Merge pull request clMathLibraries#210 from TimmyLiu/master

merge develop branch into master branch. bump the version number to 2.10

.travis.yml

@@ -15,6 +15,7 @@ language: cpp
 # sudo: false instructs travis to build our project in a docker VM (faster)
 # Can not yet install fglrx packages with 'false'
 sudo: required # false
+dist: trusty
 
 # os: expands the build matrix to include multiple os's
 # disable linux, as we get sporadic failures on building boost, needs investigation
@@ -43,6 +44,7 @@ addons:
       # boost-latest contains boost v1.55
       - boost-latest
     packages:
+      - gfortran
       # g++-4.8 is minimum version considered to be the first good c++11 gnu compiler
       # - g++-4.8
       # - clang-3.6
@@ -66,6 +68,8 @@ addons:
 env:
   global:
     - CLBLAS_ROOT=${TRAVIS_BUILD_DIR}/bin/make/release
+    - OPENCL_REGISTRY=https://www.khronos.org/registry/cl
+    - OPENCL_ROOT=${TRAVIS_BUILD_DIR}/bin/opencl
 
 # The following filters our build matrix; we are interested in linux-gcc & osx-clang
 matrix:
@@ -77,13 +81,13 @@ matrix:
 
 before_install:
   # Remove the following linux clause when fglrx can be installed with sudo: false
-  - if [ ${TRAVIS_OS_NAME} == "linux" ]; then
-      sudo apt-get update -qq &&
-      sudo apt-get install -qq fglrx=2:13.350.1-0ubuntu0.0.1;
-    fi
-  - if [ ${TRAVIS_OS_NAME} == "linux" ]; then
-      export OPENCL_ROOT="${TRAVIS_BUILD_DIR}/opencl-headers";
-    fi
+  #- if [ ${TRAVIS_OS_NAME} == "linux" ]; then
+  #    sudo apt-get update -qq &&
+  #    sudo apt-get install -qq fglrx=2:13.350.1-0ubuntu0.0.1;
+  #  fi
+  #- if [ ${TRAVIS_OS_NAME} == "linux" ]; then
+  #    export OPENCL_ROOT="${TRAVIS_BUILD_DIR}/opencl-headers";
+  #  fi
   - if [ ${TRAVIS_OS_NAME} == "osx" ]; then
       brew update;
       brew outdated boost || brew upgrade boost;
@@ -97,10 +101,33 @@ before_install:
 install:
   # 'Precise' only distributes v1.1 opencl headers; download 1.2 headers from khronos website
   # Remove when the travis VM upgrades to 'trusty' or beyond
+  #- if [ ${TRAVIS_OS_NAME} == "linux" ]; then
+  #    mkdir -p ${OPENCL_ROOT}/include/CL;
+  #    pushd ${OPENCL_ROOT}/include/CL;
+  #    wget -w 1 -r -np -nd -nv -A h,hpp https://www.khronos.org/registry/cl/api/1.2/;
+  #    popd;
+  #  fi
+  # The following linux logic is necessary because of Travis's move to the GCE platform, which does not
+  # currently contain packages for fglrx: https://github.com/travis-ci/travis-ci/issues/5221
+  # We build our own linkable .so file
   - if [ ${TRAVIS_OS_NAME} == "linux" ]; then
-      mkdir -p ${OPENCL_ROOT}/include/CL;
-      pushd ${OPENCL_ROOT}/include/CL;
-      wget -w 1 -r -np -nd -nv -A h,hpp https://www.khronos.org/registry/cl/api/1.2/;
+      mkdir -p ${OPENCL_ROOT};
+      pushd ${OPENCL_ROOT};
+      wget ${OPENCL_REGISTRY}/specs/opencl-icd-1.2.11.0.tgz;
+      tar -xf opencl-icd-1.2.11.0.tgz;
+      mv ./icd/* .;
+      mkdir -p inc/CL;
+      pushd inc/CL;
+      wget -r -w 1 -np -nd -nv -A h,hpp https://www.khronos.org/registry/cl/api/1.2/;
+      wget -w 1 -np -nd -nv -A h,hpp https://www.khronos.org/registry/cl/api/2.1/cl.hpp;
+      popd;
+      mkdir -p lib;
+      pushd lib;
+      cmake -G "Unix Makefiles" ..;
+      make;
+      cp ../bin/libOpenCL.so .;
+      popd;
+      mv inc/ include/;
       popd;
     fi
   # osx image does not contain cl.hpp file; download from Khronos
@@ -114,7 +141,7 @@ install:
 before_script:
   - mkdir -p ${CLBLAS_ROOT}
   - pushd ${CLBLAS_ROOT}
-  - cmake -DCMAKE_BUILD_TYPE=Release -DBUILD_TEST=OFF -DBUILD_CLIENT=ON -DOCL_VERSION=2.0 -DOPENCL_ROOT=${OPENCL_ROOT} ${TRAVIS_BUILD_DIR}/src
+  - cmake -DCMAKE_BUILD_TYPE=Release -DBUILD_TEST=OFF -DBUILD_CLIENT=OFF -DOCL_VERSION=2.0 -DOPENCL_ROOT=${OPENCL_ROOT} ${TRAVIS_BUILD_DIR}/src
 
 # use script: to execute build steps
 script:

src/CMakeLists.txt

@@ -82,8 +82,8 @@ message( STATUS "You have confirmed OpenCL ${OPENCL_VERSION} is supported in you
 if ( APPLE )
 	set(CORR_TEST_WITH_ACML OFF CACHE BOOL "Use ACML library in correctness tests")
 else ( )
-	message(STATUS "CORR_TEST_WITH_ACML set to ON")
-	set(CORR_TEST_WITH_ACML ON CACHE BOOL "Use ACML library in correctness tests")
+	message(STATUS "CORR_TEST_WITH_ACML set to OFF. Try link with libblas.so")
+	set(CORR_TEST_WITH_ACML OFF CACHE BOOL "Use ACML library in correctness tests")
 endif( )
 
 if( CMAKE_GENERATOR MATCHES "NMake" )
@@ -95,8 +95,8 @@ if( CMAKE_GENERATOR MATCHES "NMake" )
   endif( )
 endif( )
 
-# If we are on linux, and we wish to link with the netlib BLAS implementation, we need to have a valid fortran compiler
-if( NOT CORR_TEST_WITH_ACML AND NOT WIN32 AND NOT APPLE )
+# If we are on linux, and we wish to link with the netlib BLAS implementation when BUILD_TEST is ON, we need to have a valid fortran compiler
+if(BUILD_TEST AND NOT CORR_TEST_WITH_ACML AND NOT WIN32 AND NOT APPLE)
   project(clBLAS Fortran C CXX )
 else( )
   project(clBLAS C CXX)
@@ -108,7 +108,7 @@ if( NOT DEFINED clBLAS_VERSION_MAJOR )
 endif( )
 
 if( NOT DEFINED clBLAS_VERSION_MINOR )
-    set( clBLAS_VERSION_MINOR 8 )
+    set( clBLAS_VERSION_MINOR 10 )
 endif( )
 
 if( NOT DEFINED clBLAS_VERSION_PATCH )

src/FindNetlib.cmake

@@ -74,10 +74,10 @@ endif( )
 
 # The library name available from Netlib has different names for 64bit and 32bit libs
 if( LIB64 )
-	set( Netlib_BLAS_LIBNAME libblas )
+	set( Netlib_BLAS_LIBNAME blas )
 #	set( Netlib_BLAS_LIBNAME BLAS )  Even though the download is named BLAS, the linker expects the .dll to be called libblas.dll
 else( )
-	set( Netlib_BLAS_LIBNAME libblas )
+	set( Netlib_BLAS_LIBNAME blas )
 endif( )
 
 list( FIND Netlib_FIND_COMPONENTS BLAS contains_BLAS )
@@ -91,6 +91,7 @@ if( NOT contains_BLAS EQUAL -1 )
 		PATHS
 			/usr/lib
 			/usr/local/lib
+			/usr/lib/libblas
 		DOC "Netlib dynamic library path"
 		PATH_SUFFIXES lib
 	)

src/library/CMakeLists.txt

@@ -14,6 +14,7 @@
 # limitations under the License.
 # ########################################################################
 
+find_package(PythonInterp REQUIRED)
 
 
 
@@ -43,6 +44,7 @@ set(AUTOGEMM_HEADERS
 )
 
 set(AUTOGEMM_SRC
+  ${CMAKE_SOURCE_DIR}/library/blas/AutoGemm/UserGemmKernelSources/UserGemmClKernels.cc
   ${CMAKE_BINARY_DIR}/include/AutoGemmIncludes/AutoGemmClKernels.cpp
   ${CMAKE_BINARY_DIR}/include/AutoGemmIncludes/AutoGemmKernelBuildOptionsBinary.cpp
   ${CMAKE_BINARY_DIR}/include/AutoGemmIncludes/AutoGemmKernelBinaries.cpp
@@ -90,6 +92,10 @@ option( PRECOMPILE_GEMM_TRANS_CN "AutoGemm: pre-compile CN transpose cases" OFF)
 option( PRECOMPILE_GEMM_TRANS_CT "AutoGemm: pre-compile CT transpose cases" OFF)
 option( PRECOMPILE_GEMM_TRANS_CC "AutoGemm: pre-compile CC transpose cases" OFF)
 
+set( AUTOGEMM_ARCHITECTURE "Hawaii" CACHE STRING "AutoGemm: device for kernel selection logic" )
+set_property( CACHE AUTOGEMM_ARCHITECTURE PROPERTY STRINGS "Hawaii" "Fiji" )
+
+
 
 # opencl compiler version
 #set( PRECOMPILE_GEMM_OPENCL_VERSION "2.0" CACHE STRING "OpenCL compiler version supported by device driver." )
@@ -197,7 +203,7 @@ set( AUTOGEMM_PRECOMPILE_HEADER_SRC ${CMAKE_SOURCE_DIR}/library/blas/AutoGemm/Ke
 set( AUTOGEMM_PRECOMPILE_HEADER_OUT ${CMAKE_BINARY_DIR}/include/AutoGemmIncludes/AutoGemmKernelsToPreCompile.h )
 add_custom_command(
   OUTPUT ${AUTOGEMM_PRECOMPILE_HEADER_OUT}
-  COMMAND python ${CMAKE_SOURCE_DIR}/library/blas/AutoGemm/KernelsToPreCompile.py ${AGPC_ARGS}
+  COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_SOURCE_DIR}/library/blas/AutoGemm/KernelsToPreCompile.py ${AGPC_ARGS}
   DEPENDS ${AUTOGEMM_PRECOMPILE_HEADER_SRC}
 )
 
@@ -238,7 +244,7 @@ endif()#endif precompile active
 ################################################################################
 add_custom_command(
   OUTPUT ${AUTOGEMM_HEADERS} ${AUTOGEMM_SRC}
-  COMMAND python ${CMAKE_SOURCE_DIR}/library/blas/AutoGemm/AutoGemm.py --output-path ${CMAKE_BINARY_DIR}/include --opencl-compiler-version ${OPENCL_VERSION}
+  COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_SOURCE_DIR}/library/blas/AutoGemm/AutoGemm.py --output-path ${CMAKE_BINARY_DIR}/include --opencl-compiler-version ${OPENCL_VERSION} --architecture ${AUTOGEMM_ARCHITECTURE}
   DEPENDS ${AUTOGEMM_SCRIPTS}
 )
 
@@ -851,6 +857,7 @@ set(CLBLAS_ALL_SOURCES
 	#${USERGEMM_SRC}
 	#${USERGEMM_HEADERS}
 )
+add_definitions(-DOPENCL_VERSION="${OPENCL_VERSION}")
 add_library(clBLAS ${CLBLAS_ALL_SOURCES})
 add_dependencies(clBLAS GENERATE_CLT)
 

src/library/blas/AutoGemm/AutoGemm.py

@@ -21,6 +21,7 @@
 import Common
 import Includes
 import KernelSelection
+import AutoGemmParameters
 import KernelOpenCL
 
 
@@ -32,16 +33,17 @@
   ap = argparse.ArgumentParser(description="AutoGemm")
   ap.add_argument("--output-path", dest="output" )
   ap.add_argument("--opencl-compiler-version", dest="clCompilerVersion", action="store", choices=["1.1", "1.2", "2.0" ])
+  ap.add_argument("--architecture", dest="architecture", action="store", choices=["Hawaii", "Fiji" ])
   args = ap.parse_args()
   if args.output:
     Common.setOutputPath(args.output)
   else:
-    print "AutoGemm.py: Warning: No output path specified; default is working directory."
+    print("AutoGemm.py: Warning: No output path specified; default is working directory.")
 
-  print "AutoGemm.py: using OpenCL " + args.clCompilerVersion + " compiler"
+  print("AutoGemm.py: using OpenCL " + args.clCompilerVersion + " compiler")
   Common.setClCompilerVersion(args.clCompilerVersion)
+  AutoGemmParameters.setArchitecture(args.architecture)
 
   KernelOpenCL.writeOpenCLKernels()
   KernelSelection.writeKernelSelection()
   Includes.writeIncludes()
-

src/library/blas/AutoGemm/AutoGemmParameters.py

@@ -5,7 +5,7 @@
 # Tile Parameters for Kernel Selection Data
 ################################################################################
 
-kernelSelectionData = {
+kernelSelectionDataHawaii = {
 # [ size, fallback tile, [ valid tiles ] ],
   "s":[
     [ 4000, [ 16, 16,  6,  6], [ [ 16, 16,  6,  6] ] ],
@@ -54,17 +54,55 @@
     ],
   }
 
-"""
-for testing all micro-tile sizes
-    [  128, [ 16, 16,  8,  8], [ [ 16, 16,  8,  8] ] ],
-    [  112, [ 16, 16,  7,  7], [ [ 16, 16,  7,  7] ] ],
-    [   96, [ 16, 16,  6,  6], [ [ 16, 16,  6,  6] ] ],
-    [   80, [ 16, 16,  5,  5], [ [ 16, 16,  5,  5] ] ],
-    [   64, [ 16, 16,  4,  4], [ [ 16, 16,  4,  4] ] ],
-    [   48, [ 16, 16,  3,  3], [ [ 16, 16,  3,  3] ] ],
-    [   32, [ 16, 16,  2,  2], [ [ 16, 16,  2,  2] ] ],
-    [    0, [ 16, 16,  1,  1], [ [ 16, 16,  1,  1] ] ],
-"""
+kernelSelectionDataFiji = {
+  "s":[
+    [ 3072, [ 16, 16, 6, 6], [ [ 16, 16, 6, 6], [ 16, 16, 5, 5], [ 16, 16, 4, 4] ] ],
+    [ 2240, [ 16, 16, 6, 6], [ [ 16, 16, 6, 6], [ 16, 16, 4, 4], [ 16, 16, 5, 5], [ 16, 16, 3, 3] ] ],
+    [ 1760, [ 16, 16, 4, 4], [ [ 16, 16, 6, 6], [ 16, 16, 4, 4], [ 16, 16, 5, 5], [ 16, 16, 3, 3] ] ],
+    [ 1600, [ 16, 16, 4, 4], [ [ 16, 16, 4, 4], [ 16, 16, 6, 6], [ 16, 16, 5, 5], [ 16, 16, 3, 3] ] ],
+    [ 1056, [ 16, 16, 4, 4], [ [ 16, 16, 4, 4], [ 16, 16, 6, 6], [ 16, 16, 5, 5], [ 16, 16, 3, 3], [ 16, 16, 2, 2] ] ],
+    [  960, [ 16, 16, 4, 4], [ [ 16, 16, 4, 4], [ 16, 16, 5, 5], [ 16, 16, 3, 3], [ 16, 16, 2, 2] ] ],
+    [  736, [ 16, 16, 3, 3], [ [ 16, 16, 4, 4], [ 16, 16, 3, 3], [ 16, 16, 5, 5], [ 16, 16, 2, 2] ] ],
+    [  528, [ 16, 16, 3, 3], [ [ 16, 16, 4, 4], [ 16, 16, 3, 3], [ 16, 16, 2, 2], [ 16, 16, 1, 1] ] ],
+    [  432, [ 16, 16, 2, 2], [ [ 16, 16, 3, 3], [ 16, 16, 2, 2], [ 16, 16, 1, 1] ] ],
+    [  320, [ 16, 16, 2, 2], [ [ 16, 16, 2, 2], [ 16, 16, 1, 1] ] ],
+    [    0, [ 16, 16, 1, 1], [ [ 16, 16, 1, 1] ] ],
+  ],
+  "d":[
+    [ 3200, [ 16, 16, 4, 4], [ [ 16, 16, 4, 4], [ 16, 16, 5, 5], [ 16, 16, 2, 2], [  8,  8, 6, 6 ] ] ],
+    [ 1632, [ 16, 16, 2, 2], [ [ 16, 16, 4, 4], [ 16, 16, 2, 2], [ 16, 16, 5, 5], [  8,  8, 6, 6 ] ] ],
+    [ 1280, [ 16, 16, 2, 2], [ [ 16, 16, 4, 4], [ 16, 16, 2, 2], [ 16, 16, 5, 5], [  8,  8, 6, 6 ], [ 16, 16, 1, 1] ] ],
+    [ 1056, [ 16, 16, 2, 2], [ [ 16, 16, 2, 2], [ 16, 16, 1, 1] ] ],
+    [  672, [ 16, 16, 2, 2], [ [ 16, 16, 1, 1] ] ],
+    [    0, [ 16, 16, 1, 1], [ [ 16, 16, 1, 1] ] ],
+  ],
+  "c":[
+    [ 2240,  [ 16, 16, 4, 4], [ [ 16, 16, 4, 4], [ 16, 16, 6, 6], ] ],
+    [ 1440,  [ 16, 16, 4, 4], [ [ 16, 16, 4, 4], [ 16, 16, 6, 6], [ 16, 16, 5, 5], [ 16, 16, 2, 2] ] ],
+    [ 1088,  [ 16, 16, 2, 2], [ [ 16, 16, 4, 4], [ 16, 16, 3, 3], [ 16, 16, 2, 2], [ 16, 16, 5, 5] ] ],
+    [  704,  [ 16, 16, 2, 2], [ [ 16, 16, 2, 2], [ 16, 16, 3, 3], [ 16, 16, 5, 5] ] ],
+    [  528,  [ 16, 16, 2, 2], [ [ 16, 16, 2, 2], [ 16, 16, 3, 3], [ 16, 16, 1, 1] ] ],
+    [  336,  [ 16, 16, 2, 2], [ [ 16, 16, 2, 2], [ 16, 16, 1, 1] ] ],
+    [    0,  [ 16, 16, 1, 1], [ [ 16, 16, 1, 1] ] ],
+  ],
+  "z":[
+    [ 2528, [ 16, 16, 2, 2], [ [ 16, 16, 4, 4], [ 16, 16, 2, 2], [ 16, 16, 3, 3] ] ],
+    [ 1872, [ 16, 16, 2, 2], [ [ 16, 16, 2, 2], [ 16, 16, 3, 3], [ 16, 16, 1, 1] ] ],
+    [ 1040, [ 16, 16, 2, 2], [ [ 16, 16, 2, 2], [ 16, 16, 1, 1] ] ],
+    [  768, [ 16, 16, 1, 1], [ [ 16, 16, 2, 2], [ 16, 16, 1, 1] ] ],
+    [    0, [ 16, 16, 1, 1], [ [ 16, 16, 1, 1] ] ],
+  ]
+}
+
+kernelSelectionData = kernelSelectionDataHawaii
+def setArchitecture(architecture):
+  global kernelSelectionData, kernelSelectionDataHawaii, kernelSelectionDataFiji
+
+  if architecture == "Fiji":
+    kernelSelectionData = kernelSelectionDataFiji
+  else:
+    kernelSelectionData = kernelSelectionDataHawaii
+
 
 ################################################################################
 # Non-Tile Parameters
@@ -89,20 +127,20 @@ def getTilesForPrecision(precision):
     validTiles = sizeData[2]
     # add valid tiles
     for tileParams in validTiles:
-      #print tileParams
+      #print(tileParams)
       tile.workGroupNumRows = tileParams[0]
       tile.workGroupNumCols = tileParams[1]
       tile.microTileNumRows = tileParams[2]
       tile.microTileNumCols = tileParams[3]
       tile.macroTileNumRows = tile.workGroupNumRows*tile.microTileNumRows
       tile.macroTileNumCols = tile.workGroupNumCols*tile.microTileNumCols
-      #print tile.getName()
+      #print(tile.getName())
       for unroll in unrolls[precision]:
         tile.unroll = unroll
         if tile.isValid():
           tiles.append( copy.copy(tile) )
         else:
-          print tile.getName() + " - SKIPPING - "
+          print(tile.getName() + " - SKIPPING - ")
 
     # add fallback tile
     tile.workGroupNumRows = fallbackTile[0]
@@ -116,7 +154,7 @@ def getTilesForPrecision(precision):
       if tile.isValid():
         tiles.append( copy.copy(tile) )
       else:
-        print tile.getName() + " - SKIPPING - "
+        print(tile.getName() + " - SKIPPING - ")
 
   setTiles = set(tiles)
   tiles = list( setTiles )

src/library/blas/AutoGemm/Includes.py

@@ -425,7 +425,7 @@ def writeToFile(self):
 # Write Includes
 ################################################################################
 def writeIncludes():
-  print "AutoGemm.py: Generating include files."
+  print("AutoGemm.py: Generating include files.")
   if not os.path.exists( Common.getIncludePath() ):
     os.makedirs( Common.getIncludePath() )
 
@@ -490,5 +490,5 @@ def writeIncludes():
   if len(sys.argv) == 2:
     Common.setOutputPath(sys.argv[1])
   else:
-    print "Warning: No output path specified; default is working directory."
+    print("Warning: No output path specified; default is working directory.")
   writeIncludes()