Readme.md
In rkeops: Kernel Operations on GPU or CPU, with Autodiff, without Memory Overflows

The files where the elementary math operators are defined.

Available math operations are : * Add : adds FA and FB functions * Subtract : subtracts FA and FB functions * Scal : product of FA (scalar valued) with FB * Mult : element-wise multiplication of FA and FB * Minus : alias for Scal,F> * Divide : alias for Scal> * Pow : Mth power of F (vectorized) ; M is an integer * Powf : alias for Exp>> * Square : Pointwise square, more efficient than Pow * Inv : Pointwise inverse, more efficient than Pow * IntInv : alias for Inv> * Sqrt : alias for Powf> * Rsqrt : inverse square root

Standard math functions : * Sum : sum of values in F * Max : max of values in F * Min : min of values in F * ArgMax : argmax of values in F * ArgMin : argmin of values in F * Abs : absolute value of F (vectorized) * Exp : exponential of F (vectorized) * Log : logarithm of F (vectorized) * XLogX : function F*log(F) (vectorized) * Sin : sine of F (vectorized) * Cos : cosine of F (vectorized) * Acos : arc-cosine of F (vectorized) * Asin : arc-sine of F (vectorized) * Atan : arc-tangent of F (vectorized) * Sign : sign of F (vectorized) * Step : step of F (vectorized) * ReLU : ReLU of F (vectorized) * Sign : sign of F (vectorized) * ClampInt : clamping of F in [A,B] (vectorized) ; F is function, A and B are integers * Clamp : clamping of F in [G,H] (vectorized) ; F, G, H are functions

Concatenation and matrix-vector products: * Concat : concatenation of FB and FB * Elem : extract Nth element of F * Extract : extract a number LENGTH starting at index START * MatVecMult : matrix-vector product (FA::DIM must be a muliple of FB::DIM) * OneHot : represents a scalar formula (rounded to an integer) as a one-hot vector of dimension D * VecMatMult : vector-matrix product (FB::DIM must be a muliple of FA::DIM) * TensorProd : tensor product (output is of dimension FA::DIM*FB::DIM) * TensorDot : tensor dot as in numpy. FA and FB are formulas and DA, DB, CA and CB are IndexSequences that store the parameters needed to perform the tensor contraction. FA::DIM (resp. FB::DIM) must be the product of elements in DA (resp. DB). CA and CB are the index of the contracted dimensions.