The provided source code is a proof-of concept implementation of the
techniques described in our paper CUDA Expression Templates. Please
find the link to the paper on this website
https://graphics.tu-bs.de/publications/wiemann2011cuda


Requirements
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

  - CUDA-enabled device with an compute capability of 1.1 or higher
  - device driver with cuda support
  - CUDA Toolkit 3.0. Higher versions should also work.
  - thrust v1.2 or higher. Available at http://code.google.com/p/thrust/
  - gcc-4.3

This library was only tested on Ubuntu 9.04/9.10 and Fedora 13



Usage
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The provides main.cpp and Makefile should give an idea of how the library is used.
You may need to adapt the paths specified in the Makefile.
Since cudaVec inherits thrust::device_vector cudaVec is fully compatible with thrust.
You may want to use this possibilities. See therefor http://code.google.com/p/thrust/.


Provided Functions
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

unary functions to apply component-wisely on a cudaVec.

  abs
  sin
  cos
  tan
  asin
  acos
  atan
  sinh
  cosh
  tanh
  asinh
  acosh
  atanh
  sqrt
  exp2
  exp10
  exp
  log2
  log10
  log
  trunc
  floor
  round
  isNan
  isInf
  signBit


binary operators to apply component-wisely on two cudaVecs or on a cudaVec and a scalar.

  +
  -
  /
  *
  &&
  ||
  ==
  !=
  <
  <=
  >
  >=


binary functions to apply component-wisely on two cudaVecs or on a cudaVec and a scalar.

  min
  max
  pow


unary functions to apply on a cudaVec
  sum
  argmax - find the smallest index of a maximum magnitude element
  argmin - find the smallest index pf a minimum magnitude element
  norm2 - compute the Euclidean norm of a vector

binary functions to apply on two cudaVec
  dot - computes the dot product of two vectors

a list of functions (maybe not complete) that are implemented in cuda but not yet in
this library.

  rsqrt
  expm1
  log1p
  ldexp
  logb
  ilogb
  rint
  nearbyint
  ceil
  fdim
  atan2
  hypot
  cbrt
  rcbrt
  sinpi
  modf
  fmod
  remainder
  remquo
  erf
  erfinv
  erfc
  erfcinv
  tgamma
  lgamma
  copysignf
  nextafterf
  finitef



TODO
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  - cudaMatrix, template specialization for CUBLAS functions like
    y = A * x + y and the like
  - optimize generated kernels 
  - implement missing cuda functions.
  - the usage of template specialization to use cublas functions if possible could
    speed up some calculations