Re-integrate some changes from NVIDIA

configure.ac

@@ -428,7 +428,7 @@ AC_MSG_RESULT([${fortran_can_check_environment}])
 
 dnl check whether GPU version is requested
 
-CUDA_INSTALL_PATH="/usr/local/cuda/"
+#CUDA_INSTALL_PATH="/usr/local/cuda/"
 #CUDA_SDK_INSTALL_PATH="/usr/local/NVIDIA_GPU_Computing_SDK"
 
 AC_MSG_CHECKING(whether GPU support is requested)

m4/ltsugar.m4

 # ltsugar.m4 -- libtool m4 base layer.                         -*-Autoconf-*-
 #
-# Copyright (C) 2004, 2005, 2007, 2008 Free Software Foundation, Inc.
+# Copyright (C) 2004-2005, 2007-2008, 2011-2015 Free Software
+# Foundation, Inc.
 # Written by Gary V. Vaughan, 2004
 #
 # This file is free software; the Free Software Foundation gives
@@ -33,7 +34,7 @@ m4_define([_lt_join],
 # ------------
 # Manipulate m4 lists.
 # These macros are necessary as long as will still need to support
-# Autoconf-2.59 which quotes differently.
+# Autoconf-2.59, which quotes differently.
 m4_define([lt_car], [[$1]])
 m4_define([lt_cdr],
 [m4_if([$#], 0, [m4_fatal([$0: cannot be called without arguments])],
@@ -44,7 +45,7 @@ m4_define([lt_unquote], $1)
 
 # lt_append(MACRO-NAME, STRING, [SEPARATOR])
 # ------------------------------------------
-# Redefine MACRO-NAME to hold its former content plus `SEPARATOR'`STRING'.
+# Redefine MACRO-NAME to hold its former content plus 'SEPARATOR''STRING'.
 # Note that neither SEPARATOR nor STRING are expanded; they are appended
 # to MACRO-NAME as is (leaving the expansion for when MACRO-NAME is invoked).
 # No SEPARATOR is output if MACRO-NAME was previously undefined (different

m4/ltversion.m4

 # ltversion.m4 -- version numbers			-*- Autoconf -*-
 #
-#   Copyright (C) 2004 Free Software Foundation, Inc.
+#   Copyright (C) 2004, 2011-2015 Free Software Foundation, Inc.
 #   Written by Scott James Remnant, 2004
 #
 # This file is free software; the Free Software Foundation gives
@@ -9,15 +9,15 @@
 
 # @configure_input@
 
-# serial 3337 ltversion.m4
+# serial 4179 ltversion.m4
 # This file is part of GNU Libtool
 
-m4_define([LT_PACKAGE_VERSION], [2.4.2])
-m4_define([LT_PACKAGE_REVISION], [1.3337])
+m4_define([LT_PACKAGE_VERSION], [2.4.6])
+m4_define([LT_PACKAGE_REVISION], [2.4.6])
 
 AC_DEFUN([LTVERSION_VERSION],
-[macro_version='2.4.2'
-macro_revision='1.3337'
+[macro_version='2.4.6'
+macro_revision='2.4.6'
 _LT_DECL(, macro_version, 0, [Which release of libtool.m4 was used?])
 _LT_DECL(, macro_revision, 0)
 ])

m4/lt~obsolete.m4

 # lt~obsolete.m4 -- aclocal satisfying obsolete definitions.    -*-Autoconf-*-
 #
-#   Copyright (C) 2004, 2005, 2007, 2009 Free Software Foundation, Inc.
+#   Copyright (C) 2004-2005, 2007, 2009, 2011-2015 Free Software
+#   Foundation, Inc.
 #   Written by Scott James Remnant, 2004.
 #
 # This file is free software; the Free Software Foundation gives
@@ -11,7 +12,7 @@
 
 # These exist entirely to fool aclocal when bootstrapping libtool.
 #
-# In the past libtool.m4 has provided macros via AC_DEFUN (or AU_DEFUN)
+# In the past libtool.m4 has provided macros via AC_DEFUN (or AU_DEFUN),
 # which have later been changed to m4_define as they aren't part of the
 # exported API, or moved to Autoconf or Automake where they belong.
 #
@@ -25,7 +26,7 @@
 # included after everything else.  This provides aclocal with the
 # AC_DEFUNs it wants, but when m4 processes it, it doesn't do anything
 # because those macros already exist, or will be overwritten later.
-# We use AC_DEFUN over AU_DEFUN for compatibility with aclocal-1.6. 
+# We use AC_DEFUN over AU_DEFUN for compatibility with aclocal-1.6.
 #
 # Anytime we withdraw an AC_DEFUN or AU_DEFUN, remember to add it here.
 # Yes, that means every name once taken will need to remain here until

src/ev_tridi_band_gpu_c_v2.cu

@@ -16,6 +16,7 @@ static __device__ __forceinline__ double shfl_xor(double r, int mask)
     return __hiloint2double(hi, lo);
 }
 
+#if 0
 static __device__ __forceinline__ cuDoubleComplex  shfl_xor_complex(cuDoubleComplex r, int mask)
 {
     double real = cuCreal(r) ;
@@ -35,7 +36,7 @@ static __device__ __forceinline__ cuDoubleComplex  shfl_xor_complex(cuDoubleComp
     return       make_cuDoubleComplex(real, imag);
 
 }
-
+#endif
 
 
 // Perform the equivalent of "__shfl_down" on an 8-byte value

src/interface_cuda.F90

@@ -67,6 +67,13 @@ module cuda_routines
       integer(C_INT)           :: istat
     end function cuda_setdevice
 
+function cuda_getdevicecount(n) result(istat) &
+        bind(C, name="cudaGetDeviceCount")
+    use iso_c_binding
+    integer, intent(out)      :: n
+    integer(C_INT)                      :: istat
+end function cuda_getdevicecount
+
     function cuda_ProfilerStart() result(istat)&
             bind (C, name="cudaProfilerStart")
 

test/test_complex2.F90

@@ -149,7 +149,7 @@ program test_complex2
 
 #ifdef WITH_GPU_VERSION
    character(len=1024)     :: envname
-   integer                 :: istat, devnum
+   integer                 :: istat, devnum, numdevs
 #endif
 
    write_to_file = .false.
@@ -165,13 +165,23 @@ program test_complex2
    !  MPI Initialization
    call setup_mpi(myid, nprocs)
 #ifdef WITH_GPU_VERSION
-   devnum = 0
+   istat = cuda_getdevicecount(numdevs)
+   if (istat .ne. 0) then
+     print *,"Error in cuda_getdevicecount"
+     stop
+   endif
+   if(myid==0) then
+       print *
+       print '(3(a,i0))','Found ', numdevs, ' GPUs'
+   endif
+   devnum = mod(myid, numdevs)
    istat = cuda_setdevice(devnum)
 
    if (istat .ne. 0) then
      print *,"Cannot set CudaDevice"
      stop
    endif
+   print '(3(a,i0))', 'MPI rank ', myid, ' uses GPU #', devnum
 #endif
 
    STATUS = 0
@@ -345,8 +355,8 @@ program test_complex2
    ! Calculate eigenvalues/eigenvectors
 
    call mpi_barrier(mpi_comm_world, mpierr) ! for correct timings only
-   success = solve_evp_complex_2stage(na, nev, a, na_rows, ev, z, na_rows, nblk, &
-                                 na_rows, na_cols, mpi_comm_rows, mpi_comm_cols, mpi_comm_world)
+   success = solve_evp_complex_2stage(na, nev, a, na_rows, ev, z, na_rows, na_cols, nblk, &
+                                 mpi_comm_rows, mpi_comm_cols, mpi_comm_world)
 
    if (.not.(success)) then
       write(error_unit,*) "solve_evp_complex_2stage produced an error! Aborting..."

test/test_real2.F90

@@ -143,7 +143,7 @@ program test_real2
 
 #ifdef WITH_GPU_VERSION
    character(len=1024)     :: envname
-   integer                 :: istat, devnum
+   integer                 :: istat, devnum, numdevs
 #endif
    write_to_file = .false.
    success = .true.
@@ -160,13 +160,26 @@ program test_real2
    call setup_mpi(myid, nprocs)
 
 #ifdef WITH_GPU_VERSION
-   devnum = 0
+   ! call getenv("CUDA_PROXY_PIPE_DIRECTORY", envname)
+   istat = cuda_getdevicecount(numdevs)
+   if (istat .ne. 0) then
+     print *,"error in cuda_getdevicecount"
+     stop
+   endif
+
+   if (myid==0) then
+     print *
+     print '(3(a,i0))','Found ', numdevs, ' GPUs'
+   endif
+
+   devnum = mod(myid, numdevs)
    istat = cuda_setdevice(devnum)
 
    if (istat .ne. 0) then
      print *,"Cannot set CudaDevice"
      stop
    endif
+   print '(3(a,i0))', 'MPI rank ', myid, ' uses GPU #', devnum
 #endif
 
    STATUS = 0
@@ -341,7 +354,7 @@ program test_real2
    end if
 
    call mpi_barrier(mpi_comm_world, mpierr) ! for correct timings only
-   success = solve_evp_real_2stage(na, nev, a, na_rows, ev, z, na_rows, nblk, na_rows, na_cols, &
+   success = solve_evp_real_2stage(na, nev, a, na_rows, ev, z, na_rows, nblk, &
                               mpi_comm_rows, mpi_comm_cols, mpi_comm_world)
 
    if (.not.(success)) then