added SSE,AVX,FMA4 tri to band real kernels for x86 based systems

ELPA_development_version/src/elpa2.f90

@@ -1527,6 +1527,7 @@ subroutine trans_ev_tridi_to_band_real(na, nev, nblk, nbw, q, ldq, mpi_comm_rows
 
     a_dim2 = max_blk_size + nbw
 
+!DEC$ ATTRIBUTES ALIGN: 64:: a
     allocate(a(stripe_width,a_dim2,stripe_count,max_threads))
     ! a(:,:,:,:) should be set to 0 in a parallel region, not here!
 
@@ -2014,8 +2015,8 @@ contains
 
         ! Private variables in OMP regions (my_thread) should better be in the argument list!
         integer, intent(in) :: off, ncols, istripe, my_thread
-        integer j, nl, noff
-        real*8 w(nbw,2), ttt
+        integer j, jj, jjj, nl, noff
+        real*8 w(nbw,6), ttt
 
         ttt = mpi_wtime()
         if(istripe<stripe_count) then
@@ -2025,12 +2026,86 @@ contains
           nl = min(my_thread*thread_width-noff, l_nev-noff)
           if(nl<=0) return
         endif
+        
+        !FORTRAN CODE
         do j = ncols, 2, -2
             w(:,1) = bcast_buffer(1:nbw,j+off)
             w(:,2) = bcast_buffer(1:nbw,j+off-1)
             call double_hh_trafo(a(1,j+off+a_off-1,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
         enddo
         if(j==1) call single_hh_trafo(a(1,1+off+a_off,istripe,my_thread),bcast_buffer(1,off+1), nbw, nl, stripe_width)
+        
+        !INTRINSIC CODE, USING 2 HOUSEHOLDER VECTORS
+        !do j = ncols, 2, -2
+        !    w(:,1) = bcast_buffer(1:nbw,j+off)
+        !    w(:,2) = bcast_buffer(1:nbw,j+off-1)
+        !    if (mod(nl,24) == 0) then
+        !        call double_hh_trafo_2hv_fast(a(1,j+off+a_off-1,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    else
+        !        call double_hh_trafo_2hv(a(1,j+off+a_off-1,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    endif
+        !enddo
+        !if(j==1) call single_hh_trafo(a(1,1+off+a_off,istripe,my_thread),bcast_buffer(1,off+1), nbw, nl, stripe_width)
+        
+        !INTRINSIC CODE, USING 4 HOUSEHOLDER VECTORS
+        !do j = ncols, 4, -4
+        !    w(:,1) = bcast_buffer(1:nbw,j+off)
+        !    w(:,2) = bcast_buffer(1:nbw,j+off-1)
+        !    w(:,3) = bcast_buffer(1:nbw,j+off-2)
+        !    w(:,4) = bcast_buffer(1:nbw,j+off-3)
+        !    if (mod(nl,12) == 0) then
+        !       call double_hh_trafo_4hv_fast(a(1,j+off+a_off-3,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    else
+        !       call double_hh_trafo_4hv(a(1,j+off+a_off-3,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    endif
+        !enddo
+        !do jj = j, 2, -2
+        !    w(:,1) = bcast_buffer(1:nbw,jj+off)
+        !    w(:,2) = bcast_buffer(1:nbw,jj+off-1)
+        !    if (mod(nl,24) == 0) then
+        !       call double_hh_trafo_2hv_fast(a(1,jj+off+a_off-1,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    else
+        !       call double_hh_trafo_2hv(a(1,jj+off+a_off-1,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    endif
+        !enddo
+        !if(jj==1) call single_hh_trafo(a(1,1+off+a_off,istripe,my_thread),bcast_buffer(1,off+1), nbw, nl, stripe_width)
+
+        !INTRINSIC CODE, USING 6 HOUSEHOLDER VECTORS
+        !do j = ncols, 6, -6
+        !    w(:,1) = bcast_buffer(1:nbw,j+off)
+        !    w(:,2) = bcast_buffer(1:nbw,j+off-1)
+        !    w(:,3) = bcast_buffer(1:nbw,j+off-2)
+        !    w(:,4) = bcast_buffer(1:nbw,j+off-3)
+        !    w(:,5) = bcast_buffer(1:nbw,j+off-4)
+        !    w(:,6) = bcast_buffer(1:nbw,j+off-5)
+        !    if (mod(nl,8) == 0) then
+        !       call double_hh_trafo_6hv_fast(a(1,j+off+a_off-5,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    else
+        !       call double_hh_trafo_6hv(a(1,j+off+a_off-5,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    endif
+        !enddo
+        !do jj = j, 4, -4
+        !    w(:,1) = bcast_buffer(1:nbw,jj+off)
+        !    w(:,2) = bcast_buffer(1:nbw,jj+off-1)
+        !    w(:,3) = bcast_buffer(1:nbw,jj+off-2)
+        !    w(:,4) = bcast_buffer(1:nbw,jj+off-3)
+        !    if (mod(nl,12) == 0) then
+        !       call double_hh_trafo_4hv_fast(a(1,jj+off+a_off-3,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    else
+        !       call double_hh_trafo_4hv(a(1,jj+off+a_off-3,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    endif
+        !enddo
+        !do jjj = jj, 2, -2
+        !    w(:,1) = bcast_buffer(1:nbw,jjj+off)
+        !    w(:,2) = bcast_buffer(1:nbw,jjj+off-1)
+        !    if (mod(nl,24) == 0) then
+        !       call double_hh_trafo_2hv_fast(a(1,jjj+off+a_off-1,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    else
+        !       call double_hh_trafo_2hv(a(1,jjj+off+a_off-1,istripe,my_thread), w, nbw, nl, stripe_width, nbw)
+        !    endif
+        !enddo
+        !if(jjj==1) call single_hh_trafo(a(1,1+off+a_off,istripe,my_thread),bcast_buffer(1,off+1), nbw, nl, stripe_width)
+        
         if(my_thread==1) then
             kernel_flops = kernel_flops + 4*int(nl,8)*int(ncols,8)*int(nbw,8)
             kernel_time  = kernel_time + mpi_wtime()-ttt

ELPA_development_version/test/Makefile

 # ------------------------------------------------------------------------------
 # Please set the variables below according to your system!
 # ------------------------------------------------------------------------------
-# Settings for Intel Fortran (Linux):
+# Settings for Intel Fortran (Linux), Intel Composer XE 2011 (ifort 12.1) with AVX for Sandy Bridge:
+#
+#F90=mpiifort -O3 -traceback -g -fpe0
+#CC=mpiicc -O3
+#F90OPT=$(F90) -mavx
+#CCOPT=$(CC) -mavx
+#MKL_HOME=/opt/intel/mkl/lib/intel64
+#LIBS = -mkl -L$(MKL_HOME) -lmkl_scalapack_lp64 -lmkl_blacs_intelmpi_lp64
+#
+# ------------------------------------------------------------------------------
+# Settings for Intel Fortran (Linux), Intel Composer XE 2011 (ifort 12.1) with SSE3:
+#
+F90=mpiifort -O3 -traceback -g -fpe0
+CC=mpiicc -O3
+F90OPT=$(F90) -msse3
+CCOPT=$(CC) -msse3
+MKL_HOME=/opt/intel/mkl/lib/intel64
+LIBS = -mkl -L$(MKL_HOME) -lmkl_scalapack_lp64 -lmkl_blacs_intelmpi_lp64
+#
+# ------------------------------------------------------------------------------
+# Settings for Intel Fortran (Linux), Intel Composer XE 2011 (ifort 12.1) and GCC 4.6 with FMA4 for AMD Bulldozer:
+#
+#F90=mpiifort -O3 -traceback -g -fpe0
+#CC=gcc -O3
+#F90OPT=$(F90) -msse3
+#CCOPT=$(CC) -mfma4 -mxop -march=bdver1 -D__USE_AVX128__
+#LIBS = -L/opt/acml5.0.0/gfortran64_fma4/lib/ -lacml -lgfortran libscalapack.a
+#
+# ------------------------------------------------------------------------------
+# Settings for Intel Fortran (Linux) old 11.x Toolchain, do not use:
 #
 #F90=mpif90 -O3 -traceback -g -fpe0
 #F90OPT=$(F90) -xSSE4.2
@@ -25,7 +54,7 @@
 #LIBS = -L/usr/local/lib -lscalapack -llapack-essl -lessl -lblacsF77init -lblacs -lblacsF77init -lblacs -lc
 #
 # ------------------------------------------------------------------------------
-# Settings for IBM BlueGene/P
+# Settings for IBM AIX BlueGene
 #
 F90 = mpixlf95_r -O3 -g -qarch=auto -qtune=auto
 F90OPT = mpixlf95_r -O4 -g -qarch=auto -qtune=auto
@@ -54,8 +83,14 @@ read_real_gen: read_real_gen.o elpa1.o
 test_complex_gen: test_complex_gen.o read_test_parameters.o elpa1.o
 	$(F90) -o $@ test_complex_gen.o read_test_parameters.o elpa1.o $(LIBS)
 
+#test_real2: test_real2.o elpa1.o elpa2.o read_test_parameters.o elpa2_kernels.o elpa_pdgeqrf.o elpa_pdlarfb.o elpa_qrkernels.o tum_utils.o elpa2_tum_kernels_sse-avx_2hv.o elpa2_tum_kernels_sse-avx_4hv.o elpa2_tum_kernels_sse-avx_6hv.o
+#	$(F90) -o $@ test_real2.o elpa1.o elpa2.o read_test_parameters.o elpa2_kernels.o elpa_pdgeqrf.o elpa_pdlarfb.o elpa_qrkernels.o tum_utils.o elpa2_tum_kernels_sse-avx_2hv.o elpa2_tum_kernels_sse-avx_4hv.o elpa2_tum_kernels_sse-avx_6hv.o $(LIBS)
+
+#test_complex2: test_complex2.o read_test_parameters.o elpa1.o elpa2.o elpa2_kernels.o elpa_pdgeqrf.o elpa_pdlarfb.o elpa_qrkernels.o tum_utils.o elpa2_tum_kernels_sse-avx_2hv.o elpa2_tum_kernels_sse-avx_4hv.o elpa2_tum_kernels_sse-avx_6hv.o
+#	$(F90) -o $@ test_complex2.o read_test_parameters.o elpa1.o elpa2.o elpa2_kernels.o elpa_pdgeqrf.o elpa_pdlarfb.o elpa_qrkernels.o tum_utils.o elpa2_tum_kernels_sse-avx_2hv.o elpa2_tum_kernels_sse-avx_4hv.o elpa2_tum_kernels_sse-avx_6hv.o $(LIBS)
+
 test_real2: test_real2.o elpa1.o elpa2.o read_test_parameters.o elpa2_kernels.o elpa_pdgeqrf.o elpa_pdlarfb.o elpa_qrkernels.o tum_utils.o
-	$(F90) -o $@ test_real2.o read_test_parameters.o elpa1.o elpa2.o elpa2_kernels.o elpa_pdgeqrf.o elpa_pdlarfb.o elpa_qrkernels.o tum_utils.o $(LIBS)
+	$(F90) -o $@ test_real2.o elpa1.o elpa2.o read_test_parameters.o elpa2_kernels.o elpa_pdgeqrf.o elpa_pdlarfb.o elpa_qrkernels.o tum_utils.o $(LIBS)
 
 test_complex2: test_complex2.o read_test_parameters.o elpa1.o elpa2.o elpa2_kernels.o elpa_pdgeqrf.o elpa_pdlarfb.o elpa_qrkernels.o tum_utils.o
 	$(F90) -o $@ test_complex2.o read_test_parameters.o elpa1.o elpa2.o elpa2_kernels.o elpa_pdgeqrf.o elpa_pdlarfb.o elpa_qrkernels.o tum_utils.o $(LIBS)
@@ -107,6 +142,15 @@ elpa2.o: ../src/elpa2.f90 elpa1.o elpa_pdgeqrf.o
 
 elpa2_kernels.o: ../src/elpa2_kernels.f90
 	$(F90OPT) -c ../src/elpa2_kernels.f90
+	
+elpa2_tum_kernels_sse-avx_2hv.o: ../src/elpa2_tum_kernels_sse-avx_2hv.c
+	$(CCOPT) -c ../src/elpa2_tum_kernels_sse-avx_2hv.c
+	
+elpa2_tum_kernels_sse-avx_4hv.o: ../src/elpa2_tum_kernels_sse-avx_4hv.c
+	$(CCOPT) -c ../src/elpa2_tum_kernels_sse-avx_4hv.c
+	
+elpa2_tum_kernels_sse-avx_6hv.o: ../src/elpa2_tum_kernels_sse-avx_6hv.c
+	$(CCOPT) -c ../src/elpa2_tum_kernels_sse-avx_6hv.c
 
 clean:
-	rm -f *.o *.mod test_real test_complex test_real_gen test_complex_gen test_real2 test_complex2 read_real read_real_gen read_test_parameters.o
+	rm -f *.o *.mod test_real test_complex test_real_gen test_complex_gen test_real2 test_complex2 read_real read_real_gen read_test_parameters.o 

ELPA_development_version/test/test_real2.f90

@@ -26,7 +26,8 @@ program test_real2
    ! nblk: Blocking factor in block cyclic distribution
    !-------------------------------------------------------------------------------
 
-   integer :: na = 4000, nev = 1500, nblk = 16
+   integer, parameter :: nblk = 16
+   integer na, nev
 
    !-------------------------------------------------------------------------------
    !  Local Variables
@@ -43,6 +44,21 @@ program test_real2
 
    integer :: iseed(4096) ! Random seed, size should be sufficient for every generator
 
+   !-------------------------------------------------------------------------------
+   !  Pharse command line argumnents, if given
+   character*16 arg1
+   character*16 arg2
+
+   na = 12059
+   nev = 3401
+
+   if (iargc() == 2) then
+      call getarg(1, arg1)
+      call getarg(2, arg2)
+      read(arg1, *) na
+      read(arg2, *) nev
+   endif
+
    !-------------------------------------------------------------------------------
    !  MPI Initialization
 
@@ -50,14 +66,6 @@ program test_real2
    call mpi_comm_rank(mpi_comm_world,myid,mpierr)
    call mpi_comm_size(mpi_comm_world,nprocs,mpierr)
 
-   !-------------------------------------------------------------------------------
-   !  Reading of test parameters (matrix size, number of requested eigenvalue/eigenvector
-   !  pairs, block size) from a file 'test_parameters.in', if that file exists. 
-   !  We only read on mpi task number myid = 0 to avoid any possible confusion. 
-   !  The parameters of interest are subsequently broadcast to all other mpi tasks.
-
-   call read_test_parameters (na,nev,nblk,myid,mpi_comm_world)
-
    !-------------------------------------------------------------------------------
    ! Selection of number of processor rows/columns
    ! We try to set up the grid square-like, i.e. start the search for possible