[SIMD] Unary Ops: Add Floor/Ceil/Round and Transcendental Vectorization

[Nucs]

Overview

Extend the IL kernel generator's SIMD unary operation support beyond the current Negate/Abs/Sqrt.

Parent issue: #545

Current State

Operation	SIMD Status	Implementation
Negate	✅ SIMD	Vector256.op_UnaryNegation
Abs	✅ SIMD	Vector256.Abs()
Sqrt	✅ SIMD	Vector256.Sqrt()
Floor	❌ Scalar	Math.Floor per-element
Ceil	❌ Scalar	Math.Ceiling per-element
Round	❌ Scalar	Math.Round per-element
Exp	❌ Scalar	Math.Exp per-element
Log	❌ Scalar	Math.Log per-element
Sin/Cos/Tan	❌ Scalar	Math.Sin/Cos/Tan per-element

SIMD eligibility check: ILKernelGenerator.cs:2086
Vector operation dispatch: ILKernelGenerator.cs:2980-3014

Task List

Tier 1: Quick Wins (Vector256 methods exist in .NET)

• SIMD Floor

  • .NET has Vector256.Floor()
  • Add to CanUseUnarySimd() eligibility
  • Add to EmitUnaryVectorOperation() dispatch
  • Expected: 2× speedup

• SIMD Ceiling

  • .NET has Vector256.Ceiling()
  • Same implementation pattern as Floor
  • Expected: 2× speedup

• SIMD Truncate

  • .NET has Vector256.Truncate()
  • Same implementation pattern
  • Expected: 2× speedup

Tier 2: Medium Effort

• SIMD Round
  • May need Vector256.Round() or composition
  • Check .NET 8+ availability
  • Expected: 1.5-2× speedup

Tier 3: Transcendentals (Complex)

• SIMD Exp/Log (research)

  • No Vector256.Exp() in .NET BCL
  • Options:
    1. Polynomial approximation (Remez/minimax)
    2. External library (MathNet.Numerics)
    3. P/Invoke to Intel SVML
    4. Wait for .NET Tensor primitives
  • Expected: 2-4× speedup if implemented

• SIMD Sin/Cos/Tan (research)

  • Same challenge as Exp/Log
  • Range reduction + polynomial approximation
  • More complex due to periodicity

Implementation Details

Floor/Ceil (Tier 1)

// In CanUseUnarySimd (line ~2086), add:
|| key.Op == UnaryOp.Floor
|| key.Op == UnaryOp.Ceil

// In EmitUnaryVectorOperation (line ~2980), add:
case UnaryOp.Floor:
    var floorMethod = typeof(Vector256).GetMethod("Floor", 
        new[] { typeof(Vector256<>).MakeGenericType(GetClrType(type)) });
    il.EmitCall(OpCodes.Call, floorMethod, null);
    break;
    
case UnaryOp.Ceil:
    var ceilMethod = typeof(Vector256).GetMethod("Ceiling", 
        new[] { typeof(Vector256<>).MakeGenericType(GetClrType(type)) });
    il.EmitCall(OpCodes.Call, ceilMethod, null);
    break;

Transcendentals (Tier 3) - Research Notes

Exp approximation approach:

// Exp(x) via range reduction + polynomial
// 1. Clamp x to avoid overflow
// 2. n = round(x / ln2), r = x - n*ln2
// 3. exp(r) ≈ polynomial (|r| < ln2/2)
// 4. result = 2^n * exp(r)
public static Vector256<float> Exp(Vector256<float> x)
{
    var ln2 = Vector256.Create(0.693147180559945f);
    var invLn2 = Vector256.Create(1.44269504088896f);
    // ... polynomial coefficients ...
}

Files to Modify

File	Changes
ILKernelGenerator.cs:2086	Add Floor/Ceil/Round to eligibility
ILKernelGenerator.cs:2980	Add vector operation dispatch
SimdKernels.cs (optional)	C# fallback implementations

Benchmarks

[Benchmark] public NDArray Floor_10M() => np.floor(_array);
[Benchmark] public NDArray Ceil_10M() => np.ceil(_array);
[Benchmark] public NDArray Exp_10M() => np.exp(_array);
[Benchmark] public NDArray Sin_10M() => np.sin(_array);

NumPy Baseline (10M float64)

Operation	NumPy Time
np.floor	~8 ms
np.ceil	~8 ms
np.exp	~20 ms
np.sin	~50 ms

Success Criteria

1. Floor/Ceil SIMD: ≥1.5× faster than current scalar
2. All existing unary tests pass
3. No accuracy regression vs scalar implementation

[Nucs]

Analysis: SIMD Transcendentals Feasibility

After thorough investigation, here's the assessment:

Current State

• Transcendentals (Sin, Cos, Tan, Exp, Log, Pow, etc.) use scalar Math.Sin/MathF.Sin calls
• Each element is processed individually in a loop
• The JIT does optimize these scalar calls, but no vectorization

The Core Problem

.NET does NOT provide vectorized transcendental functions.

Vector256<float> has: Add, Multiply, Min, Max, Abs, Sqrt, Floor, Ceiling
Vector256<float> does NOT have: Sin, Cos, Tan, Exp, Log, Pow

Available Options

Option	Pros	Cons
Intel SVML	Very fast (4-8x), Intel-optimized	x86/x64 only, native interop, proprietary
Polynomial approx	Pure .NET, cross-platform	Accuracy risk, significant effort
Intel MKL/VML	Optimized, supports SIMD	+100MB dependency, licensing
Wait for .NET	Clean integration	Uncertain timeline

NumPy's Approach

NumPy uses Intel SVML on x86 (see numpy/_core/src/umath/svml/). This is why NumPy is fast for transcendentals - but it's x86-only and adds native dependencies.

Recommendation: Defer / Close as Won't Fix

Reasons:

1. No built-in .NET solution exists
2. Adding native dependencies (SVML/MKL) breaks cross-platform purity
3. Polynomial approximations require extensive accuracy validation
4. Current scalar Math.Sin() is already JIT-optimized
5. Most NumSharp use cases don't bottleneck on transcendentals

The effort-to-benefit ratio is unfavorable - significant implementation work for a feature that would either:

• Only work on x86 (SVML)
• Add large dependencies (MKL)
• Risk accuracy issues (polynomials)

Future Possibilities

• .NET may add Vector.Sin() etc. in future versions
• If added, implementation would be straightforward using existing IL infrastructure

Marking as deferred/won't fix for now. The PR #573 can proceed without this.

[Nucs]

Shipped in 0.70.0 (#628). Floor/ceil/trunc/round are vectorized (bit-level AVX2 for float16, SIMD for float32/float64), and the float32 transcendentals (exp/log/sin/cos/tanh/exp2) are bit-exact NumPy kernel ports:

Console.WriteLine(np.floor(np.array(new[] {1.5f, 2.5f, -1.5f})));   // [ 1.  2. -2.]
Console.WriteLine(np.ceil(np.array(new[] {1.1f, 2.9f})));           // [2. 3.]
Console.WriteLine(np.around(np.array(new[] {1.5, 2.5, 3.5})));      // [2. 2. 4.]  (half-to-even)

Closing.