blob: f5b8825558c9a86b25e8cd2f1dc39639a9fe3ec9 [file] [log] [blame]
Jenkinsb9abeae2018-11-22 11:58:08 +00001///
Jenkins36ccc902020-02-21 11:10:48 +00002/// Copyright (c) 2017-2020 ARM Limited.
Jenkinsb9abeae2018-11-22 11:58:08 +00003///
4/// SPDX-License-Identifier: MIT
5///
6/// Permission is hereby granted, free of charge, to any person obtaining a copy
7/// of this software and associated documentation files (the "Software"), to
8/// deal in the Software without restriction, including without limitation the
9/// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10/// sell copies of the Software, and to permit persons to whom the Software is
11/// furnished to do so, subject to the following conditions:
12///
13/// The above copyright notice and this permission notice shall be included in all
14/// copies or substantial portions of the Software.
15///
16/// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17/// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18/// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19/// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20/// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21/// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22/// SOFTWARE.
23///
Jenkinsc3f34a42018-03-02 12:38:09 +000024namespace arm_compute
25{
Anthony Barbierdbdab852017-06-23 15:42:00 +010026/** @mainpage Introduction
27
28@tableofcontents
29
30The Computer Vision and Machine Learning library is a set of functions optimised for both ARM CPUs and GPUs using SIMD technologies.
31
32Several builds of the library are available using various configurations:
33 - OS: Linux, Android or bare metal.
34 - Architecture: armv7a (32bit) or arm64-v8a (64bit)
Anthony Barbier8140e1e2017-12-14 23:48:46 +000035 - Technology: NEON / OpenCL / GLES_COMPUTE / NEON and OpenCL and GLES_COMPUTE
Anthony Barbierdbdab852017-06-23 15:42:00 +010036 - Debug / Asserts / Release: Use a build with asserts enabled to debug your application and enable extra validation. Once you are sure your application works as expected you can switch to a release build of the library for maximum performance.
37
38@section S0_1_contact Contact / Support
39
40Please email developer@arm.com
41
42In order to facilitate the work of the support team please provide the build information of the library you are using. To get the version of the library you are using simply run:
43
44 $ strings android-armv7a-cl-asserts/libarm_compute.so | grep arm_compute_version
45 arm_compute_version=v16.12 Build options: {'embed_kernels': '1', 'opencl': '1', 'arch': 'armv7a', 'neon': '0', 'asserts': '1', 'debug': '0', 'os': 'android', 'Werror': '1'} Git hash=f51a545d4ea12a9059fe4e598a092f1fd06dc858
46
Anthony Barbier8140e1e2017-12-14 23:48:46 +000047@section S0_2_prebuilt_binaries Pre-built binaries
48
49For each release we provide some pre-built binaries of the library [here](https://github.com/ARM-software/ComputeLibrary/releases)
50
51These binaries have been built using the following toolchains:
Jenkins6a7771e2020-05-28 11:28:36 +010052 - Linux armv7a: gcc-linaro-6.3.1-2017.05-x86_64_arm-linux-gnueabihf
53 - Linux arm64-v8a: gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu
54 - Android armv7a: clang++ / libc++ NDK r18b
55 - Android am64-v8a: clang++ / libc++ NDK r18b
Anthony Barbier8140e1e2017-12-14 23:48:46 +000056
57@warning Make sure to use a compatible toolchain to build your application or you will get some std::bad_alloc errors at runtime.
58
Anthony Barbierdbdab852017-06-23 15:42:00 +010059@section S1_file_organisation File organisation
60
61This archive contains:
62 - The arm_compute header and source files
63 - The latest Khronos OpenCL 1.2 C headers from the <a href="https://www.khronos.org/registry/cl/">Khronos OpenCL registry</a>
64 - The latest Khronos cl2.hpp from the <a href="https://www.khronos.org/registry/cl/">Khronos OpenCL registry</a> (API version 2.1 when this document was written)
Anthony Barbier8140e1e2017-12-14 23:48:46 +000065 - The latest Khronos OpenGL ES 3.1 C headers from the <a href="https://www.khronos.org/registry/gles/">Khronos OpenGL ES registry</a>
66 - The latest Khronos EGL 1.5 C headers from the <a href="https://www.khronos.org/registry/gles/">Khronos EGL registry</a>
67 - The sources for a stub version of libOpenCL.so, libGLESv1_CM.so, libGLESv2.so and libEGL.so to help you build your application.
Anthony Barbierdbdab852017-06-23 15:42:00 +010068 - An examples folder containing a few examples to compile and link against the library.
69 - A @ref utils folder containing headers with some boiler plate code used by the examples.
70 - This documentation.
71
72You should have the following file organisation:
73
74 .
75 ├── arm_compute --> All the arm_compute headers
Jenkins4ba87db2019-05-23 17:11:51 +010076 │ ├── graph.h --> Includes all the Graph headers at once.
Anthony Barbierdbdab852017-06-23 15:42:00 +010077 │   ├── core
78 │   │   ├── CL
Kaizen8938bd32017-09-28 14:38:23 +010079 │   │   │   ├── CLKernelLibrary.h --> Manages all the OpenCL kernels compilation and caching, provides accessors for the OpenCL Context.
Anthony Barbierdbdab852017-06-23 15:42:00 +010080 │   │   │   ├── CLKernels.h --> Includes all the OpenCL kernels at once
Jenkins6a7771e2020-05-28 11:28:36 +010081 │   │   │   ├── CL specialisation of all the generic interfaces (ICLTensor, ICLArray, etc.)
82 │   │   │   ├── gemm --> Folder containing all the configuration files for GEMM
Anthony Barbierdbdab852017-06-23 15:42:00 +010083 │   │   │   ├── kernels --> Folder containing all the OpenCL kernels
84 │   │   │   │   └── CL*Kernel.h
85 │   │   │   └── OpenCL.h --> Wrapper to configure the Khronos OpenCL C++ header
86 │   │ ├── CPP
Kaizen8938bd32017-09-28 14:38:23 +010087 │   │   │   ├── CPPKernels.h --> Includes all the CPP kernels at once
Anthony Barbierdbdab852017-06-23 15:42:00 +010088 │   │ │   └── kernels --> Folder containing all the CPP kernels
Kaizen8938bd32017-09-28 14:38:23 +010089 │   │   │      └── CPP*Kernel.h
Anthony Barbier8140e1e2017-12-14 23:48:46 +000090 │   │   ├── GLES_COMPUTE
91 │   │   │   ├── GCKernelLibrary.h --> Manages all the GLES kernels compilation and caching, provides accessors for the GLES Context.
92 │   │   │   ├── GCKernels.h --> Includes all the GLES kernels at once
Jenkins6a7771e2020-05-28 11:28:36 +010093 │   │   │   ├── GLES specialisation of all the generic interfaces (IGCTensor etc.)
Anthony Barbier8140e1e2017-12-14 23:48:46 +000094 │   │   │   ├── kernels --> Folder containing all the GLES kernels
95 │   │   │   │   └── GC*Kernel.h
96 │   │   │   └── OpenGLES.h --> Wrapper to configure the Khronos EGL and OpenGL ES C header
Anthony Barbierdbdab852017-06-23 15:42:00 +010097 │   │   ├── NEON
98 │   │   │   ├── kernels --> Folder containing all the NEON kernels
Jenkinsb3a371b2018-05-23 11:36:53 +010099 │   │   │   │ ├── assembly --> headers for assembly optimised NEON kernels.
100 │   │   │   │ ├── convolution --> headers for convolution assembly optimised NEON kernels.
101 │   │   │   │   │   ├── common --> headers for code which is common to several convolution implementations.
Jenkins6a7771e2020-05-28 11:28:36 +0100102 │   │   │   │   │   ├── depthwise --> headers for Depthwise convolution assembly implementation
Jenkinsb3a371b2018-05-23 11:36:53 +0100103 │   │   │   │   │   └── winograd --> headers for Winograd convolution assembly implementation
104 │   │   │   │ ├── detail --> Common code for several intrinsics implementations.
Anthony Barbierdbdab852017-06-23 15:42:00 +0100105 │   │   │   │   └── NE*Kernel.h
Jenkins36ccc902020-02-21 11:10:48 +0000106 │   │   │   ├── wrapper --> NEON wrapper used to simplify code
Jenkins6a7771e2020-05-28 11:28:36 +0100107 │   │   │   │ ├── intrinsics --> NEON intrinsics wrappers
Jenkins36ccc902020-02-21 11:10:48 +0000108 │   │   │   │ ├── scalar --> Scalar operations
109 │   │   │   │ ├── traits.h --> Traits defined on NEON vectors
110 │   │   │   │   └── wrapper.h --> Includes all wrapper headers at once
Anthony Barbierdbdab852017-06-23 15:42:00 +0100111 │   │   │   └── NEKernels.h --> Includes all the NEON kernels at once
112 │   │   ├── All common basic types (Types.h, Window, Coordinates, Iterator, etc.)
Jenkins6a7771e2020-05-28 11:28:36 +0100113 │   │   ├── All generic interfaces (ITensor, IArray, etc.)
Jenkins36ccc902020-02-21 11:10:48 +0000114 │   │   └── Objects metadata classes (TensorInfo, MultiImageInfo)
Kaizen8938bd32017-09-28 14:38:23 +0100115 │   ├── graph
Jenkins6a7771e2020-05-28 11:28:36 +0100116 │   │   ├── algorithms --> Generic algorithms used by the graph backend (e.g Order of traversal)
Jenkinsb3a371b2018-05-23 11:36:53 +0100117 │   │   ├── backends --> The backend specific code
118 │   │   │   ├── CL --> OpenCL specific operations
119 │   │   │   ├── GLES --> OpenGLES Compute Shaders specific operations
120 │   │   │   └── NEON --> NEON specific operations
Jenkins6a7771e2020-05-28 11:28:36 +0100121 │   │   ├── detail --> Collection of internal utilities.
122 │   │   ├── frontend --> Code related to the stream frontend interface.
123 │   │   ├── mutators --> Used to modify / optimise the Graph intermediate representation(Operator fusion, in place operations, etc.)
124 │   │   ├── nodes --> The various nodes supported by the graph API
125 │   │   ├── printers --> Debug printers
126 │   │   └── Graph objects interfaces (INode, ITensorAccessor, Graph, etc.)
Anthony Barbierdbdab852017-06-23 15:42:00 +0100127 │   └── runtime
Jenkins36ccc902020-02-21 11:10:48 +0000128 │   ├── common
129 │ │ └── Common utility code used by all backends
Anthony Barbierdbdab852017-06-23 15:42:00 +0100130 │   ├── CL
Jenkins36ccc902020-02-21 11:10:48 +0000131 │   │   ├── CL objects & allocators (CLArray, CLTensor, etc.)
Anthony Barbierdbdab852017-06-23 15:42:00 +0100132 │   │   ├── functions --> Folder containing all the OpenCL functions
133 │   │   │   └── CL*.h
Kaizen8938bd32017-09-28 14:38:23 +0100134 │   │   ├── CLScheduler.h --> Interface to enqueue OpenCL kernels and get/set the OpenCL CommandQueue and ICLTuner.
Jenkinsb3a371b2018-05-23 11:36:53 +0100135 │   │   ├── CLFunctions.h --> Includes all the OpenCL functions at once
Jenkins36ccc902020-02-21 11:10:48 +0000136 │   │   ├── ICLTuner.h --> Interface used to tune the local work-group size of OpenCL kernels
Jenkinsb3a371b2018-05-23 11:36:53 +0100137 │   │   └── tuners
138 │   │      └── Local workgroup size tuners for specific architectures / GPUs
Anthony Barbierdbdab852017-06-23 15:42:00 +0100139 │   ├── CPP
Kaizen8938bd32017-09-28 14:38:23 +0100140 │      │   ├── CPPKernels.h --> Includes all the CPP functions at once.
Jenkinsb3a371b2018-05-23 11:36:53 +0100141 │   │   ├── CPPScheduler.h --> Basic pool of threads to execute CPP/NEON code on several cores in parallel
142 │   │   └── functions --> Folder containing all the CPP functions
143 │   │      └── CPP*.h
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000144 │   ├── GLES_COMPUTE
Jenkins36ccc902020-02-21 11:10:48 +0000145 │   │   ├── GLES objects & allocators (GCArray, GCTensor, etc.)
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000146 │   │   ├── functions --> Folder containing all the GLES functions
147 │   │   │   └── GC*.h
148 │   │   ├── GCScheduler.h --> Interface to enqueue GLES kernels and get/set the GLES CommandQueue.
149 │   │   └── GCFunctions.h --> Includes all the GLES functions at once
Anthony Barbierdbdab852017-06-23 15:42:00 +0100150 │   ├── NEON
151 │   │ ├── functions --> Folder containing all the NEON functions
152 │   │ │   └── NE*.h
153 │   │ └── NEFunctions.h --> Includes all the NEON functions at once
Kaizen8938bd32017-09-28 14:38:23 +0100154 │   ├── OMP
155 │   │   └── OMPScheduler.h --> OpenMP scheduler (Alternative to the CPPScheduler)
Jenkins36ccc902020-02-21 11:10:48 +0000156 │ ├── Memory & weights manager files (LifetimeManager, PoolManager, etc.)
157 │   └── Basic implementations of the generic object interfaces (Array, Tensor, etc.)
158 ├── data --> Contains test images and reference data dumps used by validation tests
Jenkins6a7771e2020-05-28 11:28:36 +0100159 ├── docs --> Contains Doxyfile and Doxygen sources used to generate the HTML pages.
Anthony Barbierdbdab852017-06-23 15:42:00 +0100160 ├── examples
Jenkins6a7771e2020-05-28 11:28:36 +0100161 │   ├── gemm_tuner
162 │   │ └── OpenCL GEMM tuner utility
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000163 │   ├── cl_*.cpp --> OpenCL examples
164 │   ├── gc_*.cpp --> GLES compute shaders examples
165 │   ├── graph_*.cpp --> Graph examples
166 │   ├── neoncl_*.cpp --> NEON / OpenCL interoperability examples
167 │   └── neon_*.cpp --> NEON examples
Anthony Barbierdbdab852017-06-23 15:42:00 +0100168 ├── include
Kaizen8938bd32017-09-28 14:38:23 +0100169 │   ├── CL
170 │   │ └── Khronos OpenCL C headers and C++ wrapper
171 │   ├── half --> FP16 library available from http://half.sourceforge.net
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000172 │   ├── libnpy --> Library to load / write npy buffers, available from https://github.com/llohse/libnpy
Jenkins36ccc902020-02-21 11:10:48 +0000173 │  ├── linux --> Headers only needed for Linux builds
174 │   │ └── Khronos EGL and OpenGLES headers
175 │ └── stb
176 │ └── stb_image.h --> Single header library to load image files, available from https://github.com/nothings/stb
Kaizen8938bd32017-09-28 14:38:23 +0100177 ├── scripts
178 │   ├── caffe_data_extractor.py --> Basic script to export weights from Caffe to npy files
179 │   └── tensorflow_data_extractor.py --> Basic script to export weights from Tensor Flow to npy files
Anthony Barbierdbdab852017-06-23 15:42:00 +0100180 ├── src
181 │   ├── core
182 │ │ └── ... (Same structure as headers)
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000183 │   │ ├── CL
184 │   │ │ └── cl_kernels --> All the OpenCL kernels
185 │   │ └── GLES_COMPUTE
186 │   │ └── cs_shaders --> All the OpenGL ES Compute Shaders
Kaizen8938bd32017-09-28 14:38:23 +0100187 │   ├── graph
188 │ │ └── ... (Same structure as headers)
Anthony Barbierdbdab852017-06-23 15:42:00 +0100189 │ └── runtime
190 │ └── ... (Same structure as headers)
Kaizen8938bd32017-09-28 14:38:23 +0100191 ├── support
192 │ └── Various headers to work around toolchains / platform issues.
Anthony Barbierdbdab852017-06-23 15:42:00 +0100193 ├── tests
194 │   ├── All test related files shared between validation and benchmark
Jenkinsb3a371b2018-05-23 11:36:53 +0100195 │   ├── benchmark --> Sources for benchmarking
196 │ │ ├── Benchmark specific files
197 │   │ ├── fixtures
198 │ │ │ └── Backend agnostic fixtures to initialise and run the functions to test.
199 │ │ ├── CL --> OpenCL benchmarking tests
200 │ │ ├── GLES_COMPUTE --> GLES benchmarking tests
201 │ │ └── NEON --> NEON benchmarking tests
Jenkins36ccc902020-02-21 11:10:48 +0000202 │ ├── benchmark_examples --> Sources needed to wrap examples to run through our benchmarking framework.
Kaizen8938bd32017-09-28 14:38:23 +0100203 │   ├── CL --> OpenCL accessors
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000204 │   ├── GLES_COMPUTE --> GLES accessors
Kaizen8938bd32017-09-28 14:38:23 +0100205 │   ├── NEON --> NEON accessors
Kaizen8938bd32017-09-28 14:38:23 +0100206 │   ├── datasets
207 │ │ └── Datasets for all the validation / benchmark tests, layer configurations for various networks, etc.
208 │   ├── framework
209 │ │ └── Boiler plate code for both validation and benchmark test suites (Command line parsers, instruments, output loggers, etc.)
Jenkins36ccc902020-02-21 11:10:48 +0000210 │   ├── instruments --> User defined instruments that can be registered to the framework.
211 │ ├── validate_examples --> Sources needed to wrap examples to run through our validation framework.
Jenkinsb3a371b2018-05-23 11:36:53 +0100212 │   └── validation --> Sources for validation
213 │ ├── Validation specific files
214 │   ├── fixtures
215 │ │ └── Backend agnostic fixtures to initialise and run the functions to test.
216 │   ├── reference
217 │ │ └── Reference implementation used to validate the results of the various backends.
218 │ ├── CL --> OpenCL validation tests
219 │ ├── GLES_COMPUTE --> GLES validation tests
220 │ ├── CPP --> C++ reference implementations
221 │ └── NEON --> NEON validation tests
Anthony Barbierdbdab852017-06-23 15:42:00 +0100222 └── utils --> Boiler plate code used by examples
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000223 └── Various utilities to print types, load / store assets, etc.
Anthony Barbierdbdab852017-06-23 15:42:00 +0100224
225@section S2_versions_changelog Release versions and changelog
226
227@subsection S2_1_versions Release versions
228
229All releases are numbered vYY.MM Where YY are the last two digits of the year, and MM the month number.
230If there is more than one release in a month then an extra sequential number is appended at the end:
231
232 v17.03 (First release of March 2017)
233 v17.03.1 (Second release of March 2017)
234 v17.04 (First release of April 2017)
235
236@note We're aiming at releasing one major public release with new features per quarter. All releases in between will only contain bug fixes.
237
238@subsection S2_2_changelog Changelog
239
Jenkins6a7771e2020-05-28 11:28:36 +0100240v20.05 Public major release
241 - Various bug fixes.
242 - Various optimisations.
243 - Updated recommended NDK version to r18b.
244 - Updated recommended gcc version to Linaro 6.3.1.
245 - Added Bfloat16 type support
246 - Added Bfloat16 support in:
247 - @ref NEWeightsReshapeKernel
248 - @ref NEConvolutionLayerReshapeWeights
249 - @ref NEIm2ColKernel
250 - @ref NEIm2Col
251 - @ref NEDepthConvertLayerKernel
252 - @ref NEDepthConvertLayer
253 - @ref NEGEMMConvolutionLayer
254 - @ref NEGEMMAssemblyDispatch
255 - Added new data type QASYMM8_SIGNED support for:
256 - @ref CLDirectConvolutionLayer
257 - @ref CLDeconvolutionLayer
258 - @ref CLDirectDeconvolutionLayer
259 - @ref CLGEMMDeconvolutionLayer
260 - @ref CLGEMMLowpMatrixMultiplyReshapedKernel
261 - @ref CLGEMMLowpQuantizeDownInt32ScaleKernel
262 - @ref CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel
263 - @ref CLReductionOperation
264 - @ref CLReduceMean
265 - @ref NEScale
266 - @ref NEScaleKernel
267 - @ref NEUpsampleLayer
268 - @ref NECast
269 - @ref NEReductionOperation
270 - @ref NEReduceMean
271 - @ref NEArgMinMaxLayer
272 - @ref NEDeconvolutionLayer
273 - @ref NEGEMMLowpQuantizeDownInt32ScaleKernel
274 - @ref CPPBoxWithNonMaximaSuppressionLimit
275 - @ref CPPDetectionPostProcessLayer
276 - @ref CPPPermuteKernel
277 - @ref CPPPermute
278 - @ref CPPTopKVKernel
279 - @ref CPPTopKV
280 - @ref CPPUpsample
281 - @ref CPPUpsampleKernel
282 - New OpenCL kernels / functions:
283 - @ref CLQLSTMLayer
284 - @ref CLQLSTMLayerNormalizationKernel
285 - New NEON kernels / functions:
286 - @ref NEQLSTMLayer
287 - @ref NEQLSTMLayerNormalizationKernel
288 - Added HARD_SWISH support in:
289 - @ref CLActivationLayerKernel
290 - @ref NEActivationLayerKernel
291 - Deprecated OpenCL kernels / functions:
292 - CLGEMMLowpQuantizeDownInt32ToUint8Scale
293 - CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat
294 - Deprecated NEON kernels / functions:
295 - NEGEMMLowpQuantizeDownInt32ToUint8Scale
296 - Removed CPP kernels / functions:
297 - CPPFlipWeightsKernel
298 - Removed PoolingLayerInfo constructors without Data Layout.
299 - Removed CLDepthwiseConvolutionLayer3x3
300 - Removed NEDepthwiseConvolutionLayerOptimized
301 - Added support for Winograd 3x3,4x4 on NEON FP16:
302 - @ref NEWinogradConvolutionLayer
303 - @ref NEWinogradLayerTransformInputKernel
304 - @ref NEWinogradLayerTransformOutputKernel
305 - @ref NEWinogradLayerTransformWeightsKernel
306 - Added CLCompileContext
307 - Added NEON GEMM kernel with 2D window support
308
Jenkins575c81f2020-03-05 16:07:35 +0000309v20.02.1 Maintenance release
310 - Added Android-NN build script.
311
Jenkins36ccc902020-02-21 11:10:48 +0000312v20.02 Public major release
313 - Various bug fixes.
314 - Various optimisations.
315 - Added new data type QASYMM8_SIGNED support for:
316 - @ref CLDepthwiseConvolutionLayer
Jenkins6a7771e2020-05-28 11:28:36 +0100317 - CLDepthwiseConvolutionLayer3x3
Jenkins36ccc902020-02-21 11:10:48 +0000318 - @ref CLGEMMConvolutionLayer
319 - @ref CLGEMMLowpMatrixMultiplyCore
320 - @ref CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel
321 - @ref CLGEMMLowpMatrixMultiplyNativeKernel
322 - @ref NEActivationLayer
323 - @ref NEComparisonOperationKernel
324 - @ref NEConvolutionLayer
325 - @ref NEDepthwiseConvolutionLayer
326 - @ref NEDepthwiseConvolutionLayer3x3Kernel
327 - @ref NEDirectConvolutionLayerOutputStageKernel
328 - @ref NEElementwiseComparison
329 - @ref NEElementwiseMax
330 - @ref NEElementwiseMin
331 - @ref NEElementwiseSquaredDiff
332 - @ref NEFullyConnectedLayer
333 - @ref NEGEMMMatrixVectorMultiplyKernel
334 - @ref NEPixelWiseMultiplication
335 - @ref NEPoolingLayer
336 - @ref NEPReluLayer
337 - Added support for QSYMM8_PER_CHANNEL in:
338 - @ref NEDepthwiseConvolutionLayer3x3Kernel
339 - Added support for split sizes in:
340 - @ref CLSplit
341 - @ref NESplit
342 - New OpenCL kernels / functions:
343 - @ref CLFill
344 - @ref CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel / @ref CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
345 - New NEON kernels / functions:
346 - @ref NEFill
347 - @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel / @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
348 - Deprecated NEON functions / interfaces:
Jenkins6a7771e2020-05-28 11:28:36 +0100349 - CLDepthwiseConvolutionLayer3x3
350 - NEDepthwiseConvolutionLayerOptimized
351 - PoolingLayerInfo constructors without Data Layout.
Jenkins36ccc902020-02-21 11:10:48 +0000352 - Added support for quantization with multiplier greater than 1 on NEON and CL.
353 - Added support for quantized inputs of type QASYMM8_SIGNED and QASYMM8 to @ref CLQuantizationLayer.
354 - Added the ability to build bootcode for bare metal.
355 - Added support for generating synthetic QASYMM8 graphs.
356 - Added support for F16 datatype in VGG16.
357 - Removed pre-built binaries for GLES.
358
Jenkins7f09cf72020-01-22 18:08:16 +0000359v19.11.1 Public maintenance release
360 - Fix offset calculation in NEReductionOperationKernel.
361 - Fix data layout in NEScaleKernel for nhwc.
362 - Retain configuration step data layout to avoid side-effects.
363 - Perform sqrt in double domain for L2 pooling.
364 - Fix output shape calculation for Reduce Mean
365 - Restrict cases where optimized NEPadLayer runs.
366
Jenkins0e205f72019-11-28 16:53:35 +0000367v19.11 Public major release
368 - Various bug fixes.
369 - Various optimisations.
370 - Updated recommended NDK version to r17c.
371 - Deprecated OpenCL kernels / functions:
372 - CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel
373 - CLDepthwiseIm2ColKernel
374 - CLDepthwiseSeparableConvolutionLayer
375 - CLDepthwiseVectorToTensorKernel
376 - CLDirectConvolutionLayerOutputStageKernel
377 - Deprecated NEON kernels / functions:
378 - NEDepthwiseWeightsReshapeKernel
379 - NEDepthwiseIm2ColKernel
380 - NEDepthwiseSeparableConvolutionLayer
381 - NEDepthwiseVectorToTensorKernel
382 - NEDepthwiseConvolutionLayer3x3
383 - New OpenCL kernels / functions:
384 - @ref CLInstanceNormalizationLayerKernel / @ref CLInstanceNormalizationLayer
385 - @ref CLDepthwiseConvolutionLayerNativeKernel to replace the old generic depthwise convolution (see Deprecated
386 OpenCL kernels / functions)
387 - @ref CLLogSoftmaxLayer
388 - New NEON kernels / functions:
389 - @ref NEBoundingBoxTransformKernel / @ref NEBoundingBoxTransform
390 - @ref NEComputeAllAnchorsKernel / @ref NEComputeAllAnchors
391 - @ref NEDetectionPostProcessLayer
392 - @ref NEGenerateProposalsLayer
393 - @ref NEInstanceNormalizationLayerKernel / @ref NEInstanceNormalizationLayer
394 - @ref NELogSoftmaxLayer
395 - @ref NEROIAlignLayerKernel / @ref NEROIAlignLayer
396 - Added QASYMM8 support for:
397 - @ref CLGenerateProposalsLayer
398 - @ref CLROIAlignLayer
399 - @ref CPPBoxWithNonMaximaSuppressionLimit
400 - Added QASYMM16 support for:
401 - @ref CLBoundingBoxTransform
402 - Added FP16 support for:
403 - @ref CLGEMMMatrixMultiplyReshapedKernel
404 - Added new data type QASYMM8_PER_CHANNEL support for:
405 - @ref CLDequantizationLayer
406 - @ref NEDequantizationLayer
407 - Added new data type QSYMM8_PER_CHANNEL support for:
408 - @ref CLConvolutionLayer
409 - @ref NEConvolutionLayer
410 - @ref CLDepthwiseConvolutionLayer
411 - @ref NEDepthwiseConvolutionLayer
412 - Added FP16 mixed-precision support for:
413 - @ref CLGEMMMatrixMultiplyReshapedKernel
414 - @ref CLPoolingLayerKernel
415 - Added FP32 and FP16 ELU activation for:
416 - @ref CLActivationLayer
417 - @ref NEActivationLayer
418 - Added asymmetric padding support for:
419 - @ref CLDirectDeconvolutionLayer
420 - @ref CLGEMMDeconvolutionLayer
421 - @ref NEDeconvolutionLayer
422 - Added SYMMETRIC and REFLECT modes for @ref CLPadLayerKernel / @ref CLPadLayer.
423 - Replaced the calls to @ref NECopyKernel and @ref NEMemsetKernel with @ref NEPadLayer in @ref NEGenerateProposalsLayer.
424 - Replaced the calls to @ref CLCopyKernel and @ref CLMemsetKernel with @ref CLPadLayer in @ref CLGenerateProposalsLayer.
425 - Improved performance for CL Inception V3 - FP16.
426 - Improved accuracy for CL Inception V3 - FP16 by enabling FP32 accumulator (mixed-precision).
427 - Improved NEON performance by enabling fusing batch normalization with convolution and depth-wise convolution layer.
428 - Improved NEON performance for MobileNet-SSD by improving the output detection performance.
429 - Optimized @ref CLPadLayer.
430 - Optimized CL generic depthwise convolution layer by introducing @ref CLDepthwiseConvolutionLayerNativeKernel.
431 - Reduced memory consumption by implementing weights sharing.
432
Jenkins7f09cf72020-01-22 18:08:16 +0000433v19.08.1 Public maintenance release
434 - Fix offset calculation in NEReductionOperationKernel.
435 - Fix data layout in NEScaleKernel for nhwc.
436 - Retain configuration step data layout to avoid side-effects.
437 - Perform sqrt in double domain for L2 pooling.
438 - Fix output shape calculation for Reduce Mean
439 - Fix broadcast CLPixelwiseMultiplication with 5D tensors
440
Jenkins975dfe12019-09-02 11:47:54 +0100441v19.08 Public major release
442 - Various bug fixes.
443 - Various optimisations.
444 - Deprecated NEON functions
445 - NEDepthConcatenateLayer
446 - NEWidthConcatenateLayer
447 - Deprecated OpenCL kernels / functions
448 - CLDepthConcatenateLayer
449 - CLGEMMInterleave4x4Kernel / CLGEMMInterleave4x4
450 - CLGEMMTranspose1xWKernel / CLGEMMTranspose1xW
451 - CLWidthConcatenateLayer
452 - New NEON kernels / functions:
453 - @ref NEAbsLayer
454 - @ref NECast
455 - @ref NEElementwisePower
456 - @ref NELogLayer
457 - @ref NELSTMLayerQuantized
458 - @ref NENegLayer
459 - @ref NEPReluLayer
460 - @ref NESinLayer
461 - @ref NEBatchConcatenateLayerKernel
462 - @ref NEDepthToSpaceLayerKernel / @ref NEDepthToSpaceLayer
463 - @ref NEDepthwiseConvolutionLayerNativeKernel
464 - @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
465 - @ref NEMeanStdDevNormalizationKernel / @ref NEMeanStdDevNormalizationLayer
466 - @ref NESpaceToDepthLayerKernel / @ref NESpaceToDepthLayer
467 - New OpenCL kernels / functions:
468 - @ref CLAbsLayer
469 - @ref CLElementwisePower
470 - @ref CLLogLayer
471 - @ref CLLSTMLayerQuantized
472 - @ref CLNegLayer
473 - @ref CLPReluLayer
474 - @ref CLSinLayer
475 - @ref CLBatchConcatenateLayerKernel
476 - @ref CLDepthToSpaceLayerKernel / @ref CLDepthToSpaceLayer
477 - @ref CLGEMMLowpMatrixMultiplyNativeKernel
478 - @ref CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
479 - @ref CLGEMMMatrixMultiplyNativeKernel
480 - @ref CLMeanStdDevNormalizationKernel / @ref CLMeanStdDevNormalizationLayer
481 - @ref CLSpaceToDepthLayerKernel / @ref CLSpaceToDepthLayer
482 - New examples:
483 - neon_opticalflow
484 - cl_cache
485 - neon_permute
486 - Added support for FP16 in @ref NEDeconvolutionLayer
487 - Added support for FP16 in @ref CLDeconvolutionLayer
488 - Added support for REDUCE_MIN and REDUCE_MAX in @ref ReductionOperation
489 - Enable the fusion of batch normalization with convolution and depthwise convolution layer for FP32 in the graph API (OpenCL only)
490 - Added support for fusing activation function and broadcast addition with the matrix multiplication for FP32 (OpenCL only)
491 - Re-factored the depthwise convolution layer kernel on NEON for generic cases
492 - Added an optimized depthwise convolution layer kernel for 5x5 filters (NEON only)
493 - Added support to enable OpenCL kernel cache. Added example showing how to load the prebuilt OpenCL kernels from a binary cache file
494 - Altered @ref QuantizationInfo interface to support per-channel quantization.
Jenkins6a7771e2020-05-28 11:28:36 +0100495 - The CLDepthwiseConvolutionLayer3x3 will be included by @ref CLDepthwiseConvolutionLayer to accommodate for future optimizations.
496 - The NEDepthwiseConvolutionLayerOptimized will be included by @ref NEDepthwiseConvolutionLayer to accommodate for future optimizations.
Jenkins975dfe12019-09-02 11:47:54 +0100497 - Removed inner_border_right and inner_border_top parameters from @ref CLDeconvolutionLayer interface
498 - Removed inner_border_right and inner_border_top parameters from @ref NEDeconvolutionLayer interface
499 - Optimized the NEON assembly kernel for GEMMLowp. The new implementation fuses the output stage and quantization with the matrix multiplication kernel
500
Jenkins4ba87db2019-05-23 17:11:51 +0100501v19.05 Public major release
502 - Various bug fixes.
503 - Various optimisations.
504 - New Neon kernels / functions:
505 - @ref NEBatchToSpaceLayerKernel / @ref NEBatchToSpaceLayer
506 - @ref NEComplexPixelWiseMultiplicationKernel / @ref NEComplexPixelWiseMultiplication
507 - @ref NECropKernel / @ref NECropResize
508 - @ref NEDepthwiseConvolutionAssemblyDispatch
509 - @ref NEFFTDigitReverseKernel
510 - @ref NEFFTRadixStageKernel
511 - @ref NEFFTScaleKernel
512 - @ref NEGEMMLowpOffsetContributionOutputStageKernel
513 - @ref NEHeightConcatenateLayerKernel
514 - @ref NESpaceToBatchLayerKernel / @ref NESpaceToBatchLayer
515 - @ref NEFFT1D
516 - @ref NEFFT2D
517 - @ref NEFFTConvolutionLayer
518 - New OpenCL kernels / functions:
519 - @ref CLComplexPixelWiseMultiplicationKernel / @ref CLComplexPixelWiseMultiplication
520 - @ref CLCropKernel / @ref CLCropResize
521 - @ref CLDeconvolutionReshapeOutputKernel
522 - @ref CLFFTDigitReverseKernel
523 - @ref CLFFTRadixStageKernel
524 - @ref CLFFTScaleKernel
525 - @ref CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel
526 - @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
527 - @ref CLHeightConcatenateLayerKernel
528 - @ref CLDirectDeconvolutionLayer
529 - @ref CLFFT1D
530 - @ref CLFFT2D
531 - @ref CLFFTConvolutionLayer
532 - @ref CLGEMMDeconvolutionLayer
533 - New OpenGLES kernels / functions:
534 - @ref GCConcatenateLayer
535 - Deprecated functions/interfaces
Jenkins975dfe12019-09-02 11:47:54 +0100536 - GCDepthConcatenateLayer
537 - NEWidthConcatenateLayer
538 - NEDepthConcatenateLayer
539 - CLWidthConcatenateLayer
540 - CLDepthConcatenateLayer
541 - CLGEMMInterleave4x4
542 - CLGEMMTranspose1xW
Jenkins4ba87db2019-05-23 17:11:51 +0100543 - Support different quantization info in CLConcatLayer.
544 - Add checks on different input/output quantization info were not supported.
545 - Tensors have different quantization information.
546 - Add FP16 support checks.
547 - Fix output quantization CLDeptwiseConv3x3 when activation is fused.
548 - New graph examples:
549 - graph_convolution
550 - graph_fully_connected
551 - graph_depthwise_convolution
552 - Deepspeech v0.4.1
553 - Add support for QASYMM8 in NEArithmeticSubtractionKernel.
554 - Add support for QASYMM8 in NEPixelWiseMultiplicationKernel.
555 - Add support for QASYMM8 NEDeconvolution.
556 - Add support for DequantizationLayer for NEON/CL.
557 - Add support for dilation in CLDepthwiseConvolution.
558 - Fuse offset contribution with the output stage when we use NEGEMMLowpMatrixMultiplyCore.
559 - Optimize CLDeconvolution.
560 - Add StackLayer to the graph API.
561 - Add support for "reflect" padding mode in NEPad.
562 - Winograd 7x7 NHWC on OpenCL.
563 - Rework CL ML layers to run exclusively on CL.
564 - Support different quantization info in PoolingLayer.
565 - Implement and test import memory interfaces.
566 - Added new tests and removed old ones.
567 - Various clang-tidy fixes.
568
Jenkins514be652019-02-28 12:25:18 +0000569v19.02 Public major release
570 - Various bug fixes.
571 - Various optimisations.
572 - New Neon kernels / functions:
573 - @ref NETileKernel / @ref NETile
574 - @ref NEFuseBatchNormalizationKernel / @ref NEFuseBatchNormalization
575 - @ref NEElementwiseOperationKernel
576 - @ref NEElementwiseMax
577 - @ref NEElementwiseMin
578 - @ref NEElementwiseSquaredDiff
579 - @ref NESelectKernel / @ref NESelect
580 - @ref NESplit
581 - @ref NESlice
582 - @ref NEUnstack
583 - @ref NEStridedSliceKernel / @ref NEStridedSlice
584 - @ref NEElementwiseUnaryKernel
585 - @ref NERsqrtLayer
586 - @ref NEExpLayer
587 - @ref NEReverseKernel / @ref NEReverse
588 - @ref NEArgMinMaxLayer
589 - @ref NEStackLayerKernel / @ref NEStackLayer
590 - @ref NERangeKernel / @ref NERange
591 - @ref NEPadLayer
592 - @ref NEMemsetKernel
593 - @ref NEGatherKernel / @ref NEGather
594 - @ref NEElementwiseComparison
595 - @ref NEElementwiseComparisonStatic
596 - @ref NEComparisonOperationKernel
597 - @ref NEElementwiseDivision
598 - New OpenCL kernels / functions:
599 - @ref CLSelectKernel / @ref CLSelect
600 - @ref CLTileKernel / @ref CLTile
601 - @ref CLComparisonKernel / @ref CLComparison
602 - @ref CLArgMinMaxLayer
603 - @ref CLElementwiseMax
604 - @ref CLElementwiseMin
605 - @ref CLElementwiseSquaredDiff
606 - @ref CLStackLayerKernel / @ref CLStackLayer
607 - @ref CLReverse / @ref CLReverseKernel
608 - @ref CLRsqrtLayer
609 - @ref CLExpLayer
610 - @ref CLElementWiseUnaryLayerKernel
611 - @ref CLGEMMReshapeLHSMatrixKernel
612 - @ref CLGEMMReshapeRHSMatrixKernel
613 - @ref CLGEMMMatrixMultiplyReshapedKernel
614 - @ref CLRangeKernel / @ref CLRange
615 - @ref CLUnstack
616 - @ref CLGatherKernel / @ref CLGather
617 - @ref CLGEMMLowpMatrixMultiplyReshapedKernel
618 - New CPP kernels / functions:
619 - @ref CPPDetectionOutputLayer
620 - @ref CPPTopKV / @ref CPPTopKVKernel
621 - Added new examples:
622 - graph_ssd_mobilenet.cpp
623 - graph_mobilenet_v2.cpp
624 - graph_resnet12.cpp
625 - graph_srcnn955.cpp
626 - graph_vgg_vdsr.cpp
627 - graph_inception_resnet_v1.cpp
628 - Add 4D tensors support to
629 - @ref NESoftmaxLayer
630 - Fused activation in @ref CLWinogradConvolutionLayer
631 - Extented @ref NEPermute to support more cases
632 - Added NEON/SVE GEMM Hybrid kernels
633 - Added u8 and s8 hybrid assembly kernels
634 - Introduced GEMM strategy name in NEGEMMAssemblyWrapper
635 - Improved @ref CLTuner
636 - Fused the bias addition within @ref CLGEMM
637 - Added support for QASYMM8 LOGISTIC activation in @ref NEActivationLayer
638 - Added NHWC data layout support to:
639 - @ref NEScale for F16
640 - @ref CLNormalizationLayer IN_MAP_2D for FP32/FP16
641 - @ref NEL2NormalizeLayer for FP32/FP16
642 - @ref NENormalizationLayer IN_MAP_2D for FP32/FP16
643 - @ref CLROIAlignLayer
644 - @ref CLGenerateProposalsLayer
645 - Added QASYMM8 support to the following kernels:
646 - @ref NEArithmeticAdditionKernel
647 - @ref NEScale
648 - Added new tests and improved validation and benchmarking suites.
649 - Deprecated functions/interfaces
650 - Usage of inner_border_right and inner_border_top has been deprecated in @ref CLDeconvolutionLayer and @ref NEDeconvolutionLayer
651
Jenkinsb9abeae2018-11-22 11:58:08 +0000652v18.11 Public major release
653 - Various bug fixes.
654 - Various optimisations.
655 - New Neon kernels / functions:
656 - @ref NEChannelShuffleLayer / @ref NEChannelShuffleLayerKernel
657 - @ref NEReduceMean
658 - @ref NEReorgLayer / @ref NEReorgLayerKernel
659 - @ref NEPriorBoxLayer / @ref NEPriorBoxLayerKernel
660 - @ref NEUpsampleLayer / @ref NEUpsampleLayerKernel
661 - @ref NEYOLOLayer / @ref NEYOLOLayerKernel
662 - New OpenCL kernels / functions:
663 - @ref CLBatchToSpaceLayer / @ref CLBatchToSpaceLayerKernel
664 - @ref CLBoundingBoxTransform / @ref CLBoundingBoxTransformKernel
665 - @ref CLComputeAllAnchorsKernel
666 - @ref CLGenerateProposalsLayer
667 - @ref CLNormalizePlanarYUVLayer / @ref CLNormalizePlanarYUVLayerKernel
668 - @ref CLReorgLayer / @ref CLReorgLayerKernel
669 - @ref CLSpaceToBatchLayer / @ref CLSpaceToBatchLayerKernel
670 - @ref CLPadLayer
671 - @ref CLReduceMean
672 - @ref CLPriorBoxLayer / @ref CLPriorBoxLayerKernel
673 - @ref CLROIAlignLayer / @ref CLROIAlignLayerKernel
674 - @ref CLSlice
675 - @ref CLSplit
676 - @ref CLStridedSlice / @ref CLStridedSliceKernel
677 - @ref CLUpsampleLayer / @ref CLUpsampleLayerKernel
678 - @ref CLYOLOLayer / @ref CLYOLOLayerKernel
679 - New CPP kernels / functions:
680 - @ref CPPBoxWithNonMaximaSuppressionLimit / @ref CPPBoxWithNonMaximaSuppressionLimitKernel
681 - Added the validate method in:
682 - @ref NEDepthConvertLayer
683 - @ref NEFloor / @ref CLFloor
684 - @ref NEGEMMMatrixAdditionKernel
685 - @ref NEReshapeLayer / @ref CLReshapeLayer
686 - @ref CLScale
687 - Added new examples:
688 - graph_shufflenet.cpp
689 - graph_yolov3.cpp
690 - Added documentation for add a new function or kernel.
691 - Improved doxygen documentation adding a list of the existing functions.
692 - Add 4D tensors support to
Jenkins975dfe12019-09-02 11:47:54 +0100693 - CLWidthConcatenateLayer
Jenkinsb9abeae2018-11-22 11:58:08 +0000694 - @ref CLFlattenLayer
695 - @ref CLSoftmaxLayer
696 - Add dot product support for @ref CLDepthwiseConvolutionLayer3x3NHWCKernel non-unit stride
697 - Add SVE support
698 - Fused batch normalization into convolution layer weights in @ref CLFuseBatchNormalization
699 - Fuses activation in @ref CLDepthwiseConvolutionLayer3x3NCHWKernel, @ref CLDepthwiseConvolutionLayer3x3NHWCKernel and @ref NEGEMMConvolutionLayer
700 - Added NHWC data layout support to:
701 - @ref CLChannelShuffleLayer
702 - @ref CLDeconvolutionLayer
703 - @ref CLL2NormalizeLayer
704 - Added QASYMM8 support to the following kernels:
705 - @ref CLScaleKernel
706 - @ref NEDepthwiseConvolutionLayer3x3Kernel
707 - @ref CLPixelWiseMultiplicationKernel
708 - Added FP16 support to the following kernels:
709 - @ref CLDepthwiseConvolutionLayer3x3NHWCKernel
710 - @ref NEDepthwiseConvolutionLayer3x3Kernel
711 - @ref CLNormalizePlanarYUVLayerKernel
712 - @ref CLWinogradConvolutionLayer (5x5 kernel)
713 - More tests added to both validation and benchmarking suites.
714
Jenkins52ba29e2018-08-29 15:32:11 +0000715v18.08 Public major release
716 - Various bug fixes.
717 - Various optimisations.
718 - Updated recommended NDK version to r17b.
719 - Removed support for QS8/QS16 data types.
720 - Added support for grouped convolution in @ref CLConvolutionLayer.
721 - Added NHWC data layout support to:
Jenkins975dfe12019-09-02 11:47:54 +0100722 - NEDepthConcatenateLayer / CLDepthConcatenateLayer
Jenkins52ba29e2018-08-29 15:32:11 +0000723 - @ref NEWinogradConvolutionLayer / @ref CLWinogradConvolutionLayer
724 - @ref CLDepthwiseConvolutionLayer
725 - @ref CLDirectConvolutionLayer
726 - @ref CLConvolutionLayer
727 - @ref CLScale
728 - @ref CLIm2ColKernel
729 - New Neon kernels / functions:
730 - @ref NERNNLayer
731 - New OpenCL kernels / functions:
732 - @ref CLArithmeticDivision
733 - Introduced prepare() stage support in the graph API for GLES.
734 - Added support for memory reusage when trying to allocate smaller CLTensors.
735 - Enabled NHWC execution on graph examples.
736 - Added JPEG accessor for validation purposes.
737 - Added validate methods to some kernels / functions.
738
739v18.05 Public major release
Jenkinsb3a371b2018-05-23 11:36:53 +0100740 - Various bug fixes.
741 - Various optimisations.
742 - Major redesign in the interface for the neon kernels implemented in assembly.
743 - Removed arm_compute::NEGEMMLowpAArch64A53Kernel / arm_compute::NEGEMMLowpAArch64Kernel / arm_compute::NEGEMMLowpAArch64V8P4Kernel / arm_compute::NEGEMMInterleavedBlockedKernel / arm_compute::NEGEMMLowpAssemblyMatrixMultiplyCore / arm_compute::NEHGEMMAArch64FP16Kernel
744 - Added NEGEMMAssemblyWrapper and AssemblyKernelGlue which are used to execute assembly kernels in neon functions.
745 - Minor changes to the CPUInfo type to make it compatible with the new assembly gemm interface.
746 - Moved neon assembly kernels to the folder src/core/NEON/kernels/arm_gemm.
747 - Improved doxygen documentation.
748 - Improved memory management for layer's transitions.
749 - Added support for NHWC data layout in tensors.
750 - Added NHWC data layout support to:
751 - @ref NEGEMMConvolutionLayer
752 - @ref NEDirectConvolutionLayer
753 - @ref NEPoolingLayer / @ref CLPoolingLayer
754 - @ref NEBatchNormalizationLayer / @ref CLBatchNormalizationLayer
755 - @ref NEDepthwiseConvolutionLayer
756 - @ref NEScale
757 - @ref NEIm2Col
758 - Added support for dilated convolutions in @ref NEConvolutionLayer and @ref CLConvolutionLayer.
759 - New OpenCL kernels / functions:
760 - @ref CLChannelShuffleLayer / @ref CLChannelShuffleLayerKernel
761 - @ref CLConvertFullyConnectedWeightsKernel / @ref CLConvertFullyConnectedWeights
762 - @ref CLCopy / @ref CLCopyKernel
763 - @ref CLLSTMLayer
764 - @ref CLRNNLayer
Jenkins975dfe12019-09-02 11:47:54 +0100765 - CLWidthConcatenateLayer / @ref CLWidthConcatenateLayerKernel
Jenkinsb3a371b2018-05-23 11:36:53 +0100766 - @ref CLWinogradFilterTransformKernel / @ref CLWinogradInputTransformKernel / @ref CLWinogradConvolutionLayer
767 - @ref CLWinogradInputTransformKernel / @ref CLWinogradInputTransform
768 - New Neon kernels / functions:
Jenkinsb3a371b2018-05-23 11:36:53 +0100769 - @ref NEConvertFullyConnectedWeightsKernel / @ref NEConvertFullyConnectedWeights.
770 - Created the validate method in @ref CLDepthwiseConvolutionLayer.
771 - Beta and gamma are no longer mandatory arguments in @ref NEBatchNormalizationLayer and @ref CLBatchNormalizationLayer.
772 - Added depth multiplier support in @ref NEDepthwiseConvolutionLayer and @ref CLDepthwiseConvolutionLayer.
773 - Added broadcast multiply support in @ref NEPixelWiseMultiplication / @ref NEPixelWiseMultiplicationKernel.
774 - Port mobilenet example to NHWC data layout.
775 - Enabled Winograd method in @ref CLConvolutionLayer.
776 - Renamed NEWinogradLayer to @ref NEWinogradConvolutionLayer.
777 - Updated @ref NEWinogradConvolutionLayer to use highly optimised assembly kernels in src/core/NEON/kernels/arm_gemm.
778 - Added memory manager support in GLES functions.
779 - Major refactoring of the graph API.
780 - Added GLES backend in the graph API.
781 - Added support for the memory manager in the graph API.
782 - Enabled Winograd Convolution method in the graph API.
783 - Added support for grouped convolutions in the graph API.
784 - Replaced NEDeconvolutionLayerUpsampleKernel with @ref NEScaleKernel in @ref NEDeconvolutionLayer.
785 - Added fast maths flag in @ref CLConvolutionLayer.
786 - Added new tests and benchmarks in validation and benchmark frameworks
787 - Merge Activation layer with Convolution Layer (NEON. CL, GLES)
788 - Added support to OpenCL 2.0 SVM
789 - Added support to import memory in OpenCL tensors.
790 - Added the prepare() method to perform any one off pre-processing before running the function.
791 - Added new examples:
792 - graph_inception_v4.cpp
793 - graph_resnext50.cpp
794 - Added memory measurement instrument for CL.
795
Jenkinsc3f34a42018-03-02 12:38:09 +0000796v18.03 Public maintenance release
797 - Various bug fixes.
798 - Fixed bug in @ref NEActivationLayer
799 - Fix in @ref CLTuner when using batches.
800 - Updated recommended NDK version to r16b (And fixed warnings).
801 - Fixed bug in validation code.
802 - Added Inception v4 graph example.
Jenkinsb3a371b2018-05-23 11:36:53 +0100803 - Renamed NEWinogradLayer.cpp to @ref NEWinogradConvolutionLayer
Jenkinsc3f34a42018-03-02 12:38:09 +0000804
Anthony Barbier06ea0482018-02-22 15:45:35 +0000805v18.02 Public major release
806 - Various NEON / OpenCL / GLES optimisations.
807 - Various bug fixes.
808 - Changed default number of threads on big LITTLE systems.
809 - Refactored examples and added:
810 - graph_mobilenet_qassym8
811 - graph_resnet
812 - graph_squeezenet_v1_1
Jenkinsc3f34a42018-03-02 12:38:09 +0000813 - Renamed @ref CLConvolutionLayer into @ref CLGEMMConvolutionLayer and created a new @ref CLConvolutionLayer to select the fastest convolution method.
814 - Renamed @ref NEConvolutionLayer into @ref NEGEMMConvolutionLayer and created a new @ref NEConvolutionLayer to select the fastest convolution method.
Anthony Barbier06ea0482018-02-22 15:45:35 +0000815 - Added in place support to:
Jenkinsc3f34a42018-03-02 12:38:09 +0000816 - @ref CLActivationLayer
817 - @ref CLBatchNormalizationLayer
Anthony Barbier06ea0482018-02-22 15:45:35 +0000818 - Added QASYMM8 support to:
Jenkinsc3f34a42018-03-02 12:38:09 +0000819 - @ref CLActivationLayer
820 - @ref CLDepthwiseConvolutionLayer
821 - @ref NEDepthwiseConvolutionLayer
822 - @ref NESoftmaxLayer
Anthony Barbier06ea0482018-02-22 15:45:35 +0000823 - Added FP16 support to:
Jenkins6a7771e2020-05-28 11:28:36 +0100824 - CLDepthwiseConvolutionLayer3x3
Jenkinsc3f34a42018-03-02 12:38:09 +0000825 - @ref CLDepthwiseConvolutionLayer
826 - Added broadcasting support to @ref NEArithmeticAddition / @ref CLArithmeticAddition / @ref CLPixelWiseMultiplication
827 - Added fused batched normalization and activation to @ref CLBatchNormalizationLayer and @ref NEBatchNormalizationLayer
828 - Added support for non-square pooling to @ref NEPoolingLayer and @ref CLPoolingLayer
Anthony Barbier06ea0482018-02-22 15:45:35 +0000829 - New OpenCL kernels / functions:
Jenkins0e205f72019-11-28 16:53:35 +0000830 - CLDirectConvolutionLayerOutputStageKernel
Anthony Barbier06ea0482018-02-22 15:45:35 +0000831 - New NEON kernels / functions
832 - Added name() method to all kernels.
833 - Added support for Winograd 5x5.
Jenkinsc3f34a42018-03-02 12:38:09 +0000834 - @ref NEPermuteKernel / @ref NEPermute
Jenkinsb3a371b2018-05-23 11:36:53 +0100835 - @ref NEWinogradLayerTransformInputKernel / NEWinogradLayer
836 - @ref NEWinogradLayerTransformOutputKernel / NEWinogradLayer
837 - @ref NEWinogradLayerTransformWeightsKernel / NEWinogradLayer
Jenkins52ba29e2018-08-29 15:32:11 +0000838 - Renamed NEWinogradLayerKernel into NEWinogradLayerBatchedGEMMKernel
Anthony Barbier06ea0482018-02-22 15:45:35 +0000839 - New GLES kernels / functions:
Jenkinsc3f34a42018-03-02 12:38:09 +0000840 - @ref GCTensorShiftKernel / @ref GCTensorShift
Anthony Barbier06ea0482018-02-22 15:45:35 +0000841
Anthony Barbierf45d5a92018-01-24 16:23:15 +0000842v18.01 Public maintenance release
843 - Various bug fixes
844 - Added some of the missing validate() methods
Jenkinsc3f34a42018-03-02 12:38:09 +0000845 - Added @ref CLDeconvolutionLayerUpsampleKernel / @ref CLDeconvolutionLayer @ref CLDeconvolutionLayerUpsample
846 - Added @ref CLPermuteKernel / @ref CLPermute
Anthony Barbierf45d5a92018-01-24 16:23:15 +0000847 - Added method to clean the programs cache in the CL Kernel library.
Jenkinsc3f34a42018-03-02 12:38:09 +0000848 - Added @ref GCArithmeticAdditionKernel / @ref GCArithmeticAddition
849 - Added @ref GCDepthwiseConvolutionLayer3x3Kernel / @ref GCDepthwiseConvolutionLayer3x3
850 - Added @ref GCNormalizePlanarYUVLayerKernel / @ref GCNormalizePlanarYUVLayer
851 - Added @ref GCScaleKernel / @ref GCScale
852 - Added @ref GCWeightsReshapeKernel / @ref GCConvolutionLayer
Anthony Barbierf45d5a92018-01-24 16:23:15 +0000853 - Added FP16 support to the following GLES compute kernels:
Jenkinsc3f34a42018-03-02 12:38:09 +0000854 - @ref GCCol2ImKernel
855 - @ref GCGEMMInterleave4x4Kernel
856 - @ref GCGEMMTranspose1xWKernel
857 - @ref GCIm2ColKernel
858 - Refactored NEON Winograd (NEWinogradLayerKernel)
859 - Added @ref NEDirectConvolutionLayerOutputStageKernel
Anthony Barbierf45d5a92018-01-24 16:23:15 +0000860 - Added QASYMM8 support to the following NEON kernels:
Jenkinsc3f34a42018-03-02 12:38:09 +0000861 - @ref NEDepthwiseConvolutionLayer3x3Kernel
862 - @ref NEFillBorderKernel
863 - @ref NEPoolingLayerKernel
Anthony Barbierf45d5a92018-01-24 16:23:15 +0000864 - Added new examples:
865 - graph_cl_mobilenet_qasymm8.cpp
866 - graph_inception_v3.cpp
867 - gc_dc.cpp
868 - More tests added to both validation and benchmarking suites.
869
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000870v17.12 Public major release
871 - Most machine learning functions on OpenCL support the new data type QASYMM8
872 - Introduced logging interface
873 - Introduced opencl timer
874 - Reworked GEMMLowp interface
875 - Added new NEON assembly kernels for GEMMLowp, SGEMM and HGEMM
876 - Added validation method for most Machine Learning kernels / functions
877 - Added new graph examples such as googlenet, mobilenet, squeezenet, vgg16 and vgg19
878 - Added sgemm example for OpenCL
879 - Added absolute difference example for GLES compute
880 - Added new tests and benchmarks in validation and benchmark frameworks
881 - Added new kernels / functions for GLES compute
882
883 - New OpenGL ES kernels / functions
Jenkinsc3f34a42018-03-02 12:38:09 +0000884 - @ref GCAbsoluteDifferenceKernel / @ref GCAbsoluteDifference
885 - @ref GCActivationLayerKernel / @ref GCActivationLayer
886 - @ref GCBatchNormalizationLayerKernel / @ref GCBatchNormalizationLayer
887 - @ref GCCol2ImKernel
Jenkins975dfe12019-09-02 11:47:54 +0100888 - @ref GCDepthConcatenateLayerKernel / GCDepthConcatenateLayer
Jenkinsc3f34a42018-03-02 12:38:09 +0000889 - @ref GCDirectConvolutionLayerKernel / @ref GCDirectConvolutionLayer
890 - @ref GCDropoutLayerKernel / @ref GCDropoutLayer
891 - @ref GCFillBorderKernel / @ref GCFillBorder
892 - @ref GCGEMMInterleave4x4Kernel / @ref GCGEMMInterleave4x4
893 - @ref GCGEMMMatrixAccumulateBiasesKernel / @ref GCGEMMMatrixAdditionKernel / @ref GCGEMMMatrixMultiplyKernel / @ref GCGEMM
894 - @ref GCGEMMTranspose1xWKernel / @ref GCGEMMTranspose1xW
895 - @ref GCIm2ColKernel
896 - @ref GCNormalizationLayerKernel / @ref GCNormalizationLayer
897 - @ref GCPixelWiseMultiplicationKernel / @ref GCPixelWiseMultiplication
898 - @ref GCPoolingLayerKernel / @ref GCPoolingLayer
899 - @ref GCLogits1DMaxKernel / @ref GCLogits1DShiftExpSumKernel / @ref GCLogits1DNormKernel / @ref GCSoftmaxLayer
900 - @ref GCTransposeKernel / @ref GCTranspose
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000901
902 - New NEON kernels / functions
Jenkinsb3a371b2018-05-23 11:36:53 +0100903 - arm_compute::NEGEMMLowpAArch64A53Kernel / arm_compute::NEGEMMLowpAArch64Kernel / arm_compute::NEGEMMLowpAArch64V8P4Kernel / arm_compute::NEGEMMInterleavedBlockedKernel / arm_compute::NEGEMMLowpAssemblyMatrixMultiplyCore
904 - arm_compute::NEHGEMMAArch64FP16Kernel
Jenkins0e205f72019-11-28 16:53:35 +0000905 - @ref NEDepthwiseConvolutionLayer3x3Kernel / NEDepthwiseIm2ColKernel / @ref NEGEMMMatrixVectorMultiplyKernel / NEDepthwiseVectorToTensorKernel / @ref NEDepthwiseConvolutionLayer
Jenkinsc3f34a42018-03-02 12:38:09 +0000906 - @ref NEGEMMLowpOffsetContributionKernel / @ref NEGEMMLowpMatrixAReductionKernel / @ref NEGEMMLowpMatrixBReductionKernel / @ref NEGEMMLowpMatrixMultiplyCore
907 - @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel / @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
Jenkinsb3a371b2018-05-23 11:36:53 +0100908 - NEWinogradLayer / NEWinogradLayerKernel
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000909
910 - New OpenCL kernels / functions
Jenkinsc3f34a42018-03-02 12:38:09 +0000911 - @ref CLGEMMLowpOffsetContributionKernel / @ref CLGEMMLowpMatrixAReductionKernel / @ref CLGEMMLowpMatrixBReductionKernel / @ref CLGEMMLowpMatrixMultiplyCore
912 - @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel / @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000913
914 - New graph nodes for NEON and OpenCL
Jenkinsb3a371b2018-05-23 11:36:53 +0100915 - graph::BranchLayer
916 - graph::DepthConvertLayer
917 - graph::DepthwiseConvolutionLayer
918 - graph::DequantizationLayer
919 - graph::FlattenLayer
920 - graph::QuantizationLayer
921 - graph::ReshapeLayer
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000922
Kaizenbf8b01d2017-10-12 14:26:51 +0100923v17.10 Public maintenance release
924 - Bug fixes:
925 - Check the maximum local workgroup size supported by OpenCL devices
926 - Minor documentation updates (Fixed instructions to build the examples)
Jenkinsc3f34a42018-03-02 12:38:09 +0000927 - Introduced a graph::GraphContext
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000928 - Added a few new Graph nodes, support for branches and grouping.
Kaizenbf8b01d2017-10-12 14:26:51 +0100929 - Automatically enable cl_printf in debug builds
930 - Fixed bare metal builds for armv7a
931 - Added AlexNet and cartoon effect examples
932 - Fixed library builds: libraries are no longer built as supersets of each other.(It means application using the Runtime part of the library now need to link against both libarm_compute_core and libarm_compute)
933
Kaizen8938bd32017-09-28 14:38:23 +0100934v17.09 Public major release
935 - Experimental Graph support: initial implementation of a simple stream API to easily chain machine learning layers.
Jenkinsc3f34a42018-03-02 12:38:09 +0000936 - Memory Manager (@ref BlobLifetimeManager, @ref BlobMemoryPool, @ref ILifetimeManager, @ref IMemoryGroup, @ref IMemoryManager, @ref IMemoryPool, @ref IPoolManager, @ref MemoryManagerOnDemand, @ref PoolManager)
Kaizen8938bd32017-09-28 14:38:23 +0100937 - New validation and benchmark frameworks (Boost and Google frameworks replaced by homemade framework).
938 - Most machine learning functions support both fixed point 8 and 16 bit (QS8, QS16) for both NEON and OpenCL.
939 - New NEON kernels / functions:
Jenkinsb3a371b2018-05-23 11:36:53 +0100940 - arm_compute::NEGEMMAssemblyBaseKernel arm_compute::NEGEMMAArch64Kernel
Jenkinsc3f34a42018-03-02 12:38:09 +0000941 - @ref NEDequantizationLayerKernel / @ref NEDequantizationLayer
942 - @ref NEFloorKernel / @ref NEFloor
943 - @ref NEL2NormalizeLayerKernel / @ref NEL2NormalizeLayer
944 - @ref NEQuantizationLayerKernel @ref NEMinMaxLayerKernel / @ref NEQuantizationLayer
945 - @ref NEROIPoolingLayerKernel / @ref NEROIPoolingLayer
946 - @ref NEReductionOperationKernel / @ref NEReductionOperation
947 - @ref NEReshapeLayerKernel / @ref NEReshapeLayer
Kaizen8938bd32017-09-28 14:38:23 +0100948
949 - New OpenCL kernels / functions:
Jenkins6a7771e2020-05-28 11:28:36 +0100950 - @ref CLDepthwiseConvolutionLayer3x3NCHWKernel @ref CLDepthwiseConvolutionLayer3x3NHWCKernel CLDepthwiseIm2ColKernel CLDepthwiseVectorToTensorKernel CLDepthwiseWeightsReshapeKernel / CLDepthwiseConvolutionLayer3x3 @ref CLDepthwiseConvolutionLayer CLDepthwiseSeparableConvolutionLayer
Jenkinsc3f34a42018-03-02 12:38:09 +0000951 - @ref CLDequantizationLayerKernel / @ref CLDequantizationLayer
952 - @ref CLDirectConvolutionLayerKernel / @ref CLDirectConvolutionLayer
953 - @ref CLFlattenLayer
954 - @ref CLFloorKernel / @ref CLFloor
Jenkins975dfe12019-09-02 11:47:54 +0100955 - CLGEMMTranspose1xW
Jenkinsc3f34a42018-03-02 12:38:09 +0000956 - @ref CLGEMMMatrixVectorMultiplyKernel
957 - @ref CLL2NormalizeLayerKernel / @ref CLL2NormalizeLayer
958 - @ref CLQuantizationLayerKernel @ref CLMinMaxLayerKernel / @ref CLQuantizationLayer
959 - @ref CLROIPoolingLayerKernel / @ref CLROIPoolingLayer
960 - @ref CLReductionOperationKernel / @ref CLReductionOperation
961 - @ref CLReshapeLayerKernel / @ref CLReshapeLayer
Kaizen8938bd32017-09-28 14:38:23 +0100962
Anthony Barbierdbdab852017-06-23 15:42:00 +0100963v17.06 Public major release
964 - Various bug fixes
965 - Added support for fixed point 8 bit (QS8) to the various NEON machine learning kernels.
966 - Added unit tests and benchmarks (AlexNet, LeNet)
967 - Added support for sub tensors.
968 - Added infrastructure to provide GPU specific optimisation for some OpenCL kernels.
Jenkinsc3f34a42018-03-02 12:38:09 +0000969 - Added @ref OMPScheduler (OpenMP) scheduler for NEON
970 - Added @ref SingleThreadScheduler scheduler for NEON (For bare metal)
971 - User can specify his own scheduler by implementing the @ref IScheduler interface.
Anthony Barbierdbdab852017-06-23 15:42:00 +0100972 - New OpenCL kernels / functions:
Jenkinsc3f34a42018-03-02 12:38:09 +0000973 - @ref CLBatchNormalizationLayerKernel / @ref CLBatchNormalizationLayer
Jenkins975dfe12019-09-02 11:47:54 +0100974 - @ref CLDepthConcatenateLayerKernel / CLDepthConcatenateLayer
Jenkinsc3f34a42018-03-02 12:38:09 +0000975 - @ref CLHOGOrientationBinningKernel @ref CLHOGBlockNormalizationKernel, @ref CLHOGDetectorKernel / @ref CLHOGDescriptor @ref CLHOGDetector @ref CLHOGGradient @ref CLHOGMultiDetection
976 - @ref CLLocallyConnectedMatrixMultiplyKernel / @ref CLLocallyConnectedLayer
977 - @ref CLWeightsReshapeKernel / @ref CLConvolutionLayerReshapeWeights
Anthony Barbierdbdab852017-06-23 15:42:00 +0100978 - New C++ kernels:
Jenkinsc3f34a42018-03-02 12:38:09 +0000979 - @ref CPPDetectionWindowNonMaximaSuppressionKernel
Anthony Barbierdbdab852017-06-23 15:42:00 +0100980 - New NEON kernels / functions:
Jenkinsc3f34a42018-03-02 12:38:09 +0000981 - @ref NEBatchNormalizationLayerKernel / @ref NEBatchNormalizationLayer
Jenkins975dfe12019-09-02 11:47:54 +0100982 - @ref NEDepthConcatenateLayerKernel / NEDepthConcatenateLayer
Jenkinsc3f34a42018-03-02 12:38:09 +0000983 - @ref NEDirectConvolutionLayerKernel / @ref NEDirectConvolutionLayer
984 - @ref NELocallyConnectedMatrixMultiplyKernel / @ref NELocallyConnectedLayer
985 - @ref NEWeightsReshapeKernel / @ref NEConvolutionLayerReshapeWeights
Anthony Barbierdbdab852017-06-23 15:42:00 +0100986
987v17.05 Public bug fixes release
988 - Various bug fixes
989 - Remaining of the functions ported to use accurate padding.
990 - Library does not link against OpenCL anymore (It uses dlopen / dlsym at runtime instead to determine whether or not OpenCL is available).
991 - Added "free" method to allocator.
992 - Minimum version of g++ required for armv7 Linux changed from 4.8 to 4.9
993
994v17.04 Public bug fixes release
995
996 The following functions have been ported to use the new accurate padding:
Jenkinsc3f34a42018-03-02 12:38:09 +0000997 - @ref CLColorConvertKernel
998 - @ref CLEdgeNonMaxSuppressionKernel
999 - @ref CLEdgeTraceKernel
1000 - @ref CLGaussianPyramidHorKernel
1001 - @ref CLGaussianPyramidVertKernel
1002 - @ref CLGradientKernel
1003 - @ref NEChannelCombineKernel
1004 - @ref NEFillArrayKernel
1005 - @ref NEGaussianPyramidHorKernel
1006 - @ref NEGaussianPyramidVertKernel
Jenkinsb9abeae2018-11-22 11:58:08 +00001007 - NEHarrisScoreFP16Kernel
Jenkinsc3f34a42018-03-02 12:38:09 +00001008 - @ref NEHarrisScoreKernel
1009 - @ref NEHOGDetectorKernel
1010 - @ref NELogits1DMaxKernel
1011 - NELogits1DShiftExpSumKernel
1012 - NELogits1DNormKernel
1013 - @ref NENonMaximaSuppression3x3FP16Kernel
1014 - @ref NENonMaximaSuppression3x3Kernel
Anthony Barbierdbdab852017-06-23 15:42:00 +01001015
Anthony Barbierdbdab852017-06-23 15:42:00 +01001016v17.03.1 First Major public release of the sources
1017 - Renamed the library to arm_compute
1018 - New CPP target introduced for C++ kernels shared between NEON and CL functions.
1019 - New padding calculation interface introduced and ported most kernels / functions to use it.
1020 - New OpenCL kernels / functions:
Jenkins6a7771e2020-05-28 11:28:36 +01001021 - CLGEMMLowpMatrixMultiplyKernel / CLGEMMLowp
Anthony Barbierdbdab852017-06-23 15:42:00 +01001022 - New NEON kernels / functions:
Jenkinsc3f34a42018-03-02 12:38:09 +00001023 - @ref NENormalizationLayerKernel / @ref NENormalizationLayer
1024 - @ref NETransposeKernel / @ref NETranspose
1025 - @ref NELogits1DMaxKernel, NELogits1DShiftExpSumKernel, NELogits1DNormKernel / @ref NESoftmaxLayer
1026 - @ref NEIm2ColKernel, @ref NECol2ImKernel, NEConvolutionLayerWeightsReshapeKernel / @ref NEConvolutionLayer
1027 - @ref NEGEMMMatrixAccumulateBiasesKernel / @ref NEFullyConnectedLayer
1028 - @ref NEGEMMLowpMatrixMultiplyKernel / NEGEMMLowp
Anthony Barbierdbdab852017-06-23 15:42:00 +01001029
1030v17.03 Sources preview
1031 - New OpenCL kernels / functions:
Jenkinsc3f34a42018-03-02 12:38:09 +00001032 - @ref CLGradientKernel, @ref CLEdgeNonMaxSuppressionKernel, @ref CLEdgeTraceKernel / @ref CLCannyEdge
Jenkins0e205f72019-11-28 16:53:35 +00001033 - GEMM refactoring + FP16 support: CLGEMMInterleave4x4Kernel, CLGEMMTranspose1xWKernel, @ref CLGEMMMatrixMultiplyKernel, CLGEMMMatrixAdditionKernel / @ref CLGEMM
Jenkinsc3f34a42018-03-02 12:38:09 +00001034 - @ref CLGEMMMatrixAccumulateBiasesKernel / @ref CLFullyConnectedLayer
1035 - @ref CLTransposeKernel / @ref CLTranspose
1036 - @ref CLLKTrackerInitKernel, @ref CLLKTrackerStage0Kernel, @ref CLLKTrackerStage1Kernel, @ref CLLKTrackerFinalizeKernel / @ref CLOpticalFlow
1037 - @ref CLNormalizationLayerKernel / @ref CLNormalizationLayer
1038 - @ref CLLaplacianPyramid, @ref CLLaplacianReconstruct
Anthony Barbierdbdab852017-06-23 15:42:00 +01001039 - New NEON kernels / functions:
Jenkinsc3f34a42018-03-02 12:38:09 +00001040 - @ref NEActivationLayerKernel / @ref NEActivationLayer
1041 - GEMM refactoring + FP16 support (Requires armv8.2 CPU): @ref NEGEMMInterleave4x4Kernel, @ref NEGEMMTranspose1xWKernel, @ref NEGEMMMatrixMultiplyKernel, @ref NEGEMMMatrixAdditionKernel / @ref NEGEMM
1042 - @ref NEPoolingLayerKernel / @ref NEPoolingLayer
Anthony Barbierdbdab852017-06-23 15:42:00 +01001043
1044v17.02.1 Sources preview
1045 - New OpenCL kernels / functions:
Jenkinsc3f34a42018-03-02 12:38:09 +00001046 - @ref CLLogits1DMaxKernel, @ref CLLogits1DShiftExpSumKernel, @ref CLLogits1DNormKernel / @ref CLSoftmaxLayer
1047 - @ref CLPoolingLayerKernel / @ref CLPoolingLayer
1048 - @ref CLIm2ColKernel, @ref CLCol2ImKernel, CLConvolutionLayerWeightsReshapeKernel / @ref CLConvolutionLayer
1049 - @ref CLRemapKernel / @ref CLRemap
1050 - @ref CLGaussianPyramidHorKernel, @ref CLGaussianPyramidVertKernel / @ref CLGaussianPyramid, @ref CLGaussianPyramidHalf, @ref CLGaussianPyramidOrb
1051 - @ref CLMinMaxKernel, @ref CLMinMaxLocationKernel / @ref CLMinMaxLocation
1052 - @ref CLNonLinearFilterKernel / @ref CLNonLinearFilter
Anthony Barbierdbdab852017-06-23 15:42:00 +01001053 - New NEON FP16 kernels (Requires armv8.2 CPU)
Jenkinsc3f34a42018-03-02 12:38:09 +00001054 - @ref NEAccumulateWeightedFP16Kernel
1055 - @ref NEBox3x3FP16Kernel
1056 - @ref NENonMaximaSuppression3x3FP16Kernel
Anthony Barbierdbdab852017-06-23 15:42:00 +01001057
1058v17.02 Sources preview
1059 - New OpenCL kernels / functions:
Jenkinsc3f34a42018-03-02 12:38:09 +00001060 - @ref CLActivationLayerKernel / @ref CLActivationLayer
1061 - @ref CLChannelCombineKernel / @ref CLChannelCombine
1062 - @ref CLDerivativeKernel / @ref CLChannelExtract
1063 - @ref CLFastCornersKernel / @ref CLFastCorners
1064 - @ref CLMeanStdDevKernel / @ref CLMeanStdDev
Anthony Barbierdbdab852017-06-23 15:42:00 +01001065 - New NEON kernels / functions:
Jenkinsc3f34a42018-03-02 12:38:09 +00001066 - HOG / SVM: @ref NEHOGOrientationBinningKernel, @ref NEHOGBlockNormalizationKernel, @ref NEHOGDetectorKernel, NEHOGNonMaximaSuppressionKernel / @ref NEHOGDescriptor, @ref NEHOGDetector, @ref NEHOGGradient, @ref NEHOGMultiDetection
1067 - @ref NENonLinearFilterKernel / @ref NENonLinearFilter
Anthony Barbierdbdab852017-06-23 15:42:00 +01001068 - Introduced a CLScheduler to manage the default context and command queue used by the runtime library and create synchronisation events.
1069 - Switched all the kernels / functions to use tensors instead of images.
1070 - Updated documentation to include instructions to build the library from sources.
1071
1072v16.12 Binary preview release
1073 - Original release
1074
1075@section S3_how_to_build How to build the library and the examples
1076
1077@subsection S3_1_build_options Build options
1078
1079scons 2.3 or above is required to build the library.
1080To see the build options available simply run ```scons -h```:
1081
1082 debug: Debug (yes|no)
1083 default: False
1084 actual: False
1085
1086 asserts: Enable asserts (this flag is forced to 1 for debug=1) (yes|no)
1087 default: False
1088 actual: False
1089
1090 arch: Target Architecture (armv7a|arm64-v8a|arm64-v8.2-a|x86_32|x86_64)
1091 default: armv7a
1092 actual: armv7a
1093
1094 os: Target OS (linux|android|bare_metal)
1095 default: linux
1096 actual: linux
1097
Anthony Barbier06ea0482018-02-22 15:45:35 +00001098 build: Build type (native|cross_compile|embed_only)
Anthony Barbierdbdab852017-06-23 15:42:00 +01001099 default: cross_compile
1100 actual: cross_compile
1101
1102 examples: Build example programs (yes|no)
1103 default: True
1104 actual: True
1105
1106 Werror: Enable/disable the -Werror compilation flag (yes|no)
1107 default: True
1108 actual: True
1109
1110 opencl: Enable OpenCL support (yes|no)
1111 default: True
1112 actual: True
1113
1114 neon: Enable Neon support (yes|no)
1115 default: False
1116 actual: False
1117
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001118 gles_compute: Enable OpenGL ES Compute Shader support (yes|no)
1119 default: False
1120 actual: False
1121
1122 embed_kernels: Embed OpenCL kernels and OpenGL ES compute shader in library binary (yes|no)
Anthony Barbierf45d5a92018-01-24 16:23:15 +00001123 default: True
1124 actual: True
Anthony Barbierdbdab852017-06-23 15:42:00 +01001125
1126 set_soname: Set the library's soname and shlibversion (requires SCons 2.4 or above) (yes|no)
1127 default: False
1128 actual: False
1129
1130 openmp: Enable OpenMP backend (yes|no)
1131 default: False
1132 actual: False
1133
1134 cppthreads: Enable C++11 threads backend (yes|no)
1135 default: True
1136 actual: True
1137
1138 build_dir: Specify sub-folder for the build ( /path/to/build_dir )
1139 default: .
1140 actual: .
1141
1142 extra_cxx_flags: Extra CXX flags to be appended to the build command
1143 default:
1144 actual:
1145
1146 pmu: Enable PMU counters (yes|no)
1147 default: False
1148 actual: False
1149
Kaizen8938bd32017-09-28 14:38:23 +01001150 mali: Enable Mali hardware counters (yes|no)
1151 default: False
1152 actual: False
1153
Anthony Barbierdbdab852017-06-23 15:42:00 +01001154 validation_tests: Build validation test programs (yes|no)
1155 default: False
1156 actual: False
1157
1158 benchmark_tests: Build benchmark test programs (yes|no)
1159 default: False
1160 actual: False
1161
1162@b debug / @b asserts:
1163 - With debug=1 asserts are enabled, and the library is built with symbols and no optimisations enabled.
1164 - With debug=0 and asserts=1: Optimisations are enabled and symbols are removed, however all the asserts are still present (This is about 20% slower than the release build)
1165 - With debug=0 and asserts=0: All optimisations are enable and no validation is performed, if the application misuses the library it is likely to result in a crash. (Only use this mode once you are sure your application is working as expected).
1166
1167@b arch: The x86_32 and x86_64 targets can only be used with neon=0 and opencl=1.
1168
1169@b os: Choose the operating system you are targeting: Linux, Android or bare metal.
1170@note bare metal can only be used for NEON (not OpenCL), only static libraries get built and NEON's multi-threading support is disabled.
1171
1172@b build: you can either build directly on your device (native) or cross compile from your desktop machine (cross-compile). In both cases make sure the compiler is available in your path.
1173
1174@note If you want to natively compile for 32bit on a 64bit ARM device running a 64bit OS then you will have to use cross-compile too.
1175
Anthony Barbier06ea0482018-02-22 15:45:35 +00001176There is also an 'embed_only' option which will generate all the .embed files for the OpenCL kernels and / or OpenGLES compute shaders. This might be useful if using a different build system to compile the library.
1177
Anthony Barbierdbdab852017-06-23 15:42:00 +01001178@b Werror: If you are compiling using the same toolchains as the ones used in this guide then there shouldn't be any warning and therefore you should be able to keep Werror=1. If with a different compiler version the library fails to build because of warnings interpreted as errors then, if you are sure the warnings are not important, you might want to try to build with Werror=0 (But please do report the issue either on Github or by an email to developer@arm.com so that the issue can be addressed).
1179
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001180@b opencl / @b neon / @b gles_compute: Choose which SIMD technology you want to target. (NEON for ARM Cortex-A CPUs or OpenCL / GLES_COMPUTE for ARM Mali GPUs)
Anthony Barbierdbdab852017-06-23 15:42:00 +01001181
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001182@b embed_kernels: For OpenCL / GLES_COMPUTE only: set embed_kernels=1 if you want the OpenCL / GLES_COMPUTE kernels to be built in the library's binaries instead of being read from separate ".cl" / ".cs" files. If embed_kernels is set to 0 then the application can set the path to the folder containing the OpenCL / GLES_COMPUTE kernel files by calling CLKernelLibrary::init() / GCKernelLibrary::init(). By default the path is set to "./cl_kernels" / "./cs_shaders".
Anthony Barbierdbdab852017-06-23 15:42:00 +01001183
1184@b set_soname: Do you want to build the versioned version of the library ?
1185
1186If enabled the library will contain a SONAME and SHLIBVERSION and some symlinks will automatically be created between the objects.
1187Example:
1188 libarm_compute_core.so -> libarm_compute_core.so.1.0.0
1189 libarm_compute_core.so.1 -> libarm_compute_core.so.1.0.0
1190 libarm_compute_core.so.1.0.0
1191
1192@note This options is disabled by default as it requires SCons version 2.4 or above.
1193
1194@b extra_cxx_flags: Custom CXX flags which will be appended to the end of the build command.
1195
1196@b build_dir: Build the library in a subfolder of the "build" folder. (Allows to build several configurations in parallel).
1197
1198@b examples: Build or not the examples
1199
1200@b validation_tests: Enable the build of the validation suite.
1201
Anthony Barbierdbdab852017-06-23 15:42:00 +01001202@b benchmark_tests: Enable the build of the benchmark tests
1203
1204@b pmu: Enable the PMU cycle counter to measure execution time in benchmark tests. (Your device needs to support it)
1205
Kaizen8938bd32017-09-28 14:38:23 +01001206@b mali: Enable the collection of Mali hardware counters to measure execution time in benchmark tests. (Your device needs to have a Mali driver that supports it)
Anthony Barbierdbdab852017-06-23 15:42:00 +01001207
1208@b openmp Build in the OpenMP scheduler for NEON.
1209
1210@note Only works when building with g++ not clang++
1211
1212@b cppthreads Build in the C++11 scheduler for NEON.
1213
Jenkinsc3f34a42018-03-02 12:38:09 +00001214@sa Scheduler::set
Anthony Barbierdbdab852017-06-23 15:42:00 +01001215
Kaizen8938bd32017-09-28 14:38:23 +01001216@subsection S3_2_linux Building for Linux
Anthony Barbierdbdab852017-06-23 15:42:00 +01001217
1218@subsubsection S3_2_1_library How to build the library ?
1219
1220For Linux, the library was successfully built and tested using the following Linaro GCC toolchain:
1221
Jenkins6a7771e2020-05-28 11:28:36 +01001222 - gcc-linaro-6.3.1-2017.05-x86_64_arm-linux-gnueabihf
1223 - gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu
Anthony Barbierdbdab852017-06-23 15:42:00 +01001224
Anthony Barbierdbdab852017-06-23 15:42:00 +01001225To cross-compile the library in debug mode, with NEON only support, for Linux 32bit:
1226
1227 scons Werror=1 -j8 debug=1 neon=1 opencl=0 os=linux arch=armv7a
1228
1229To cross-compile the library in asserts mode, with OpenCL only support, for Linux 64bit:
1230
1231 scons Werror=1 -j8 debug=0 asserts=1 neon=0 opencl=1 embed_kernels=1 os=linux arch=arm64-v8a
1232
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001233To cross-compile the library in asserts mode, with GLES_COMPUTE only support, for Linux 64bit:
1234
1235 scons Werror=1 -j8 debug=0 asserts=1 neon=0 opencl=0 gles_compute=1 embed_kernels=1 os=linux arch=arm64-v8a
1236
Anthony Barbierdbdab852017-06-23 15:42:00 +01001237You can also compile the library natively on an ARM device by using <b>build=native</b>:
1238
1239 scons Werror=1 -j8 debug=0 neon=1 opencl=0 os=linux arch=arm64-v8a build=native
1240 scons Werror=1 -j8 debug=0 neon=1 opencl=0 os=linux arch=armv7a build=native
1241
1242@note g++ for ARM is mono-arch, therefore if you want to compile for Linux 32bit on a Linux 64bit platform you will have to use a cross compiler.
1243
1244For example on a 64bit Debian based system you would have to install <b>g++-arm-linux-gnueabihf</b>
1245
1246 apt-get install g++-arm-linux-gnueabihf
1247
1248Then run
1249
1250 scons Werror=1 -j8 debug=0 neon=1 opencl=0 os=linux arch=armv7a build=cross_compile
1251
1252or simply remove the build parameter as build=cross_compile is the default value:
1253
1254 scons Werror=1 -j8 debug=0 neon=1 opencl=0 os=linux arch=armv7a
1255
Anthony Barbierdbdab852017-06-23 15:42:00 +01001256@subsubsection S3_2_2_examples How to manually build the examples ?
1257
1258The examples get automatically built by scons as part of the build process of the library described above. This section just describes how you can build and link your own application against our library.
1259
Jenkinsb3a371b2018-05-23 11:36:53 +01001260@note The following command lines assume the arm_compute binaries are present in the current directory or in the system library path. If this is not the case you can specify the location of the pre-built library with the compiler option -L. When building the OpenCL example the commands below assume that the CL headers are located in the include folder where the command is executed.
Anthony Barbierdbdab852017-06-23 15:42:00 +01001261
1262To cross compile a NEON example for Linux 32bit:
1263
Kaizenbf8b01d2017-10-12 14:26:51 +01001264 arm-linux-gnueabihf-g++ examples/neon_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -mfpu=neon -L. -larm_compute -larm_compute_core -o neon_convolution
Anthony Barbierdbdab852017-06-23 15:42:00 +01001265
1266To cross compile a NEON example for Linux 64bit:
1267
Kaizenbf8b01d2017-10-12 14:26:51 +01001268 aarch64-linux-gnu-g++ examples/neon_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -L. -larm_compute -larm_compute_core -o neon_convolution
Anthony Barbierdbdab852017-06-23 15:42:00 +01001269
1270(notice the only difference with the 32 bit command is that we don't need the -mfpu option and the compiler's name is different)
1271
1272To cross compile an OpenCL example for Linux 32bit:
1273
Jenkinsb3a371b2018-05-23 11:36:53 +01001274 arm-linux-gnueabihf-g++ examples/cl_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -mfpu=neon -L. -larm_compute -larm_compute_core -o cl_convolution -DARM_COMPUTE_CL
Anthony Barbierdbdab852017-06-23 15:42:00 +01001275
1276To cross compile an OpenCL example for Linux 64bit:
1277
Jenkinsb3a371b2018-05-23 11:36:53 +01001278 aarch64-linux-gnu-g++ examples/cl_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -L. -larm_compute -larm_compute_core -o cl_convolution -DARM_COMPUTE_CL
Kaizenbf8b01d2017-10-12 14:26:51 +01001279
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001280To cross compile a GLES example for Linux 32bit:
1281
1282 arm-linux-gnueabihf-g++ examples/gc_absdiff.cpp utils/Utils.cpp -I. -Iinclude/ -L. -larm_compute -larm_compute_core -std=c++11 -mfpu=neon -DARM_COMPUTE_GC -Iinclude/linux/ -o gc_absdiff
1283
1284To cross compile a GLES example for Linux 64bit:
1285
1286 aarch64-linux-gnu-g++ examples/gc_absdiff.cpp utils/Utils.cpp -I. -Iinclude/ -L. -larm_compute -larm_compute_core -std=c++11 -DARM_COMPUTE_GC -Iinclude/linux/ -o gc_absdiff
1287
Kaizenbf8b01d2017-10-12 14:26:51 +01001288(notice the only difference with the 32 bit command is that we don't need the -mfpu option and the compiler's name is different)
1289
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001290To cross compile the examples with the Graph API, such as graph_lenet.cpp, you need to link the examples against arm_compute_graph.so too.
1291
Kaizenbf8b01d2017-10-12 14:26:51 +01001292i.e. to cross compile the "graph_lenet" example for Linux 32bit:
1293
Jenkins52ba29e2018-08-29 15:32:11 +00001294 arm-linux-gnueabihf-g++ examples/graph_lenet.cpp utils/Utils.cpp utils/GraphUtils.cpp utils/CommonGraphOptions.cpp -I. -Iinclude -std=c++11 -mfpu=neon -L. -larm_compute_graph -larm_compute -larm_compute_core -Wl,--allow-shlib-undefined -o graph_lenet
Kaizenbf8b01d2017-10-12 14:26:51 +01001295
1296i.e. to cross compile the "graph_lenet" example for Linux 64bit:
1297
Jenkins52ba29e2018-08-29 15:32:11 +00001298 aarch64-linux-gnu-g++ examples/graph_lenet.cpp utils/Utils.cpp utils/GraphUtils.cpp utils/CommonGraphOptions.cpp -I. -Iinclude -std=c++11 -L. -larm_compute_graph -larm_compute -larm_compute_core -Wl,--allow-shlib-undefined -o graph_lenet
Anthony Barbierdbdab852017-06-23 15:42:00 +01001299
1300(notice the only difference with the 32 bit command is that we don't need the -mfpu option and the compiler's name is different)
1301
giorgio-arena869d4242017-10-23 16:58:59 +01001302@note If compiling using static libraries, this order must be followed when linking: arm_compute_graph_static, arm_compute, arm_compute_core
1303
Anthony Barbierdbdab852017-06-23 15:42:00 +01001304To compile natively (i.e directly on an ARM device) for NEON for Linux 32bit:
1305
Kaizenbf8b01d2017-10-12 14:26:51 +01001306 g++ examples/neon_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -mfpu=neon -larm_compute -larm_compute_core -o neon_convolution
Anthony Barbierdbdab852017-06-23 15:42:00 +01001307
1308To compile natively (i.e directly on an ARM device) for NEON for Linux 64bit:
1309
Kaizenbf8b01d2017-10-12 14:26:51 +01001310 g++ examples/neon_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -larm_compute -larm_compute_core -o neon_convolution
Anthony Barbierdbdab852017-06-23 15:42:00 +01001311
1312(notice the only difference with the 32 bit command is that we don't need the -mfpu option)
1313
1314To compile natively (i.e directly on an ARM device) for OpenCL for Linux 32bit or Linux 64bit:
1315
Jenkinsb3a371b2018-05-23 11:36:53 +01001316 g++ examples/cl_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -larm_compute -larm_compute_core -o cl_convolution -DARM_COMPUTE_CL
Anthony Barbierdbdab852017-06-23 15:42:00 +01001317
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001318To compile natively (i.e directly on an ARM device) for GLES for Linux 32bit or Linux 64bit:
Kaizenbf8b01d2017-10-12 14:26:51 +01001319
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001320 g++ examples/gc_absdiff.cpp utils/Utils.cpp -I. -Iinclude/ -L. -larm_compute -larm_compute_core -std=c++11 -DARM_COMPUTE_GC -Iinclude/linux/ -o gc_absdiff
Kaizenbf8b01d2017-10-12 14:26:51 +01001321
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001322To compile natively the examples with the Graph API, such as graph_lenet.cpp, you need to link the examples against arm_compute_graph.so too.
Kaizenbf8b01d2017-10-12 14:26:51 +01001323
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001324i.e. to natively compile the "graph_lenet" example for Linux 32bit:
Kaizenbf8b01d2017-10-12 14:26:51 +01001325
Jenkins52ba29e2018-08-29 15:32:11 +00001326 g++ examples/graph_lenet.cpp utils/Utils.cpp utils/GraphUtils.cpp utils/CommonGraphOptions.cpp -I. -Iinclude -std=c++11 -mfpu=neon -L. -larm_compute_graph -larm_compute -larm_compute_core -Wl,--allow-shlib-undefined -o graph_lenet
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001327
1328i.e. to natively compile the "graph_lenet" example for Linux 64bit:
1329
Jenkins52ba29e2018-08-29 15:32:11 +00001330 g++ examples/graph_lenet.cpp utils/Utils.cpp utils/GraphUtils.cpp utils/CommonGraphOptions.cpp -I. -Iinclude -std=c++11 L. -larm_compute_graph -larm_compute -larm_compute_core -Wl,--allow-shlib-undefined -o graph_lenet
Kaizenbf8b01d2017-10-12 14:26:51 +01001331
1332(notice the only difference with the 32 bit command is that we don't need the -mfpu option)
Anthony Barbierdbdab852017-06-23 15:42:00 +01001333
giorgio-arena869d4242017-10-23 16:58:59 +01001334@note If compiling using static libraries, this order must be followed when linking: arm_compute_graph_static, arm_compute, arm_compute_core
1335
Anthony Barbierdbdab852017-06-23 15:42:00 +01001336@note These two commands assume libarm_compute.so is available in your library path, if not add the path to it using -L
Jenkins6a7771e2020-05-28 11:28:36 +01001337@note You might need to export the path to OpenCL library as well in your LD_LIBRARY_PATH if Compute Library was built with OpenCL enabled.
Anthony Barbierdbdab852017-06-23 15:42:00 +01001338
1339To run the built executable simply run:
1340
1341 LD_LIBRARY_PATH=build ./neon_convolution
1342
1343or
1344
1345 LD_LIBRARY_PATH=build ./cl_convolution
1346
Jenkins52ba29e2018-08-29 15:32:11 +00001347@note Examples accept different types of arguments, to find out what they are run the example with \a --help as an argument. If no arguments are specified then random values will be used to execute the graph.
Jenkinsc3f34a42018-03-02 12:38:09 +00001348
1349For example:
Jenkinsb3a371b2018-05-23 11:36:53 +01001350
Jenkins52ba29e2018-08-29 15:32:11 +00001351 LD_LIBRARY_PATH=. ./graph_lenet --help
Jenkinsc3f34a42018-03-02 12:38:09 +00001352
Jenkins52ba29e2018-08-29 15:32:11 +00001353Below is a list of the common parameters among the graph examples :
1354@snippet utils/CommonGraphOptions.h Common graph examples parameters
Jenkinsc3f34a42018-03-02 12:38:09 +00001355
Kaizen8938bd32017-09-28 14:38:23 +01001356@subsection S3_3_android Building for Android
Anthony Barbierdbdab852017-06-23 15:42:00 +01001357
1358For Android, the library was successfully built and tested using Google's standalone toolchains:
Jenkins6a7771e2020-05-28 11:28:36 +01001359 - clang++ from NDK r18b for armv7a
1360 - clang++ from NDK r18b for arm64-v8a
1361 - clang++ from NDK r18b for arm64-v8.2-a with FP16 support
Anthony Barbierdbdab852017-06-23 15:42:00 +01001362
1363Here is a guide to <a href="https://developer.android.com/ndk/guides/standalone_toolchain.html">create your Android standalone toolchains from the NDK</a>
1364
Jenkins6a7771e2020-05-28 11:28:36 +01001365- Download the NDK r18b from here: https://developer.android.com/ndk/downloads/index.html
Jenkins4ba87db2019-05-23 17:11:51 +01001366- Make sure you have Python 2.7 installed on your machine.
Anthony Barbierdbdab852017-06-23 15:42:00 +01001367- Generate the 32 and/or 64 toolchains by running the following commands:
1368
Jenkinsb3a371b2018-05-23 11:36:53 +01001369
Jenkins6a7771e2020-05-28 11:28:36 +01001370 $NDK/build/tools/make_standalone_toolchain.py --arch arm64 --install-dir $MY_TOOLCHAINS/aarch64-linux-android-ndk-r18b --stl libc++ --api 21
1371 $NDK/build/tools/make_standalone_toolchain.py --arch arm --install-dir $MY_TOOLCHAINS/arm-linux-android-ndk-r18b --stl libc++ --api 21
Anthony Barbierdbdab852017-06-23 15:42:00 +01001372
Jenkins52ba29e2018-08-29 15:32:11 +00001373@attention We used to use gnustl but as of NDK r17 it is deprecated so we switched to libc++
Anthony Barbierdbdab852017-06-23 15:42:00 +01001374
Jenkinsb3a371b2018-05-23 11:36:53 +01001375@note Make sure to add the toolchains to your PATH:
1376
Jenkins6a7771e2020-05-28 11:28:36 +01001377 export PATH=$PATH:$MY_TOOLCHAINS/aarch64-linux-android-ndk-r18b/bin:$MY_TOOLCHAINS/arm-linux-android-ndk-r18b/bin
Anthony Barbierdbdab852017-06-23 15:42:00 +01001378
1379@subsubsection S3_3_1_library How to build the library ?
1380
Anthony Barbierdbdab852017-06-23 15:42:00 +01001381To cross-compile the library in debug mode, with NEON only support, for Android 32bit:
1382
1383 CXX=clang++ CC=clang scons Werror=1 -j8 debug=1 neon=1 opencl=0 os=android arch=armv7a
1384
1385To cross-compile the library in asserts mode, with OpenCL only support, for Android 64bit:
1386
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001387 CXX=clang++ CC=clang scons Werror=1 -j8 debug=0 asserts=1 neon=0 opencl=1 embed_kernels=1 os=android arch=arm64-v8a
1388
1389To cross-compile the library in asserts mode, with GLES_COMPUTE only support, for Android 64bit:
1390
1391 CXX=clang++ CC=clang scons Werror=1 -j8 debug=0 asserts=1 neon=0 opencl=0 gles_compute=1 embed_kernels=1 os=android arch=arm64-v8a
Anthony Barbierdbdab852017-06-23 15:42:00 +01001392
1393@subsubsection S3_3_2_examples How to manually build the examples ?
1394
1395The examples get automatically built by scons as part of the build process of the library described above. This section just describes how you can build and link your own application against our library.
1396
Jenkinsb3a371b2018-05-23 11:36:53 +01001397@note The following command lines assume the arm_compute binaries are present in the current directory or in the system library path. If this is not the case you can specify the location of the pre-built library with the compiler option -L. When building the OpenCL example the commands below assume that the CL headers are located in the include folder where the command is executed.
Anthony Barbierdbdab852017-06-23 15:42:00 +01001398
1399Once you've got your Android standalone toolchain built and added to your path you can do the following:
1400
1401To cross compile a NEON example:
1402
1403 #32 bit:
Kaizenbf8b01d2017-10-12 14:26:51 +01001404 arm-linux-androideabi-clang++ examples/neon_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -larm_compute-static -larm_compute_core-static -L. -o neon_convolution_arm -static-libstdc++ -pie
Anthony Barbierdbdab852017-06-23 15:42:00 +01001405 #64 bit:
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001406 aarch64-linux-android-clang++ examples/neon_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -larm_compute-static -larm_compute_core-static -L. -o neon_convolution_aarch64 -static-libstdc++ -pie
Anthony Barbierdbdab852017-06-23 15:42:00 +01001407
1408To cross compile an OpenCL example:
1409
1410 #32 bit:
Jenkinsb3a371b2018-05-23 11:36:53 +01001411 arm-linux-androideabi-clang++ examples/cl_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -larm_compute-static -larm_compute_core-static -L. -o cl_convolution_arm -static-libstdc++ -pie -DARM_COMPUTE_CL
Anthony Barbierdbdab852017-06-23 15:42:00 +01001412 #64 bit:
Jenkinsb3a371b2018-05-23 11:36:53 +01001413 aarch64-linux-android-clang++ examples/cl_convolution.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -larm_compute-static -larm_compute_core-static -L. -o cl_convolution_aarch64 -static-libstdc++ -pie -DARM_COMPUTE_CL
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001414
1415To cross compile a GLES example:
Anthony Barbierf45d5a92018-01-24 16:23:15 +00001416
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001417 #32 bit:
1418 arm-linux-androideabi-clang++ examples/gc_absdiff.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -larm_compute-static -larm_compute_core-static -L. -o gc_absdiff_arm -static-libstdc++ -pie -DARM_COMPUTE_GC
1419 #64 bit:
1420 aarch64-linux-android-clang++ examples/gc_absdiff.cpp utils/Utils.cpp -I. -Iinclude -std=c++11 -larm_compute-static -larm_compute_core-static -L. -o gc_absdiff_aarch64 -static-libstdc++ -pie -DARM_COMPUTE_GC
Kaizenbf8b01d2017-10-12 14:26:51 +01001421
1422To cross compile the examples with the Graph API, such as graph_lenet.cpp, you need to link the library arm_compute_graph also.
Kaizenbf8b01d2017-10-12 14:26:51 +01001423
1424 #32 bit:
Jenkins52ba29e2018-08-29 15:32:11 +00001425 arm-linux-androideabi-clang++ examples/graph_lenet.cpp utils/Utils.cpp utils/GraphUtils.cpp utils/CommonGraphOptions.cpp -I. -Iinclude -std=c++11 -Wl,--whole-archive -larm_compute_graph-static -Wl,--no-whole-archive -larm_compute-static -larm_compute_core-static -L. -o graph_lenet_arm -static-libstdc++ -pie -DARM_COMPUTE_CL
Kaizenbf8b01d2017-10-12 14:26:51 +01001426 #64 bit:
Jenkins52ba29e2018-08-29 15:32:11 +00001427 aarch64-linux-android-clang++ examples/graph_lenet.cpp utils/Utils.cpp utils/GraphUtils.cpp utils/CommonGraphOptions.cpp -I. -Iinclude -std=c++11 -Wl,--whole-archive -larm_compute_graph-static -Wl,--no-whole-archive -larm_compute-static -larm_compute_core-static -L. -o graph_lenet_aarch64 -static-libstdc++ -pie -DARM_COMPUTE_CL
Anthony Barbierdbdab852017-06-23 15:42:00 +01001428
1429@note Due to some issues in older versions of the Mali OpenCL DDK (<= r13p0), we recommend to link arm_compute statically on Android.
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001430@note When linked statically the arm_compute_graph library currently needs the --whole-archive linker flag in order to work properly
Anthony Barbierdbdab852017-06-23 15:42:00 +01001431
1432Then you need to do is upload the executable and the shared library to the device using ADB:
1433
1434 adb push neon_convolution_arm /data/local/tmp/
1435 adb push cl_convolution_arm /data/local/tmp/
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001436 adb push gc_absdiff_arm /data/local/tmp/
Anthony Barbierdbdab852017-06-23 15:42:00 +01001437 adb shell chmod 777 -R /data/local/tmp/
1438
1439And finally to run the example:
1440
1441 adb shell /data/local/tmp/neon_convolution_arm
1442 adb shell /data/local/tmp/cl_convolution_arm
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001443 adb shell /data/local/tmp/gc_absdiff_arm
Anthony Barbierdbdab852017-06-23 15:42:00 +01001444
1445For 64bit:
1446
1447 adb push neon_convolution_aarch64 /data/local/tmp/
1448 adb push cl_convolution_aarch64 /data/local/tmp/
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001449 adb push gc_absdiff_aarch64 /data/local/tmp/
Anthony Barbierdbdab852017-06-23 15:42:00 +01001450 adb shell chmod 777 -R /data/local/tmp/
1451
1452And finally to run the example:
1453
1454 adb shell /data/local/tmp/neon_convolution_aarch64
1455 adb shell /data/local/tmp/cl_convolution_aarch64
Anthony Barbier8140e1e2017-12-14 23:48:46 +00001456 adb shell /data/local/tmp/gc_absdiff_aarch64
Anthony Barbierdbdab852017-06-23 15:42:00 +01001457
Jenkins52ba29e2018-08-29 15:32:11 +00001458@note Examples accept different types of arguments, to find out what they are run the example with \a --help as an argument. If no arguments are specified then random values will be used to execute the graph.
Jenkinsc3f34a42018-03-02 12:38:09 +00001459
1460For example:
Jenkins52ba29e2018-08-29 15:32:11 +00001461 adb shell /data/local/tmp/graph_lenet --help
Jenkinsc3f34a42018-03-02 12:38:09 +00001462
1463In this case the first argument of LeNet (like all the graph examples) is the target (i.e 0 to run on NEON, 1 to run on OpenCL if available, 2 to run on OpenCL using the CLTuner), the second argument is the path to the folder containing the npy files for the weights and finally the third argument is the number of batches to run.
1464
Kaizenbf8b01d2017-10-12 14:26:51 +01001465@subsection S3_4_bare_metal Building for bare metal
1466
Jenkins6a7771e2020-05-28 11:28:36 +01001467For bare metal, the library was successfully built using linaro's latest (gcc-linaro-6.3.1-2017.05) bare metal toolchains:
Kaizenbf8b01d2017-10-12 14:26:51 +01001468 - arm-eabi for armv7a
1469 - aarch64-elf for arm64-v8a
1470
1471Download linaro for <a href="https://releases.linaro.org/components/toolchain/binaries/6.3-2017.05/arm-eabi/">armv7a</a> and <a href="https://releases.linaro.org/components/toolchain/binaries/6.3-2017.05/aarch64-elf/">arm64-v8a</a>.
1472
1473@note Make sure to add the toolchains to your PATH: export PATH=$PATH:$MY_TOOLCHAINS/gcc-linaro-6.3.1-2017.05-x86_64_aarch64-elf/bin:$MY_TOOLCHAINS/gcc-linaro-6.3.1-2017.05-x86_64_arm-eabi/bin
1474
1475@subsubsection S3_4_1_library How to build the library ?
1476
1477To cross-compile the library with NEON support for baremetal arm64-v8a:
1478
1479 scons Werror=1 -j8 debug=0 neon=1 opencl=0 os=bare_metal arch=arm64-v8a build=cross_compile cppthreads=0 openmp=0 standalone=1
1480
1481@subsubsection S3_4_2_examples How to manually build the examples ?
1482
1483Examples are disabled when building for bare metal. If you want to build the examples you need to provide a custom bootcode depending on the target architecture and link against the compute library. More information about bare metal bootcode can be found <a href="http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dai0527a/index.html">here</a>.
1484
1485@subsection S3_5_windows_host Building on a Windows host system
Kaizen8938bd32017-09-28 14:38:23 +01001486
1487Using `scons` directly from the Windows command line is known to cause
1488problems. The reason seems to be that if `scons` is setup for cross-compilation
1489it gets confused about Windows style paths (using backslashes). Thus it is
1490recommended to follow one of the options outlined below.
1491
Kaizenbf8b01d2017-10-12 14:26:51 +01001492@subsubsection S3_5_1_ubuntu_on_windows Bash on Ubuntu on Windows
Kaizen8938bd32017-09-28 14:38:23 +01001493
Jenkins975dfe12019-09-02 11:47:54 +01001494The best and easiest option is to use
1495<a href="https://msdn.microsoft.com/en-gb/commandline/wsl/about">Ubuntu on Windows</a>.
Kaizen8938bd32017-09-28 14:38:23 +01001496This feature is still marked as *beta* and thus might not be available.
1497However, if it is building the library is as simple as opening a *Bash on
1498Ubuntu on Windows* shell and following the general guidelines given above.
1499
Kaizenbf8b01d2017-10-12 14:26:51 +01001500@subsubsection S3_5_2_cygwin Cygwin
Kaizen8938bd32017-09-28 14:38:23 +01001501
Jenkins975dfe12019-09-02 11:47:54 +01001502If the Windows subsystem for Linux is not available <a href="https://www.cygwin.com/">Cygwin</a>
1503can be used to install and run `scons`, the minimum Cygwin version must be 3.0.7 or later. In addition
1504to the default packages installed by Cygwin `scons` has to be selected in the installer. (`git` might
Kaizen8938bd32017-09-28 14:38:23 +01001505also be useful but is not strictly required if you already have got the source
Jenkins975dfe12019-09-02 11:47:54 +01001506code of the library.) Linaro provides pre-built versions of
1507<a href="http://releases.linaro.org/components/toolchain/binaries/">GCC cross-compilers</a>
Kaizen8938bd32017-09-28 14:38:23 +01001508that can be used from the Cygwin terminal. When building for Android the
1509compiler is included in the Android standalone toolchain. After everything has
1510been set up in the Cygwin terminal the general guide on building the library
1511can be followed.
1512
Jenkins6a7771e2020-05-28 11:28:36 +01001513@subsection S3_6_cl_requirements OpenCL DDK Requirements
Anthony Barbierdbdab852017-06-23 15:42:00 +01001514
Jenkins6a7771e2020-05-28 11:28:36 +01001515@subsubsection S3_6_1_cl_hard_requirements Hard Requirements
Jenkins52ba29e2018-08-29 15:32:11 +00001516
1517Compute Library requires OpenCL 1.1 and above with support of non uniform workgroup sizes, which is officially supported in the Mali OpenCL DDK r8p0 and above as an extension (respective extension flag is \a -cl-arm-non-uniform-work-group-size).
1518
1519Enabling 16-bit floating point calculations require \a cl_khr_fp16 extension to be supported. All Mali GPUs with compute capabilities have native support for half precision floating points.
1520
1521Use of @ref CLMeanStdDev function requires 64-bit atomics support, thus \a cl_khr_int64_base_atomics should be supported in order to use.
1522
Jenkins6a7771e2020-05-28 11:28:36 +01001523@subsubsection S3_6_2_cl_performance_requirements Performance improvements
Jenkins52ba29e2018-08-29 15:32:11 +00001524
1525Integer dot product built-in function extensions (and therefore optimized kernels) are available with Mali OpenCL DDK r22p0 and above for the following GPUs : G71, G76. The relevant extensions are \a cl_arm_integer_dot_product_int8, \a cl_arm_integer_dot_product_accumulate_int8 and \a cl_arm_integer_dot_product_accumulate_int16.
1526
1527OpenCL kernel level debugging can be simplified with the use of printf, this requires the \a cl_arm_printf extension to be supported.
1528
1529SVM allocations are supported for all the underlying allocations in Compute Library. To enable this OpenCL 2.0 and above is a requirement.
1530
Jenkins6a7771e2020-05-28 11:28:36 +01001531@subsection S3_7_cl_tuner OpenCL Tuner
Jenkins52ba29e2018-08-29 15:32:11 +00001532
1533The OpenCL tuner, a.k.a. CLTuner, is a module of Arm Compute Library that can improve the performance of the OpenCL kernels tuning the Local-Workgroup-Size (LWS).
1534The optimal LWS for each unique OpenCL kernel configuration is stored in a table. This table can be either imported or exported from/to a file.
Jenkins4ba87db2019-05-23 17:11:51 +01001535The OpenCL tuner runs the same OpenCL kernel for a range of local workgroup sizes and keeps the local workgroup size of the fastest run to use in subsequent calls to the kernel. It supports three modes of tuning with different trade-offs between the time taken to tune and the kernel execution time achieved using the best LWS found. In the Exhaustive mode, it searches all the supported values of LWS. This mode takes the longest time to tune and is the most likely to find the optimal LWS. Normal mode searches a subset of LWS values to yield a good approximation of the optimal LWS. It takes less time to tune than Exhaustive mode. Rapid mode takes the shortest time to tune and finds an LWS value that is at least as good or better than the default LWS value. The mode affects only the search for the optimal LWS and has no effect when the LWS value is imported from a file.
Jenkins52ba29e2018-08-29 15:32:11 +00001536In order for the performance numbers to be meaningful you must disable the GPU power management and set it to a fixed frequency for the entire duration of the tuning phase.
1537
1538If you wish to know more about LWS and the important role on improving the GPU cache utilization, we suggest having a look at the presentation "Even Faster CNNs: Exploring the New Class of Winograd Algorithms available at the following link:
1539
1540https://www.embedded-vision.com/platinum-members/arm/embedded-vision-training/videos/pages/may-2018-embedded-vision-summit-iodice
1541
1542Tuning a network from scratch can be long and affect considerably the execution time for the first run of your network. It is recommended for this reason to store the CLTuner's result in a file to amortize this time when you either re-use the same network or the functions with the same configurations. The tuning is performed only once for each OpenCL kernel.
1543
1544CLTuner looks for the optimal LWS for each unique OpenCL kernel configuration. Since a function (i.e. Convolution Layer, Pooling Layer, Fully Connected Layer ...) can be called multiple times but with different parameters, we associate an "id" (called "config_id") to each kernel to distinguish the unique configurations.
1545
1546 #Example: 2 unique Matrix Multiply configurations
1547@code{.cpp}
1548 TensorShape a0 = TensorShape(32,32);
1549 TensorShape b0 = TensorShape(32,32);
1550 TensorShape c0 = TensorShape(32,32);
1551 TensorShape a1 = TensorShape(64,64);
1552 TensorShape b1 = TensorShape(64,64);
1553 TensorShape c1 = TensorShape(64,64);
1554
1555 Tensor a0_tensor;
1556 Tensor b0_tensor;
1557 Tensor c0_tensor;
1558 Tensor a1_tensor;
1559 Tensor b1_tensor;
1560 Tensor c1_tensor;
1561
1562 a0_tensor.allocator()->init(TensorInfo(a0, 1, DataType::F32));
1563 b0_tensor.allocator()->init(TensorInfo(b0, 1, DataType::F32));
1564 c0_tensor.allocator()->init(TensorInfo(c0, 1, DataType::F32));
1565 a1_tensor.allocator()->init(TensorInfo(a1, 1, DataType::F32));
1566 b1_tensor.allocator()->init(TensorInfo(b1, 1, DataType::F32));
1567 c1_tensor.allocator()->init(TensorInfo(c1 1, DataType::F32));
1568
1569 CLGEMM gemm0;
1570 CLGEMM gemm1;
1571
1572 // Configuration 0
1573 gemm0.configure(&a0, &b0, nullptr, &c0, 1.0f, 0.0f);
1574
1575 // Configuration 1
1576 gemm1.configure(&a1, &b1, nullptr, &c1, 1.0f, 0.0f);
1577@endcode
1578
Jenkins6a7771e2020-05-28 11:28:36 +01001579@subsubsection S3_7_1_cl_tuner_how_to How to use it
Jenkins52ba29e2018-08-29 15:32:11 +00001580
1581All the graph examples in the ACL's folder "examples" and the arm_compute_benchmark accept an argument to enable the OpenCL tuner and an argument to export/import the LWS values to/from a file
1582
1583 #Enable CL tuner
1584 ./graph_mobilenet --enable-tuner –-target=CL
1585 ./arm_compute_benchmark --enable-tuner
1586
1587 #Export/Import to/from a file
1588 ./graph_mobilenet --enable-tuner --target=CL --tuner-file=acl_tuner.csv
1589 ./arm_compute_benchmark --enable-tuner --tuner-file=acl_tuner.csv
1590
1591If you are importing the CLTuner'results from a file, the new tuned LWS values will be appended to it.
1592
1593Either you are benchmarking the graph examples or the test cases in the arm_compute_benchmark remember to:
1594
1595 -# Disable the power management
1596 -# Keep the GPU frequency constant
1597 -# Run multiple times the network (i.e. 10).
1598
1599If you are not using the graph API or the benchmark infrastructure you will need to manually pass a CLTuner object to CLScheduler before configuring any function.
1600
1601@code{.cpp}
1602CLTuner tuner;
1603
1604// Setup Scheduler
1605CLScheduler::get().default_init(&tuner);
1606@endcode
1607
1608After the first run, the CLTuner's results can be exported to a file using the method "save_to_file()".
1609- tuner.save_to_file("results.csv");
1610
1611This file can be also imported using the method "load_from_file("results.csv")".
1612- tuner.load_from_file("results.csv");
Anthony Barbierdbdab852017-06-23 15:42:00 +01001613*/
Jenkinsc3f34a42018-03-02 12:38:09 +00001614} // namespace arm_compute