[maleadt]

The BLAS-1 methods (norm, dot, BLAS.asum, axpy!, axpby!, rmul!, rotate!, reflect!) and the storage-level mul! and triangular methods pass any strided view to oneMKL with incx = stride(x, 1) and lda = max(1, stride(A, 2)), without checking that oneMKL can address that layout. Reversed views get a negative increment, which BLAS walks backwards from the lowest address while the view's pointer is at its highest element, so the result is silently wrong. Views that skip rows get passed as length(x) elements with a single increment, which reads and writes outside the view. Reversed columns give a negative lda, and oneMKL's invalid_argument escapes the support library's C interface and aborts the process:

x = oneArray(Float32[3, 4]); r = view(x, 2:-1:1)
norm(r)                    # 4.0, expected 5.0
rmul!(r, 2f0); x           # [3, 8], expected [6, 8]
mul!(C, 'N', 'N', view(A, :, 4:-1:1), B, true, false)
# terminate called after throwing an instance of 'oneapi::mkl::invalid_argument'

The triangular case is reachable through public API, e.g. lmul!(UpperTriangular(A), view(x, n:-1:1)).

This ports JuliaGPU/CUDA.jl#3411 and JuliaGPU/CUDA.jl#3409. The level-1 wrappers now take strided vectors or dense arrays only, and pass a pointer to the lowest address together with the signed increment, like LinearAlgebra's vec_pointer_stride; nrm2, asum, scal, iamax and iamin use the absolute increment since the order doesn't matter (oneMKL rejects a negative one for iamax/iamin, so their index is flipped back). The LinearAlgebra BLAS-1 methods fall back to broadcasts for strided arrays that aren't vectors or dense. Storage-level mul! only calls oneMKL when every matrix has a unit row stride and a leading dimension of at least its row count, and every vector a positive stride, falling back to GPUArrays.generic_matmatmul! otherwise as it already did for unsupported element types. The triangular methods run on contiguous copies and copy the result back.

Not covered: the level-2 wrappers (gemv!, symv!, ...) still take raw strides when called directly, as in CUDA.jl, and the support library should turn oneapi::mkl::exception into an error code instead of aborting, which needs a JLL rebuild.

Tested on an Iris Xe (Julia 1.12): the reproducers from both issues fail on main and pass here, and the new tests cover reversed and row-skipping views for every BLAS-1 method, the storage-level mul! paths (gemm, gemv, symm, hemm, symv, hemv) and the triangular products and solves.

Fixes #712
Fixes #700

[codecov]

Codecov Report

❌ Patch coverage is 97.91667% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 83.98%. Comparing base (a0adaa1) to head (804dd6b).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
lib/mkl/linalg.jl	97.56%	1 Missing ⚠️
lib/mkl/wrappers_blas.jl	98.18%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #724      +/-   ##
==========================================
+ Coverage   83.13%   83.98%   +0.85%     
==========================================
  Files          52       51       -1     
  Lines        4376     4322      -54     
==========================================
- Hits         3638     3630       -8     
+ Misses        738      692      -46     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.