Merge pull request #119 from TimmyLiu/master

merge develop branch into master branch. bump the version number to 2.6

.gitignore

@@ -17,3 +17,6 @@
 
 # Generated kernel template files
 *.clT
+
+# flags.txt file
+*flags.txt

.travis.yml

@@ -5,16 +5,34 @@ compiler:
 
 before_install:
   - sudo apt-get update -qq
-  - sudo apt-get install -qq fglrx opencl-headers libboost-program-options-dev
+  - sudo apt-get install -qq fglrx libboost-program-options-dev
 # Uncomment below to help verify the installs above work
 #  - ls -la /usr/lib/libboost*
 #  - ls -la /usr/include/boost
 
 before_script:
   - cd ${TRAVIS_BUILD_DIR}
+# download OpenCL 1.2 header files since Travis CI only provides 1.1
+  - mkdir -p OpenCLInclude/CL
+  - cd OpenCLInclude/CL
+  #- wget -r --no-parent -nH --cut-dirs=4 --reject="index.html*" https://www.khronos.org/registry/cl/api/1.2/
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl.h
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl.hpp
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl_d3d10.h
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl_d3d11.h
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl_dx9_media_sharing.h
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl_egl.h
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl_ext.h
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl_gl.h
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl_gl_ext.h
+  - wget https://www.khronos.org/registry/cl/api/1.2/cl_platform.h
+  - wget https://www.khronos.org/registry/cl/api/1.2/opencl.h
+  - ls
+  - pwd
+  - cd ../..
   - mkdir -p bin/clBLAS
   - cd bin/clBLAS
-  - cmake -DCMAKE_BUILD_TYPE=Release -DBUILD_TEST=OFF -DBUILD_CLIENT=ON -DCMAKE_INSTALL_PREFIX:PATH=$PWD/package ../../src
+  - cmake -DCMAKE_BUILD_TYPE=Release -DBUILD_TEST=OFF -DBUILD_CLIENT=ON -DOPENCL_INCLUDE_DIRS:PATH=$PWD/../../OpenCLInclude -DCMAKE_INSTALL_PREFIX:PATH=$PWD/package ../../src
 
 script: 
   - make install

LICENSE

@@ -175,28 +175,3 @@
       of your accepting any such warranty or additional liability.
 
    END OF TERMS AND CONDITIONS
-
-   APPENDIX: How to apply the Apache License to your work.
-
-      To apply the Apache License to your work, attach the following
-      boilerplate notice, with the fields enclosed by brackets "[]"
-      replaced with your own identifying information. (Don't include
-      the brackets!)  The text should be enclosed in the appropriate
-      comment syntax for the file format. We also recommend that a
-      file or class name and description of purpose be included on the
-      same "printed page" as the copyright notice for easier
-      identification within third-party archives.
-
-   Copyright [yyyy] [name of copyright owner]
-
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
-
-       http://www.apache.org/licenses/LICENSE-2.0
-
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License.

README.md

@@ -20,6 +20,24 @@ library does generate and enqueue optimized OpenCL kernels, relieving
 the user from the task of writing, optimizing and maintaining kernel
 code themselves.
 
+## clBLAS update notes 04/2015
+-   A subset of GEMM and TRSM can be off-line compiled for Hawaii, Bonaire and Tahiti device at compile-time. This feature
+    eliminates the overhead of calling clBuildProgram() at run-time.
+-   Off-line compilation can be done with OpenCL 1.1, OpenCL 1.2 and OpenCl 2.0 runtime. However, for better
+    performance OpenCL 2.0 is recommended. Library user can select "OCL_VERSION" from CMake to ensure the library with
+    OpenCL version. It is library user's responsibility to ensure compatible hardware and driver.
+-   Added flags_public.txt file that contains OpenCL compiler flags used by off-line compilation. The flags_public.txt
+    will only be loaded when OCL_VERSION is 2.0.
+-   User can off-line compile one or more supported device by selecting 
+    OCL_OFFLINE_BUILD_BONAIRE_KERNEL
+    OCL_OFFLINE_BUILD_HAWII_KERNEL
+    OCL_OFFLINE_BUILD_TAHITI_KERNEL.
+    However, compile for more than one device at a time might result in running out of heap memory. Thus, compile for
+    one device at a time is recommended.
+-   User may also supply specific OpenCL compiler path with OCL_COMPILER_DIR or the library will load default OpenCL compiler.
+-   The minimum driver requirement for off-line compilation is 14.502.
+
+
 ## clBLAS library user documentation
 
 [Library and API documentation][] for developers is available online as
@@ -48,15 +66,12 @@ how to contribute code to this open source project. The code in the
 be made against the /develop branch.
 
 ## License
-
-The source for clBLAS is licensed under the [Apache License, Version
-2.0][]
+The source for clBLAS is licensed under the [Apache License, Version 2.0]( http://www.apache.org/licenses/LICENSE-2.0 )
 
 ## Example
+The simple example below shows how to use clBLAS to compute an OpenCL accelerated SGEMM
 
-The simple example below shows how to use clBLAS to compute an OpenCL
-accelerated SGEMM
-
+```c
     #include <sys/types.h>
     #include <stdio.h>
 
@@ -170,42 +185,30 @@ accelerated SGEMM
 
     return ret;
     }
+```
 
 ## Build dependencies
-
 ### Library for Windows
-
--   Windows® 7/8
-
--   Visual Studio 2010 SP1, 2012
-
--   An OpenCL SDK, such as APP SDK 2.9
-
--   Latest CMake
+*  Windows® 7/8
+*  Visual Studio 2010 SP1, 2012
+*  An OpenCL SDK, such as APP SDK 2.8
+*  Latest CMake
 
 ### Library for Linux
-
--   GCC 4.6 and onwards
-
--   An OpenCL SDK, such as APP SDK 2.9
-
--   Latest CMake
+*  GCC 4.6 and onwards
+*  An OpenCL SDK, such as APP SDK 2.9
+*  Latest CMake
 
 ### Library for Mac OSX
-
--   Recommended to generate Unix makefiles with cmake
+*  Recommended to generate Unix makefiles with cmake
 
 ### Test infrastructure
-
--   Googletest v1.6
-
--   ACML on windows/linux; Accelerate on Mac OSX
-
--   Latest Boost
+*  Googletest v1.6
+*  ACML on windows/linux; Accelerate on Mac OSX
+*  Latest Boost
 
 ### Performance infrastructure
-
--   Python
+* Python
 
   [Library and API documentation]: http://clmathlibraries.github.io/clBLAS/
   [[email protected]]: https://groups.google.com/forum/#!forum/clmath

doc/README-BinaryCacheOnDisk.txt

@@ -0,0 +1,69 @@
+S. Chauveau
+CAPS Entreprise
+clBLAS Project
+------------------------------
+April 30,2014
+
+
+The implementation of a binary cache for CL programs can be found in 
+files src/include/binary_lookup.h and src/library/blas/generic/binary_lookup.cc
+
+The cache is currently disabled by default. It can be enabled by
+setting the environment variable 'CLBLAS_CACHE_PATH' to the directory
+containing the cache entries.
+
+In the code itself, accesses to the cache are controlled by the
+BinaryLookup class. A typical cache query looks as follow:
+
+   (1) Create a local instance of BinaryLookup 
+
+   (2) Specify the additional characteristics (i.e. variants) of the
+       requested program. That information combined with the program name
+       and the OpenCL context and device shall form a unique signature
+       for the binary program.
+
+   (3) Perform the effective search by calling the 'found' method
+
+   (4a) If the search was successful then cl_program can be retrieved 
+       by a call to the 'getProgram' method
+
+   (4b) If the search was not successful then a cl_program 
+       must be created  and populated in the cache by a call 
+       to the 'setProgram' method.
+
+   (5) Destroy the BinaryLookup local instance.
+
+
+So in practice a typical query shall looks as follow:
+
+  cl_program program  ;
+
+  // The program name is part of the signature and shall be unique 
+  const char * program_name = "... my unique program name ... " ;
+
+  BinaryLookup bl(context, device, program_name);
+
+  // Specify some additional information used to build a 
+  // unique signature for that cache entry
+
+  bl.variantInt( vectorSize );
+  bl.variantInt( hasBorder );
+  ... 
+
+  // Perform the query 
+  if ( bl.found() ) 
+  {
+     // Success! use the cl_program retrieved from the cache
+     program = bl.getProgram();
+  }
+  else 
+  {
+     // Failure! we need to build the program 
+     program = build_my_program(context,device,vectorSize,...) ; 
+     // and inform the lookup object of the program
+     bl.setProgram(program);  
+     // and finally populate the cache
+     bl.populateCache() 
+  }
+
+  // The BinaryLookup shall now be destroyed 

doc/README-FunctorConcepts.txt

@@ -0,0 +1,100 @@
+S. Chauveau
+CAPS Entreprise
+April 30, 2014
+
+The Functor concept was introduced in clBLAS to simplify the creation 
+of specialized versions for dedicated architectures.
+
+The original system, referred as the 'Solver' system in this document, 
+is very centralized and not flexible enough to insert customized kernels.
+
+The Functor 
+===========
+
+A functor is simply a C++ object that provides an implementation of 
+a function. In the current case, that function is one of the BLAS calls 
+implemented in OpenCL. 
+
+The base class of all functors is clblasFunctor 
+  - see src/library/blas/functor/include/functor.h
+  - see src/library/blas/functor/functor.cc
+
+That class does not provide much by itself but it is supposed to be derived
+once for each BLAS function to be implemented. 
+
+For instance the clblasSgemmFunctor class will be the base class of all
+functors providing a generic or specific implementation of SGEMM.
+
+A generic functor is one that is applicable to all possible arguments of the
+function it implements. In most cases, there will be at least one generic
+functor that will simply call the existing Solver-based implementation of the
+function. For SGEMM, that is the class clblasSgemmFunctorFallback.
+
+A specific functor is one that is applicable to only a subset of the possible
+arguments of the function it implements. For instance, a SGEMM functor could
+only implement it for matrices of a given block size or only for square
+matrices or only for a specific device architecture (e.g. AMD Hawai) etc
+
+The Functor Selector 
+====================
+
+Multiple generic and specific functors may be available to implement each
+clBLAS call. The selection of the proper functor is delegated to the class
+clblasFunctorSelector whose default implementation typically returns the
+fallback functors.
+
+  - see src/library/blas/functor/include/functor_selector.h
+  - see src/library/blas/functor/functor_selector.cc
+
+So clblasFunctorSelector provides a large set of virtual selection methods.
+Typically, a method to select a specific functor will be provided for each
+supported BLAS function. Another method may be provided to select a generic 
+functor but that is not mandatory.
+
+The default implementation of clblasFunctorSelector is typically that the 
+specific selector is redirected to the generic one returning the fallback 
+functor (so using the existing Solver-based  implementation).  
+
+
+The class clblasFunctorSelector is supposed to be derived once for each
+supported architecture (e.g. Hawai, Tahiti, ...) and a single global instance
+of each of those derived classes shall be created. This is important because
+those instances register themselves in a global data structure that is later
+used to find the proper clblasFunctorSelector according to the architecture
+(see clblasFunctorSelector::find() )
+
+
+Functor Management & Cache
+==========================
+
+Each functor contains a reference counter that, when it reaches zero, causes
+the functor destruction. See the members clblasFunctor::retain() and
+clblasFunctor::release().
+
+Of course, to be efficient, functors must be reusable between BLAS calls so 
+some mechanisms must be implemented to manage the functors. 
+
+Some functors, such as the fallback functors, are independent of the 
+arguments and of the opencl context & device. Those can typically be 
+implemented using a single global instance that will never be destroyed.
+
+Other functors, such as those that manage a cl_program internally, are 
+dependent of the opencl context & device and sometimes of some arguments. 
+They need to be stored in caches using some information as keys. 
+
+In the current implementation, we propose that each functor class shall 
+implement its own private cache. Such functors shall not be created directly 
+using its constructor but via a dedicated 'provide' function (the name 'provide'
+is not mandatory) that will take care of managing the internal cache.
+
+The template class clblasFunctorCache<F> is provided as a simple
+implementation of a cache of functors of type F. Use of that cache is not a
+mandatory part of the functor design. Another strategies could be to keep a
+single instance of the functor and implement a cache for the cl_program or to
+implement a global cache shared by multiple functor classes.
+
+
+
+
+
+