[API Proposal]: SIMD/Vectorial operations on arrays/Spans

[vpenades]

Background and motivation

In image processing and probably many other fields, it is fairly common to perform basic operations on large arrays, for example:

var left = new float[10000];
var right = new float[10000];

var result = new float[10000];

for(int i=0; i < 10000; ++i) result[i] = left[i] * right[i];

These operations can be made a lot faster if the operands are casted to Vector4, which make the whole operation much faster. But probably there's even more performance improvements that can be done at low level.

API Proposal

Something like this:

static class BulkOperations
{
// sum 'value' to all destination elements
public static void Sum(float value, Span<float> destination);

// sum all 'values' to destination
public static void Sum(ReadOnlySpan<float> values, Span<float> destination);

// sum left and right into destination
public static void Sum(ReadOnlySpan<float> left, float right, Span<float> destination);

// sum left and right into destination
public static void Sum(ReadOnlySpan<float> left, ReadOnlySpan<float> right, Span<float> destination);

// multiply left and right into destination
public static void Multiply(ReadOnlySpan<float> left, ReadOnlySpan<float> right, Span<float> destination);

// cast source into destination, applying a mean and scale factor
// this is typically used to convert float values in the 0-1 range to bytes in the 0-255 range
public static void Cast(ReadOnlySpan<float> source, float mean, float scale, Span<byte> destination);
}

This is just a small example of the functions the API would have, obviously there would be much more: multiply, division, absolute values, casting, byte shifting, and all kinds of bulk operations that can be greatly enhanced with simd/vectorial instrinsics.

API Usage

var left = new float[10000];
var right = new float[10000];

var result = new float[10000];

BulkOperations.Sum(left, right, result);

Alternative Designs

No response

Risks

No response

[AaronRobinsonMSFT]

/cc @tannergooding

[jeffhandley]

Tagging @dotnet/area-system-numerics since the bot didn't.

[huoyaoyuan]

This is covered by TensorPrimitives in #89639, for example TensorPrimitives.Add(left, right, result).

[vpenades]

then I think we can close this issue as resolved.