Single precision assembly kernel for ELPA2 real and complex

The single precision version of the SSE assembly kernel is about 1.8
times faster than the double precision version

Makefile.am

@@ -81,10 +81,17 @@ if WITH_REAL_BGQ_KERNEL
 endif
 
 if WITH_REAL_SSE_KERNEL
-  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64.s
+  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64_double_precision.s
+if WANT_SINGLE_PRECISION_REAL
+  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64_single_precision.s
+endif
 else
 if WITH_COMPLEX_SSE_KERNEL
-  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64.s
+  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64_double_precision.s
+if WANT_SINGLE_PRECISION_COMPLEX
+  libelpa@SUFFIX@_la_SOURCES += src/elpa2_kernels/elpa2_kernels_asm_x86_64_single_precision.s
+endif
+
 endif
 endif
 
@@ -332,6 +339,7 @@ elpa2_test_real_single_precision@SUFFIX@_LDADD = $(build_lib)
 
 elpa2_test_real_default_kernel_single_precision@SUFFIX@_SOURCES = test/fortran_test_programs/test_real2_default_kernel_single_precision.F90 $(shared_sources) $(redirect_sources)
 elpa2_test_real_default_kernel_single_precision@SUFFIX@_LDADD = $(build_lib)
+#elpa2_test_real_default_kernel_single_precision@SUFFIX@_LDFLAGS = -static
 
 elpa2_test_real_default_kernel_qr_decomposition_single_precision@SUFFIX@_SOURCES = test/fortran_test_programs/test_real2_default_kernel_qr_decomposition_single_precision.F90 $(shared_sources) $(redirect_sources)
 elpa2_test_real_default_kernel_qr_decomposition_single_precision@SUFFIX@_LDADD = $(build_lib)
@@ -349,6 +357,7 @@ elpa2_test_complex_single_precision@SUFFIX@_LDADD = $(build_lib)
 
 elpa2_test_complex_default_kernel_single_precision@SUFFIX@_SOURCES = test/fortran_test_programs/test_complex2_default_kernel_single_precision.F90 $(shared_sources) $(redirect_sources)
 elpa2_test_complex_default_kernel_single_precision@SUFFIX@_LDADD = $(build_lib)
+#elpa2_test_complex_default_kernel_single_precision@SUFFIX@_LDFLAGS = -static
 
 elpa2_test_complex_choose_kernel_with_api_single_precision@SUFFIX@_SOURCES = test/fortran_test_programs/test_complex2_choose_kernel_with_api_single_precision.F90 $(shared_sources) $(redirect_sources)
 elpa2_test_complex_choose_kernel_with_api_single_precision@SUFFIX@_LDADD = $(build_lib)

configure.ac

@@ -101,6 +101,14 @@ AM_PROG_AR
 AM_PROG_AS
 
 # Fortran
+
+dnl check whether single precision is requested
+AC_MSG_CHECKING(whether ELPA library should contain also single precision functions)
+AC_ARG_ENABLE(single-precision,[AS_HELP_STRING([--enable-single-precision],
+	                                   [build with single precision])],
+              want_single_precision="yes", want_single_precision="no")
+AC_MSG_RESULT([${want_single_precision}])
+
 AC_LANG([Fortran])
 m4_include([m4/ax_prog_fc_mpi.m4])
 AX_PROG_FC_MPI([test x"$enable_shared_memory_only" = xno],[use_mpi=yes],[use_mpi=no])
@@ -190,9 +198,9 @@ if test x"${with_ftimings}" = x"yes"; then
 fi
 AM_CONDITIONAL([HAVE_DETAILED_TIMINGS],[test x"$with_ftimings" = x"yes"])
 
-AC_MSG_CHECKING(whether SSE assembler kernel can be compiled)
+AC_MSG_CHECKING(whether double-precision SSE assembler kernel can be compiled)
 
-$CC -c $srcdir/src/elpa2_kernels/elpa2_kernels_asm_x86_64.s -o test.o 2>/dev/null
+$CC -c $srcdir/src/elpa2_kernels/elpa2_kernels_asm_x86_64_double_precision.s -o test.o 2>/dev/null
 if test "$?" == 0; then
   can_compile_sse=yes
   install_real_sse=yes
@@ -205,6 +213,26 @@ fi
 rm -f ./test.o
 AC_MSG_RESULT([${can_compile_sse}])
 
+if test x"${want_single_precision}" = x"yes" ; then
+  AC_MSG_CHECKING(whether single-precision SSE assembler kernel can be compiled)
+
+  $CC -c $srcdir/src/elpa2_kernels/elpa2_kernels_asm_x86_64_double_precision.s -o test.o 2>/dev/null
+  if test "$?" == 0; then
+    can_compile_sse=yes
+    install_real_sse=yes
+    install_complex_sse=yes
+  else
+    can_compile_sse=no
+    install_real_sse=no
+    install_complex_sse=no
+  fi
+  rm -f ./test.o
+  AC_MSG_RESULT([${can_compile_sse}])
+  if test x"${can_compile_sse}" = x"no" ; then
+    AC_MSG_WARN([Cannot compile single-precision SSE kernel: disabling SSE kernels alltogether])
+  fi
+fi
+
 dnl check whether one can compile with avx - gcc intrinsics
 
 dnl first pass: try with specified CFLAGS and CXXFLAGS
@@ -522,14 +550,6 @@ if test x"${want_gpu}" = x"yes" ; then
   can_compile_gpu=yes
 fi
 
-dnl check whether single precision is requested
-AC_MSG_CHECKING(whether ELPA library should contain also single precision functions)
-AC_ARG_ENABLE(single-precision,[AS_HELP_STRING([--enable-single-precision],
-	                                   [build with single precision])],
-              want_single_precision="yes", want_single_precision="no")
-AC_MSG_RESULT([${want_single_precision}])
-
-
 dnl now check which kernels can be compiled
 
 dnl the checks for SSE were already done before
@@ -829,3 +849,7 @@ if test "${can_compile_avx2}" = "no" ; then
     AC_MSG_WARN([Could not compile AVX2 instructions])
   fi
 fi
+
+if test "${can_compile_sse}" = "no" ; then
+  AC_MSG_WARN([Could not compile SSE instructions])
+fi

src/elpa2.F90

@@ -301,12 +301,14 @@ contains
 
     ! some temporarilly checks until single precision works with all kernels
 #ifndef DOUBLE_PRECISION_REAL
-    if ( (THIS_REAL_ELPA_KERNEL .ne. REAL_ELPA_KERNEL_GENERIC) .or. &
-         (THIS_REAL_ELPA_KERNEL .ne. REAL_ELPA_KERNEL_GENERIC_SIMPLE) .or. &
-         (THIS_REAL_ELPA_KERNEL .ne. REAL_ELPA_KERNEL_GPU) ) then
-       print *,"At the moment single precision only works with the generic kernels"
-       stop
-     endif
+    if ( (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC) .or. &
+         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC_SIMPLE) .or. &
+         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE) .or. &
+         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GPU) ) then
+    else
+      print *,"At the moment single precision only works with the generic kernels"
+      stop
+    endif
 #endif
 
     ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32
@@ -564,7 +566,6 @@ contains
     call mpi_comm_rank(mpi_comm_cols,my_pcol,mpierr)
     call mpi_comm_size(mpi_comm_cols,np_cols,mpierr)
 
-
     wantDebug = .false.
     if (firstCall) then
       ! are debug messages desired?
@@ -641,14 +642,7 @@ contains
         stop
       endif
     ! some temporarilly checks until single precision works with all kernels
-#ifndef DOUBLE_PRECISION_REAL
-    if ( (THIS_REAL_ELPA_KERNEL .ne. REAL_ELPA_KERNEL_GENERIC) .or. &
-         (THIS_REAL_ELPA_KERNEL .ne. REAL_ELPA_KERNEL_GENERIC_SIMPLE) .or. &
-         (THIS_REAL_ELPA_KERNEL .ne. REAL_ELPA_KERNEL_GPU) ) then
-       print *,"At the moment single precision only works with the generic kernels"
-       stop
-     endif
-#endif
+
       ! set the neccessary parameters
       cudaMemcpyHostToDevice   = cuda_memcpyHostToDevice()
       cudaMemcpyDeviceToHost   = cuda_memcpyDeviceToHost()
@@ -657,6 +651,16 @@ contains
       cudaHostRegisterMapped   = cuda_hostRegisterMapped()
     endif
 
+#ifndef DOUBLE_PRECISION_REAL
+    if ( (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC) .or. &
+         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC_SIMPLE) .or. &
+         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE) .or. &
+         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GPU) ) then
+    else
+      print *,"At the moment single precision only works with the generic kernels"
+      stop
+    endif
+#endif
     ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32
     ! On older systems (IBM Bluegene/P, Intel Nehalem) a value of 32 was optimal.
     ! For Intel(R) Xeon(R) E5 v2 and v3, better use 64 instead of 32!
@@ -943,12 +947,15 @@ function solve_evp_complex_2stage_single(na, nev, a, lda, ev, q, ldq, nblk, &
       endif
       THIS_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
     endif
+
 #ifndef DOUBLE_PRECISION_COMPLEX
-    if ( (THIS_COMPLEX_ELPA_KERNEL .ne. COMPLEX_ELPA_KERNEL_GENERIC) .or. &
-         (THIS_COMPLEX_ELPA_KERNEL .ne. COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE) ) then
-       print *,"At the moment single precision only works with the generic kernels"
-       stop
-     endif
+    if ( (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_GENERIC) .or. &
+         (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE) .or. &
+         (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_SSE) ) then
+    else
+      print *,"At the moment single precision only works with the generic kernels"
+      stop
+    endif
 #endif
     if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_GPU) then
       if (check_for_gpu(my_pe, numberOfGPUDevices, wantDebug=wantDebug)) then
@@ -1266,11 +1273,13 @@ function solve_evp_complex_2stage_single(na, nev, a, lda, ev, q, ldq, nblk, &
       THIS_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
     endif
 #ifndef DOUBLE_PRECISION_COMPLEX
-    if ( (THIS_COMPLEX_ELPA_KERNEL .ne. COMPLEX_ELPA_KERNEL_GENERIC) .or. &
-         (THIS_COMPLEX_ELPA_KERNEL .ne. COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE) ) then
-       print *,"At the moment single precision only works with the generic kernels"
-       stop
-     endif
+    if ( (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_GENERIC) .or. &
+        (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE)  .or. &
+        (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_SSE) ) then
+    else
+      print *,"At the moment single precision only works with the generic kernels"
+      stop
+    endif
 #endif
     if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_GPU) then
       if (check_for_gpu(my_pe, numberOfGPUDevices, wantDebug=wantDebug)) then

src/elpa2_kernels/elpa2_kernels_real.F90

@@ -770,8 +770,8 @@ module real_generic_kernel
     real(kind=rk4)                   :: x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, &
                                        y1, y2, y3, y4, y5, y6, y7, y8, y9, y10, y11, y12
 #endif
-    real(kind=rk8)                   :: h1, h2, tau1, tau2
-    integer(kind=ik)                :: i
+    real(kind=rk4)                   :: h1, h2, tau1, tau2
+    integer(kind=ik)                 :: i
 
 #ifdef HAVE_DETAILED_TIMINGS
     call timer%start("kernel generic: hh_trafo_kernel_12_generic_single")
@@ -984,8 +984,8 @@ module real_generic_kernel
     real(kind=rk4)                    :: x1, x2, x3, x4, x5, x6, x7, x8, &
                                         y1, y2, y3, y4, y5, y6, y7, y8
 #endif
-    real(kind=rk8)                    :: h1, h2, tau1, tau2
-    integer(kind=ik)                 :: i
+    real(kind=rk4)                    :: h1, h2, tau1, tau2
+    integer(kind=ik)                  :: i
 
 #ifdef HAVE_DETAILED_TIMINGS
     call timer%start("kernel generic: hh_trafo_kernel_8_generic_single")
@@ -1153,7 +1153,7 @@ module real_generic_kernel
     real(kind=rk4)                   :: x1, x2, x3, x4, y1, y2, y3, y4
 #endif
     real(kind=rk4)                   :: h1, h2, tau1, tau2
-    integer(kind=ik)                :: i
+    integer(kind=ik)                 :: i
 
 #ifdef HAVE_DETAILED_TIMINGS
     call timer%start("kernel generic: hh_trafo_kernel_4_generic_single")

src/mod_compute_hh_trafo_complex.F90

@@ -492,13 +492,11 @@ module compute_hh_trafo_complex
              ttt = mpi_wtime()
              do j = ncols, 1, -1
 #ifdef WITH_OPENMP
- print *,"at the moment no sse single-precision kernel"
- stop
-!              call single_hh_trafo_complex_single(a(1,j+off+a_off,istripe,my_thread), &
-!                                           bcast_buffer(1,j+off),nbw,nl,stripe_width)
+              call single_hh_trafo_complex_single(a(1,j+off+a_off,istripe,my_thread), &
+                                           bcast_buffer(1,j+off),nbw,nl,stripe_width)
 #else
-!              call single_hh_trafo_complex_single(a(1,j+off+a_off,istripe), &
-!                                           bcast_buffer(1,j+off),nbw,nl,stripe_width)
+              call single_hh_trafo_complex_single(a(1,j+off+a_off,istripe), &
+                                           bcast_buffer(1,j+off),nbw,nl,stripe_width)
 #endif
             enddo
 #if defined(WITH_NO_SPECIFIC_COMPLEX_KERNEL)

src/mod_compute_hh_trafo_real.F90

@@ -708,14 +708,12 @@ module compute_hh_trafo_real
                do j = ncols, 2, -2
                  w(:,1) = bcast_buffer(1:nbw,j+off)
                  w(:,2) = bcast_buffer(1:nbw,j+off-1)
-print *,"at the moment no single-precision sse kernel"
-stop
 #ifdef WITH_OPENMP
-!                 call double_hh_trafo_single(a(1,j+off+a_off-1,istripe,my_thread), w, nbw, nl, &
-!                                      stripe_width, nbw)
+                 call double_hh_trafo_single(a(1,j+off+a_off-1,istripe,my_thread), w, nbw, nl, &
+                                      stripe_width, nbw)
 #else
-!                 call double_hh_trafo_single(a(1,j+off+a_off-1,istripe), w, nbw, nl, &
-!                                      stripe_width, nbw)
+                 call double_hh_trafo_single(a(1,j+off+a_off-1,istripe), w, nbw, nl, &
+                                      stripe_width, nbw)
 #endif
                enddo
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)

test/shared_sources/read_input_parameters.F90

@@ -94,7 +94,7 @@ module mod_read_input_parameters
         read(arg3, *) nblk
 
         if (arg4 .eq. "output_eigenvalues") then
-          write_to_file%eigenvectors = .true.
+          write_to_file%eigenvalues = .true.
         else
           write(error_unit, *) "Invalid value for output flag! Must be ""output_eigenvalues"" or omitted"
           stop 1