[derek-gerstmann]

Issue

make_buffer_copy() in src/runtime/device_buffer_utils.h skips extent-1 dimensions, but previously bounded both the sorted insertion search loop and the element-shift loop by the source dimension index i instead of the number of dimensions inserted so far. When an extent-1 dimension preceded a non-trivial dimension, an unused {extent = 1, stride = 0} slot remained at position 0 (or between inserted dimensions), preventing the contiguous-dimension folding loop (chunk_size == src_stride_bytes[0] && chunk_size == dst_stride_bytes[0]) from coalescing dense copies into a single chunk.

Changes

• src/runtime/device_buffer_utils.h:
  • Track the count of inserted non-trivial dimensions (inserted) in make_buffer_copy().
  • Compute copy_extent = min(src->dim[i].extent, dst->dim[i].extent) and skip dimensions where copy_extent == 1.
  • Bound the insertion search loop (insert < inserted) and shift loop (j = inserted; j > insert; j--) by inserted rather than i.
• test/runtime/device_buffer_utils.cpp:
  • Add regression tests covering dense buffers with leading, interior, and transposed extent-1 dimensions, strided buffers with skipped extent-1 dimensions, and overlapping crop copies with extent-1 dimensions.
• test/performance/device_copy.cpp:
  • Add benchmark shapes with a leading extent-1 dimension ({1, 1024, 1024, 6}).

Fixes #9563

Breaking changes

None.

Checklist

• Tests added or updated (not required for docs, CI config, or typo fixes)
• Documentation updated (if public API changed)
• Python bindings updated (if public API changed)
• Benchmarks are included here if the change is intended to affect performance.
• Commits include AI attribution where applicable (see Code of Conduct)

Co-authored-by: Copilot App 223556219+Copilot@users.noreply.github.com