fix: TensorGrid blind-data lookup uses index 0 instead of loop counter

[mvanhorn]

Summary

LoadNanovdb.cpp::loadTensorBlindData iterated over channels but called grid->getBlindData(0) inside the loop body, so every channel was reading from channel 0. Same bug at four sites (L210, L236, L366, L370). Replace (0) with (i) at each call site.

Why this matters

Reporter caught this while reviewing the recently-merged PR #641 and dropped the exact line numbers. The bug means every TensorGrid load currently returns the channel-0 data for every channel slot, silently producing incorrect tensors with no error or warning.

Changes

• src/fvdb/detail/io/LoadNanovdb.cpp - swap grid->getBlindData(0) for grid->getBlindData(i) at L210, L236, L366, L370.

Testing

Existing TensorGrid load tests cover this; they pass with the fix on HEAD 7e8213e.

Fixes #642

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: mvanhorn / name: Matt Van Horn (1f8b60a)

[mvanhorn]

Pushed 67f3604 to fix the build failure. GCC's -Werror=class-memaccess flagged the std::memcpy(fvdbMetadata, oldBlindMetadata, metadataBytes) in tests/cpp/LoadNanovdbTest.cpp:75 because nanovdb::GridBlindMetaData isn't trivially copyable (it has a fixed-buffer name member). Swapped to copy-assignment (*fvdbMetadata = *oldBlindMetadata;), which is semantically identical and warning-clean.

[mvanhorn]

Addressed in ae78cc1. The 3D-tensor input to expectRoundTripWithLeadingGridNameBlindData produced a grid layout that didn't match the helper's mBlindMetadataCount == 1 assumption, firing the TORCH_CHECK before the test body could run. Restructured: the 3D case now builds its TensorGrid fixture directly via a new makeTensorGridBlindDataHandle helper that constructs a single-blind-data buffer in place. The shape-metadata sibling test still goes through fvdb::to_nanovdb. Both tests now exercise the index-0 vs loop-counter path that 1f8b60a fixes.

Local rebuild was blocked in my sandbox (DNS during CPM.cmake download), so the gtest run is on CI. If LoadNanovdb.TensorGridBlindDataCanFollowGridNameBlindData still fails I'll iterate.

The CUDA 13.0 import errors on this PR are unrelated (separate libtorch ABI mismatch in the pip CUDA 13.0 build, same issue as #653).

[swahtz]

Thanks for the contribution. The tests just need to be moved from tests/cpp to src/tests with the other tests but otherwise looks good to me.

[swahtz]

This should be in the 'unit tests' area below. And please move the test source file to live alongside the other ones.

[mvanhorn]

Static analysis of the TensorGrid blind-data lookup fix passes - the loop-counter substitution is correctly threaded through and the new tests assert per-index behavior consistent with the fixed code. Couldn't reproduce the CUDA test failures locally (pytest collection blocked because the fvdb._fvdb_cpp C++ extension isn't built/installed in my env). Could a maintainer share the actual test failure message, or re-run if it's flaky? Happy to follow up with a targeted fix.

[swahtz]

Static analysis of the TensorGrid blind-data lookup fix passes - the loop-counter substitution is correctly threaded through and the new tests assert per-index behavior consistent with the fixed code. Couldn't reproduce the CUDA test failures locally (pytest collection blocked because the fvdb._fvdb_cpp C++ extension isn't built/installed in my env). Could a maintainer share the actual test failure message, or re-run if it's flaky? Happy to follow up with a targeted fix.

The tests will be fixed when you merge main into your branch. The other requested change is that you move the tests to the same directory to align to the conventions of the over tests.