LinAlg HLK: Add layout, offset, stride, alignment and bounds coverage

Goal

Expand LinAlg HLK execution coverage for matrix and vector data moving between public dx::linalg objects and descriptor or group-shared memory.

The tests must verify observable driver behaviour for layouts, non-zero offsets, padded strides, alignment requirements and specification-permitted descriptor bounds handling without conflating those axes with arithmetic coverage.

Why

Current tests primarily use row-major data at offset zero with tightly packed strides. That misses common implementation defects in address calculation, layout conversion, padding, source preservation and writes near resource boundaries.

Descriptor and group-shared operations also have different layout and alignment rules, and Thread-scope matrices use optimal layouts that may require conversion before results can be compared.

Required coverage

• Exercise RowMajor and ColumnMajor descriptor and group-shared paths where permitted.
• Exercise ByteAddressBuffer and RWByteAddressBuffer descriptor-load paths independently, retaining the resource-flag dependency on #8378.
• Exercise applicable optimal layouts, including transpose variants, through the public API.
• Cover offset zero and representative non-zero aligned offsets.
• Cover tightly packed and padded row/column strides.
• Verify stride-zero behaviour for layouts where an explicit row or column stride is not used.
• Exercise the public default alignment argument. Add non-default alignment guarantees only when their legality is explicitly defined by the public contract; do not infer a minimum alignment from the default value or treat alignment as a capability-query axis.
• Place guard data before and after source and destination regions to detect unintended reads or writes.
• Use bounded descriptor-table resources for runtime bounds cases; do not use root descriptors as a bounds oracle.
• For descriptor loads crossing a boundary, accept either whole-operation zeroing or per-element zero substitution as permitted by the specification.
• For descriptor stores and accumulation crossing a boundary, accept either a whole-operation no-op or per-element suppression of out-of-range writes.
• Keep group-shared execution in bounds. Invalid or undersized group-shared targets remain validator coverage under #7840.
• Cover exact-end and first-crossing descriptor accesses using an expected-result set containing both specification-permitted bounds strategies, while retaining guard regions around the tested range.
• Verify operations that promise source preservation do not modify their source region.
• Cover descriptor and group-shared load, store and accumulation separately.
• Cover vector accumulation addressing independently from matrix addressing.
• Use #8647 for capability handling and #8649 for independent byte-level expectations.
• Exercise the public dx::linalg API; document any operation that has no public wrapper.

Acceptance criteria

• RowMajor and ColumnMajor paths have execution coverage for both reading and writing where supported.
• At least one valid optimal-layout round trip is covered for each applicable Thread-scope operation family.
• Every memory operation has a non-zero-offset case.
• Every strided operation has both tightly packed and padded-stride coverage.
• Default alignment paths are covered; non-default values are added only when publicly specified as legal.
• Guard regions prove that successful operations do not overwrite unrelated data.
• Descriptor bounds cases use resources with known dimensions and exclude root descriptors.
• Load expectations accept both whole-operation and per-element zeroing where either is permitted.
• Store and accumulation expectations accept both whole-operation and per-element write suppression where either is permitted.
• Group-shared runtime OOB is not used as a conformance case.
• SRV and UAV descriptor-load paths remain independently diagnosable.
• Descriptor and group-shared paths remain independently diagnosable.
• Every tested scope/layout/alignment combination is legal under the public contract. Capability gating is used only where #8647 exposes a corresponding query; a failure of a legal unqueryable combination is not NotApplicable.
• Failure logs include layout, offset, stride, alignment and the first mismatching byte or element.

Initial operation mapping

• CopyConvert and source preservation: #8546.
• Descriptor load/store: #8547 and #8553.
• Group-shared load/store: #8548 and #8554.
• Descriptor and group-shared matrix accumulation: #8561 and #8562.
• Vector accumulation: #8565.
• Resource flag framework dependency: #8378.
• Completed matrix-layout conversion precedent: #8386.

Existing negative-validation work

Invalid constant arguments and validator rules remain under #7840, including #8496-#8499, #8505, #8507, #8635, #8636 and #8645. This issue covers runtime execution of legal inputs and specification-defined runtime bounds behaviour only.

Public references

• HLSL proposal 0035: Linear Algebra Matrix
• D3D12 Linear Algebra Runtime Feature Support
• Parent tracker #7841
• Capability handling #8647
• Independent oracle #8649

Out of scope

• Duplicating DXIL-negative validation under #7840.
• Exhaustive component-type coverage.
• Arithmetic correctness unrelated to memory representation.
• Publishing private resource, driver or hardware details.
• Root-descriptor bounds assertions, because D3D does not attach dimensions to root descriptors.
• Group-shared runtime out-of-bounds behaviour and invalid target sizing.
• Inventing non-default alignment validity or capability policy not defined by the public contract.

Assisted-by: GitHub Copilot