Merge branch 'develop' for 2.9.00

Part of Kokkos C++ Performance Portability Programming EcoSystem 2.9

CHANGELOG.md

@@ -1,5 +1,22 @@
 # Change Log
 
+## [2.9.00](https://github.com/kokkos/kokkos-kernels/tree/2.9.00) (2019-06-24)
+[Full Changelog](https://github.com/kokkos/kokkos-kernels/compare/2.8.00...2.9.00)
+
+**Implemented enhancements:**
+
+- KokkosBatched: Add specialization for float2, float4 and double4 [\#427](https://github.com/kokkos/kokkos-kernels/pull/427)
+- KokkosBatched: Reduce VectorLength (16 to 8) [\#432](https://github.com/kokkos/kokkos-kernels/pull/432)
+- KokkosBatched: Remove experimental name space for batched blas [\#371](https://github.com/kokkos/kokkos-kernels/issues/371)
+- Capability: Initial sparse triangular solve capability [\#435](https://github.com/kokkos/kokkos-kernels/pull/435)
+- Capability: Add support for MAGMA GESV TPL [\#409](https://github.com/kokkos/kokkos-kernels/pull/409)
+- cuBLAS: Add CudaUVMSpace specializations for GEMM [\#397](https://github.com/kokkos/kokkos-kernels/issues/397)
+
+**Fixed bugs:**
+
+- Deprecated Code Fixes [\#411](https://github.com/kokkos/kokkos-kernels/issues/411)
+- BuildSystem: Compilation error on rzansel [\#401](https://github.com/kokkos/kokkos-kernels/issues/401)
+
 ## [2.8.00](https://github.com/kokkos/kokkos-kernels/tree/2.8.00) (2019-02-05)
 [Full Changelog](https://github.com/kokkos/kokkos-kernels/compare/2.7.24...2.8.00)
 

CMakeLists.txt

@@ -290,23 +290,64 @@ ENDIF()
 # Enable Third Party Libraries
 # ==================================================================
 
-IF (TPL_ENABLE_BLAS)
-  SET(KOKKOSKERNELS_ENABLE_TPL_BLAS ${TPL_ENABLE_BLAS})
+IF (DEFINED KOKKOSKERNELS_ENABLE_TPL_BLAS)
+  # user overriding kokkoskernels
+  IF (KOKKOSKERNELS_ENABLE_TPL_BLAS)
+    IF (NOT TPL_ENABLE_BLAS)
+      MESSAGE( WARNING "KOKKOSKERNELS_ENABLE_TPL_BLAS is ON but TPL_ENABLE_BLAS is OFF. Please set TPL_ENABLE_BLAS:BOOL=ON")
+      SET(KOKKOSKERNELS_ENABLE_TPL_BLAS OFF)  
+    ENDIF()
+  ENDIF()
+ELSE()
+  # default behavior
+  IF (TPL_ENABLE_BLAS)
+    SET(KOKKOSKERNELS_ENABLE_TPL_BLAS ${TPL_ENABLE_BLAS})
+  ENDIF()
 ENDIF()
-IF (TPL_ENABLE_MKL)
-  SET(KOKKOSKERNELS_ENABLE_TPL_MKL ${TPL_ENABLE_MKL})
+
+IF (DEFINED KOKKOSKERNELS_ENABLE_TPL_MKL)
+  # user overriding kokkoskernels
+  IF (KOKKOSKERNELS_ENABLE_TPL_MKL)
+    IF (NOT TPL_ENABLE_MKL)
+      MESSAGE( WARNING "KOKKOSKERNELS_ENABLE_TPL_MKL is ON but TPL_ENABLE_MKL is OFF. Please set TPL_ENABLE_MKL:BOOL=ON")
+      SET(KOKKOSKERNELS_ENABLE_TPL_MKL OFF)  
+    ENDIF()
+  ENDIF()
+ELSE()
+  IF (TPL_ENABLE_MKL)
+    SET(KOKKOSKERNELS_ENABLE_TPL_MKL ${TPL_ENABLE_MKL})
+  ENDIF()
 ENDIF()
 
 IF(${Kokkos_ENABLE_Cuda})
-  IF (NOT KOKKOSKERNELS_ENABLE_TPL_BLAS)
-    SET(KOKKOSKERNELS_ENABLE_TPL_BLAS ON)
-    LIST(APPEND TPL_LIST "BLAS")
-  ENDIF()
   # CUBLAS is ON by default when CUDA is enabled
-  SET(KOKKOSKERNELS_ENABLE_TPL_CUBLAS ON)
+  IF (NOT DEFINED KOKKOSKERNELS_ENABLE_TPL_CUBLAS)  
+    SET(KOKKOSKERNELS_ENABLE_TPL_CUBLAS ON)
+  ENDIF()
+
   # Tribit provides TPL mechanism for CUSPARSE; thus, use it
-  IF (TPL_ENABLE_CUSPARSE)
-    SET(KOKKOSKERNELS_ENABLE_TPL_CUSPARSE ${TPL_ENABLE_CUSPARSE})
+  IF (DEFINED KOKKOSKERNELS_ENABLE_TPL_CUSPARSE)  
+    IF (NOT TPL_ENABLE_CUSPARSE)
+      MESSAGE( WARNING "KOKKOSKERNELS_ENABLE_TPL_CUSPARSE is ON but TPL_ENABLE_CUSPARSE is OFF. Please set TPL_ENABLE_CUSPARSE:BOOL=ON")
+      SET(KOKKOSKERNELS_ENABLE_TPL_CUSPARSE OFF)  
+    ENDIF()
+  ELSE()
+    IF (TPL_ENABLE_CUSPARSE)
+      SET(KOKKOSKERNELS_ENABLE_TPL_CUSPARSE ${TPL_ENABLE_CUSPARSE})
+    ENDIF()
+  ENDIF()
+
+  IF (DEFINED KOKKOSKERNELS_ENABLE_TPL_MAGMA)
+    IF (KOKKOSKERNELS_ENABLE_TPL_MAGMA)  
+      IF (NOT TPL_ENABLE_MAGMA)
+        MESSAGE( WARNING "KOKKOSKERNELS_ENABLE_TPL_MAGMA is ON but TPL_ENABLE_MAGMA is OFF. Please set TPL_ENABLE_MAGMA:BOOL=ON")
+        SET(KOKKOSKERNELS_ENABLE_TPL_MAGMA OFF)  
+      ENDIF()
+    ENDIF()
+  ELSE()
+    IF (TPL_ENABLE_MAGMA)
+      SET(KOKKOSKERNELS_ENABLE_TPL_MAGMA ${TPL_ENABLE_MAGMA})
+    ENDIF()
   ENDIF()
 ENDIF()
 
@@ -322,6 +363,18 @@ ENDIF()
 IF (KOKKOSKERNELS_ENABLE_TPL_CUBLAS)
   LIST(APPEND TPL_LIST "CUBLAS")
 ENDIF()
+IF (KOKKOSKERNELS_ENABLE_TPL_MAGMA)
+  LIST(APPEND TPL_LIST "MAGMA")
+  IF (F77_BLAS_MANGLE STREQUAL "(name,NAME) name ## _")
+    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DADD_ -fopenmp -lgfortran")
+  ELSEIF (F77_BLAS_MANGLE STREQUAL "(name,NAME) NAME")
+    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUPCASE -fopenmp -lgfortran")
+  ELSEIF (F77_BLAS_MANGLE STREQUAL "(name,NAME) name")
+    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DNOCHANGE -fopenmp -lgfortran")
+  ELSE ()
+    MESSAGE(FATAL_ERROR "F77_BLAS_MANGLE ${F77_BLAS_MANGLE} detected while MAGMA only accepts Fortran mangling that is one of single underscore (-DADD_), uppercase (-DUPCASE), and no change (-DNOCHANGE)")
+  ENDIF()
+ENDIF()
 
 # ==================================================================
 # Fortran Complex BLAS

Makefile.kokkos-kernels

@@ -12,7 +12,7 @@ KOKKOSKERNELS_SPACES ?= $(KOKKOS_DEVICES),HostSpace
 KOKKOSKERNELS_ORDINALS ?= int,int64_t
 KOKKOSKERNELS_OFFSETS ?= int,size_t
 
-# mkl, blas, yaml
+# mkl, blas, yaml, magma
 KOKKOSKERNELS_ENABLE_TPLS ?= ""
 
 # eti-only
@@ -23,6 +23,7 @@ KOKKOSKERNELS_INTERNAL_ENABLE_BLAS := $(strip $(shell echo $(KOKKOSKERNELS_ENABL
 KOKKOSKERNELS_INTERNAL_ENABLE_MKL := $(strip $(shell echo $(KOKKOSKERNELS_ENABLE_TPLS) | grep "mkl" | wc -l))
 KOKKOSKERNELS_INTERNAL_ENABLE_CUSPARSE := $(strip $(shell echo $(KOKKOSKERNELS_ENABLE_TPLS) | grep "cusparse" | wc -l))
 KOKKOSKERNELS_INTERNAL_ENABLE_CUBLAS := $(strip $(shell echo $(KOKKOSKERNELS_ENABLE_TPLS) | grep "cublas" | wc -l))
+KOKKOSKERNELS_INTERNAL_ENABLE_MAGMA := $(strip $(shell echo $(KOKKOSKERNELS_ENABLE_TPLS) | grep "magma" | wc -l))
 
 YAML_PATH ?= ${HOME}/yaml-cpp/install
 YAML_LIBS ?= -L${YAML_PATH}/lib -lyaml-cpp
@@ -33,6 +34,8 @@ BLAS_LIBS ?= -L${BLAS_PATH}/lib64 $(BLAS_LIBRARIES) -lgfortran -lgomp
 MKL_PATH ?= ${SEMS_INTEL_ROOT}/mkl
 MKL_LIBS ?= -L${MKL_PATH}/lib/intel64 -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -liomp5 -lpthread -lm -ldl -L${MKL_PATH}/../compiler/lib/intel64
 
+MAGMA_PATH     ?= ${MAGMADIR}
+
 default-kokkos-kernels: kokkoskernels-build-lib
 	echo "Start Build"
 
@@ -77,17 +80,30 @@ tmp := $(shell echo "\#define KOKKOSKERNELS_CONFIG_H_" >> KokkosKernels_config.t
 
 # default mangling scheme with a single under score
 KOKKOSKERNELS_FORTRAN_GLOBAL = name\#\#_
+ifeq (${KOKKOSKERNELS_INTERNAL_ENABLE_MAGMA}, 1)
+  MAGMA_FORTRAN_MANGLING = ADD_
+endif
+
 KOKKOSKERNELS_INTERNAL_OVERRIDE_FORTRAN_MANGLING_WITH_DOUBLE_UNDERSCORES := $(strip $(shell echo $(KOKKOSKERNELS_OPTIONS) | grep "blas-mangle__" | wc -l))
 ifeq ($(KOKKOSKERNELS_INTERNAL_OVERRIDE_FORTRAN_MANGLING_WITH_DOUBLE_UNDERSCORES), 1)
   KOKKOSKERNELS_FORTRAN_GLOBAL = name\#\#__
+  ifeq (${KOKKOSKERNELS_INTERNAL_ENABLE_MAGMA}, 1)
+    $(error MAGMA does not prefer Fortran mangling with double underscore)
+  endif
 else
   KOKKOSKERNELS_INTERNAL_OVERRIDE_FORTRAN_MANGLING_WITH_SINGLE_UNDERSCORE := $(strip $(shell echo $(KOKKOSKERNELS_OPTIONS) | grep "blas-mangle_" | wc -l))
   ifeq ($(KOKKOSKERNELS_INTERNAL_OVERRIDE_FORTRAN_MANGLING_WITH_SINGLE_UNDERSCORE), 1)
     KOKKOSKERNELS_FORTRAN_GLOBAL = name\#\#_
+	ifeq (${KOKKOSKERNELS_INTERNAL_ENABLE_MAGMA}, 1)
+      MAGMA_FORTRAN_MANGLING = ADD_
+    endif
   else 
     KOKKOSKERNELS_INTERNAL_OVERRIDE_FORTRAN_MANGLING_WITH_NO_UNDERSCORE := $(strip $(shell echo $(KOKKOSKERNELS_OPTIONS) | grep "blas-mangle" | wc -l))
     ifeq ($(KOKKOSKERNELS_INTERNAL_OVERRIDE_FORTRAN_MANGLING_WITH_NO_UNDERSCORE), 1)
       KOKKOSKERNELS_FORTRAN_GLOBAL = name\#\#
+      ifeq (${KOKKOSKERNELS_INTERNAL_ENABLE_MAGMA}, 1)
+        MAGMA_FORTRAN_MANGLING = NOCHANGE
+      endif
     endif
   endif
 endif
@@ -306,6 +322,16 @@ ifeq (${KOKKOSKERNELS_INTERNAL_ENABLE_CUBLAS}, 1)
   tmp := $(shell echo "\#define KOKKOSKERNELS_ENABLE_TPL_CUBLAS" >> KokkosKernels_config.tmp )
 endif
 
+ifeq (${KOKKOSKERNELS_INTERNAL_ENABLE_MAGMA}, 1)
+  KOKKOS_INTERNAL_NVCC_PATH := $(shell which nvcc)
+  CUDA_PATH ?= $(KOKKOS_INTERNAL_NVCC_PATH:/bin/nvcc=)
+  KOKKOSKERNELS_CPPFLAGS += -D${MAGMA_FORTRAN_MANGLING} -I${MAGMA_PATH}/include -I${MAGMA_PATH}/sparse/include -I${CUDA_PATH}/include
+  KOKKOSKERNELS_LDFLAGS += -L${MAGMA_PATH}/lib -L${CUDA_PATH}/lib64
+  KOKKOSKERNELS_LIBS += -lmagma_sparse -lmagma -lcublas -lcudart -lcusparse -fopenmp
+
+  tmp := $(shell echo "\#define KOKKOSKERNELS_ENABLE_TPL_MAGMA" >> KokkosKernels_config.tmp )
+endif
+
   tmp := $(shell echo "\#if defined(KOKKOSKERNELS_ENABLE_TPL_MKL)" >> KokkosKernels_config.tmp )
   tmp := $(shell echo "\#if !defined(KOKKOSKERNELS_ENABLE_TPL_BLAS)" >> KokkosKernels_config.tmp )  
   tmp := $(shell echo "\#define KOKKOSKERNELS_ENABLE_TPL_BLAS" >> KokkosKernels_config.tmp )
@@ -435,6 +461,10 @@ ifeq (${KOKKOSKERNELS_INTERNAL_ENABLE_CUBLAS}, 1)
   KOKKOSKERNELS_INTERNAL_SRC_BLAS += ${KOKKOSKERNELS_PATH}/src/impl/tpls/KokkosBlas_Cuda_tpl.cpp
 endif
 
+ifeq (${KOKKOSKERNELS_INTERNAL_ENABLE_MAGMA}, 1)
+  KOKKOSKERNELS_INTERNAL_SRC_BLAS += ${KOKKOSKERNELS_PATH}/src/impl/tpls/KokkosBlas_Cuda_tpl.cpp
+endif
+
 KOKKOSKERNELS_INTERNAL_ENABLE_HOST_BLAS=0
 ifeq (${KOKKOSKERNELS_INTERNAL_ENABLE_BLAS}, 1)
   KOKKOSKERNELS_INTERNAL_ENABLE_HOST_BLAS=1
@@ -456,15 +486,24 @@ endif
 
 DEPFLAGS = -M
 
+CP = cp
+
 ifeq ($(KOKKOS_OS),CYGWIN)
   COPY_FLAG = -u
 endif
 ifeq ($(KOKKOS_OS),Linux)
   COPY_FLAG = -u
 endif
-ifeq ($(KOKKOS_OS),Darwin)
-  COPY_FLAG =
-endif
+#ifeq ($(KOKKOS_OS),Darwin)
+  # If using Homebrew on OSX with the 'coreutils' package installed
+  # we have a gnu version of cp which has the -u option
+#  ifeq (,$(wildcard "/usr/local/opt/coreutils/libexec/gnubin/cp"))
+#    CP = /usr/local/opt/coreutils/libexec/gnubin/cp
+#    COPY_FLAG = -u
+#  else
+#    COPY_FLAG =
+#  endif
+#endif
 
 KOKKOSKERNELS_INTERNAL_SRC_BLAS_NODIR = $(notdir $(KOKKOSKERNELS_INTERNAL_SRC_BLAS))
 KOKKOSKERNELS_INTERNAL_SRC_SPARSE_NODIR = $(notdir $(KOKKOSKERNELS_INTERNAL_SRC_SPARSE))

cmake/Dependencies.cmake

@@ -1,5 +1,5 @@
 TRIBITS_PACKAGE_DEFINE_DEPENDENCIES(
   LIB_REQUIRED_PACKAGES KokkosCore KokkosContainers KokkosAlgorithms 
-  LIB_OPTIONAL_TPLS quadmath MKL BLAS LAPACK CUSPARSE
+  LIB_OPTIONAL_TPLS quadmath MKL BLAS LAPACK CUSPARSE MAGMA
   TEST_OPTIONAL_TPLS yaml-cpp
 )

cmake/KokkosKernels_config.h.in

@@ -87,6 +87,8 @@
 #cmakedefine KOKKOSKERNELS_ENABLE_TPL_CUSPARSE
 /* CUBLAS */
 #cmakedefine KOKKOSKERNELS_ENABLE_TPL_CUBLAS
+/* MAGMA */
+#cmakedefine KOKKOSKERNELS_ENABLE_TPL_MAGMA
 
 /* if MKL, BLAS is also defined */
 #if defined(KOKKOSKERNELS_ENABLE_TPL_MKL)