arrayfire / arrayfire Goto Github PK

• mean
• standard deviation
• variance
• co-variance
• median
• correlation coefficient

Right now the interpolation method used is nearest. Add the option to perform linear interpolation.

The following lines of cuda/CMakeLists.txt depend on the architecture

Line 33:

FILE(GLOB cuda_ptx "ptx/PTX64/sm_20/*.ptx")

Line 55:

CUDA_ADD_LIBRARY(afcuda SHARED ${cuda_src} OPTIONS "-arch=sm_20")

Possible libraries for CPU:

• netlib's LAPACK
• netlib's PLASMA
• Intel's MKL

Possible libraries for CUDA:

• MAGMA
• CULA

Possible libraries for OpenCL:

• clMAGMA
• ViennaCL

• Move constant / randu / randn to data.h
• Move reductions / scan / sort etc into algorithm.h
• Move approx1, approx2 to signal.h

Remove ArrayInfo from src/array

• How to find boost
• How to find fftw
• How to find OpenCL
• How to find CUDA
• How to find clBLAS and clFFT

Add 3D and 4D tests for indexing and assignment functions.

Make sure all backends build properly on windows + visual studio >= 2013

cblas based functions are currently commented out on cpu backend for windows platform until this issue is resolved.

Implement Canny Edge Detector for all backends.

• Change separable convolution to be convolve instead of convolve2
• Add a new function convolve that calls the appropriate function based on number of dimensions.

Bug appears when using floating point types

Write inline wrappers for the following to not break older API

- deviceset --> setDevice
- devicecount --> getDeviceCount
- deviceget --> getDevice
- loadimage --> loadImage
- saveimage --> saveImage

Current sort, sort_index and sort_by_key exist in C++ API. Simply overload these using sort.

• Fixed width formatting
• Make sure positive and negative numbers have same width
• Display dims and strides only in debug mode.

• Make arrayfire_data a submodule
• Move test data directory from build/data/src/arrayfire_data to build/test/arrayfire_data.
• Change TEST_DIR to be relative

• setunique
• setunion
• setintersect

Some options to consider:

• Sort based on platform
• Sort based on type (GPU, Accelerator, CPU)
• Hardcoded order of preference ?
• Sort based on amount of memory available
• Sort based on OpenCL compute capability

dim4 is struct in include/af/dim4.hpp but is being defined as class in array.h. Change it to class.

Add new fft / ifft functions without the normalize parameters. forward fft's don't normalize. Inverse fft's normalize by the number of elements.

Fix compile errors and get it to build on OSX.

Not sure if it is a defective GPU or a problem with the code. Investigate on other CUDA 5.0 cards.

Implement hooks that enable opengl inter operation in CUDA backend.

Adding below new interfaces

• moddims(af::array const& input, af::dim4 const& dims)
• moddims(af::array const& input, dim_type d0, dim_type d1=1, dim_type d2=1, dim_type d3=1)

af::isnan and af::isinf are causing compiler warning on tegra. This is because those symbols are defined in c (http://en.cppreference.com/w/c/numeric/math/isinf / isnan).
Fix this by changing the API to isNaN and isInf.

Implement Infinite impulse response and Finite impulse response filters in ArrayFire.

CUDA random number generator values are not same from run to run. Expected behavior is that it generates same values across runs.

Fix CUDA 6.0 compile issues for boost shared ptr.

Make sure all the functions in C++ wrapper have tests associated with them.

Implement ORB in ArrayFire

Implement template matching algorithm in ArrayFire.

Currently the libraries are being created in the build/src/backend/<cpu/cuda/opencl>/ directories.
They should be created in build/lib64 directory.

Change cmake to make freeimage optional. Disable imageio if unavailable.

Instead of allocating and freeing memory using native calls for every array, implement a memory manager that uses a pool of buffers with fairly simple garbage collection.

Implement hooks that enable opengl inter operation in OpenCL backend.

Implement histogram equilazation for all backends.

Add FAST feature detection to ArrayFire.

Implement seq class for C++ API. Move C-API seq to af_seq.
Implement span and end.
Refactor code from v2.1.