geotiff: read minimal source window for VRT resample sources#1821
Merged
Conversation
…ple (closes #1704) Previously when a VRT SimpleSource had a SrcRect/DstRect size mismatch and the caller passed a small ``window=``, ``read_vrt`` decoded the full SrcRect from disk and built a full DstRect-sized resample intermediate before slicing out the clip subregion. For large source rects this forced multi-gigabyte decodes and intermediates on tiny output windows. Now the resample path inverts the nearest-neighbour mapping: for the clipped destination sub-window it computes the smallest SrcRect-relative range of rows and columns that ``_resample_nearest`` would gather from, reads only that sub-rect, and resamples directly into the sub-window output. The new ``_resample_nearest_window`` helper applies the same per-pixel ``floor((d + 0.5) * src / dst)`` mapping the full-rect path uses, so output is byte-identical to a full read sliced after the fact. The per-source pixel-budget guard from #1737 now applies to the clipped sub-window rather than the raw DstRect. The huge-DstRect attack vector is already neutralised by the windowed read; the guard is retained as defence in depth. Regression coverage in test_vrt_resample_window_inverse_1704.py: byte-identical parity for 4x upsample, 4x downsample, and the 7-to-11 non-integer ratio across several window offsets; edge-aligned windows at origin and last pixel; a window crossing two SimpleSources; nodata round-trip; and a read-bound check asserting the source read shrinks under the new path.
Contributor
There was a problem hiding this comment.
Pull request overview
Optimizes read_vrt so that VRT SimpleSources requiring nearest-neighbour resampling (mismatched SrcRect/DstRect sizes) decode only the minimal source sub-rect that feeds the caller's clipped destination window, instead of the full SrcRect. The per-source pixel-budget guard from #1737 is rescoped to the clipped sub-window since the windowed read already neutralizes the huge-DstRect DoS vector.
Changes:
- Add
_nn_src_indexand_resample_nearest_windowhelpers that invert the existing nearest-neighbour mapping for arbitrary sub-windows while remaining byte-identical to "resample full, then slice." - Replace the full-
SrcRectread in theneeds_resamplebranch ofread_vrtwith a windowed read driven by the inverse mapping; rescope themax_pixelsresample-intermediate guard to the clipped sub-window. - Update issue #1737 regression tests to reflect new bounded-by-window semantics and add new #1704 regression tests asserting parity and bounded reads.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.
| File | Description |
|---|---|
| xrspatial/geotiff/_vrt.py | New _nn_src_index/_resample_nearest_window helpers; needs_resample branch now reads minimal source sub-rect and resamples directly into sub-window; per-source pixel-budget cap measured against sub-window. |
| xrspatial/geotiff/tests/test_vrt_resample_window_inverse_1704.py | New regression suite locking byte-identical parity with full-then-slice across upsample/downsample/non-integer ratios, edge alignment, multi-source clipping, nodata round-trip, and a read-bound assertion via read_to_array spy. |
| xrspatial/geotiff/tests/test_vrt_dstrect_resample_cap_1737.py | Updates #1737 tests for new clipped-sub-window cap semantics: huge DstRect no longer pre-rejected; per-source cap now exercised by oversized sub-windows; non-tiled source write to avoid tile-size collision with small max_pixels. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Copilot AI
added a commit
that referenced
this pull request
May 13, 2026
Keep _read_vrt_chunked dispatch (handles gpu=True + chunks=) over the non-GPU-capable _read_vrt_dask added in #1807. Remove the now-dead _read_vrt_dask and _vrt_effective_dtype functions that were only reachable via the superseded dispatch branch. Auto-merged from main: _vrt.py (VRT resample window inverse #1704 + XML size cap #1815 + minimal source window #1821), test files test_read_vrt_lazy_chunks_1798.py, test_vrt_dstrect_resample_cap_1737.py, test_vrt_resample_window_inverse_1704.py, test_vrt_xml_size_cap_1815.py. Co-authored-by: brendancol <433221+brendancol@users.noreply.github.com>
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 free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Closes #1704.
Computes the inverse of the nearest-neighbour mapping from the clipped destination window back to the minimal source sub-window, then reads only that sub-window. Cuts decode and resample-intermediate memory from full-SrcRect-sized to user-window-sized.
The new
_resample_nearest_windowhelper preserves the per-pixel mapping used by the existing full-rect path, so output for the windowed case is byte-identical to a full read sliced after the fact. Regression tests in test_vrt_resample_window_inverse_1704.py lock that contract in across upsample, downsample, and non-integer ratios.The per-source pixel-budget guard from #1737 now applies to the clipped sub-window rather than the raw DstRect. The huge-DstRect attack vector is already neutralised by the windowed read; the guard is retained as defence in depth. The existing #1737 tests are updated to reflect this.
Tested: pytest xrspatial/geotiff/tests/test_vrt_resample_window_inverse_1704.py