grain.functions.binary

A module for binary autograd functions.

TODO: - support add tensor with cudnnAddTensor and mir.ndslice - support opBinary(add, mul, min, max) cudnnOpTensor - convolution - batchnorm

Members

dim2lin int dim2lin(int[length] ids, int[] strides): Undocumented in source. Be warned that the author may not have intended to support it.
doEpilog void doEpilog(float[] o, int idx, float alphaAcc, float beta): Undocumented in source. Be warned that the author may not have intended to support it.
generateStrides void generateStrides(int* dimA, int* strideA, int nbDims, bool isNchw): Undocumented in source. Be warned that the author may not have intended to support it.
lin2dim void lin2dim(int id, int[length] ids, int[length] dims): Convert a linear index i = d_1 s_1 ... s_n + d_2 s_2 ... s_n + d_n-1 s_n + d_n into a multidimensional index (d_1, d_2, ..., d_n)

AddBias struct AddBias(T): Add bias vector to matrix used inside grain.chain.Linear TODO: generalize to broadcastable addition (use cudnnAddTensor)
Convolution struct Convolution(T, size_t imDims, bool isConv = false, bool isNchw = true): Convolution/Cross-correration function
ConvolutionRefImpl struct ConvolutionRefImpl(T, size_t imDims, bool isConv = false, bool isNchw = true): Reference CPU implementation of Convolution function
Embedding struct Embedding(T): Emebedding ID into vector. TODO: support N-dim input. support sparse weight matrix
MatMul struct MatMul(T): Matrix-Matrix multiplication (using cuBLAS)
OpBinary struct OpBinary(T, size_t dim, string ops): c = op(alpha1 * a + alpha2 * b) + beta * c;