IGpuHalfPrecisionBackend

Interfaces · AiDotNet.Tensors.Engines.Gpu

No summary documentation available yet.

Known implementations: CudaBackend, HipBackend, MetalBackend, OpenClBackend, VulkanBackend, WebGpuBackend

Properties

Property Summary
bool Fp16Im2colAvailable { get; }
bool SupportsFp16FusedBackward { get; }
bool SupportsFp16NativeOps { get; }
bool SupportsHgemm { get; }

Methods

Method Summary
void Fp16LayerNorm(IGpuBuffer input, IGpuBuffer gamma, IGpuBuffer beta, IGpuBuffer output, IGpuBuffer meanFp32, IGpuBuffer varFp32, int rows, int cols, float eps)
void Fp16Softmax(IGpuBuffer input, IGpuBuffer output, int rows, int cols)
void GemmFp16In32fOut(IGpuBuffer aFp16, IGpuBuffer bFp16, IGpuBuffer cFp32, int m, int n, int k)
void Hgemm(IGpuBuffer aFp16, IGpuBuffer bFp16, IGpuBuffer cFp16, int m, int n, int k)
void Im2colKNFp16(IGpuBuffer input, IGpuBuffer outputHalf, int batch, int channels, int height, int width, int kernelH, int kernelW, int strideH, int strideW, int padH, int padW, int dilationH, int dilationW)
void MatMulBackwardFp16Fused(IGpuBuffer gradCFp16, IGpuBuffer aFp16, IGpuBuffer bFp16, IGpuBuffer gradAOut, IGpuBuffer gradBOut, int m, int n, int k, bool gradOutHalf)