[Refactor] IL-generated kernels for UnmanagedMemoryBlock

[Nucs]

Problem

UnmanagedMemoryBlock.Casting.cs contains 2,228 lines of repetitive type-dispatch code with 144 nested switch cases (12 input types × 12 output types), each containing nearly identical for-loops:

case NPTypeCode.Boolean:
{
    var src = (bool*)source.Address;
    switch (InfoOf<TOut>.NPTypeCode)
    {
        case NPTypeCode.Int32:
            var dst = (int*)ret.Address;
            for (int i = 0; i < len; i++)
                *(dst + i) = Converts.ToInt32(*(src + i));
            break;
        // ... 11 more output types
    }
    break;
}
// ... 11 more input types (144 total combinations)

Issues

Problem	Impact
Code bloat	2,228 lines for a simple operation
Maintenance burden	Changes must be replicated across 144 branches
Regen dependency	Uses #if _REGEN template generation
No SIMD	Scalar loops where vectorization is possible
Cache pollution	144 code paths = poor instruction cache utilization

Proposed Solution

Replace with IL-generated kernels using the established ILKernelGenerator pattern:

New API (~20 lines)

public static partial class UnmanagedMemoryBlock
{
    [MethodImpl(MethodImplOptions.AggressiveInlining)]
    public static IMemoryBlock CastTo(this IMemoryBlock source, NPTypeCode to)
    {
        if (source.TypeCode == to)
            return source.Clone();
        return CastKernelGenerator.Execute(source, to);
    }
}

Kernel Generator (~300 lines)

public static class CastKernelGenerator
{
    private delegate void CastKernel(IntPtr src, IntPtr dst, int count);
    private static readonly ConcurrentDictionary<(NPTypeCode, NPTypeCode), CastKernel> _cache = new();

    public static IMemoryBlock Execute(IMemoryBlock source, NPTypeCode dstType)
    {
        var kernel = _cache.GetOrAdd(
            (source.TypeCode, dstType), 
            key => GenerateKernel(key.Item1, key.Item2));
        
        var dst = AllocateBlock(dstType, source.Count);
        kernel((IntPtr)source.Address, (IntPtr)dst.Address, source.Count);
        return dst;
    }

    private static CastKernel GenerateKernel(NPTypeCode srcType, NPTypeCode dstType)
    {
        var method = new DynamicMethod($"Cast_{srcType}_{dstType}", ...);
        var il = method.GetILGenerator();
        
        // Try SIMD for compatible types (widening, float<->double)
        if (TryEmitSimdCast(il, srcType, dstType))
            return (CastKernel)method.CreateDelegate(typeof(CastKernel));
        
        // Fallback: scalar loop with IL conversion opcodes
        EmitScalarCast(il, srcType, dstType);
        return (CastKernel)method.CreateDelegate(typeof(CastKernel));
    }
}

IL Emission (uses native conversion opcodes)

private static void EmitConversion(ILGenerator il, NPTypeCode srcType, NPTypeCode dstType)
{
    switch (dstType)
    {
        case NPTypeCode.Byte:    il.Emit(OpCodes.Conv_U1); break;
        case NPTypeCode.Int16:   il.Emit(OpCodes.Conv_I2); break;
        case NPTypeCode.Int32:   il.Emit(OpCodes.Conv_I4); break;
        case NPTypeCode.Int64:   il.Emit(OpCodes.Conv_I8); break;
        case NPTypeCode.Single:  il.Emit(OpCodes.Conv_R4); break;
        case NPTypeCode.Double:  il.Emit(OpCodes.Conv_R8); break;
        // ... etc
    }
}

Expected Outcome

Metric	Before	After	Change
Lines of code	2,228	~320	-86%
Type switches	144	2	-99%
For-loops in source	291	0	-100%
SIMD support	None	Yes	New
Regen dependency	Yes	No	Removed

SIMD Opportunities

Conversion	SIMD Method
int32 → int64	Avx2.ConvertToVector256Int64(Vector128<int>)
float → double	Avx.ConvertToVector256Double(Vector128<float>)
byte → int32	Avx2.ConvertToVector256Int32(Vector64<byte>)
Same-size reinterpret	Buffer.MemoryCopy

Implementation Plan

• Create ILKernelGenerator.Cast.cs with scalar conversion loop
• Add kernel caching with (srcType, dstType) key
• Implement SIMD paths for widening conversions
• Implement SIMD paths for float↔double
• Update UnmanagedMemoryBlock.CastTo to use new generator
• Add unit tests for all 144 type pairs
• Remove old UnmanagedMemoryBlock.Casting.cs
• Update ArrayConvert.cs to reuse cast kernels

Complexity Assessment

Aspect	Difficulty	Notes
IL emission basics	Easy	Copy patterns from ILKernelGenerator.Binary.cs
Conversion opcodes	Easy	IL has native Conv_* opcodes
Decimal handling	Medium	Requires Convert.ToDecimal() call
SIMD widening	Medium	Well-documented intrinsics
Testing 144 pairs	Tedious	Straightforward but time-consuming

Related Files

Will be deleted:

• src/NumSharp.Core/Backends/Unmanaged/UnmanagedMemoryBlock.Casting.cs (2,228 lines)

Will be simplified:

• src/NumSharp.Core/Utilities/ArrayConvert.cs (can reuse cast kernels)

New file:

• src/NumSharp.Core/Backends/Kernels/ILKernelGenerator.Cast.cs (~300 lines)

References

• Existing pattern: ILKernelGenerator.Binary.cs, ILKernelGenerator.Unary.cs
• Design doc: docs/examples/CastKernel_Proposal.cs
• Parent tracking issue: docs/ISSUE_IL_MIGRATION.md