1 Release Notes {#releasenotes}
8 - Add oneAPI backend \PR{3296}
9 - Add support to directly access arrays on other devices \PR{3447}
10 - Add broadcast support \PR{2871}
11 - Improve OpenCL CPU JIT performance \PR{3257} \PR{3392}
12 - Optimize thread/block calculations of several kernels \PR{3144}
13 - Add support for fast math compiliation when building ArrayFire \PR{3334 \PR{3337}
14 - Optimize performance of fftconvolve when using floats \PR{3338}
15 - Add support for CUDA 12.1 and 12.2
16 - Better handling of empty arrays \PR{3398}
17 - Better handling of memory in linear algebra functions in OpenCL \PR{3423}
18 - Better logging with JIT kernels \PR{3468}
19 - Optimize memory manager/JIT interactions for small number of buffers \PR{3468}
20 - Documentation improvements \PR{3485}
21 - Optimize reorder function \PR{3488}
24 - Improve Errors when creating OpenCL contexts from devices \PR{3257}
25 - Improvements to vcpkg builds \PR{3376 \PR{3476}
26 - Fix reduce by key when nan's are present \PR{3261}
27 - Fix error in convolve where the ndims parameter was forced to be equal to 2 \PR{3277}
28 - Make constructors that accept dim_t to be explicit to avoid invalid conversions \PR{3259}
29 - Fix error in randu when compiling against clang 14 \PR{3333}
30 - Fix bug in OpenCL linear algebra functions \PR{3398}
31 - Fix bug with thread local variables when device was changed \PR{3420} \PR{3421}
32 - Fix bug in qr related to uninitialized memory \PR{3422}
33 - Fix bug in shift where the array had an empty middle dimension \PR{3488}
38 Special thanks to our contributors:
39 [Willy Born](https://github.com/willyborn)
40 [Mike Mullen](https://github.com/mfzmullen)
47 - Add support for CUDA 12 \PR{3352}
48 - Modernize documentation style and content \PR{3351}
49 - memcpy performance improvements \PR{3144}
50 - JIT performance improvements \PR{3144}
51 - join performance improvements \PR{3144}
52 - Improve support for Intel and newer Clang compilers \PR{3334}
53 - CCache support on Windows \PR{3257}
57 - Fix issue with some locales with OpenCL kernel generation \PR{3294}
58 - Internal improvements
59 - Fix leak in clfft on exit.
60 - Fix some cases where ndims was incorrectly used ot calculate shape \PR{3277}
61 - Fix issue when setDevice was not called in new threads \PR{3269}
62 - Restrict initializer list to just fundamental types \PR{3264}
66 Special thanks to our contributors:
67 [Carlo Cabrera](https://github.com/carlocab)
68 [Guillaume Schmid](https://github.com/GuillaumeSchmid)
69 [Willy Born](https://github.com/willyborn)
70 [ktdq](https://github.com/ktdq)
78 - Optimize JIT by removing some consecutive cast operations \PR{3031}
79 - Add driver checks checks for CUDA 11.5 and 11.6 \PR{3203}
80 - Improve the timing algorithm used for timeit \PR{3185}
81 - Dynamically link against CUDA numeric libraries by default \PR{3205}
82 - Add support for pruning CUDA binaries to reduce static binary sizes \PR{3234} \PR{3237}
83 - Remove unused cuDNN libraries from installations \PR{3235}
84 - Add support to staticly link NVRTC libraries after CUDA 11.5 \PR{3236}
85 - Add support for compiling with ccache when building the CUDA backend \PR{3241}
86 - Make cuSparse an optional runtime dependency \PR{3240}
90 - Fix issue with consecutive moddims operations in the CPU backend \PR{3232}
91 - Better floating point comparisons for tests \PR{3212}
92 - Fix several warnings and inconsistencies with doxygen and documentation \PR{3226}
93 - Fix issue when passing empty arrays into join \PR{3211}
94 - Fix default value for the `AF_COMPUTE_LIBRARY` when not set \PR{3228}
95 - Fix missing symbol issue when MKL is staticly linked \PR{3244}
96 - Remove linking of OpenCL's library to the unified backend \PR{3244}
100 Special thanks to our contributors:
101 [Jacob Kahn](https://github.com/jacobkahn)
102 [Willy Born](https://github.com/willyborn)
109 - moddims now uses JIT approach for certain special cases - \PR{3177}
110 - Embed Version Info in Windows DLLs - \PR{3025}
111 - OpenCL device max parameter is now queries from device properties - \PR{3032}
112 - JIT Performance Optimization: Unique funcName generation sped up - \PR{3040}
113 - Improved readability of log traces - \PR{3050}
114 - Use short function name in non-debug build error messages - \PR{3060}
115 - SIFT/GLOH are now available as part of website binaries - \PR{3071}
116 - Short-circuit zero elements case in detail::copyArray backend function - \PR{3059}
117 - Speedup of kernel caching mechanism - \PR{3043}
118 - Add short-circuit check for empty Arrays in JIT evalNodes - \PR{3072}
119 - Performance optimization of indexing using dynamic thread block sizes - \PR{3111}
120 - 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 - \PR{3120}
121 - Add shortcut check for zero elements in af_write_array - \PR{3130}
122 - Speedup join by eliminating temp buffers for cascading joins - \PR{3145}
123 - Added batch support for solve - \PR{1705}
124 - Use pinned memory to copy device pointers in CUDA solve - \PR{1705}
125 - Added package manager instructions to docs - \PR{3076}
126 - CMake Build Improvements - \PR{3027} , \PR{3089} , \PR{3037} , \PR{3072} , \PR{3095} , \PR{3096} , \PR{3097} , \PR{3102} , \PR{3106} , \PR{3105} , \PR{3120} , \PR{3136} , \PR{3135} , \PR{3137} , \PR{3119} , \PR{3150} , \PR{3138} , \PR{3156} , \PR{3139} , \PR{1705} , \PR{3162}
127 - CPU backend improvements - \PR{3010} , \PR{3138} , \PR{3161}
128 - CUDA backend improvements - \PR{3066} , \PR{3091} , \PR{3093} , \PR{3125} , \PR{3143} , \PR{3161}
129 - OpenCL backend improvements - \PR{3091} , \PR{3068} , \PR{3127} , \PR{3010} , \PR{3039} , \PR{3138} , \PR{3161}
130 - General(including JIT) performance improvements across backends - \PR{3167}
131 - Testing improvements - \PR{3072} , \PR{3131} , \PR{3151} , \PR{3141} , \PR{3153} , \PR{3152} , \PR{3157} , \PR{1705} , \PR{3170} , \PR{3167}
132 - Update CLBlast to latest version - \PR{3135} , \PR{3179}
133 - Improved Otsu threshold computation helper in canny algorithm - \PR{3169}
134 - Modified default parameters for fftR2C and fftC2R C++ API from 0 to 1.0 - \PR{3178}
135 - Use appropriate MKL getrs_batch_strided API based on MKL Versions - \PR{3181}
139 - Fixed a bug JIT kernel disk caching - \PR{3182}
140 - Fixed stream used by thrust(CUDA backend) functions - \PR{3029}
141 - Added workaround for new cuSparse API that was added by CUDA amid fix releases - \PR{3057}
142 - Fixed `const` array indexing inside `gfor` - \PR{3078}
143 - Handle zero elements in copyData to host - \PR{3059}
144 - Fixed double free regression in OpenCL backend - \PR{3091}
145 - Fixed an infinite recursion bug in NaryNode JIT Node - \PR{3072}
146 - Added missing input validation check in sparse-dense arithmetic operations - \PR{3129}
147 - Fixed bug in `getMappedPtr` in OpenCL due to invalid lambda capture - \PR{3163}
148 - Fixed bug in `getMappedPtr` on Arrays that are not ready - \PR{3163}
149 - Fixed edgeTraceKernel for CPU devices on OpenCL backend - \PR{3164}
150 - Fixed windows build issue(s) with VS2019 - \PR{3048}
151 - API documentation fixes - \PR{3075} , \PR{3076} , \PR{3143} , \PR{3161}
152 - CMake Build Fixes - \PR{3088}
153 - Fixed the tutorial link in README - \PR{3033}
154 - Fixed function name typo in timing tutorial - \PR{3028}
155 - Fixed couple of bugs in CPU backend canny implementation - \PR{3169}
156 - 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 - \PR{3167}
157 - Added assert that checks if topk is called with a negative value for k - \PR{3176}
158 - Fixed an Issue where countByKey would give incorrect results for any n > 128 - \PR{3175}
162 Special thanks to our contributors:
163 [HO-COOH][https://github.com/HO-COOH]
164 [Willy Born][https://github.com/willyborn]
165 [Gilad Avidov][https://github.com/avidov]
166 [Pavan Yalamanchili][https://github.com/pavanky]
173 - Non-uniform(ragged) reductions \PR{2786}
174 - Bit-wise not operator support for array and C API (af\_bitnot) \PR{2865}
175 - Initialization list constructor for array class \PR{2829} \PR{2987}
179 - New API for following statistics function: cov, var and stdev - \PR{2986}
180 - allocV2 and freeV2 which return cl\_mem on OpenCL backend \PR{2911}
181 - Move constructor and move assignment operator for Dim4 class \PR{2946}
182 - Support for CUDA 11.1 and Compute 8.6 \PR{3023}
183 - Fix af::feature copy constructor for multi-threaded sceanarios \PR{3022}
190 - Add f16 support for histogram - \PR{2984}
191 - Update confidence connected components example with better illustration - \PR{2968}
192 - Enable disk caching of OpenCL kernel binaries - \PR{2970}
193 - Refactor extension of kernel binaries stored to disk `.bin` - \PR{2970}
194 - Add minimum driver versions for CUDA toolkit 11 in internal map - \PR{2982}
195 - Improve warnings messages from run-time kernel compilation functions - \PR{2996}
199 - Fix bias factor of variance in var_all and cov functions - \PR{2986}
200 - Fix a race condition in confidence connected components function for OpenCL backend - \PR{2969}
201 - Safely ignore disk cache failures in CUDA backend for compiled kernel binaries - \PR{2970}
202 - Fix randn by passing in correct values to Box-Muller - \PR{2980}
203 - Fix rounding issues in Box-Muller function used for RNG - \PR{2980}
204 - Fix problems in RNG for older compute architectures with fp16 - \PR{2980} \PR{2996}
205 - Fix performance regression of approx functions - \PR{2977}
206 - Remove assert that check that signal/filter types have to be the same - \PR{2993}
207 - Fix `checkAndSetDevMaxCompute` when the device cc is greater than max - \PR{2996}
208 - Fix documentation errors and warnings - \PR{2973} , \PR{2987}
209 - Add missing opencl-arrayfire interoperability functions in unified backend - \PR{2981}
213 Special thanks to our contributors:
214 [P. J. Reed](https://github.com/pjreed)
221 - Cache CUDA kernels to disk to improve load times(Thanks to \@cschreib-ibex) \PR{2848}
222 - Staticly link against cuda libraries \PR{2785}
223 - Make cuDNN an optional build dependency \PR{2836}
224 - Improve support for different compilers and OS \PR{2876} \PR{2945} \PR{2925} \PR{2942} \PR{2943} \PR{2945} \PR{2958}
225 - Improve performance of join and transpose on CPU \PR{2849}
226 - Improve documentation \PR{2816} \PR{2821} \PR{2846} \PR{2918} \PR{2928} \PR{2947}
227 - Reduce binary size using NVRTC and template reducing instantiations \PR{2849} \PR{2861} \PR{2890} \PR{2957}
228 - reduceByKey performance improvements \PR{2851} \PR{2957}
229 - Improve support for Intel OpenCL GPUs \PR{2855}
230 - Allow staticly linking against MKL \PR{2877} (Sponsered by SDL)
231 - Better support for older CUDA toolkits \PR{2923}
232 - Add support for CUDA 11 \PR{2939}
233 - Add support for ccache for faster builds \PR{2931}
234 - Add support for the conan package manager on linux \PR{2875}
235 - Propagate build errors up the stack in AFError exceptions \PR{2948} \PR{2957}
236 - Improve runtime dependency library loading \PR{2954}
237 - Improved cuDNN runtime checks and warnings \PR{2960}
238 - Document af\_memory\_manager\_* native memory return values \PR{2911}
242 - Bug crash when allocating large arrays \PR{2827}
243 - Fix various compiler warnings \PR{2827} \PR{2849} \PR{2872} \PR{2876}
244 - Fix minor leaks in OpenCL functions \PR{2913}
245 - Various continuous integration related fixes \PR{2819}
246 - Fix zero padding with convolv2NN \PR{2820}
247 - Fix af_get_memory_pressure_threshold return value \PR{2831}
248 - Increased the max filter length for morph
249 - Handle empty array inputs for LU, QR, and Rank functions \PR{2838}
250 - Fix FindMKL.cmake script for sequential threading library \PR{2840} \PR{2952}
251 - Various internal refactoring \PR{2839} \PR{2861} \PR{2864} \PR{2873} \PR{2890} \PR{2891} \PR{2913} \PR{2959}
252 - Fix OpenCL 2.0 builtin function name conflict \PR{2851}
253 - Fix error caused when releasing memory with multiple devices \PR{2867}
254 - Fix missing set stacktrace symbol from unified API \PR{2915}
255 - Fix zero padding issue in convolve2NN \PR{2820}
256 - Fixed bugs in ReduceByKey \PR{2957}
260 Special thanks to our contributors:
261 [Corentin Schreiber](https://github.com/cschreib-ibex)
262 [Jacob Kahn](https://github.com/jacobkahn)
263 [Paul Jurczak](https://github.com/pauljurczak)
264 [Christoph Junghans](https://github.com/junghans)
272 - Improve mtx download for test data \PR{2742}
273 - Documentation improvements \PR{2754} \PR{2792} \PR{2797}
274 - Remove verbose messages in older CMake versions \PR{2773}
275 - Reduce binary size with the use of nvrtc \PR{2790}
276 - Use texture memory to load LUT in orb and fast \PR{2791}
277 - Add missing print function for f16 \PR{2784}
278 - Add checks for f16 support in the CUDA backend \PR{2784}
279 - Create a thrust policy to intercept tmp buffer allocations \PR{2806}
284 - Fix segfault on exit when ArrayFire is not initialized in the main thread
285 - Fix support for CMake 3.5.1 \PR{2771} \PR{2772} \PR{2760}
286 - Fix evalMultiple if the input array sizes aren't the same \PR{2766}
287 - Fix error when AF_BACKEND_DEFAULT is passed directly to backend \PR{2769}
288 - Workaround name collision with AMD OpenCL implementation \PR{2802}
289 - Fix on-exit errors with the unified backend \PR{2769}
290 - Fix check for f16 compatibility in OpenCL \PR{2773}
291 - Fix matmul on Intel OpenCL when passing same array as input \PR{2774}
292 - Fix CPU OpenCL blas batching \PR{2774}
293 - Fix memory pressure in the default memory manager \PR{2801}
297 Special thanks to our contributors:
298 [padentomasello](https://github.com/padentomasello)
299 [glavaux2](https://github.com/glavaux2)
307 - Added the ability to customize the memory manager(Thanks jacobkahn and flashlight) \PR{2461}
308 - Added 16-bit floating point support for several functions \PR{2413} \PR{2587} \PR{2585} \PR{2587} \PR{2583}
309 - Added sumByKey, productByKey, minByKey, maxByKey, allTrueByKey, anyTrueByKey, countByKey \PR{2254}
310 - Added confidence connected components \PR{2748}
311 - Added neural network based convolution and gradient functions \PR{2359}
312 - Added a padding function \PR{2682}
313 - Added pinverse for pseudo inverse \PR{2279}
314 - Added support for uniform ranges in approx1 and approx2 functions. \PR{2297}
315 - Added support to write to preallocated arrays for some functions \PR{2599} \PR{2481} \PR{2328} \PR{2327}
316 - Added meanvar function \PR{2258}
317 - Add support for sparse-sparse arithmetic support
318 - Added rsqrt function for reciprocal square root
319 - Added a lower level af_gemm function for general matrix multiplication \PR{2481}
320 - Added a function to set the cuBLAS math mode for the CUDA backend \PR{2584}
321 - Separate debug symbols into separate files \PR{2535}
322 - Print stacktraces on errors \PR{2632}
323 - Support move constructor for af::array \PR{2595}
324 - Expose events in the public API \PR{2461}
325 - Add setAxesLabelFormat to format labels on graphs \PR{2495}
330 - Better error messages for systems with driver or device incompatibilities \PR{2678} \PR{2448}
331 - Optimized unified backend function calls
332 - Optimized anisotropic smoothing \PR{2713}
333 - Optimized canny filter for CUDA and OpenCL
334 - Better MKL search script
335 - Better logging of different submodules in ArrayFire \PR{2670} \PR{2669}
336 - Improve documentation \PR{2665} \PR{2620} \PR{2615} \PR{2639} \PR{2628} \PR{2633} \PR{2622} \PR{2617} \PR{2558} \PR{2326} \PR{2515}
337 - Optimized af::array assignment \PR{2575}
338 - Update the k-means example to display the result \PR{2521}
344 - Fix multi-config generators
345 - Fix access errors in canny
346 - Fix segfault in the unified backend if no backends are available
347 - Fix access errors in scan-by-key
349 - Fix an issue with the random number generator and s16
350 - Fix issue with boolean product reduction
351 - Fix array_proxy move constructor
352 - Fix convolve3 launch configuration
353 - Fix an issue where the fft function modified the input array \PR{2520}
357 Special thanks to our contributors:
358 [Jacob Khan](https://github.com/jacobkahn)
359 [William Tambellini](https://github.com/WilliamTambellini)
360 [Alexey Kuleshevich](https://github.com/lehins)
361 [Richard Barnes](https://github.com/r-barnes)
362 [Gaika](https://github.com/gaika)
363 [ShalokShalom](https://github.com/ShalokShalom)
371 - Address a JIT performance regression due to moving kernel arguments to shared memory \PR{2501}
372 - Fix the default parameter for setAxisTitle \PR{2491}
379 - Graphics are now a runtime dependency instead of a link time dependency \PR{2365}
380 - Reduce the CUDA backend binary size using runtime compilation of kernels \PR{2437}
381 - Improved batched matrix multiplication on the CPU backend by using Intel MKL's
382 `cblas_Xgemm_batched`\PR{2206}
383 - Print JIT kernels to disk or stream using the `AF_JIT_KERNEL_TRACE`
384 environment variable \PR{2404}
385 - `void*` pointers are now allowed as arguments to `af::array::write()` \PR{2367}
386 - Slightly improve the efficiency of JITed tile operations \PR{2472}
387 - Make the random number generation on the CPU backend to be consistent with
388 CUDA and OpenCL \PR{2435}
389 - Handled very large JIT tree generations \PR{2484} \PR{2487}
393 - Fixed `af::array::array_proxy` move assignment operator \PR{2479}
394 - Fixed input array dimensions validation in svdInplace() \PR{2331}
395 - Fixed the typedef declaration for window resource handle \PR{2357}.
396 - Increase compatibility with GCC 8 \PR{2379}
397 - Fixed `af::write` tests \PR{2380}
398 - Fixed a bug in broadcast step of 1D exclusive scan \PR{2366}
399 - Fixed OpenGL related build errors on OSX \PR{2382}
400 - Fixed multiple array evaluation. Performance improvement. \PR{2384}
401 - Fixed buffer overflow and expected output of kNN SSD small test \PR{2445}
402 - Fixed MKL linking order to enable threaded BLAS \PR{2444}
403 - Added validations for forge module plugin availability before calling
404 resource cleanup \PR{2443}
405 - Improve compatibility on MSVC toolchain(_MSC_VER > 1914) with the CUDA
407 - Fixed BLAS gemm func generators for newest MSVC 19 on VS 2017 \PR{2464}
408 - Fix errors on exits when using the cuda backend with unified \PR{2470}
412 - Updated svdInplace() documentation following a bugfix \PR{2331}
413 - Fixed a typo in matrix multiplication documentation \PR{2358}
414 - Fixed a code snippet demostrating C-API use \PR{2406}
415 - Updated hamming matcher implementation limitation \PR{2434}
416 - Added illustration for the rotate function \PR{2453}
420 - Use cudaMemcpyAsync instead of cudaMemcpy throughout the codebase \PR{2362}
421 - Display a more informative error message if CUDA driver is incomptible
423 - Changed forge resource managemenet to use smart pointers \PR{2452}
424 - Deprecated intl and uintl typedefs in API \PR{2360}
425 - Enabled graphics by default for all builds starting with v3.6.3 \PR{2365}
426 - Fixed several warnings \PR{2344} \PR{2356} \PR{2361}
427 - Refactored initArray() calls to use createEmptyArray(). initArray() is for
428 internal use only by Array class. \PR{2361}
429 - Refactored `void*` memory allocations to use unsigned char type \PR{2459}
430 - Replaced deprecated MKL API with in-house implementations for sparse
431 to sparse/dense conversions \PR{2312}
432 - Reorganized and fixed some internal backend API \PR{2356}
433 - Updated compilation order of cuda files to speed up compile time \PR{2368}
434 - Removed conditional graphics support builds after enabling runtime
435 loading of graphics dependencies \PR{2365}
436 - Marked graphics dependencies as optional in CPack RPM config \PR{2365}
437 - Refactored a sparse arithmetic backend API \PR{2379}
438 - Fixed const correctness of `af_device_array` API \PR{2396}
439 - Update Forge to v1.0.4 \PR{2466}
440 - Manage Forge resources from the DeviceManager class \PR{2381}
441 - Fixed non-mkl & non-batch blas upstream call arguments \PR{2401}
442 - Link MKL with OpenMP instead of TBB by default
443 - use clang-format to format source code
447 Special thanks to our contributors:
448 [Alessandro Bessi](https://github.com/alessandrobessi)
449 [zhihaoy](https://github.com/zhihaoy)
450 [Jacob Khan](https://github.com/jacobkahn)
451 [William Tambellini](https://github.com/WilliamTambellini)
458 - Added support for batching on the `cond` argument in select() \PR{2243}
459 - Added support for broadcasting batched matmul() \PR{2315}
460 - Added support for multiple nearest neighbors in nearestNeighbour() \PR{2280}
461 - Added support for clamp-to-edge padding as an `af_border_type` option \PR{2333}
465 - Improved performance of morphological operations \PR{2238}
466 - Fixed linking errors when compiling without Freeimage/Graphics \PR{2248}
467 - Improved the usage of ArrayFire as a CMake subproject \PR{2290}
468 - Enabled configuration of custom library path for loading dynamic backend
473 - Fixed LAPACK definitions and linking errors \PR{2239}
474 - Fixed overflow in dim4::ndims() \PR{2289}
475 - Fixed pow() precision for integral types \PR{2305}
476 - Fixed issues with tile() with a large repeat dimension \PR{2307}
477 - Fixed svd() sub-array output on OpenCL \PR{2279}
478 - Fixed grid-based indexing calculation in histogram() \PR{2230}
479 - Fixed bug in indexing when used after reorder \PR{2311}
480 - Fixed errors when exiting on Windows when using
481 [CLBlast](https://github.com/CNugteren/CLBlast) \PR{2222}
482 - Fixed fallthrough error in medfilt1 \PR{2349}
486 - Improved unwrap() documentation \PR{2301}
487 - Improved wrap() documentation \PR{2320}
488 - Improved accum() documentation \PR{2298}
489 - Improved tile() documentation \PR{2293}
490 - Clarified approx1() and approx2() indexing in documentation \PR{2287}
491 - Updated examples of [select()](@ref data_func_select) in detailed documentation
493 - Updated lookup() examples \PR{2288}
494 - Updated set operations' documentation \PR{2299}
498 - `af*` libraries and dependencies directory changed to `lib64` \PR{2186}
499 - Added new arrayfire ASSERT utility functions \PR{2249} \PR{2256} \PR{2257} \PR{2263}
500 - Improved error messages in JIT \PR{2309}
504 Special thanks to our contributors: [Jacob Kahn](https://github.com/jacobkahn),
505 [Vardan Akopian](https://github.com/vakopian)
512 - FreeImage is now a run-time dependency [#2164]
513 - Reduced binary size by setting the symbol visibility to hidden [#2168]
514 - Add memory manager logging using the AF_TRACE=mem environment variable [#2169]
515 - Improved CPU Anisotropic Diffusion performance [#2174]
516 - Perform normalization after FFT for improved accuracy [#2185][#2192]
517 - Updated CLBlast to v1.4.0 [#2178]
518 - Added additional validation when using af::seq for indexing [#2153]
519 - Perform checks for unsupported cards by the CUDA implementation [#2182]
523 - Fixed region when all pixels were the foreground or background [#2152]
524 - Fixed several memory leaks [#2202][#2201][#2180][#2179][#2177][#2175]
525 - Fixed bug in setDevice which didn't allow you to select the last device [#2189]
526 - Fixed bug in min/max where the first element of the array was a NaN value [#2155]
527 - Fixed window cell indexing for graphics [#2207]
532 The source code with submodules can be downloaded directly from the following link:
533 http://arrayfire.com/arrayfire_source/arrayfire-full-3.6.0.tar.bz2
538 - Added the `topk()` function
539 [Documentation](http://arrayfire.org/docs/group__stat__func__topk.htm).
540 <sup>[1](https://github.com/arrayfire/arrayfire/pull/2061)</sup>
541 - Added batched matrix multiply support.
542 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1898)</sup>
543 <sup>[3](https://github.com/arrayfire/arrayfire/pull/2059)</sup>
544 - Added anisotropic diffusion, `anisotropicDiffusion()`.
545 [Documentation](http://arrayfire.org/docs/group__image__func__anisotropic__diffusion.htm)
546 <sup>[4](https://github.com/arrayfire/arrayfire/pull/1850)</sup>.
551 - Added support for batched matrix multiply.
552 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1898)</sup>
553 <sup>[2](https://github.com/arrayfire/arrayfire/pull/2059)</sup>
554 - New anisotropic diffusion function, `anisotropicDiffusion()`.
555 [Documentation](http://arrayfire.org/docs/group__image__func__anisotropic__diffusion.htm)
556 <sup>[3](https://github.com/arrayfire/arrayfire/pull/1850)</sup>.
557 - New `topk()` function, which returns the top k elements along a given
558 dimension of the input.
559 [Documentation](http://arrayfire.org/docs/group__stat__func__topk.htm).
560 <sup>[4](https://github.com/arrayfire/arrayfire/pull/2061)</sup>
561 - New gradient diffusion
562 [example](https://github.com/arrayfire/arrayfire/blob/master/examples/image_processing/gradient_diffusion.cpp).
567 - JITted `select()` and `shift()` functions for CUDA and OpenCL backends.
568 <sup>[1](https://github.com/arrayfire/arrayfire/pull/2047)</sup>
569 - Significant CMake improvements.
570 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1861)</sup>
571 <sup>[3](https://github.com/arrayfire/arrayfire/pull/2070)</sup>
572 <sup>[4](https://github.com/arrayfire/arrayfire/pull/2018)</sup>
573 - Improved the quality of the random number generator, thanks to Ralf Stubner.
574 <sup>[5](https://github.com/arrayfire/arrayfire/pull/2122)</sup>
575 - Modified `af_colormap` struct to match forge's definition.
576 <sup>[6](https://github.com/arrayfire/arrayfire/pull/2082)</sup>
577 - Improved Black Scholes example.
578 <sup>[7](https://github.com/arrayfire/arrayfire/pull/2079)</sup>
579 - Using CPack to generate installers.
580 <sup>[8](https://github.com/arrayfire/arrayfire/pull/1861)</sup>
582 [black_scholes_options](https://github.com/arrayfire/arrayfire/blob/master/examples/financial/black_scholes_options.cpp)
583 example to use built-in `af::erfc` function for cumulative normal
584 distribution.<sup>[9](https://github.com/arrayfire/arrayfire/pull/2079)</sup>.
585 - Reduced the scope of mutexes in memory manager
586 <sup>[10](https://github.com/arrayfire/arrayfire/pull/2125)</sup>
587 - Official installers do not require the CUDA toolkit to be installed
588 - Significant CMake improvements have been made. Using CPack to generate
589 installers. <sup>[11](https://github.com/arrayfire/arrayfire/pull/1861)</sup>
590 <sup>[12](https://github.com/arrayfire/arrayfire/pull/2070)</sup>
591 <sup>[13](https://github.com/arrayfire/arrayfire/pull/2018)</sup>
592 - Corrected assert function calls in select() tests.
593 <sup>[14](https://github.com/arrayfire/arrayfire/pull/2058)</sup>
598 - Fixed `shfl_down()` warnings with CUDA 9.
599 <sup>[1](https://github.com/arrayfire/arrayfire/pull/2040)</sup>
600 - Disabled CUDA JIT debug flags on ARM
601 architecture.<sup>[2](https://github.com/arrayfire/arrayfire/pull/2037)</sup>
602 - Fixed CLBLast install lib dir for linux platform where `lib` directory has
603 arch(64) suffix.<sup>[3](https://github.com/arrayfire/arrayfire/pull/2094)</sup>
604 - Fixed assert condition in 3d morph opencl
605 kernel.<sup>[4](https://github.com/arrayfire/arrayfire/pull/2033)</sup>
606 - Fix JIT errors with large non-linear
607 kernels<sup>[5](https://github.com/arrayfire/arrayfire/pull/2127)</sup>
608 - Fix bug in CPU jit after moddims was called
609 <sup>[5](https://github.com/arrayfire/arrayfire/pull/2127)</sup>
610 - Fixed deadlock caused by calls to from the worker thread
611 <sup>[6](https://github.com/arrayfire/arrayfire/pull/2124)</sup>
616 - Fixed variable name typo in `vectorization.md`.
617 <sup>[1](https://github.com/arrayfire/arrayfire/pull/2032)</sup>
618 - Fixed `AF_API_VERSION` value in Doxygen config file.
619 <sup>[2](https://github.com/arrayfire/arrayfire/pull/2053)</sup>
624 - Several OpenCL tests failing on OSX:
625 - `canny_opencl, fft_opencl, gen_assign_opencl, homography_opencl,
626 reduce_opencl, scan_by_key_opencl, solve_dense_opencl,
627 sparse_arith_opencl, sparse_convert_opencl, where_opencl`
629 Community contributions
630 -----------------------
632 Special thanks to our contributors:
633 [Adrien F. Vincent](https://github.com/afvincent), [Cedric
634 Nugteren](https://github.com/CNugteren),
635 [Felix](https://github.com/fzimmermann89), [Filip
636 Matzner](https://github.com/FloopCZ),
637 [HoneyPatouceul](https://github.com/HoneyPatouceul), [Patrick
638 Lavin](https://github.com/plavin), [Ralf Stubner](https://github.com/rstub),
639 [William Tambellini](https://github.com/WilliamTambellini)
645 The source code with submodules can be downloaded directly from the following
646 link: http://arrayfire.com/arrayfire_source/arrayfire-full-3.5.1.tar.bz2
648 Installer CUDA Version: 8.0 (Required) Installer OpenCL Version: 1.2 (Minimum)
652 - Relaxed `af::unwrap()` function's arguments.
653 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1853)</sup>
654 - Changed behavior of af::array::allocated() to specify memory allocated.
655 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1877)</sup>
656 - Removed restriction on the number of bins for `af::histogram()` on CUDA and
657 OpenCL kernels. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1895)</sup>
663 - Improved JIT performance.
664 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1864)</sup>
665 - Improved CPU element-wise operation performance.
666 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1890)</sup>
667 - Improved regions performance using texture objects. <sup>
668 [1](https://github.com/arrayfire/arrayfire/pull/1903)</sup>
673 - Fixed overflow issues in mean.
674 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1849)</sup>
675 - Fixed memory leak when chaining indexing operations.
676 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1879)</sup>
677 - Fixed bug in array assignment when using an empty array to index.
678 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1897)</sup>
679 - Fixed bug with `af::matmul()` which occured when its RHS argument was an
681 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1883)</sup>
682 - Fixed bug deadlock bug when sparse array was used with a JIT Array.
683 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1889)</sup>
684 - Fixed pixel tests for FAST kernels.
685 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1891)</sup>
686 - Fixed `af::replace` so that it is now copy-on-write.
687 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1892)</sup>
688 - Fixed launch configuration issues in CUDA JIT.
689 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1893)</sup>
690 - Fixed segfaults and "Pure Virtual Call" error warnings when exiting on
691 Windows. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1899)
692 [2](https://github.com/arrayfire/arrayfire/pull/1924)</sup>
693 - Workaround for `clEnqueueReadBuffer` bug on OSX.
694 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1888)</sup>
699 - Fixed issues when compiling with GCC 7.1.
700 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1872)</sup>
701 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1876)</sup>
702 - Eliminated unnecessary Boost dependency from CPU and CUDA backends.
703 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1857)</sup>
708 - Updated support links to point to Slack instead of Gitter.
709 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1905)</sup>
719 * ArrayFire now supports threaded applications.
720 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1706)</sup>
721 * Added Canny edge detector.
722 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1743)</sup>
723 * Added Sparse-Dense arithmetic operations.
724 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1696)</sup>
729 * ArrayFire Threading
730 * \ref af::array can be read by multiple threads
731 * All ArrayFire functions can be executed concurrently by multiple threads
732 * Threads can operate on different devices to simplify Muli-device workloads
733 * New Canny edge detector function, \ref af::canny().
734 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1743)</sup>
735 * Can automatically calculate high threshold with `AF_CANNY_THRESHOLD_AUTO_OTSU`
736 * Supports both L1 and L2 Norms to calculate gradients
737 * New tuned OpenCL BLAS backend,
738 [CLBlast](https://github.com/arrayfire/arrayfire/pull/1727).
743 * Converted CUDA JIT to use
744 [NVRTC](http://docs.nvidia.com/cuda/nvrtc/index.html) instead of
745 [NVVM](http://docs.nvidia.com/cuda/nvvm-ir-spec/index.html).
746 * Performance improvements in \ref af::reorder().
747 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1766)</sup>
748 * Performance improvements in \ref af::array::scalar<T>().
749 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1809)</sup>
750 * Improved unified backend performance.
751 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1770)</sup>
752 * ArrayFire now depends on Forge
753 v1.0. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1800)</sup>
754 * Can now specify the FFT plan cache size using the
755 \ref af::setFFTPlanCacheSize() function.
756 * Get the number of physical bytes allocated by the memory manager
757 \ref af_get_allocated_bytes(). <sup>[1](https://github.com/arrayfire/arrayfire/pull/1630)</sup>
758 * \ref af::dot() can now return a scalar value to the
759 host. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1628)</sup>
764 * Fixed improper release of default Mersenne random
765 engine. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1716)</sup>
766 * Fixed \ref af::randu() and \ref af::randn() ranges for floating point
767 types. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1784)</sup>
768 * Fixed assignment bug in CPU
769 backend. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1765)</sup>
770 * Fixed complex (`c32`,`c64`) multiplication in OpenCL convolution
771 kernels. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1816)</sup>
772 * Fixed inconsistent behavior with \ref af::replace() and \ref
773 af_replace_scalar(). <sup>[1](https://github.com/arrayfire/arrayfire/pull/1773)</sup>
774 * Fixed memory leak in \ref
775 af_fir(). <sup>[1](https://github.com/arrayfire/arrayfire/pull/1765)</sup>
776 * Fixed memory leaks in \ref af_cast for sparse arrays.
777 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1826)</sup>
778 * Fixing correctness of \ref af_pow for complex numbers by using Cartesian
779 form. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1765)</sup>
780 * Corrected \ref af::select() with indexing in CUDA and OpenCL
781 backends. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1731)</sup>
782 * Workaround for VS2015 compiler ternary
783 bug. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1771)</sup>
784 * Fixed memory corruption in
785 `cuda::findPlan()`. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1793)</sup>
786 * Argument checks in \ref af_create_sparse_array avoids inputs of type
787 int64. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1747)</sup>
788 * Fixed issue with indexing an array with a step size != 1. <sup>[1](https://github.com/arrayfire/arrayfire/issues/1846)</sup>
793 * On OSX, utilize new GLFW package from the brew package
794 manager. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1720)</sup>
795 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1775)</sup>
796 * Fixed CUDA PTX names generated by CMake
797 v3.7. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1689)</sup>
798 * Support `gcc` > 5.x for
799 CUDA. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1708)</sup>
804 * New genetic algorithm example.
805 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1695)</sup>
810 * Updated `README.md` to improve readability and
811 formatting. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1726)</sup>
812 * Updated `README.md` to mention Julia and Nim
813 wrappers. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1714)</sup>
814 * Improved installation instructions -
815 `docs/pages/install.md`. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1740)</sup>
820 * A few improvements for ROCm
821 support. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1710)</sup>
822 * Removed CUDA 6.5 support.
823 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1687)</sup>
829 * The Windows NVIDIA driver version `37x.xx` contains a bug which causes
830 `fftconvolve_opencl` to fail. Upgrade or downgrade to a different version of
831 the driver to avoid this failure.
832 * The following tests fail on Windows with NVIDIA hardware:
833 `threading_cuda`,`qr_dense_opencl`, `solve_dense_opencl`.
835 * The Accelerate framework, used by the CPU backend on macOS, leverages Intel
836 graphics cards (Iris) when there are no discrete GPUs available. This OpenCL
837 implementation is known to give incorrect results on the following tests:
838 `lu_dense_{cpu,opencl}`, `solve_dense_{cpu,opencl}`,
839 `inverse_dense_{cpu,opencl}`.
840 * Certain tests intermittently fail on macOS with NVIDIA GPUs apparently due
841 to inconsistent driver behavior: `fft_large_cuda` and `svd_dense_cuda`.
842 * The following tests are currently failing on macOS with AMD GPUs:
843 `cholesky_dense_opencl` and `scan_by_key_opencl`.
849 Deprecation Announcement
850 ------------------------
852 This release supports CUDA 6.5 and higher. The next ArrayFire relase will
853 support CUDA 7.0 and higher, dropping support for CUDA 6.5. Reasons for no
854 longer supporting CUDA 6.5 include:
856 * CUDA 7.0 NVCC supports the C++11 standard (whereas CUDA 6.5 does not), which
857 is used by ArrayFire's CPU and OpenCL backends.
858 * Very few ArrayFire users still use CUDA 6.5.
860 As a result, the older Jetson TK1 / Tegra K1 will no longer be supported in
861 the next ArrayFire release. The newer Jetson TX1 / Tegra X1 will continue to
862 have full capability with ArrayFire.
866 * [ArrayFire has been Dockerized](https://github.com/arrayfire/arrayfire-docker).
870 * Implemented sparse storage format conversions between \ref AF_STORAGE_CSR
871 and \ref AF_STORAGE_COO.
872 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1642)</sup>
873 * Directly convert between \ref AF_STORAGE_COO <--> \ref AF_STORAGE_CSR
874 using the af::sparseConvertTo() function.
875 * af::sparseConvertTo() now also supports converting to dense.
876 * Added cast support for [sparse arrays](\ref sparse_func).
877 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1653)</sup>
878 * Casting only changes the values array and the type. The row and column
879 index arrays are not changed.
880 * Reintroduced automated computation of chart axes limits for graphics functions.
881 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1639)</sup>
882 * The axes limits will always be the minimum/maximum of the current and new
884 * The user can still set limits from API calls. If the user sets a limit
885 from the API call, then the automatic limit setting will be disabled.
886 * Using `boost::scoped_array` instead of `boost::scoped_ptr` when managing
888 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1637)</sup>
889 * Internal performance improvements to getInfo() by using `const` references
890 to avoid unnecessary copying of `ArrayInfo` objects.
891 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1665)</sup>
892 * Added support for scalar af::array inputs for af::convolve() and
893 [set functions](\ref set_mat).
894 <sup>[1](https://github.com/arrayfire/arrayfire/issues/1660)</sup>
895 <sup>[2](https://github.com/arrayfire/arrayfire/issues/1675)</sup>
896 <sup>[3](https://github.com/arrayfire/arrayfire/pull/1668)</sup>
897 * Performance fixes in af::fftConvolve() kernels.
898 <sup>[1](https://github.com/arrayfire/arrayfire/issues/1679)</sup>
899 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1680)</sup>
903 * Support for Visual Studio 2015 compilation.
904 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1632)</sup>
905 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1640)</sup>
906 * Fixed `FindCBLAS.cmake` when PkgConfig is used.
907 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1657)</sup>
911 * Fixes to JIT when tree is large.
912 <sup>[1](https://github.com/arrayfire/arrayfire/issues/1646)</sup>
913 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1638)</sup>
914 * Fixed indexing bug when converting dense to sparse af::array as \ref
916 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1642)</sup>
917 * Fixed af::bilateral() OpenCL kernel compilation on OS X.
918 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1638)</sup>
919 * Fixed memory leak in af::regions() (CPU) and af::rgb2ycbcr().
920 <sup>[1](https://github.com/arrayfire/arrayfire/issues/1664)</sup>
921 <sup>[2](https://github.com/arrayfire/arrayfire/issues/1664)</sup>
922 <sup>[3](https://github.com/arrayfire/arrayfire/pull/1666)</sup>
926 * Major OS X installer fixes.
927 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1629)</sup>
928 * Fixed installation scripts.
929 * Fixed installation symlinks for libraries.
930 * Windows installer now ships with more pre-built examples.
934 * Added af::choleskyInPlace() calls to `cholesky.cpp` example.
935 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1671)</sup>
939 * Added `u8` as supported data type in `getting_started.md`.
940 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1661)</sup>
942 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1652)</sup>
946 * [CUDA 8.0.55](https://developer.nvidia.com/cuda-toolkit) supports Xcode 8.
947 <sup>[1](https://github.com/arrayfire/arrayfire/issues/1664)</sup>
951 * Known failures with CUDA 6.5. These include all functions that use
952 sorting. As a result, sparse storage format conversion between \ref
953 AF_STORAGE_COO and \ref AF_STORAGE_CSR has been disabled for CUDA 6.5.
960 * Installers for Linux, OS X and Windows
961 * CUDA backend now uses [CUDA 8.0](https://developer.nvidia.com/cuda-toolkit).
962 * Uses [Intel MKL 2017](https://software.intel.com/en-us/intel-mkl).
963 * CUDA Compute 2.x (Fermi) is no longer compiled into the library.
965 * The libraries shipping in the OS X Installer are now compiled with Apple
966 Clang v7.3.1 (previously v6.1.0).
967 * The OS X version used is 10.11.6 (previously 10.10.5).
968 * Installer for Jetson TX1 / Tegra X1
969 * Requires [JetPack for L4T 2.3](https://developer.nvidia.com/embedded/jetpack)
970 (containing Linux for Tegra r24.2 for TX1).
971 * CUDA backend now uses [CUDA 8.0](https://developer.nvidia.com/cuda-toolkit) 64-bit.
972 * Using CUDA's cusolver instead of CPU fallback.
973 * Uses OpenBLAS for CPU BLAS.
974 * All ArrayFire libraries are now 64-bit.
978 * Add [sparse array](\ref sparse_func) support to \ref af::eval().
979 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1598)</sup>
980 * Add OpenCL-CPU fallback support for sparse \ref af::matmul() when running on
981 a unified memory device. Uses MKL Sparse BLAS.
982 * When using CUDA libdevice, pick the correct compute version based on device.
983 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1612)</sup>
984 * OpenCL FFT now also supports prime factors 7, 11 and 13.
985 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1383)</sup>
986 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1619)</sup>
990 * Allow CUDA libdevice to be detected from custom directory.
991 * Fix `aarch64` detection on Jetson TX1 64-bit OS.
992 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1593)</sup>
993 * Add missing definition of `af_set_fft_plan_cache_size` in unified backend.
994 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1591)</sup>
995 * Fix intial values for \ref af::min() and \ref af::max() operations.
996 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1594)</sup>
997 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1595)</sup>
998 * Fix distance calculation in \ref af::nearestNeighbour for CUDA and OpenCL backend.
999 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1596)</sup>
1000 <sup>[2](https://github.com/arrayfire/arrayfire/pull/1595)</sup>
1001 * Fix OpenCL bug where scalars where are passed incorrectly to compile options.
1002 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1595)</sup>
1003 * Fix bug in \ref af::Window::surface() with respect to dimensions and ranges.
1004 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1604)</sup>
1005 * Fix possible double free corruption in \ref af_assign_seq().
1006 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1605)</sup>
1007 * Add missing eval for key in \ref af::scanByKey in CPU backend.
1008 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1605)</sup>
1009 * Fixed creation of sparse values array using \ref AF_STORAGE_COO.
1010 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1620)</sup>
1011 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1621)</sup>
1015 * Add a [Conjugate Gradient solver example](\ref benchmarks/cg.cpp)
1016 to demonstrate sparse and dense matrix operations.
1017 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1599)</sup>
1021 * When using [CUDA 8.0](https://developer.nvidia.com/cuda-toolkit),
1022 compute 2.x are no longer in default compute list.
1023 * This follows [CUDA 8.0](https://developer.nvidia.com/cuda-toolkit)
1024 deprecating computes 2.x.
1025 * Default computes for CUDA 8.0 will be 30, 50, 60.
1026 * When using CUDA pre-8.0, the default selection remains 20, 30, 50.
1027 * CUDA backend now uses `-arch=sm_30` for PTX compilation as default.
1028 * Unless compute 2.0 is enabled.
1032 * \ref af::lu() on CPU is known to give incorrect results when built run on
1033 OS X 10.11 or 10.12 and compiled with Accelerate Framework.
1034 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1617)</sup>
1035 * Since the OS X Installer libraries uses MKL rather than Accelerate
1036 Framework, this issue does not affect those libraries.
1044 * [Sparse Matrix and BLAS](\ref sparse_func). <sup>[1](https://github.com/arrayfire/arrayfire/issues/821)
1045 [2](https://github.com/arrayfire/arrayfire/pull/1319)</sup>
1046 * Faster JIT for CUDA and OpenCL. <sup>[1](https://github.com/arrayfire/arrayfire/issues/1472)
1047 [2](https://github.com/arrayfire/arrayfire/pull/1462)</sup>
1048 * Support for [random number generator engines](\ref af::randomEngine).
1049 <sup>[1](https://github.com/arrayfire/arrayfire/issues/868)
1050 [2](https://github.com/arrayfire/arrayfire/pull/1551)</sup>
1051 * Improvements to graphics. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1555)
1052 [2](https://github.com/arrayfire/arrayfire/pull/1566)</sup>
1056 * **[Sparse Matrix and BLAS](\ref sparse_func)** <sup>[1](https://github.com/arrayfire/arrayfire/issues/821)
1057 [2](https://github.com/arrayfire/arrayfire/pull/1319)</sup>
1058 * Support for [CSR](\ref AF_STORAGE_CSR) and [COO](\ref AF_STORAGE_COO)
1059 [storage types](\ref af_storage).
1060 * Sparse-Dense Matrix Multiplication and Matrix-Vector Multiplication as a
1061 part of af::matmul() using \ref AF_STORAGE_CSR format for sparse.
1062 * Conversion to and from [dense](\ref AF_STORAGE_DENSE) matrix to [CSR](\ref AF_STORAGE_CSR)
1063 and [COO](\ref AF_STORAGE_COO) [storage types](\ref af_storage).
1064 * **Faster JIT** <sup>[1](https://github.com/arrayfire/arrayfire/issues/1472)
1065 [2](https://github.com/arrayfire/arrayfire/pull/1462)</sup>
1066 * Performance improvements for CUDA and OpenCL JIT functions.
1067 * Support for evaluating multiple outputs in a single kernel. See af::array::eval() for more.
1068 * **[Random Number Generation](\ref af::randomEngine)**
1069 <sup>[1](https://github.com/arrayfire/arrayfire/issues/868)
1070 [2](https://github.com/arrayfire/arrayfire/pull/1551)</sup>
1071 * af::randomEngine(): A random engine class to handle setting the [type](af_random_type) and seed
1072 for random number generator engines.
1073 * Supported engine types are (\ref af_random_engine_type):
1074 * [Philox](http://www.thesalmons.org/john/random123/)
1075 * [Threefry](http://www.thesalmons.org/john/random123/)
1076 * [Mersenne Twister](http://www.math.sci.hiroshima-u.ac.jp/~m-mat/MT/MTGP/)
1077 * **Graphics** <sup>[1](https://github.com/arrayfire/arrayfire/pull/1555)
1078 [2](https://github.com/arrayfire/arrayfire/pull/1566)</sup>
1079 * Using [Forge v0.9.0](https://github.com/arrayfire/forge/releases/tag/v0.9.0)
1080 * [Vector Field](\ref af::Window::vectorField) plotting functionality.
1081 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1566)</sup>
1082 * Removed [GLEW](http://glew.sourceforge.net/) and replaced with [glbinding](https://github.com/cginternals/glbinding).
1083 * Removed usage of GLEW after support for MX (multithreaded) was dropped in v2.0.
1084 <sup>[1](https://github.com/arrayfire/arrayfire/issues/1540)</sup>
1085 * Multiple overlays on the same window are now possible.
1086 * Overlays support for same type of object (2D/3D)
1087 * Supported by af::Window::plot, af::Window::hist, af::Window::surface,
1088 af::Window::vectorField.
1089 * New API to set axes limits for graphs.
1090 * Draw calls do not automatically compute the limits. This is now under user control.
1091 * af::Window::setAxesLimits can be used to set axes limits automatically or manually.
1092 * af::Window::setAxesTitles can be used to set axes titles.
1093 * New API for plot and scatter:
1094 * af::Window::plot() and af::Window::scatter() now can handle 2D and 3D and determine appropriate order.
1098 * af_draw_scatter_nd()
1099 * af_draw_scatter_2d()
1100 * af_draw_scatter_3d()
1101 * **New [interpolation methods](\ref af_interp_type)**
1102 <sup>[1](https://github.com/arrayfire/arrayfire/issues/1562)</sup>
1105 * \ref af::transform()
1106 * \ref af::approx1()
1107 * \ref af::approx2()
1108 * **Support for [complex mathematical functions](\ref mathfunc_mat)**
1109 <sup>[1](https://github.com/arrayfire/arrayfire/issues/1507)</sup>
1110 * Add complex support for \ref trig_mat, \ref af::sqrt(), \ref af::log().
1111 * **af::medfilt1(): Median filter for 1-d signals** <sup>[1](https://github.com/arrayfire/arrayfire/pull/1479)</sup>
1112 * <b>Generalized scan functions: \ref scan_func_scan and \ref scan_func_scanbykey</b>
1113 * Now supports inclusive or exclusive scans
1114 * Supports binary operations defined by \ref af_binary_op.
1115 <sup>[1](https://github.com/arrayfire/arrayfire/issues/388)</sup>
1116 * **[Image Moments](\ref moments_mat) functions**
1117 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1453)</sup>
1118 * <b>Add af::getSizeOf() function for \ref af_dtype</b>
1119 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1404)</sup>
1120 * <b>Explicitly extantiate \ref af::array::device() for `void *</b>
1121 <sup>[1](https://github.com/arrayfire/arrayfire/issues/1503)</sup>
1125 * Fixes to edge-cases in \ref morph_mat. <sup>[1](https://github.com/arrayfire/arrayfire/issues/1564)</sup>
1126 * Makes JIT tree size consistent between devices. <sup>[1](https://github.com/arrayfire/arrayfire/issues/1457)</sup>
1127 * Delegate higher-dimension in \ref convolve_mat to correct dimensions. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1445)</sup>
1128 * Indexing fixes with C++11. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1426) [2](https://github.com/arrayfire/arrayfire/pull/1426)</sup>
1129 * Handle empty arrays as inputs in various functions. <sup>[1](https://github.com/arrayfire/arrayfire/issues/799)</sup>
1130 * Fix bug when single element input to af::median. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1423)</sup>
1131 * Fix bug in calculation of time from af::timeit(). <sup>[1](https://github.com/arrayfire/arrayfire/pull/1414)</sup>
1132 * Fix bug in floating point numbers in af::seq. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1404)</sup>
1133 * Fixes for OpenCL graphics interop on NVIDIA devices.
1134 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1408/commits/e1f16e6)</sup>
1135 * Fix bug when compiling large kernels for AMD devices.
1136 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1465)</sup>
1137 * Fix bug in af::bilateral when shared memory is over the limit.
1138 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1478)</sup>
1139 * Fix bug in kernel header compilation tool `bin2cpp`.
1140 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1544)</sup>
1141 * Fix inital values for \ref morph_mat functions.
1142 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1547)</sup>
1143 * Fix bugs in af::homography() CPU and OpenCL kernels.
1144 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1584)</sup>
1145 * Fix bug in CPU TNJ.
1146 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1587)</sup>
1151 * CUDA 8 and compute 6.x(Pascal) support, current installer ships with CUDA 7.5. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1432) [2](https://github.com/arrayfire/arrayfire/pull/1487) [3](https://github.com/arrayfire/arrayfire/pull/1539)</sup>
1152 * User controlled FFT plan caching. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1448)</sup>
1153 * CUDA performance improvements for \ref image_func_wrap, \ref image_func_unwrap and \ref approx_mat.
1154 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1411)</sup>
1155 * Fallback for CUDA-OpenGL interop when no devices does not support OpenGL.
1156 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1415)</sup>
1157 * Additional forms of batching with the \ref transform_func_transform functions.
1158 [New behavior defined here](https://github.com/arrayfire/arrayfire/pull/1412).
1159 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1412)</sup>
1160 * Update to OpenCL2 headers. <sup>[1](https://github.com/arrayfire/arrayfire/issues/1344)</sup>
1161 * Support for integration with external OpenCL contexts. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1140)</sup>
1162 * Performance improvements to interal copy in CPU Backend.
1163 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1440)</sup>
1164 * Performance improvements to af::select and af::replace CUDA kernels.
1165 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1587)</sup>
1166 * Enable OpenCL-CPU offload by default for devices with Unified Host Memory.
1167 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1521)</sup>
1168 * To disable, use the environment variable `AF_OPENCL_CPU_OFFLOAD=0`.
1172 * Compilation speedups. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1526)</sup>
1173 * Build fixes with MKL. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1526)</sup>
1174 * Error message when CMake CUDA Compute Detection fails. <sup>[1](https://github.com/arrayfire/arrayfire/issues/1535)</sup>
1175 * Several CMake build issues with Xcode generator fixed.
1176 <sup>[1](https://github.com/arrayfire/arrayfire/pull/1493) [2](https://github.com/arrayfire/arrayfire/pull/1499)</sup>
1177 * Fix multiple OpenCL definitions at link time. <sup>[1](https://github.com/arrayfire/arrayfire/issues/1429)</sup>
1178 * Fix lapacke detection in CMake. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1423)</sup>
1179 * Update build tags of
1180 * [clBLAS](https://github.com/clMathLibraries/clBLAS)
1181 * [clFFT](https://github.com/clMathLibraries/clFFT)
1182 * [Boost.Compute](https://github.com/boostorg/compute)
1183 * [Forge](https://github.com/arrayfire/forge)
1184 * [glbinding](https://github.com/cginternals/glbinding)
1185 * Fix builds with GCC 6.1.1 and GCC 5.3.0. <sup>[1](https://github.com/arrayfire/arrayfire/pull/1409)</sup>
1189 * All installers now ship with ArrayFire libraries build with MKL 2016.
1190 * All installers now ship with Forge development files and examples included.
1191 * CUDA Compute 2.0 has been removed from the installers. Please contact us
1192 directly if you have a special need.
1196 * Added [example simulating gravity](\ref graphics/field.cpp) for
1197 demonstration of vector field.
1198 * Improvements to \ref financial/black_scholes_options.cpp example.
1199 * Improvements to \ref graphics/gravity_sim.cpp example.
1200 * Fix graphics examples to use af::Window::setAxesLimits and
1201 af::Window::setAxesTitles functions.
1203 Documentation & Licensing
1204 -------------------------
1205 * [ArrayFire copyright and trademark policy](http://arrayfire.com/trademark-policy)
1206 * Fixed grammar in license.
1207 * Add license information for glbinding.
1208 * Remove license infomation for GLEW.
1209 * Random123 now applies to all backends.
1210 * Random number functions are now under \ref random_mat.
1214 The following functions have been deprecated and may be modified or removed
1215 permanently from future versions of ArrayFire.
1216 * \ref af::Window::plot3(): Use \ref af::Window::plot instead.
1217 * \ref af_draw_plot(): Use \ref af_draw_plot_nd or \ref af_draw_plot_2d instead.
1218 * \ref af_draw_plot3(): Use \ref af_draw_plot_nd or \ref af_draw_plot_3d instead.
1219 * \ref af::Window::scatter3(): Use \ref af::Window::scatter instead.
1220 * \ref af_draw_scatter(): Use \ref af_draw_scatter_nd or \ref af_draw_scatter_2d instead.
1221 * \ref af_draw_scatter3(): Use \ref af_draw_scatter_nd or \ref af_draw_scatter_3d instead.
1225 Certain CUDA functions are known to be broken on Tegra K1. The following ArrayFire tests are currently failing:
1240 * Family of [Sort](\ref sort_mat) functions now support
1241 [higher order dimensions](https://github.com/arrayfire/arrayfire/pull/1373).
1242 * Improved performance of batched sort on dim 0 for all [Sort](\ref sort_mat) functions.
1243 * [Median](\ref stat_func_median) now also supports higher order dimensions.
1248 * Fixes to [error handling](https://github.com/arrayfire/arrayfire/issues/1352) in C++ API for binary functions.
1249 * Fixes to [external OpenCL context management](https://github.com/arrayfire/arrayfire/issues/1350).
1250 * Fixes to [JPEG_GREYSCALE](https://github.com/arrayfire/arrayfire/issues/1360) for FreeImage versions <= 3.154.
1251 * Fixed for [non-float inputs](https://github.com/arrayfire/arrayfire/issues/1386) to \ref af::rgb2gray().
1255 * [Disable CPU Async](https://github.com/arrayfire/arrayfire/issues/1378) when building with GCC < 4.8.4.
1256 * Add option to [disable CPUID](https://github.com/arrayfire/arrayfire/issues/1369) from CMake.
1257 * More verbose message when [CUDA Compute Detection fails](https://github.com/arrayfire/arrayfire/issues/1362).
1258 * Print message to use [CUDA library stub](https://github.com/arrayfire/arrayfire/issues/1363)
1259 from CUDA Toolkit if CUDA Library is not found from default paths.
1260 * [Build Fixes](https://github.com/arrayfire/arrayfire/pull/1385) on Windows.
1261 * For compiling tests our of source.
1262 * For compiling ArrayFire with static MKL.
1263 * [Exclude <sys/sysctl.h>](https://github.com/arrayfire/arrayfire/pull/1368) when building on GNU Hurd.
1264 * Add [manual CMake options](https://github.com/arrayfire/arrayfire/pull/1389) to build DEB and RPM packages.
1268 * Fixed documentation for \ref af::replace().
1269 * Fixed images in [Using on OSX](\ref using_on_osx) page.
1273 * Linux x64 installers will now be compiled with GCC 4.9.2.
1274 * OSX installer gives better error messages on brew failures and
1275 now includes link to [Fixing OS X Installer Failures] (https://github.com/arrayfire/arrayfire/wiki/Fixing-Common-OS-X-Installer-Failures)
1276 for brew installation failures.
1284 * Fixes to \ref af::array::device()
1285 * CPU Backend: [evaluate arrays](https://github.com/arrayfire/arrayfire/issues/1316)
1286 before returning pointer with asynchronous calls in CPU backend.
1287 * OpenCL Backend: [fix segfaults](https://github.com/arrayfire/arrayfire/issues/1324)
1288 when requested for device pointers on empty arrays.
1289 * Fixed \ref af::operator%() from using [rem to mod](https://github.com/arrayfire/arrayfire/issues/1318).
1290 * Fixed [array destruction](https://github.com/arrayfire/arrayfire/issues/1321)
1291 when backends are switched in Unified API.
1292 * Fixed [indexing](https://github.com/arrayfire/arrayfire/issues/1331) after
1293 \ref af::moddims() is called.
1294 * Fixes FFT calls for CUDA and OpenCL backends when used on
1295 [multiple devices](https://github.com/arrayfire/arrayfire/issues/1332).
1296 * Fixed [unresolved external](https://github.com/arrayfire/arrayfire/commit/32965ef)
1297 for some functions from \ref af::array::array_proxy class.
1301 * CMake compiles files in alphabetical order.
1302 * CMake fixes for BLAS and LAPACK on some Linux distributions.
1306 * Fixed [OpenCL FFT performance](https://github.com/arrayfire/arrayfire/issues/1323) regression.
1307 * \ref af::array::device() on OpenCL backend [returns](https://github.com/arrayfire/arrayfire/issues/1311)
1308 `cl_mem` instead of `(void*)cl::Buffer*`.
1309 * In Unified backend, [load versioned libraries](https://github.com/arrayfire/arrayfire/issues/1312)
1314 * Reorganized, cleaner README file.
1315 * Replaced non-free lena image in assets with free-to-distribute lena image.
1323 * CPU backend supports aysnchronous execution.
1324 * Performance improvements to OpenCL BLAS and FFT functions.
1325 * Improved performance of memory manager.
1326 * Improvements to visualization functions.
1327 * Improved sorted order for OpenCL devices.
1328 * Integration with external OpenCL projects.
1333 * \ref af::getActiveBackend(): Returns the current backend being used.
1334 * [Scatter plot](https://github.com/arrayfire/arrayfire/pull/1116) added to graphics.
1335 * \ref af::transform() now supports perspective transformation matrices.
1336 * \ref af::infoString(): Returns `af::info()` as a string.
1337 * \ref af::printMemInfo(): Print a table showing information about buffer from the memory manager
1338 * The \ref AF_MEM_INFO macro prints numbers and total sizes of all buffers (requires including af/macros.h)
1339 * \ref af::allocHost(): Allocates memory on host.
1340 * \ref af::freeHost(): Frees host side memory allocated by arrayfire.
1341 * OpenCL functions can now use CPU implementation.
1342 * Currently limited to Unified Memory devices (CPU and On-board Graphics).
1343 * Functions: af::matmul() and all [LAPACK](\ref linalg_mat) functions.
1344 * Takes advantage of optimized libraries such as MKL without doing memory copies.
1345 * Use the environment variable `AF_OPENCL_CPU_OFFLOAD=1` to take advantage of this feature.
1346 * Functions specific to OpenCL backend.
1347 * \ref afcl::addDevice(): Adds an external device and context to ArrayFire's device manager.
1348 * \ref afcl::deleteDevice(): Removes an external device and context from ArrayFire's device manager.
1349 * \ref afcl::setDevice(): Sets an external device and context from ArrayFire's device manager.
1350 * \ref afcl::getDeviceType(): Gets the device type of the current device.
1351 * \ref afcl::getPlatform(): Gets the platform of the current device.
1352 * \ref af::createStridedArray() allows [array creation user-defined strides](https://github.com/arrayfire/arrayfire/issues/1177) and device pointer.
1353 * [Expose functions](https://github.com/arrayfire/arrayfire/issues/1131) that provide information
1354 about memory layout of Arrays.
1355 * \ref af::getStrides(): Gets the strides for each dimension of the array.
1356 * \ref af::getOffset(): Gets the offsets for each dimension of the array.
1357 * \ref af::getRawPtr(): Gets raw pointer to the location of the array on device.
1358 * \ref af::isLinear(): Returns true if all elements in the array are contiguous.
1359 * \ref af::isOwner(): Returns true if the array owns the raw pointer, false if it is a sub-array.
1360 * \ref af::getStrides(): Gets the strides of the array.
1361 * \ref af::getStrides(): Gets the strides of the array.
1362 * \ref af::getDeviceId(): Gets the device id on which the array resides.
1363 * \ref af::isImageIOAvailable(): Returns true if ArrayFire was compiled with Freeimage enabled
1364 * \ref af::isLAPACKAvailable(): Returns true if ArrayFire was compiled with LAPACK functions enabled
1369 * Fixed [errors when using 3D / 4D arrays](https://github.com/arrayfire/arrayfire/pull/1251) in select and replace
1370 * Fixed [JIT errors on AMD devices](https://github.com/arrayfire/arrayfire/pull/1238) for OpenCL backend.
1371 * Fixed [imageio bugs](https://github.com/arrayfire/arrayfire/pull/1229) for 16 bit images.
1372 * Fixed [bugs when loading and storing images](https://github.com/arrayfire/arrayfire/pull/1228) natively.
1373 * Fixed [bug in FFT for NVIDIA GPUs](https://github.com/arrayfire/arrayfire/issues/615) when using OpenCL backend.
1374 * Fixed [bug when using external context](https://github.com/arrayfire/arrayfire/pull/1241) with OpenCL backend.
1375 * Fixed [memory leak](https://github.com/arrayfire/arrayfire/issues/1269) in \ref af_median_all().
1376 * Fixed [memory leaks and performance](https://github.com/arrayfire/arrayfire/pull/1274) in graphics functions.
1377 * Fixed [bugs when indexing followed by moddims](https://github.com/arrayfire/arrayfire/issues/1275).
1378 * \ref af_get_revision() now returns actual commit rather than AF_REVISION.
1379 * Fixed [releasing arrays](https://github.com/arrayfire/arrayfire/issues/1282) when using different backends.
1380 * OS X OpenCL: [LAPACK functions](\ref linalg_mat) on CPU devices use OpenCL offload (previously threw errors).
1381 * [Add support for 32-bit integer image types](https://github.com/arrayfire/arrayfire/pull/1287) in Image IO.
1382 * Fixed [set operations for row vectors](https://github.com/arrayfire/arrayfire/issues/1300)
1383 * Fixed [bugs](https://github.com/arrayfire/arrayfire/issues/1243) in \ref af::meanShift() and af::orb().
1388 * Optionally [offload BLAS and LAPACK](https://github.com/arrayfire/arrayfire/pull/1221) functions to CPU implementations to improve performance.
1389 * Performance improvements to the memory manager.
1390 * Error messages are now more detailed.
1391 * Improved sorted order for OpenCL devices.
1392 * JIT heuristics can now be tweaked using environment variables. See
1393 [Environment Variables](\ref configuring_environment) tutorial.
1394 * Add `BUILD_<BACKEND>` [options to examples and tests](https://github.com/arrayfire/arrayfire/issues/1286)
1395 to toggle backends when compiling independently.
1400 * New visualization [example simulating gravity](\ref graphics/gravity_sim.cpp).
1405 * Support for Intel `icc` compiler
1406 * Support to compile with Intel MKL as a BLAS and LAPACK provider
1407 * Tests are now available for building as standalone (like examples)
1408 * Tests can now be built as a single file for each backend
1409 * Better handling of NONFREE build options
1410 * [Searching for GLEW in CMake default paths](https://github.com/arrayfire/arrayfire/pull/1292)
1411 * Fixes for compiling with MKL on OSX.
1415 * Improvements to OSX Installer
1416 * CMake config files are now installed with libraries
1417 * Independent options for installing examples and documentation components
1422 * `af_lock_device_arr` is now deprecated to be removed in v4.0.0. Use \ref af_lock_array() instead.
1423 * `af_unlock_device_arr` is now deprecated to be removed in v4.0.0. use \ref af_unlock_array() instead.
1428 * Fixes to documentation for \ref af::matchTemplate().
1429 * Improved documentation for deviceInfo.
1430 * Fixes to documentation for \ref af::exp().
1435 * [Solve OpenCL fails on NVIDIA Maxwell devices](https://github.com/arrayfire/arrayfire/issues/1246)
1436 for f32 and c32 when M > N and K % 4 is 1 or 2.
1445 * Fixed [memory leak](https://github.com/arrayfire/arrayfire/pull/1145) in
1446 CUDA Random number generators
1447 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1157) in
1448 af::select() and af::replace() tests
1449 * Fixed [exception](https://github.com/arrayfire/arrayfire/issues/1164)
1450 thrown when printing empty arrays with af::print()
1451 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1170) in CPU
1452 random number generation. Changed the generator to
1453 [mt19937](http://en.cppreference.com/w/cpp/numeric/random)
1454 * Fixed exception handling (internal)
1455 * [Exceptions](https://github.com/arrayfire/arrayfire/issues/1188)
1456 now show function, short file name and line number
1457 * Added [AF_RETURN_ERROR](https://github.com/arrayfire/arrayfire/issues/1186)
1458 macro to handle returning errors.
1459 * Removed THROW macro, and renamed AF_THROW_MSG to AF_THROW_ERR.
1460 * Fixed [bug](https://github.com/arrayfire/arrayfire/commit/9459c6)
1461 in \ref af::identity() that may have affected CUDA Compute 5.2 cards
1466 * Added a [MIN_BUILD_TIME](https://github.com/arrayfire/arrayfire/issues/1193)
1467 option to build with minimum optimization compiler flags resulting in faster
1469 * Fixed [issue](https://github.com/arrayfire/arrayfire/issues/1143) in CBLAS
1471 * Fixed tests failing for builds without optional components
1472 [FreeImage](https://github.com/arrayfire/arrayfire/issues/1143) and
1473 [LAPACK](https://github.com/arrayfire/arrayfire/issues/1167)
1474 * Added a [test](https://github.com/arrayfire/arrayfire/issues/1192)
1476 * Only [info and backend tests](https://github.com/arrayfire/arrayfire/issues/1192)
1477 are now built for unified backend
1478 * [Sort tests](https://github.com/arrayfire/arrayfire/issues/1199)
1479 execution alphabetically
1480 * Fixed compilation flags and errors in tests and examples
1481 * [Moved AF_REVISION and AF_COMPILER_STR](https://github.com/arrayfire/arrayfire/commit/2287c5)
1482 into src/backend. This is because as revision is updated with every commit,
1483 entire ArrayFire would have to be rebuilt in the old code.
1484 * v3.3 will add a af_get_revision() function to get the revision string.
1485 * [Clean up examples](https://github.com/arrayfire/arrayfire/pull/1158)
1486 * Remove getchar for Windows (this will be handled by the installer)
1487 * Other miscellaneous code cleanup
1488 * Fixed bug in [plot3.cpp](\ref graphics/plot3.cpp) example
1489 * [Rename](https://github.com/arrayfire/arrayfire/commit/35f0fc2) clBLAS/clFFT
1490 external project suffix from external -> ext
1491 * [Add OpenBLAS](https://github.com/arrayfire/arrayfire/pull/1197) as a
1492 lapack/lapacke alternative
1496 * Added \ref AF_MEM_INFO macro to print memory info from ArrayFire's memory
1497 manager ([cross issue](https://github.com/arrayfire/arrayfire/issues/1172))
1498 * Added [additional paths](https://github.com/arrayfire/arrayfire/issues/1184)
1499 for searching for `libaf*` for Unified backend on unix-style OS.
1500 * Note: This still requires dependencies such as forge, CUDA, NVVM etc to be
1501 in `LD_LIBRARY_PATH` as described in [Unified Backend](\ref unifiedbackend)
1502 * [Create streams](https://github.com/arrayfire/arrayfire/commit/ed0373f)
1503 for devices only when required in CUDA Backend
1507 * [Hide scrollbars](https://github.com/arrayfire/arrayfire/commit/9d218a5)
1508 appearing for pre and code styles
1509 * Fix [documentation](https://github.com/arrayfire/arrayfire/commit/ac09f91) for af::replace
1510 * Add [code sample](https://github.com/arrayfire/arrayfire/commit/4e06483)
1511 for converting the output of af::getAvailableBackends() into bools
1512 * Minor fixes in documentation
1520 * Fixed [bug](https://github.com/arrayfire/arrayfire/pull/1136) in homography()
1521 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1135) in behavior
1522 of af::array::device()
1523 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1129) when
1524 indexing with span along trailing dimension
1525 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1127) when
1526 indexing in [GFor](\ref gfor)
1527 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1122) in CPU
1528 information fetching
1529 * Fixed compilation [bug](https://github.com/arrayfire/arrayfire/issues/1117)
1530 in unified backend caused by missing link library
1531 * Add [missing symbol](https://github.com/arrayfire/arrayfire/pull/1114) for
1536 * Tests can now be used as a [standalone project](https://github.com/arrayfire/arrayfire/pull/1120)
1537 * Tests can now be built using pre-compiled libraries
1538 * Similar to how the examples are built
1539 * The install target now installs the examples source irrespective of the
1540 BUILD_EXAMPLES value
1541 * Examples are not built if BUILD_EXAMPLES is off
1545 * HTML documentation is now [built and installed](https://github.com/arrayfire/arrayfire/pull/1109)
1547 * Added documentation for \ref af::seq class
1548 * Updated [Matrix Manipulation](\ref matrixmanipulation) tutorial
1549 * Examples list is now generated by CMake
1550 * <a href="examples.htm">Examples</a> are now listed as dir/example.cpp
1551 * Removed dummy groups used for indexing documentation (affcted doxygen < 1.8.9)
1559 * Added Unified backend
1560 * Allows switching backends at runtime
1561 * Read [Unified Backend](\ref unifiedbackend) for more.
1562 * Support for 16-bit integers (\ref s16 and \ref u16)
1563 * All functions that support 32-bit interger types (\ref s32, \ref u32),
1564 now also support 16-bit interger types
1569 * \ref af::setBackend() - Sets a backend as active
1570 * \ref af::getBackendCount() - Gets the number of backends available for use
1571 * \ref af::getAvailableBackends() - Returns information about available backends
1572 * \ref af::getBackendId() - Gets the backend enum for an array
1575 * \ref af::homography() - Homography estimation
1576 * \ref af::gloh() - GLOH Descriptor for SIFT
1579 * \ref af::loadImageNative() - Load an image as native data without modification
1580 * \ref af::saveImageNative() - Save an image without modifying data or type
1583 * \ref af::Window::plot3() - 3-dimensional line plot
1584 * \ref af::Window::surface() - 3-dimensional curve plot
1587 * \ref af_create_indexers()
1588 * \ref af_set_array_indexer()
1589 * \ref af_set_seq_indexer()
1590 * \ref af_set_seq_param_indexer()
1591 * \ref af_release_indexers()
1593 * CUDA Backend Specific
1594 * \ref afcu::setNativeId() - Set the CUDA device with given native id as active
1595 * ArrayFire uses a modified order for devices. The native id for a
1596 device can be retreived using `nvidia-smi`
1598 * OpenCL Backend Specific
1599 * \ref afcl::setDeviceId() - Set the OpenCL device using the `clDeviceId`
1602 ------------------------
1603 * Added \ref c32 and \ref c64 support for \ref af::isNaN(), \ref af::isInf() and \ref af::iszero()
1604 * Added CPU information for `x86` and `x86_64` architectures in CPU backend's \ref af::info()
1605 * Batch support for \ref af::approx1() and \ref af::approx2()
1606 * Now can be used with gfor as well
1607 * Added \ref s64 and \ref u64 support to:
1608 * \ref af::sort() (along with sort index and sort by key)
1609 * \ref af::setUnique(), \ref af::setUnion(), \ref af::setIntersect()
1610 * \ref af::convolve() and \ref af::fftConvolve()
1611 * \ref af::histogram() and \ref af::histEqual()
1614 * Added \ref AF_MSG macro
1618 * Submodules update is now automatically called if not cloned recursively
1619 * [Fixes for compilation](https://github.com/arrayfire/arrayfire/issues/766) on Visual Studio 2015
1620 * Option to use [fallback to CPU LAPACK](https://github.com/arrayfire/arrayfire/pull/1053)
1621 for linear algebra functions in case of CUDA 6.5 or older versions.
1625 * Fixed [memory leak](https://github.com/arrayfire/arrayfire/pull/1096) in \ref af::susan()
1626 * Fixed [failing test](https://github.com/arrayfire/arrayfire/commit/144a2db)
1627 in \ref af::lower() and \ref af::upper() for CUDA compute 53
1628 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1092) in CUDA for indexing out of bounds
1629 * Fixed [dims check](https://github.com/arrayfire/arrayfire/commit/6975da8) in \ref af::iota()
1630 * Fixed [out-of-bounds access](https://github.com/arrayfire/arrayfire/commit/7fc3856) in \ref af::sift()
1631 * Fixed [memory allocation](https://github.com/arrayfire/arrayfire/commit/5e88e4a) in \ref af::fast() OpenCL
1632 * Fixed [memory leak](https://github.com/arrayfire/arrayfire/pull/994) in image I/O functions
1633 * \ref af::dog() now returns float-point type arrays
1635 Documentation Updates
1636 ---------------------
1637 * Improved tutorials documentation
1638 * More detailed Using on [Linux](\ref using_on_linux), [OSX](\ref using_on_osx),
1639 [Windows](\ref using_on_windows) pages.
1640 * Added return type information for functions that return different type
1646 * [Plot3](\ref graphics/plot3.cpp)
1647 * [Surface](\ref graphics/surface.cpp)
1648 * [Shallow Water Equation](\ref pde/swe.cpp)
1649 * [Basic](\ref unified/basic.cpp) as a Unified backend example
1653 * All installers now include the Unified backend and corresponding CMake files
1654 * Visual Studio projects include Unified in the Platform Configurations
1655 * Added installer for Jetson TX1
1656 * SIFT and GLOH do not ship with the installers as SIFT is protected by
1657 patents that do not allow commercial distribution without licensing.
1665 * Fixed [bugs](https://github.com/arrayfire/arrayfire/issues/1042) in various OpenCL kernels without offset additions
1666 * Remove ARCH_32 and ARCH_64 flags
1667 * Fix [missing symbols](https://github.com/arrayfire/arrayfire/issues/1040) when freeimage is not found
1668 * Use CUDA driver version for Windows
1669 * Improvements to SIFT
1670 * Fixed [memory leak](https://github.com/arrayfire/arrayfire/issues/1045) in median
1671 * Fixes for Windows compilation when not using MKL [#1047](https://github.com/arrayfire/arrayfire/issues/1047)
1672 * Fixed for building without LAPACK
1677 * Documentation: Fixed documentation for select and replace
1678 * Documentation: Fixed documentation for af_isnan
1686 * Fixed [bug](https://github.com/arrayfire/arrayfire/commit/4698f12) in assign that was causing test to fail
1687 * Fixed bug in convolve. Frequency condition now depends on kernel size only
1688 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1005) in indexed reductions for complex type in OpenCL backend
1689 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1006) in kernel name generation in ireduce for OpenCL backend
1690 * Fixed non-linear to linear indices in ireduce
1691 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1011) in reductions for small arrays
1692 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/1010) in histogram for indexed arrays
1693 * Fixed [compiler error](https://github.com/arrayfire/arrayfire/issues/1015) CPUID for non-compliant devices
1694 * Fixed [failing tests](https://github.com/arrayfire/arrayfire/issues/1008) on i386 platforms
1700 * Documentation: Added missing examples and other corrections
1701 * Documentation: Fixed warnings in documentation building
1702 * Installers: Send error messages to log file in OSX Installer
1710 * CUDA backend now depends on CUDA 7.5 toolkit
1711 * OpenCL backend now require OpenCL 1.2 or greater
1716 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/981) in reductions after indexing
1717 * Fixed [bug](https://github.com/arrayfire/arrayfire/issues/976) in indexing when using reverse indices
1722 * `cmake` now includes `PKG_CONFIG` in the search path for CBLAS and LAPACKE libraries
1723 * [heston_model.cpp](\ref financial/heston_model.cpp) example now builds with the default ArrayFire cmake files after installation
1728 * Fixed bug in [image_editing.cpp](\ref image_processing/image_editing.cpp)
1735 * Computer Vision Functions
1736 * \ref af::nearestNeighbour() - Nearest Neighbour with SAD, SSD and SHD distances
1737 * \ref af::harris() - Harris Corner Detector
1738 * \ref af::susan() - Susan Corner Detector
1739 * \ref af::sift() - Scale Invariant Feature Transform (SIFT)
1740 * Method and apparatus for identifying scale invariant features"
1741 "in an image and use of same for locating an object in an image,\" David"
1742 "G. Lowe, US Patent 6,711,293 (March 23, 2004). Provisional application"
1743 "filed March 8, 1999. Asignee: The University of British Columbia. For"
1744 "further details, contact David Lowe (lowe@cs.ubc.ca) or the"
1745 "University-Industry Liaison Office of the University of British"
1747 * SIFT is available for compiling but does not ship with ArrayFire
1748 hosted installers/pre-built libraries
1749 * \ref af::dog() - Difference of Gaussians
1751 * Image Processing Functions
1752 * \ref ycbcr2rgb() and \ref rgb2ycbcr() - RGB <->YCbCr color space conversion
1753 * \ref wrap() and \ref unwrap() Wrap and Unwrap
1754 * \ref sat() - Summed Area Tables
1755 * \ref loadImageMem() and \ref saveImageMem() - Load and Save images to/from memory
1756 * \ref af_image_format - Added imageFormat (af_image_format) enum
1758 * Array & Data Handling
1759 * \ref copy() - Copy
1760 * array::lock() and array::unlock() - Lock and Unlock
1761 * \ref select() and \ref replace() - Select and Replace
1762 * Get array reference count (af_get_data_ref_count)
1765 * \ref fftInPlace() - 1D in place FFT
1766 * \ref fft2InPlace() - 2D in place FFT
1767 * \ref fft3InPlace() - 3D in place FFT
1768 * \ref ifftInPlace() - 1D in place Inverse FFT
1769 * \ref ifft2InPlace() - 2D in place Inverse FFT
1770 * \ref ifft3InPlace() - 3D in place Inverse FFT
1771 * \ref fftR2C() - Real to complex FFT
1772 * \ref fftC2R() - Complex to Real FFT
1775 * \ref svd() and \ref svdInPlace() - Singular Value Decomposition
1778 * \ref sigmoid() - Sigmoid
1779 * Sum (with option to replace NaN values)
1780 * Product (with option to replace NaN values)
1783 * Window::setSize() - Window resizing using Forge API
1786 * Allow users to set print precision (print, af_print_array_gen)
1787 * \ref saveArray() and \ref readArray() - Stream arrays to binary files
1788 * \ref toString() - toString function returns the array and data as a string
1790 * CUDA specific functionality
1791 * \ref getStream() - Returns default CUDA stream ArrayFire uses for the current device
1792 * \ref getNativeId() - Returns native id of the CUDA device
1797 * Allow complex inputs with conjugate option
1798 * AF_INTERP_LOWER interpolation
1799 * For resize, rotate and transform based functions
1800 * 64-bit integer support
1801 * For reductions, random, iota, range, diff1, diff2, accum, join, shift
1804 * Support for non-overlapping batched convolutions
1806 * Fix binary ops on complex inputs of mixed types
1807 * Complex type support for exp
1809 * Performance improvements by using JIT when possible.
1810 * Add AF_API_VERSION macro
1811 * Allows disabling of API to maintain consistency with previous versions
1812 * Other Performance Improvements
1813 * Use reference counting to reduce unnecessary copies
1815 * Device properties for CPU
1816 * Improved performance when all buffers are indexed linearly
1818 * Use streams in CUDA (no longer using default stream)
1819 * Using async cudaMem ops
1820 * Add 64-bit integer support for JIT functions
1821 * Performance improvements for CUDA JIT for non-linear 3D and 4D arrays
1823 * Improve compilation times for OpenCL backend
1824 * Performance improvements for non-linear JIT kernels on OpenCL
1825 * Improved shared memory load/store in many OpenCL kernels (PR 933)
1826 * Using cl.hpp v1.2.7
1831 * Fix compatibility of c32/c64 arrays when operating with scalars
1832 * Fix median for all values of an array
1833 * Fix double free issue when indexing (30cbbc7)
1834 * Fix [bug](https://github.com/arrayfire/arrayfire/issues/901) in rank
1835 * Fix default values for scale throwing exception
1836 * Fix conjg raising exception on real input
1837 * Fix bug when using conjugate transpose for vector input
1838 * Fix issue with const input for array_proxy::get()
1840 * Fix randn generating same sequence for multiple calls
1841 * Fix setSeed for randu
1842 * Fix casting to and from complex
1843 * Check NULL values when allocating memory
1844 * Fix [offset issue](https://github.com/arrayfire/arrayfire/issues/923) for CPU element-wise operations
1850 * Heston Model (contributed by Michael Nowotny)
1854 * Fixed bug in automatic detection of ArrayFire when using with CMake in Windows
1855 * The Linux libraries are now compiled with static version of FreeImage
1859 * OpenBlas can cause issues with QR factorization in CPU backend
1860 * FreeImage older than 3.10 can cause issues with loadImageMem and
1862 * OpenCL backend issues on OSX
1863 * AMD GPUs not supported because of driver issues
1864 * Intel CPUs not supported
1865 * Linear algebra functions do not work on Intel GPUs.
1866 * Stability and correctness issues with open source OpenCL implementations such as Beignet, GalliumCompute.
1874 * Added missing symbols from the compatible API
1875 * Fixed a bug affecting corner rows and elements in \ref af::grad()
1876 * Fixed linear interpolation bugs affecting large images in the following:
1877 - \ref af::approx1()
1878 - \ref af::approx2()
1883 - \ref af::transform()
1888 * Added missing documentation for \ref af::constant()
1889 * Added missing documentation for `array::scalar()`
1890 * Added supported input types for functions in `arith.h`
1898 * Fixed header to work in Visual Studio 2015
1899 * Fixed a bug in batched mode for FFT based convolutions
1900 * Fixed graphics issues on OSX
1901 * Fixed various bugs in visualization functions
1906 * Improved fractal example
1908 * Improved Windows installer
1909 * Default install path has been changed
1910 * Fixed bug in machine learning examples
1920 * ArrayFire is now open source
1921 * Major changes to the visualization library
1922 * Introducing handle based C API
1923 * New backend: CPU fallback available for systems without GPUs
1924 * Dense linear algebra functions available for all backends
1925 * Support for 64 bit integers
1929 * Data generation functions
1933 * Computer Vision Algorithms
1935 * A data structure to hold features
1937 * FAST feature detector
1939 * ORB A feature descriptor extractor
1942 * convolve1(), convolve2(), convolve3()
1943 * Specialized versions of convolve() to enable better batch support
1944 * fftconvolve1(), fftconvolve2(), fftconvolve3()
1945 * Convolutions in frequency domain to support larger kernel sizes
1947 * Unified functions for calling multi dimensional ffts.
1949 * Match a kernel in an image
1951 * Get sobel gradients of an image
1952 * rgb2hsv(), hsv2rgb(), rgb2gray(), gray2rgb()
1953 * Explicit function calls to colorspace conversions
1954 * erode3d(), dilate3d()
1955 * Explicit erode and dilate calls for image morphing
1958 * matmulNT(), matmulTN(), matmulTT()
1959 * Specialized versions of matmul() for transposed inputs
1960 * luInPlace(), choleskyInPlace(), qrInPlace()
1961 * In place factorizations to improve memory requirements
1963 * Specialized solve routines to improve performance
1964 * OpenCL backend now Linear Algebra functions
1967 * lookup() - lookup indices from a table
1968 * batchFunc() - helper function to perform batch operations
1970 * Visualization functions
1971 * Support for multiple windows
1973 * Visualize the output of the histogram
1976 * Removed old pointer based C API
1977 * Introducing handle base C API
1978 * Just In Time compilation available in C API
1979 * C API has feature parity with C++ API
1980 * bessel functions removed
1981 * cross product functions removed
1982 * Kronecker product functions removed
1984 Performance Improvements
1985 ------------------------
1986 * Improvements across the board for OpenCL backend
1989 ---------------------
1990 * `print` is now af_print()
1991 * seq(): The step parameter is now the third input
1992 * seq(start, step, end) changed to seq(start, end, step)
1993 * gfor(): The iterator now needs to be seq()
1995 Deprecated Function APIs
1996 ------------------------
1997 Deprecated APIs are in af/compatible.h
1999 * devicecount() changed to getDeviceCount()
2000 * deviceset() changed to setDevice()
2001 * deviceget() changed to getDevice()
2002 * loadimage() changed to loadImage()
2003 * saveimage() changed to saveImage()
2004 * gaussiankernel() changed to gaussianKernel()
2005 * alltrue() changed to allTrue()
2006 * anytrue() changed to anyTrue()
2007 * setunique() changed to setUnique()
2008 * setunion() changed to setUnion()
2009 * setintersect() changed to setIntersect()
2010 * histequal() changed to histEqual()
2011 * colorspace() changed to colorSpace()
2012 * filter() deprecated. Use convolve1() and convolve2()
2013 * mul() changed to product()
2014 * deviceprop() changed to deviceProp()
2017 ----------------------
2018 * OpenCL backend issues on OSX
2019 * AMD GPUs not supported because of driver issues
2020 * Intel CPUs not supported
2021 * Linear algebra functions do not work on Intel GPUs.
2022 * Stability and correctness issues with open source OpenCL implementations such as Beignet, GalliumCompute.