v3.9.0

Improvements

• Add oneAPI backend #3296
• Add support to directly access arrays on other devices #3447
• Add asynchronous reduce all functions that return an af_array #3199
• Add broadcast support #2871
• Improve OpenCL CPU JIT performance #3257 #3392
• Optimize thread/block calculations of several kernels #3144
• Add support for fast math compiliation when building ArrayFire #3334 #3337
• Optimize performance of fftconvolve when using floats #3338
• Add support for CUDA 12.1 and 12.2
• Better handling of empty arrays #3398
• Better handling of memory in linear algebra functions in OpenCL #3423
• Better logging with JIT kernels #3468
• Optimize memory manager/JIT interactions for small number of buffers #3468
• Documentation improvements #3485
• Optimize reorder function #3488

Fixes

• Improve Errors when creating OpenCL contexts from devices #3257
• Improvements to vcpkg builds #3376 #3476
• Fix reduce by key when nan's are present #3261
• Fix error in convolve where the ndims parameter was forced to be equal to 2 #3277
• Make constructors that accept dim_t to be explicit to avoid invalid conversions #3259
• Fix error in randu when compiling against clang 14 #3333
• Fix bug in OpenCL linear algebra functions #3398
• Fix bug with thread local variables when device was changed #3420 #3421
• Fix bug in qr related to uninitialized memory #3422
• Fix bug in shift where the array had an empty middle dimension #3488

Contributions

Special thanks to our contributors:
Willy Born
Mike Mullen

v3.8.3

Improvements

• Add support for CUDA 12 #3352
• Modernize documentation style and content #3351
• memcpy performance improvements #3144
• JIT performance improvements #3144
• join performance improvements #3144
• Improve support for Intel and newer Clang compilers #3334
• CCache support on Windows #3257

Fixes

• Fix issue with some locales with OpenCL kernel generation #3294
• Internal improvements
• Fix leak in clfft on exit.
• Fix some cases where ndims was incorrectly used ot calculate shape #3277
• Fix issue when setDevice was not called in new threads #3269
• Restrict initializer list to just fundamental types #3264

Contributions

Special thanks to our contributors:
Carlo Cabrera
Guillaume Schmid
Willy Born
ktdq

v3.8.2

Improvements

• Optimize JIT by removing some consecutive cast operations #3031
• Add driver checks checks for CUDA 11.5 and 11.6 #3203
• Improve the timing algorithm used for timeit #3185
• Dynamically link against CUDA numeric libraries by default #3205
• Add support for pruning CUDA binaries to reduce static binary sizes #3234 #3237
• Remove unused cuDNN libraries from installations #3235
• Add support to staticly link NVRTC libraries after CUDA 11.5 #3236
• Add support for compiling with ccache when building the CUDA backend #3241

Fixes

• Fix issue with consecutive moddims operations in the CPU backend #3232
• Better floating point comparisons for tests #3212
• Fix several warnings and inconsistencies with doxygen and documentation #3226
• Fix issue when passing empty arrays into join #3211
• Fix default value for the AF_COMPUTE_LIBRARY when not set #3228
• Fix missing symbol issue when MKL is staticly linked #3244
• Remove linking of OpenCL's library to the unified backend #3244

Contributions

Special thanks to our contributors:
Jacob Kahn
Willy Born

v3.8.1

Improvements

• moddims now uses JIT approach for certain special cases - #3177
• Embed Version Info in Windows DLLs - #3025
• OpenCL device max parameter is now queries from device properties - #3032
• JIT Performance Optimization: Unique funcName generation sped up - #3040
• Improved readability of log traces - #3050
• Use short function name in non-debug build error messages - #3060
• SIFT/GLOH are now available as part of website binaries - #3071
• Short-circuit zero elements case in detail::copyArray backend function - #3059
• Speedup of kernel caching mechanism - #3043
• Add short-circuit check for empty Arrays in JIT evalNodes - #3072
• Performance optimization of indexing using dynamic thread block sizes - #3111
• ArrayFire starting with this release will use Intel MKL single dynamic library which resolves lot of linking issues unified library had when user applications used MKL themselves - #3120
• Add shortcut check for zero elements in af_write_array - #3130
• Speedup join by eliminating temp buffers for cascading joins - #3145
• Added batch support for solve - #1705
• Use pinned memory to copy device pointers in CUDA solve - #1705
• Added package manager instructions to docs - #3076
• CMake Build Improvements - #3027 , #3089 , #3037 , #3072 , #3095 , #3096 , #3097 , #3102 , #3106 , #3105 , #3120 , #3136 , #3135 , #3137 , #3119 , #3150 , #3138 , #3156 , #3139 , #1705 , #3162
• CPU backend improvements - #3010 , #3138 , #3161
• CUDA backend improvements - #3066 , #3091 , #3093 , #3125 , #3143 , #3161
• OpenCL backend improvements - #3091 , #3068 , #3127 , #3010 , #3039 , #3138 , #3161
• General(including JIT) performance improvements across backends - #3167
• Testing improvements - #3072 , #3131 , #3151 , #3141 , #3153 , #3152 , #3157 , #1705 , #3170 , #3167
• Update CLBlast to latest version - #3135 , #3179
• Improved Otsu threshold computation helper in canny algorithm - #3169
• Modified default parameters for fftR2C and fftC2R C++ API from 0 to 1.0 - #3178
• Use appropriate MKL getrs_batch_strided API based on MKL Versions - #3181

Fixes

• Fixed a bug JIT kernel disk caching - #3182
• Fixed stream used by thrust(CUDA backend) functions - #3029
• Added workaround for new cuSparse API that was added by CUDA amid fix releases - #3057
• Fixed const array indexing inside gfor - #3078
• Handle zero elements in copyData to host - #3059
• Fixed double free regression in OpenCL backend - #3091
• Fixed an infinite recursion bug in NaryNode JIT Node - #3072
• Added missing input validation check in sparse-dense arithmetic operations - #3129
• Fixed bug in getMappedPtr in OpenCL due to invalid lambda capture - #3163
• Fixed bug in getMappedPtr on Arrays that are not ready - #3163
• Fixed edgeTraceKernel for CPU devices on OpenCL backend - #3164
• Fixed windows build issue(s) with VS2019 - #3048
• API documentation fixes - #3075 , #3076 , #3143 , #3161
• CMake Build Fixes - #3088
• Fixed the tutorial link in README - #3033
• Fixed function name typo in timing tutorial - #3028
• Fixed couple of bugs in CPU backend canny implementation - #3169
• Fixed reference count of array(s) used in JIT operations. It is related to arrayfire's internal memory book keeping. The behavior/accuracy of arrayfire code wasn't broken earlier. It corrected the reference count to be of optimal value in the said scenarios. This may potentially reduce memory usage in some narrow cases - #3167
• Added assert that checks if topk is called with a negative value for k - #3176
• Fixed an Issue where countByKey would give incorrect results for any n > 128 - #3175

Contributions

Special thanks to our contributors: HO-COOH, Willy Born, Gilad Avidov, Pavan Yalamanchili

v3.8.0

New Functions

• Ragged max reduction - #2786
• Initialization list constructor for array class - #2829 , #2987
• New API for following statistics function: cov, var and stdev - #2986
• Bit-wise operator support for array and C API (af_bitnot) - #2865
• allocV2 and freeV2 which return cl_mem on OpenCL backend - #2911
• Move constructor and move assignment operator for Dim4 class - #2946

Improvements

• Add f16 support for histogram - #2984
• Update confidence connected components example for better illustration - #2968
• Enable disk caching of OpenCL kernel binaries - #2970
• Refactor extension of kernel binaries stored to disk .bin - #2970
• Add minimum driver versions for CUDA toolkit 11 in internal map - #2982
• Improve warnings messages from run-time kernel compilation functions - #2996

Fixes

• Fix bias factor of variance in var_all and cov functions - #2986
• Fix a race condition in confidence connected components function for OpenCL backend - #2969
• Safely ignore disk cache failures in CUDA backend for compiled kernel binaries - #2970
• Fix randn by passing in correct values to Box-Muller - #2980
• Fix rounding issues in Box-Muller function used for RNG - #2980
• Fix problems in RNG for older compute architectures with fp16 - #2980 #2996
• Fix performance regression of approx functions - #2977
• Remove assert that check that signal/filter types have to be the same - #2993
• Fix checkAndSetDevMaxCompute when the device cc is greater than max - #2996
• Fix documentation errors and warnings - #2973 , #2987
• Add missing opencl-arrayfire interoperability functions in unified back - #2981

Contributions

Special thanks to our contributors: P. J. Reed

v3.7.3

Improvements

• Add f16 support for histogram - #2984
• Update confidence connected components example for better illustration - #2968
• Enable disk caching of OpenCL kernel binaries - #2970
• Refactor extension of kernel binaries stored to disk .bin - #2970
• Add minimum driver versions for CUDA toolkit 11 in internal map - #2982
• Improve warnings messages from run-time kernel compilation functions - #2996

Fixes

• Fix bias factor of variance in var_all and cov functions - #2986
• Fix a race condition in confidence connected components function for OpenCL backend - #2969
• Safely ignore disk cache failures in CUDA backend for compiled kernel binaries - #2970
• Fix randn by passing in correct values to Box-Muller - #2980
• Fix rounding issues in Box-Muller function used for RNG - #2980
• Fix problems in RNG for older compute architectures with fp16 - #2980 #2996
• Fix performance regression of approx functions - #2977
• Remove assert that check that signal/filter types have to be the same - #2993
• Fix checkAndSetDevMaxCompute when the device cc is greater than max - #2996
• Fix documentation errors and warnings - #2973 , #2987
• Add missing opencl-arrayfire interoperability functions in unified back - #2981
• Fix constexpr relates compilation error with VS2019 and Clang Compilers - #3049

Contributions

Special thanks to our contributors: P. J. Reed

v3.8.0 Release Candidate

New Functions

• Ragged max reduction - #2786
• Initialization list constructor for array class - #2829 , #2987
• New API for following statistics function: cov, var and stdev - #2986
• Bit-wise operator support for array and C API (af_bitnot) - #2865
• allocV2 and freeV2 which return cl_mem on OpenCL backend - #2911
• Move constructor and move assignment operator for Dim4 class - #2946

Improvements

• Add f16 support for histogram - #2984
• Update confidence connected components example for better illustration - #2968
• Enable disk caching of OpenCL kernel binaries - #2970
• Refactor extension of kernel binaries stored to disk .bin - #2970
• Add minimum driver versions for CUDA toolkit 11 in internal map - #2982
• Improve warnings messages from run-time kernel compilation functions - #2996

Fixes

• Fix bias factor of variance in var_all and cov functions - #2986
• Fix a race condition in confidence connected components function for OpenCL backend - #2969
• Safely ignore disk cache failures in CUDA backend for compiled kernel binaries - #2970
• Fix randn by passing in correct values to Box-Muller - #2980
• Fix rounding issues in Box-Muller function used for RNG - #2980
• Fix problems in RNG for older compute architectures with fp16 - #2980 #2996
• Fix performance regression of approx functions - #2977
• Remove assert that check that signal/filter types have to be the same - #2993
• Fix checkAndSetDevMaxCompute when the device cc is greater than max - #2996
• Fix documentation errors and warnings - #2973 , #2987
• Add missing opencl-arrayfire interoperability functions in unified back - #2981

Contributions

Special thanks to our contributors: P. J. Reed

v3.7.2

Improvements

• Cache CUDA kernels to disk to improve load times(Thanks to @cschreib-ibex) #2848
• Staticly link against cuda libraries #2785
• Make cuDNN an optional build dependency #2836
• Improve support for different compilers and OS #2876 #2945 #2925 #2942 #2943 #2945
• Improve performance of join and transpose on CPU #2849
• Improve documentation #2816 #2821 #2846 #2918 #2928 #2947
• Reduce binary size using NVRTC and template reducing instantiations #2849 #2861 #2890
• Improve reduceByKey performance on OpenCL by using builtin functions #2851
• Improve support for Intel OpenCL GPUs #2855
• Allow staticly linking against MKL #2877 (Sponsered by SDL)
• Better support for older CUDA toolkits #2923
• Add support for CUDA 11 #2939
• Add support for ccache for faster builds #2931
• Add support for the conan package manager on linux #2875
• Propagate build errors up the stack in AFError exceptions #2948 #2957
• Improve runtime dependency library loading #2954
• Improved cuDNN runtime checks and warnings #2960
• Document af_memory_manager_* native memory return values #2911
• Add support for cuDNN 8 #2963

Fixes

• Bug crash when allocating large arrays #2827
• Fix various compiler warnings #2827 #2849 #2872 #2876
• Fix minor leaks in OpenCL functions #2913
• Various continuous integration related fixes #2819
• Fix zero padding with convolv2NN #2820
• Fix af_get_memory_pressure_threshold return value #2831
• Increased the max filter length for morph
• Handle empty array inputs for LU, QR, and Rank functions #2838
• Fix FindMKL.cmake script for sequential threading library #2840
• Various internal refactoring #2839 #2861 #2864 #2873 #2890 #2891 #2913
• Fix OpenCL 2.0 builtin function name conflict #2851
• Fix error caused when releasing memory with multiple devices #2867
• Fix missing set stacktrace symbol from unified API #2915
• Fix zero padding issue in convolve2NN #2820
• Fixed bugs in ReduceByKey #2957
• Add clblast patch to handle custom context with multiple devices #2967

Contributions

Special thanks to our contributors:
Corentin Schreiber
Jacob Kahn
Paul Jurczak
Christoph Junghans

v3.7.1

Improvements

• Improve mtx download for test data #2742
• Improve Documentation #2754 #2792 #2797
• Remove verbose messages in older CMake versions #2773
• Reduce binary size with the use of NVRTC #2790
• Use texture memory to load LUT in orb and fast #2791
• Add missing print function for f16 #2784
• Add checks for f16 support in the CUDA backend #2784
• Create a thrust policy to intercept temporary buffer allocations #2806

Fixes

• Fix segfault on exit when ArrayFire is not initialized in the main thread
• Fix support for CMake 3.5.1 #2771 #2772 #2760
• Fix evalMultiple if the input array sizes aren't the same #2766
• Fix error when AF_BACKEND_DEFAULT is passed directly to backend #2769
• Workaround name collision with AMD OpenCL implementation #2802
• Fix on-exit errors with the unified backend #2769
• Fix check for f16 compatibility in OpenCL #2773
• Fix matmul on Intel OpenCL when passing same array as input #2774
• Fix CPU OpenCL blas batching #2774
• Fix memory pressure in the default memory manager #2801

Contributions

Special thanks to our contributors:
padentomasello
glavaux2

v3.7.0

Major Updates

• Added the ability to customize the memory manager(Thanks jacobkahn and flashlight) [#2461]
• Added 16-bit floating point support for several functions [#2413] [#2587] [#2585] [#2587] [#2583]
• Added sumByKey, productByKey, minByKey, maxByKey, allTrueByKey, anyTrueByKey, countByKey [#2254]
• Added confidence connected components [#2748]
• Added neural network based convolution and gradient functions [#2359]
• Added a padding function [#2682]
• Added pinverse for pseudo inverse [#2279]
• Added support for uniform ranges in approx1 and approx2 functions. [#2297]
• Added support to write to preallocated arrays for some functions [#2599] [#2481] [#2328] [#2327]
• Added meanvar function [#2258]
• Add support for sparse-sparse arithmetic support [#2312]
• Added rsqrt function for reciprocal square root [#2500]
• Added a lower level af_gemm function for general matrix multiplication [#2481]
• Added a function to set the cuBLAS math mode for the CUDA backend [#2584]
• Separate debug symbols into separate files [#2535]
• Print stacktraces on errors [#2632]
• Support move constructor for af::array [#2595]
• Expose events in the public API [#2461]
• Add setAxesLabelFormat to format labels on graphs [#2495]
• Added deconvolution functions [#1881]

Improvements

• Better error messages for systems with driver or device incompatibilities [#2678] [#2448][#2761]
• Optimized unified backend function calls [#2695]
• Optimized anisotropic smoothing [#2713]
• Optimized canny filter for CUDA and OpenCL [#2727]
• Better MKL search script [#2738][#2743][#2745]
• Better logging of different submodules in ArrayFire [#2670] [#2669]
• Improve documentation [#2665] [#2620] [#2615] [#2639] [#2628] [#2633] [#2622] [#2617] [#2558] [#2326][#2515]
• Optimized af::array assignment [#2575]
• Update the k-means example to display the result [#2521]

Fixes

• Fix multi-config generators [#2736]
• Fix access errors in canny [#2727]
• Fix segfault in the unified backend if no backends are available [#2720]
• Fix access errors in scan-by-key [#2693]
• Fix sobel operator [#2600]
• Fix an issue with the random number generator and s16 [#2587]
• Fix issue with boolean product reduction [#2544]
• Fix array_proxy move constructor [#2537]
• Fix convolve3 launch configuration [#2519]
• Fix an issue where the fft function modified the input array [#2520]
• Added a work around for nvidia-opencl runtime if forge dependencies are missing [#2761]

Contributions

Special thanks to our contributors:
@jacobkahn
@WilliamTambellini
@lehins
@r-barnes
@gaika
@ShalokShalom

v3.6.4

The source code with sub-modules can be downloaded directly from the following link:

http://arrayfire.com/arrayfire_source/arrayfire-full-3.6.4.tar.bz2

Fixes

• Address a JIT performance regression due to moving kernel arguments to shared memory #2501
• Fix the default parameter for setAxisTitle #2491

v3.6.3

The source code with sub-modules can be downloaded directly from the following link:

http://arrayfire.com/arrayfire_source/arrayfire-full-3.6.3.tar.bz2

Improvements

• Graphics are now a runtime dependency instead of a link time dependency #2365
• Reduce the CUDA backend binary size using runtime compilation of kernels #2437
• Improved batched matrix multiplication on the CPU backend by using Intel MKL's cblas_Xgemm_batched#2206
• Print JIT kernels to disk or stream using the AF_JIT_KERNEL_TRACE environment variable #2404
• void* pointers are now allowed as arguments to af::array::write() #2367
• Slightly improve the efficiency of JITed tile operations #2472
• Make the random number generation on the CPU backend to be consistent with CUDA and OpenCL #2435
• Handled very large JIT tree generations #2484 #2487

Bug Fixes

• Fixed af::array::array_proxy move assignment operator #2479
• Fixed input array dimensions validation in svdInplace() #2331
• Fixed the typedef declaration for window resource handle #2357.
• Increase compatibility with GCC 8 #2379
• Fixed af::write tests #2380
• Fixed a bug in broadcast step of 1D exclusive scan #2366
• Fixed OpenGL related build errors on OSX #2382
• Fixed multiple array evaluation. Performance improvement. #2384
• Fixed buffer overflow and expected output of kNN SSD small test #2445
• Fixed MKL linking order to enable threaded BLAS #2444
• Added validations for forge module plugin availability before calling resource cleanup #2443
• Improve compatibility on MSVC toolchain(_MSC_VER > 1914) with the CUDA backend #2443
• Fixed BLAS gemm func generators for newest MSVC 19 on VS 2017 #2464
• Fix errors on exits when using the cuda backend with unified #2470

Documentation

• Updated svdInplace() documentation following a bugfix #2331
• Fixed a typo in matrix multiplication documentation #2358
• Fixed a code snippet demonstrating C-API use #2406
• Updated hamming matcher implementation limitation #2434
• Added illustration for the rotate function #2453

Misc

• Use cudaMemcpyAsync instead of cudaMemcpy throughout the codebase #2362
• Display a more informative error message if CUDA driver is incompatible #2421 #2448
• Changed forge resource management to use smart pointers #2452
• Deprecated intl and uintl typedefs in API #2360
• Enabled graphics by default for all builds starting with v3.6.3 #2365
• Fixed several warnings #2344 #2356 #2361
• Refactored initArray() calls to use createEmptyArray(). initArray() is for internal use only by Array class. #2361
• Refactored void* memory allocations to use unsigned char type #2459
• Replaced deprecated MKL API with in-house implementations for sparse to sparse/dense conversions #2312
• Reorganized and fixed some internal backend API #2356
• Updated compilation order of CUDA files to speed up compile time #2368
• Removed conditional graphics support builds after enabling runtime loading of graphics dependencies #2365
• Marked graphics dependencies as optional in CPack RPM config #2365
• Refactored a sparse arithmetic backend API #2379
• Fixed const correctness of af_device_array API #2396
• Update Forge to v1.0.4 #2466
• Manage Forge resources from the DeviceManager class #2381
• Fixed non-mkl & non-batch blas upstream call arguments #2401
• Link MKL with OpenMP instead of TBB by default
• use clang-format to format source code

Contributions

Special thanks to our contributors:
Alessandro Bessi
zhihaoy
Jacob Khan
William Tambellini

v3.6.2

The source code with sub-modules can be downloaded directly from the following link:

http://arrayfire.com/arrayfire_source/arrayfire-full-3.6.2.tar.bz2

Features

• Batching support for cond argument in select() [#2243]
• Broadcast batching for matmul [#2315]
• Add support for multiple nearest neighbours from nearestNeighbour() [#2280]

Improvements

• Performance improvements in morph() [#2238]
• Fix linking errors when compiling without Freeimage/Graphics [#2248]
• Fixes to improve the usage of ArrayFire as a sub-project [#2290]
• Allow custom library path for loading dynamic backend libraries [#2302]

Bug fixes

• Fix overflow in dim4::ndims. [#2289]
• Remove setDevice from af::array destructor [#2319]
• Fix pow precision for integral types [#2305]
• Fix issues with tile with a large repeat dimension [#2307]
• Fix grid based indexing calculation in af_draw_hist [#2230]
• Fix bug when using an af::array for indexing [#2311]
• Fix CLBlast errors on exit on Windows [#2222]

Documentation

• Improve unwrap documentation [#2301]
• Improve wrap documentation [#2320]
• Fix and improve accum documentation [#2298]
• Improve tile documentation [#2293]
• Clarify approx* indexing in documentation [#2287]
• Update examples of select in detailed documentation [#2277]
• Update lookup examples [#2288]
• Update set documentation [#2299]

Misc

• New ArrayFire ASSERT utility functions [#2249][#2256][#2257][#2263]
• Improve error messages in JIT [#2309]
• af* library and dependencies directory changed to lib64 [#2186]

Contributions

Thank you to our contributors:
Jacob Kahn
Vardan Akopian

v 3.6.1

The source code for this release can be downloaded here:
http://arrayfire.com/arrayfire_source/arrayfire-full-3.6.1.tar.bz2

Improvements

• FreeImage is now a run-time dependency [#2164]
• Reduced binary size by setting the symbol visibility to hidden [#2168]
• Add logging to memory manager and unified loader using the AF_TRACE environment variable [#2169][#2216]
• Improved CPU Anisotropic Diffusion performance [#2174]
• Perform normalization after FFT for improved accuracy [#2185, #2192]
• Updated CLBlast to v1.4.0 [#2178]
• Added additional validation when using af::seq for indexing [#2153]
• Perform checks for unsupported cards by the CUDA implementation [#2182]
• Avoid selecting backend if no devices are found. [#2218]

Bug Fixes

• Fixed region when all pixels were the foreground or background [#2152]
• Fixed several memory leaks [#2202, #2201, #2180, #2179, #2177, #2175]
• Fixed bug in setDevice which didn't allow you to select the last device [#2189]
• Fixed bug in min/max where the first element of the array was a NaN value [#2155]
• Fixed graphics window indexing [#2207]
• Fixed renaming issue when installing cuda libraries on OSX [#2221]
• Fixed NSIS installer PATH variable [#2223]

v3.6.0

The source code with submodules can be downloaded directly from the following link:
http://arrayfire.com/arrayfire_source/arrayfire-full-3.6.0.tar.bz2

Major Updates

• Added the topk() function. 1
• Added batched matrix multiply support.2 3
• Added anisotropic diffusion, anisotropicDiffusion().Documentation 3.

Features

• Added support for batched matrix multiply. 1 2
• New anisotropic diffusion function, anisotropicDiffusion(). Documentation 3.
• New topk() function, which returns the top k elements along a given dimension of the input. Documentation. 4
• New gradient diffusion example.

Improvements

• JITed select() and shift() functions for CUDA and OpenCL backends. 1
• Significant CMake improvements. 2 3 4
• Improved the quality of the random number generator 5
• Corrected assert function calls in select() tests. 5
• Modified af_colormap struct to match forge's definition. 6
• Improved Black Scholes example. 7
• Used CPack to generate installers. 8. We will be using CPack to generate installers beginning with this release.
• Refactored black_scholes_options example to use built-in af::erfc function for cumulative normal distribution.9.
• Reduced the scope of mutexes in memory manager 10
• Official installers do not require the CUDA toolkit to be installed starting with v3.6.0.

Bug fixes

• Fixed shfl_down() warnings with CUDA 9. 1
• Disabled CUDA JIT debug flags on ARM architecture.2
• Fixed CLBLast install lib dir for linux platform where lib directory has arch(64) suffix.3
• Fixed assert condition in 3d morph opencl kernel.4
• Fixed JIT errors with large non-linear kernels5
• Fixed bug in CPU JIT after moddims was called 5
• Fixed a deadlock scenario caused by the method MemoryManager::nativeFree6

Documentation

• Fixed variable name typo in vectorization.md. 1
• Fixed AF_API_VERSION value in Doxygen config file. 2

Known issues

• NVCC does not currently support platform toolset v141 (Visual Studio 2017 R15.6). Use the v140 platform toolset, instead. You may pass in the toolset version to CMake via the -T flag like so cmake -G "Visual Studio 15 2017 Win64" -T v140.
  • To download and install other platform toolsets, visit https://blogs.msdn.microsoft.com/vcblog/2017/11/15/side-by-side-minor-version-msvc-toolsets-in-visual-studio-2017
• Several OpenCL tests failing on OSX:
  • canny_opencl, fft_opencl, gen_assign_opencl, homography_opencl, reduce_opencl, scan_by_key_opencl, solve_dense_opencl, sparse_arith_opencl, sparse_convert_opencl, where_opencl

Contributions

Special thanks to our contributors:
Adrien F. Vincent, Cedric Nugteren, Felix, Filip Matzner, HoneyPatouceul, Patrick Lavin, Ralf Stubner, William Tambellini

v3.5.1

The source code with submodules can be downloaded directly from the following
link: http://arrayfire.com/arrayfire_source/arrayfire-full-3.5.1.tar.bz2

Installer CUDA Version: 8.0 (Required)
Installer OpenCL Version: 1.2 (Minimum)

Improvements

• Relaxed af::unwrap() function's arguments. 1
• Changed behavior of af::array::allocated() to specify memory allocated. 1
• Removed restriction on the number of bins for af::histogram() on CUDA and
  OpenCL kernels. 1

Performance

• Improved JIT performance. 1
• Improved CPU element-wise operation performance. 1
• Improved regions performance using texture objects. 1

Bug fixes

• Fixed overflow issues in mean. 1
• Fixed memory leak when chaining indexing operations. 1
• Fixed bug in array assignment when using an empty array to index. 1
• Fixed bug with af::matmul() which occured when its RHS argument was an
  indexed vector. 1
• Fixed bug deadlock bug when sparse array was used with a JIT Array. 1
• Fixed pixel tests for FAST kernels. 1
• Fixed af::replace so that it is now copy-on-write. 1
• Fixed launch configuration issues in CUDA JIT. 1
• Fixed segfaults and "Pure Virtual Call" error warnings when exiting on
  Windows. 1 2
• Workaround for clEnqueueReadBuffer bug on OSX.
  1

Build

• Fixed issues when compiling with GCC 7.1. 1 2
• Eliminated unnecessary Boost dependency from CPU and CUDA backends. 1

Misc

• Updated support links to point to Slack instead of Gitter. 1

v3.5.0

The source code with submodules can be downloaded directly from the following link:
http://arrayfire.com/arrayfire_source/arrayfire-full-3.5.0.tar.bz2

Installer CUDA Version: 8.0 (Required)
Installer OpenCL Version: 1.2 (Minimum)

Major Updates

• ArrayFire now supports threaded applications. 1
• Added Canny edge detector. 1
• Added Sparse-Dense arithmetic operations. 1

Features

• ArrayFire Threading
  • af::array can be read by multiple threads
  • All ArrayFire functions can be executed concurrently by multiple threads
  • Threads can operate on different devices to simplify Muli-device workloads
• New Canny edge detector function, af::canny(). 1
  • Can automatically calculate high threshold with AF_CANNY_THRESHOLD_AUTO_OTSU
  • Supports both L1 and L2 Norms to calculate gradients
• New tuned OpenCL BLAS backend, CLBlast.

Improvements

• Converted CUDA JIT to use NVRTC instead of NVVM.
• Performance improvements in af::reorder(). 1
• Performance improvements in array::scalar(). 1
• Improved unified backend performance. 1
• ArrayFire now depends on Forge v1.0. 1
• Can now specify the FFT plan cache size using the af::setFFTPlanCacheSize() function.
• Get the number of physical bytes allocated by the memory manager af_get_allocated_bytes(). 1
• af::dot() can now return a scalar value to the host. 1

Bug Fixes

• Fixed improper release of default Mersenne random engine. 1
• Fixed af::randu() and af::randn() ranges for floating point types. 1
• Fixed assignment bug in CPU backend. 1
• Fixed complex (c32,c64) multiplication in OpenCL convolution kernels. 1
• Fixed inconsistent behavior with af::replace() and replace_scalar(). 1
• Fixed memory leak in af_fir(). 1
• Fixed memory leaks in af_cast for sparse arrays. 1
• Fixing correctness of af_pow for complex numbers by using Cartesian form. 1
• Corrected af::select() with indexing in CUDA and OpenCL backends. 1
• Workaround for VS2015 compiler ternary bug. 1
• Fixed memory corruption in cuda::findPlan(). 1
• Argument checks in af_create_sparse_array avoids inputs of type int64. 1

Build fixes

• On OSX, utilize new GLFW package from the brew package manager. 1 2
• Fixed CUDA PTX names generated by CMake v3.7. 1
• Support gcc > 5.x for CUDA. 1

Examples

• New genetic algorithm example. 1

Documentation

• Updated README.md to improve readability and formatting. 1
• Updated README.md to mention Julia and Nim wrappers. 1
• Improved installation instructions - docs/pages/install.md. 1

Miscellaneous

• A few improvements for ROCm support. 1
• Removed CUDA 6.5 support. 1

Known issues

• Windows
  • The Windows NVIDIA driver version 37x.xx contains a bug which causes fftconvolve_opencl to fail. Upgrade or downgrade to a different version of the driver to avoid this failure.
  • The following tests fail on Windows with NVIDIA hardware: threading_cuda,qr_dense_opencl, solve_dense_opencl.
• macOS
  • The Accelerate framework, used by the CPU backend on macOS, leverages Intel graphics cards (Iris) when there are no discrete GPUs available. This OpenCL implementation is known to give incorrect results on the following tests: lu_dense_{cpu,opencl}, solve_dense_{cpu,opencl}, inverse_dense_{cpu,opencl}.
  • Certain tests intermittently fail on macOS with NVIDIA GPUs apparently due to inconsistent driver behavior: fft_large_cuda and svd_dense_cuda.
  • The following tests are currently failing on macOS with AMD GPUs: cholesky_dense_opencl and scan_by_key_opencl.

v3.4.2

The source code with submodules can be downloaded directly from the following link:
http://arrayfire.com/arrayfire_source/arrayfire-full-3.4.2.tar.bz2

Installer CUDA Version: 8.0 (Required)
Installer OpenCL Version: 1.2 (Minimum)

Deprecation Announcement

This release supports CUDA 6.5 and higher. The next ArrayFire release will
support CUDA 7.0 and higher, dropping support for CUDA 6.5. Reasons for no
longer supporting CUDA 6.5 include:

• CUDA 7.0 NVCC supports the C++11 standard (whereas CUDA 6.5 does not), which
  is used by ArrayFire's CPU and OpenCL backends.
• Very few ArrayFire users still use CUDA 6.5.

As a result, the older Jetson TK1 / Tegra K1 will no longer be supported in
the next ArrayFire release. The newer Jetson TX1 / Tegra X1 will continue to
have full capability with ArrayFire.

Docker

• ArrayFire has been Dockerized.

Improvements

• Implemented sparse storage format conversions between AF_STORAGE_CSR
  and AF_STORAGE_COO.
  1
  • Directly convert between AF_STORAGE_COO <--> AF_STORAGE_CSR
    using the af::sparseConvertTo() function.
  • af::sparseConvertTo() now also supports converting to dense.
• Added cast support for sparse arrays.
  1
  • Casting only changes the values array and the type. The row and column
    index arrays are not changed.
• Reintroduced automated computation of chart axes limits for graphics functions.
  1
  • The axes limits will always be the minimum/maximum of the current and new
    limit.
  • The user can still set limits from API calls. If the user sets a limit
    from the API call, then the automatic limit setting will be disabled.
• Using boost::scoped_array instead of boost::scoped_ptr when managing
  array resources.
  1
• Internal performance improvements to getInfo() by using const references
  to avoid unnecessary copying of ArrayInfo objects.
  1
• Added support for scalar af::array inputs for af::convolve() and
  set functions.
  1
  2
  3
• Performance fixes in af::fftConvolve() kernels.
  1
  2

Build

• Support for Visual Studio 2015 compilation.
  1
  2
• Fixed FindCBLAS.cmake when PkgConfig is used.
  1

Bug fixes

• Fixes to JIT when tree is large.
  1
  2
• Fixed indexing bug when converting dense to sparse af::array as
  AF_STORAGE_COO.
  1
• Fixed af::bilateral() OpenCL kernel compilation on OS X.
  1
• Fixed memory leak in af::regions() (CPU) and af::rgb2ycbcr().
  1
  2
  3

Installers

• Major OS X installer fixes.
  1
  • Fixed installation scripts.
  • Fixed installation symlinks for libraries.
• Windows installer now ships with more pre-built examples.

Examples

• Added af::choleskyInPlace() calls to cholesky.cpp example.
  1

Documentation

• Added u8 as supported data type in getting_started.md.
  1
• Fixed typos.
  1

CUDA 8 on OSX

• CUDA 8.0.55 supports Xcode 8.
  1

Known Issues

• Known failures with CUDA 6.5. These include all functions that use
  sorting. As a result, sparse storage format conversion between
  AF_STORAGE_COO and AF_STORAGE_CSR has been disabled for CUDA 6.5.

v3.4.1

The source code with submodules can be downloaded directly from the following link:
http://arrayfire.com/arrayfire_source/arrayfire-full-3.4.1.tar.bz2

Installer CUDA Version: 8.0 (Required)
Installer OpenCL Version: 1.2 (Minimum)

Installers

• Installers for Linux, OS X and Windows
  • CUDA backend now uses CUDA 8.0.
  • Uses Intel MKL 2017.
  • CUDA Compute 2.x (Fermi) is no longer compiled into the library.
• Installer for OS X
  • The libraries shipping in the OS X Installer are now compiled with Apple
    Clang v7.3.1 (previouly v6.1.0).
  • The OS X version used is 10.11.6 (previously 10.10.5).
• Installer for Jetson TX1 / Tegra X1
  • Requires JetPack for L4T 2.3
    (containing Linux for Tegra r24.2 for TX1).
  • CUDA backend now uses CUDA 8.0 64-bit.
  • Using CUDA's cusolver instead of CPU fallback.
  • Uses OpenBLAS for CPU BLAS.
  • All ArrayFire libraries are now 64-bit.

Improvements

• Add sparse array support to af::eval().
  1
• Add OpenCL-CPU fallback support for sparse af::matmul() when running on
  a unified memory device. Uses MKL Sparse BLAS.
• When using CUDA libdevice, pick the correct compute version based on device.
  1
• OpenCL FFT now also supports prime factors 7, 11 and 13.
  1
  2

Bug Fixes

• Allow CUDA libdevice to be detected from custom directory.
• Fix aarch64 detection on Jetson TX1 64-bit OS.
  1
• Add missing definition of af_set_fft_plan_cache_size in unified backend.
  1
• Fix intial values for af::min() and af::max() operations.
  1
  2
• Fix distance calculation in af::nearestNeighbour for CUDA and OpenCL backend.
  1
  2
• Fix OpenCL bug where scalars where are passed incorrectly to compile options.
  1
• Fix bug in af::Window::surface() with respect to dimensions and ranges.
  1
• Fix possible double free corruption in af_assign_seq().
  1
• Add missing eval for key in af::scanByKey in CPU backend.
  1
• Fixed creation of sparse values array using AF_STORAGE_COO.
  1
  1

Examples

• Add a Conjugate Gradient solver example
  to demonstrate sparse and dense matrix operations.
  1

CUDA Backend

• When using CUDA 8.0,
  compute 2.x are no longer in default compute list.
  • This follows CUDA 8.0
    deprecating computes 2.x.
  • Default computes for CUDA 8.0 will be 30, 50, 60.
• When using CUDA pre-8.0, the default selection remains 20, 30, 50.
• CUDA backend now uses -arch=sm_30 for PTX compilation as default.
  • Unless compute 2.0 is enabled.

Known Issues

• af::lu() on CPU is known to give incorrect results when built run on
  OS X 10.11 or 10.12 and compiled with Accelerate Framework.
  1
  • Since the OS X Installer libraries uses MKL rather than Accelerate
    Framework, this issue does not affect those libraries.

v3.4.0

The source code with submodules can be downloaded directly from the following link:
http://arrayfire.com/arrayfire_source/arrayfire-full-3.4.0.tar.bz2

Installer CUDA Version: 7.5 (Required)
Installer OpenCL Version: 1.2 (Minimum)

Major Updates

• [Sparse Matrix and BLAS](ref sparse_func). 1
  2
• Faster JIT for CUDA and OpenCL. 1
  2
• Support for [random number generator engines](ref af::randomEngine).
  1
  2
• Improvements to graphics. 1
  2

Features

• [Sparse Matrix and BLAS](ref sparse_func) 1
  2
  • Support for [CSR](ref AF_STORAGE_CSR) and [COO](ref AF_STORAGE_COO)
    [storage types](ref af_storage).
  • Sparse-Dense Matrix Multiplication and Matrix-Vector Multiplication as a
    part of af::matmul() using AF_STORAGE_CSR format for sparse.
  • Conversion to and from [dense](ref AF_STORAGE_DENSE) matrix to [CSR](ref AF_STORAGE_CSR)
    and [COO](ref AF_STORAGE_COO) [storage types](ref af_storage).
• Faster JIT 1
  2
  • Performance improvements for CUDA and OpenCL JIT functions.
  • Support for evaluating multiple outputs in a single kernel. See af::array::eval() for more.
• [Random Number Generation](ref af::randomEngine)
  1
  2
  • af::randomEngine(): A random engine class to handle setting the type and seed
    for random number generator engines.
  • Supported engine types are:
    • Philox
    • Threefry
    • Mersenne Twister
• Graphics 1
  2
  • Using Forge v0.9.0
  • [Vector Field](ref af::Window::vectorField) plotting functionality.
    1
  • Removed GLEW and replaced with glbinding.
    • Removed usage of GLEW after support for MX (multithreaded) was dropped in v2.0.
      1
  • Multiple overlays on the same window are now possible.
    • Overlays support for same type of object (2D/3D)
    • Supported by af::Window::plot, af::Window::hist, af::Window::surface,
      af::Window::vectorField.
  • New API to set axes limits for graphs.
    • Draw calls do not automatically compute the limits. This is now under user control.
    • af::Window::setAxesLimits can be used to set axes limits automatically or manually.
    • af::Window::setAxesTitles can be used to set axes titles.
  • New API for plot and scatter:
    • af::Window::plot() and af::Window::scatter() now can handle 2D and 3D and determine appropriate order.
    • af_draw_plot_nd()
    • af_draw_plot_2d()
    • af_draw_plot_3d()
    • af_draw_scatter_nd()
    • af_draw_scatter_2d()
    • af_draw_scatter_3d()
• New [interpolation methods](ref af_interp_type)
  1
  • Applies to
    • af::resize()
    • af::transform()
    • af::approx1()
    • af::approx2()
• Support for [complex mathematical functions](ref mathfunc_mat)
  1
  • Add complex support for trig_mat, af::sqrt(), af::log().
• af::medfilt1(): Median filter for 1-d signals 1
• Generalized scan functions: scan_func_scan and scan_func_scanbykey
  • Now supports inclusive or exclusive scans
  • Supports binary operations defined by af_binary_op.
    1
• [Image Moments](ref moments_mat) functions
  1
• Add af::getSizeOf() function for af_dtype
  1
• Explicitly extantiate af::array::device() for `void *
  1

Bug Fixes

• Fixes to edge-cases in morph_mat. 1
• Makes JIT tree size consistent between devices. 1
• Delegate higher-dimension in convolve_mat to correct dimensions. 1
• Indexing fixes with C++11. 1 2
• Handle empty arrays as inputs in various functions. 1
• Fix bug when single element input to af::median. 1
• Fix bug in calculation of time from af::timeit(). 1
• Fix bug in floating point numbers in af::seq. 1
• Fixes for OpenCL graphics interop on NVIDIA devices.
  1
• Fix bug when compiling large kernels for AMD devices.
  1
• Fix bug in af::bilateral when shared memory is over the limit.
  1
• Fix bug in kernel header compilation tool bin2cpp.
  1
• Fix inital values for morph_mat functions.
  1
• Fix bugs in af::homography() CPU and OpenCL kernels.
  1
• Fix bug in CPU TNJ.
  1

Improvements

• CUDA 8 and compute 6.x(Pascal) support, current installer ships with CUDA 7.5. 1 2 3
• User controlled FFT plan caching. 1
• CUDA performance improvements for image_func_wrap, image_func_unwrap and approx_mat.
  1
• Fallback for CUDA-OpenGL interop when no devices does not support OpenGL.
  1
• Additional forms of batching with the transform_func_transform functions.
  New behavior defined here.
  1
• Update to OpenCL2 headers. 1
• Support for integration with external OpenCL contexts. 1
• Performance improvements to interal copy in CPU Backend.
  1
• Performance improvements to af::select and af::replace CUDA kernels.
  1
• Enable OpenCL-CPU offload by default for devices with Unified Host Memory.
  1
  • To disable, use the environment variable AF_OPENCL_CPU_OFFLOAD=0.

Build

• Compilation speedups. 1
• Build fixes with MKL. 1
• Error message when CMake CUDA Compute Detection fails. 1
• Several CMake build issues with Xcode generator fixed.
  1 2
• Fix multiple OpenCL definitions at link time. 1
• Fix lapacke detection in CMake. 1
• Update build tags of
  • clBLAS
  • clFFT
  • Boost.Compute
  • Forge
  • glbinding
• Fix builds with GCC 6.1.1 and GCC 5.3.0. 1

Installers

• All installers now ship with ArrayFire libraries build with MKL 2016.
• All installers now ship with Forge development files and examples included.
• CUDA Compute 2.0 has been removed from the installers. Please contact us
  directly if you have a special need.

Examples

• Added [example simulating gravity](ref graphics/field.cpp) for
  demonstration of vector field.
• Improvements to financial/black_scholes_options.cpp example.
• Improvements to graphics/gravity_sim.cpp example.
• Fix graphics examples to use af::Window::setAxesLimits and
  af::Window::setAxesTitles functions.

Documentation & Licensing

• ArrayFire copyright and trademark policy
• Fixed grammar in license.
• Add license information for glbinding.
• Remove license infomation for GLEW.
• Random123 now applies to all backends.
• Random number functions are now under random_mat.

Deprecations

The following functions have been deprecated and may be modified or removed
permanently from future versions of ArrayFire.

• af::Window::plot3(): Use af::Window::plot instead.
• af_draw_plot(): Use af_draw_plot_nd or af_draw_plot_2d instead.
• af_draw_plot3(): Use af_draw_plot_nd or af_draw_plot_3d instead.
• af::Window::scatter3(): Use af::Window::scatter instead.
• af_draw_scatter(): Use af_draw_scatter_nd or af_draw_scatter_2d instead.
• af_draw_scatter3(): Use af_draw_scatter_nd or af_draw_scatter_3d instead.

Known Issues

Certain CUDA functions are known to be broken on Tegra K1. The following ArrayFire tests are currently failing:

• assign_cuda
• harris_cuda
• homography_cuda
• median_cuda
• orb_cudasort_cuda
• sort_by_key_cuda
• sort_index_cuda