in cuda-impl and forwards the CUDA stream.Why a separate module
Like cpu-lib, the adjoints get their own
pushpull_backward.cpp(sharing the new privatepushpull_dispatch.hwithpushpull.cpp) so the second copy of the instantiation matrix is a separate, parallelisable translation unit.That matters more here than on the CPU side:
pushpullis already this build's heaviest module andptxasmemory, not just wall time, is the binding constraint — exactly whyreg_field/reg_field_rlswere split in #28/#30. Folding the adjoints intopushpull.cppwould have put both copies of the matrix into oneptxasjob, which is the failure mode that split was introduced to avoid.Verification
Compile+link only — no GPU in CI. The math is validated on the CPU side against the shared kernels (cpu-lib#49: 6381 finite-difference checks, 0 failures); this file is validated by nvcc accepting it.
Measured locally with nvcc 12.0:
BOUNDFLAGS/SPLINEFLAGS(the CI legs)I have not run the full
makeacross every CI matrix leg locally (this sandbox has 4 cores / 15 GB and gets reclaimed on idle), so the peakptxasmemory of the new module under the fully static leg is the one number a reviewer may want to confirm in CI.Workstream: claude-jitfields-to-fastfields
🤖 Generated with Claude Code
Generated by Claude Code