[mcencini]

Summary

tools.igrog now grids with compiled loops instead of batched torch:

• Host (src/csrc/ops/grog.c, bartorch_grog_grid): ported from fast-encoding's grog.cpp, in single precision, with OpenMP across targets. For each target it interpolates Q (cubic Lagrange in the table), factors R_ss + mu I by Cholesky, applies W to every set and returns the noise gain.
• Card (src/csrc/ops/grog.cu, bartorch_grog_grid_cuda): one target per warp. The packed lower triangle of the system, one row of W and the offsets sit in shared memory (1.6 KB per warp at 18). Lanes split the coil-pair interpolation (coalesced table rows), the Cholesky rows, the solves and the per-set reductions. The card holds only the samples and one batch of targets.
• igrog(..., device=None) grids on the samples' device. double=True keeps the torch double-precision path for checks, and the torch path is also the fallback on a build without the card gridder.
• also= moves extra sample sets (e.g. a noise draw) with the same kernels.
• The torch kernels interpolate only the pairs i < j, using Q(-d) = Q(d)^H.

Two more changes for the gridded-positions normal (CartesianSense(positions=)):

• Packed kernel. The per-place R x R kernel is Hermitian, so only r <= c is stored (packed upper triangle). When every entry is real (a real subspace basis) it is stored as floats. At rank 4 that is 10 floats per place instead of 16 complex. The host normal and kern_contract_grid read the packed layout.
• bartorch.empty_cuda_memcache() returns BART's per-stream cache of device blocks, so a caller can free VRAM between stages without turning the cache off.

Measured (Deli-CS case, 256³, cycled R8, 6 coils, 3 sources, A40)

path	igrog
torch on card, fp64 (0.0.21)	37 s
torch on card, fp32	23 s
card, thread per target (75d5917)	5.3 s (kernel 3.5 s)
card, warp per target (885e1f1)	2.7 s (kernel 0.95 s, peak 1.1 GB)
host loop, 96 cores	3.0 s

Card and host agree to 2e-6 on the values. The noise draw differs by 2.5e-4, which is single-precision rounding: host and card are each 2.3e-4 from double.

Full recon (mri2qmap, at e7cf777): the grid kernel goes from 0.35 to 0.11 GiB (2.92 M places, real basis). Emptying the memcache after each CG solve frees about 2.5 GB during the other stages, and CG is not slower (10.1 s vs 11.8 s). The process peak drops from 6.5 to 6.0 GB, or to 5.55 GB with coil kernels on the caller's side. The maps match the previous run within run-to-run noise.

Tests

• test_single_precision_matches_double, test_also_moves_extra_samples_by_the_same_kernels, test_the_card_gridder_matches_the_host_loop (skips without a card).
• test_points_normal_with_a_real_basis_keeps_a_real_kernel_and_is_the_two_applications, test_emptying_barts_device_cache_is_safe_with_or_without_a_device.
• On an A40 CUDA build of e7cf777: test_linop_mri, test_igrog, test_cuda, test_abi give 237 passed, with 1 skip (the no-card refusal test).
• Note for local CUDA builds made with clang: set CUDAHOSTCXX=clang++, otherwise cuFINUFFT's templates mangle differently.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MFB6QLjhTxFMFDEzathbEm

[codecov]

Codecov Report

❌ Patch coverage is 61.24031% with 50 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/bartorch/tools/_igrog.py	57.75%	49 Missing ⚠️
src/bartorch/_cuda.py	50.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[mcencini]

CodeFactor reports 2 minor issues on 75d5917. Its page renders them client-side only and the check run carries no text, so I can't see which lines they are. They are most likely complexity notes on the new gridders (grog.cu's per-target kernel, tools/_igrog.py). I split the C loop into helpers in 1114d7d already. I'll restructure further once the build and test jobs (including cuda-build, which compiles grog.cu) are in, if the findings turn out to be actionable.

---

Generated by Claude Code

[mcencini]

codecov/patch reports 60 % of the diff hit (target 93 %) on 037c83b. The lines it misses are the card path: grog.cu, _card_grid in tools/_igrog.py, and the torch fallback a card takes when the build has no gridder. The CI test runners have no GPU, so nothing there can reach them; cuda-build only compiles. test_the_card_gridder_matches_the_host_loop exercises that path. It ran on an A40 build of this branch and failed only on the noise draw: 2.5e-4 against the 1e-4 tolerance, single-precision rounding, with the host and card equally far from double. 037c83b loosens that tolerance to 1e-3; the test has not been re-run since. The host loop (grog.c) and the torch double path run in CI. I'm leaving coverage as it is rather than adding no cover pragmas.

---

Generated by Claude Code