Class MetalPerformanceShadersBackend

Namespace

DotCompute.Backends.Metal.MPS

Assembly

DotCompute.Backends.Metal.dll

Metal Performance Shaders backend for optimized ML and linear algebra operations. Provides 3x+ speedup for matrix operations compared to custom kernels.

public sealed class MetalPerformanceShadersBackend : IDisposable

Inheritance

object

MetalPerformanceShadersBackend

Implements

IDisposable

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Extension Methods

BufferExtensions.IsSafeToDispose(IDisposable)

BufferExtensions.SafeDispose(IDisposable)

Constructors

MetalPerformanceShadersBackend(nint, ILogger<MetalPerformanceShadersBackend>)

public MetalPerformanceShadersBackend(nint device, ILogger<MetalPerformanceShadersBackend> logger)

Parameters

device nint

logger ILogger<MetalPerformanceShadersBackend>

Properties

Capabilities

Gets the capabilities supported by MPS on this device.

public MPSCapabilities Capabilities { get; }

Property Value

MPSCapabilities

Methods

BatchNormalization(ReadOnlySpan<float>, ReadOnlySpan<float>, ReadOnlySpan<float>, ReadOnlySpan<float>, ReadOnlySpan<float>, Span<float>, float)

Performs batch normalization. Uses MPSCNNBatchNormalization for optimal performance.

public void BatchNormalization(ReadOnlySpan<float> input, ReadOnlySpan<float> gamma, ReadOnlySpan<float> beta, ReadOnlySpan<float> mean, ReadOnlySpan<float> variance, Span<float> output, float epsilon = 1E-05)

Parameters

input ReadOnlySpan<float>

gamma ReadOnlySpan<float>

beta ReadOnlySpan<float>

mean ReadOnlySpan<float>

variance ReadOnlySpan<float>

output Span<float>

epsilon float

Convolution2D(ReadOnlySpan<float>, int, int, int, ReadOnlySpan<float>, int, int, int, Span<float>, int, int, int, int, int, int)

Performs 2D convolution operation. Uses MPSCNNConvolution for optimal performance.

public void Convolution2D(ReadOnlySpan<float> input, int inputHeight, int inputWidth, int inputChannels, ReadOnlySpan<float> kernel, int kernelHeight, int kernelWidth, int outputChannels, Span<float> output, int outputHeight, int outputWidth, int strideY = 1, int strideX = 1, int paddingY = 0, int paddingX = 0)

Parameters

input ReadOnlySpan<float>

inputHeight int

inputWidth int

inputChannels int

kernel ReadOnlySpan<float>

kernelHeight int

kernelWidth int

outputChannels int

output Span<float>

outputHeight int

outputWidth int

strideY int

strideX int

paddingY int

paddingX int

Dispose()

Performs application-defined tasks associated with freeing, releasing, or resetting unmanaged resources.

public void Dispose()

MatrixMultiply(ReadOnlySpan<float>, int, int, ReadOnlySpan<float>, int, int, Span<float>, int, int, float, float, bool, bool)

Performs matrix multiplication: C = alpha * (A * B) + beta * C Uses MPSMatrixMultiplication for optimal performance.

public void MatrixMultiply(ReadOnlySpan<float> a, int rowsA, int colsA, ReadOnlySpan<float> b, int rowsB, int colsB, Span<float> c, int rowsC, int colsC, float alpha = 1, float beta = 0, bool transposeA = false, bool transposeB = false)

Parameters

a ReadOnlySpan<float>

rowsA int

colsA int

b ReadOnlySpan<float>

rowsB int

colsB int

c Span<float>

rowsC int

colsC int

alpha float

beta float

transposeA bool

transposeB bool

MatrixVectorMultiply(ReadOnlySpan<float>, int, int, ReadOnlySpan<float>, Span<float>, float, float, bool)

Performs matrix-vector multiplication: y = alpha * (A * x) + beta * y Uses MPSMatrixVectorMultiplication for optimal performance.

public void MatrixVectorMultiply(ReadOnlySpan<float> matrix, int rows, int cols, ReadOnlySpan<float> vector, Span<float> result, float alpha = 1, float beta = 0, bool transpose = false)

Parameters

matrix ReadOnlySpan<float>

rows int

cols int

vector ReadOnlySpan<float>

result Span<float>

alpha float

beta float

transpose bool

ReLU(ReadOnlySpan<float>, Span<float>)

Applies ReLU activation function: y = max(0, x) Uses MPSCNNNeuronReLU for optimal performance.

public void ReLU(ReadOnlySpan<float> input, Span<float> output)

Parameters

input ReadOnlySpan<float>

output Span<float>

ShouldUseMPS(MPSOperationType, int, MPSCapabilities)

Determines if MPS should be used for a given operation type and size. Custom kernels may be faster for small operations.

public static bool ShouldUseMPS(MPSOperationType operationType, int dataSize, MPSCapabilities capabilities)

Parameters

operationType MPSOperationType

dataSize int

capabilities MPSCapabilities

Returns

bool