Simplify mixed precision: compute types on demand instead of caching

[ChrisRackauckas-Claude]

Summary

• Simplified mixed precision implementations by computing types on demand rather than caching them
• Reduces complexity while maintaining zero allocations for subsequent solves
• Cleaner implementation as requested in review feedback

Changes

Modified all 6 mixed precision implementations to compute T32 and Torig types on demand in solve! functions instead of storing them in the cache:

• MKL32MixedLUFactorization
• OpenBLAS32MixedLUFactorization
• AppleAccelerate32MixedLUFactorization
• RF32MixedLUFactorization
• CUDAOffload32MixedLUFactorization
• MetalOffload32MixedLUFactorization

Before

# In init_cacheval:
T32 = eltype(A) <: Complex ? ComplexF32 : Float32
Torig = eltype(u)
return (luinst, ipiv, A_32, b_32, u_32, T32, Torig)

# In solve!:
fact, ipiv, A_32, b_32, u_32, T32, Torig = @get_cacheval(cache, :MKL32MixedLUFactorization)

After

# In init_cacheval:
return (luinst, ipiv, A_32, b_32, u_32)

# In solve!:
fact, ipiv, A_32, b_32, u_32 = @get_cacheval(cache, :MKL32MixedLUFactorization)
# Compute types on demand
T32 = eltype(A) <: Complex ? ComplexF32 : Float32
Torig = eltype(cache.u)

Performance

The type computations (eltype(A) <: Complex and eltype(cache.u)) are simple operations that don't allocate, so computing them on demand has negligible performance impact while making the code cleaner and easier to understand.

Related

This is a cleaner reimplementation of #758 based on review feedback.

🤖 Generated with Claude Code