Merge pull request #912 from omarkahmed/omarkahmed/openmp-intel-gpu

Add Intel Xe Max Support for the CCSD module

src/ccsd/GNUmakefile

@@ -2,6 +2,10 @@
 
 include ../config/makefile.h
 
+ifdef USE_IMAX_OPENMP_TRPDRV
+SUBDIRS += module
+endif
+
  OBJ_OPTIMIZE = \
         ccden_driver.o \
         ccden_interm2.o \
@@ -93,17 +97,16 @@ endif
      ccsd_trpdrv_bgp2.F \
      ccsd_trpdrv_offload.F \
      ccsd_trpdrv_openacc.F \
+     ccsd_trpdrv_openmp_imax.F \
      moints_trp.F
 
 ifeq ($(TARGET),BGP)
   OBJ_OPTIMIZE += ccsd_trpdrv_bgp2.o ccsd_tengy_bgp2.o ccsd_tengy_bgp.o
-  USES_BLAS += ccsd_trpdrv_bgp2.F
   LIB_DEFINES += -DBGP
 endif
 
 ifdef USE_MIC_TRPDRV
   OBJ_OPTIMIZE += ccsd_trpdrv_offload.o
-  USES_BLAS    += ccsd_trpdrv_offload.F
   LIB_DEFINES += -DUSE_MIC_TRPDRV
 endif
 ifeq ($(_FC),xlf)
@@ -119,9 +122,23 @@ ifeq ($(HAVE_SET_GA_PROPERTY),Y)
       LIB_DEFINES += -DHAVE_SET_GA_PROPERTY
 endif
 
+ifdef USE_IMAX_OPENMP_TRPDRV
+
+  OBJ_OPTIMIZE += ccsd_trpdrv_openmp_imax.o
+
+  OBJ_OPTIMIZE += ccsd_trpdrv_omp_reduce_f.o
+
+  FOPTIONS += -O3 -fiopenmp  -fopenmp-targets=spir64="-mllvm -vpo-paropt-opt-data-sharing-for-reduction=false -mllvm -vpo-paropt-atomic-free-reduction-par-global=false" -switch offload_modvars -mllvm -vpo-paropt-atomic-free-reduction-slm=true -qmkl -DMKL_ILP64 -I"${MKLROOT}/include" -I ${NWCHEM_TOP}/src/ccsd/module -mllvm -vpo-paropt-dispatch-codegen-version=1 -switch -use-host-usm-for-implicit-reduction-map
+
+  COPTIONS:=$(filter-out -fopenmp,$(COPTIONS))
+  COPTIONS:=$(filter-out -O1,$(COPTIONS))
+
+  COPTIONS += -O3 -fiopenmp  -fopenmp-targets=spir64="-mllvm -vpo-paropt-opt-data-sharing-for-reduction=false -mllvm -vpo-paropt-atomic-free-reduction-par-global=false" -mllvm -vpo-paropt-atomic-free-reduction-slm=true -qmkl -DMKL_ILP64 -I"${MKLROOT}/include" -mllvm -vpo-paropt-dispatch-codegen-version=1
+
+endif
+
 ifdef USE_OPENACC_TRPDRV
   OBJ_OPTIMIZE += ccsd_trpdrv_openacc.o
-  USES_BLAS    += ccsd_trpdrv_openacc.F
   FOPTIONS += -DUSE_OPENACC_TRPDRV
   ifeq ($(_FC),pgf90)
       FOPTIONS += -Mextend -acc -cuda -cudalib=cublas

src/ccsd/aoccsd2.F

@@ -2,7 +2,11 @@ c#define DOGRAD
       subroutine aoccsd(basis,ncor,nocc,nvir,ndel,nact,nbf,maxit,
      &                  convi,iprt,cmo,eorb,blen,
      &                  g_ncoul, g_nexch, RefEner,
-     $   CC_Theory, rtdb, mxvec, geom, Tol2e, occd, oconverged)
+     $                  CC_Theory, rtdb, mxvec, geom, Tol2e, occd,
+     &                  oconverged)
+#if defined(USE_IMAX_OPENMP_TRPDRV)
+      use omp_lib, only: omp_interop_kind, omp_interop_none
+#endif
       implicit none
 #include "errquit.fh"
 C     $Id$
@@ -16,6 +20,7 @@ c
       logical oconverged, occd, use_trpdrv_nb
       logical use_trpdrv_omp, use_trpdrv_bgp2
       logical use_trpdrv_omp_mp
+      logical use_trpdrv_openmp_imax
       logical use_trpdrv_openacc
       logical use_trpdrv_offload
 c
@@ -31,7 +36,10 @@ c
 #include "msgids.fh"
 #include "ccsdps.fh"
 c
+#if defined(USE_IMAX_OPENMP_TRPDRV)
 c ccsd
+      integer(kind = omp_interop_kind) :: dummy_obj = omp_interop_none
+#endif
       Integer i
       logical stat
       integer nsh,maxbfsh,max2e,mem2
@@ -721,6 +729,9 @@ C
       if (.not. rtdb_get(rtdb, 'ccsd:use_trpdrv_openacc', mt_log, 1,
      1                   use_trpdrv_openacc))
      2    use_trpdrv_openacc=.false.
+      if (.not. rtdb_get(rtdb, 'ccsd:use_trpdrv_openmp_imax', mt_log,
+     1               1, use_trpdrv_openmp_imax))
+     2    use_trpdrv_openmp_imax=.false.
       if (.not. rtdb_get(rtdb, 'ccsd:use_trpdrv_offload', mt_log, 1,
      1                   use_trpdrv_offload))
      2    use_trpdrv_offload=.false.
@@ -955,6 +966,20 @@ c
      &      blen, cmo, ncor, nocc, nvir, ndel, Tol2E)
 c        print *,'call trpdrv ',nvpass
          call ga_sync()
+#if defined(USE_IMAX_OPENMP_TRPDRV)
+!     Dummy parallel construct
+      !$omp parallel num_threads(8)
+      !$omp end parallel
+
+!     Dummy target construct
+      !$omp target
+      !$omp end target
+
+!     Dummy interop object
+      !$omp interop init(prefer_type("sycl"),targetsync: dummy_obj)
+      !$omp interop destroy(dummy_obj)
+#endif
+
          tx(2)=tcgtime()
 c
          if (use_trpdrv_omp) then
@@ -992,8 +1017,23 @@ c
          call ccsd_trpdrv_openacc(dbl_mb(k_t1),eorb,
      $        g_objo,g_objv,g_coul,g_exch,ncor,nocc,nvir,iprt,
      $        empt(1),empt(2),oseg_lo,oseg_hi,kchunk)
+!!#else
+!!         call errquit('aoccsd: trpdrv_openacc disabled ',0,0)
+!!#endif
+! use_trpdrv_openmp_imax
+         else if (use_trpdrv_openmp_imax) then
+#elif defined(USE_IMAX_OPENMP_TRPDRV)
+         if (iam.eq.0.and.oprint) then
+            write(luout,1818) nvpass,util_wallsec()
+            call util_flush(luout)
+         endif
+ 1818    format(' commencing triples evaluation - OpenMP',
+     I        'MAX GPU version',i8,' at ',f20.2,' secs')
+         call ccsd_trpdrv_offload_xe(dbl_mb(k_t1),eorb,
+     $        g_objo,g_objv,g_coul,g_exch,ncor,nocc,nvir,iprt,
+     $        empt(1),empt(2),oseg_lo,oseg_hi,kchunk)
 #else
-         call errquit('aoccsd: trpdrv_openacc disabled ',0,0)
+         call errquit('aoccsd: trpdrv_openmp_gpu disabled ',0,0)
 #endif
 c
          elseif (use_trpdrv_omp_mp) then

src/ccsd/ccsd_trpdrv_omp_reduce_f.F

@@ -0,0 +1,121 @@
+      subroutine ccsd_trpdrv_omp_fbody_reduce_new (f1n, f1t, f2n, f2t,
+     &                                  f3n, f3t, f4n, f4t,
+     &                                  eorb,
+     &                                  ncor, nocc, nvir,
+     &                                  emp4, emp5,
+     &                                  i, k,
+     &                                  eaijk,
+     &                                  dintc1, dintx1, t1v1,
+     &                                  dintc2, dintx2, t1v2)
+
+      use omp_lib
+      use iso_fortran_env
+      implicit none
+
+      double precision, intent(inout) :: emp4, emp5
+      double precision, intent(inout) :: f1n(nvir,nvir), f1t(nvir,nvir)
+      double precision, intent(inout) :: f2n(nvir,nvir), f2t(nvir,nvir)
+      double precision, intent(inout) :: f3n(nvir,nvir), f3t(nvir,nvir)
+      double precision, intent(inout) :: f4n(nvir,nvir), f4t(nvir,nvir)
+      double precision, intent(in)    :: eorb(*)
+      double precision, intent(in)    :: eaijk
+      double precision, intent(in)    :: dintc1(nvir), dintx1(nvir)
+      double precision, intent(in)    :: dintc2(nvir), dintx2(nvir)
+      double precision, intent(in)    :: t1v1(nvir), t1v2(nvir)
+      integer, intent(in) :: ncor, nocc, nvir
+      integer, intent(in) :: i, k
+
+      double precision  :: emp4i,emp5i,emp4k,emp5k, denom
+      double precision  :: f1nbc,f1tbc,f1ncb,f1tcb
+      double precision  :: f2nbc,f2tbc,f2ncb,f2tcb
+      double precision  :: f3nbc,f3tbc,f3ncb,f3tcb
+      double precision  :: f4nbc,f4tbc,f4ncb,f4tcb
+      double precision  :: t1v1b,t1v2b,dintx1c,dintx2c,dintc1c,dintc2c
+      integer :: b,c
+
+      emp4i = 0.0
+      emp5i = 0.0
+      emp4k = 0.0
+      emp5k = 0.0
+
+      !$omp target teams distribute parallel do collapse(2)
+     &      reduction(+:emp5i,emp4i,emp5k,emp4k)
+     &      private(f1nbc,f1tbc,f1ncb,f1tcb,f2nbc,f2tbc,f2ncb,f2tcb)
+     &      private(f3nbc,f3tbc,f3ncb,f3tcb,f4nbc,f4tbc,f4ncb,f4tcb)
+     &      private(t1v1b,t1v2b,dintx1c,dintx2c,dintc1c,dintc2c)
+     &      private(denom) firstprivate(eaijk,nvir,ncor,nocc)
+      do b=1,nvir
+         do c=1,nvir
+            denom=-1.0/( eorb(ncor+nocc+b)+eorb(ncor+nocc+c)+eaijk )
+
+            f1nbc = f1n(b,c);
+            f1tbc = f1t(b,c);
+            f1ncb = f1n(c,b);
+            f1tcb = f1t(c,b);
+
+            f2nbc = f2n(b,c);
+            f2tbc = f2t(b,c);
+            f2ncb = f2n(c,b);
+            f2tcb = f2t(c,b);
+
+            f3nbc = f3n(b,c);
+            f3tbc = f3t(b,c);
+            f3ncb = f3n(c,b);
+            f3tcb = f3t(c,b);
+
+            f4nbc = f4n(b,c);
+            f4tbc = f4t(b,c);
+            f4ncb = f4n(c,b);
+            f4tcb = f4t(c,b);
+
+            t1v1b = t1v1(b);
+            t1v2b = t1v2(b);
+
+            dintx1c = dintx1(c);
+            dintx2c = dintx2(c);
+            dintc1c = dintc1(c);
+            dintc2c = dintc2(c);
+
+            emp4i = emp4i
+     &             + (denom * (f1tbc+f1ncb+f2tcb+f3nbc+f4ncb)
+     &             * (f1tbc-f2tbc*2-f3tbc*2+f4tbc)
+     &             - denom * (f1nbc+f1tcb+f2ncb+f3ncb)
+     &             * (f1tbc*2-f2tbc-f3tbc+f4tbc*2)
+     &             + denom * 3 * (f1nbc*(f1nbc+f3ncb+f4tcb*2)
+     &             + f2nbc*f2tcb+f3nbc*f4tbc))
+
+            emp4k = emp4k
+     &             + (denom * (f1nbc+f1tcb+f2ncb+f3tbc+f4tcb)
+     &             * (f1nbc-f2nbc*2-f3nbc*2+f4nbc)
+     &             - denom * (f1tbc+f1ncb+f2tcb+f3tcb)
+     &             * (f1nbc*2-f2nbc-f3nbc+f4nbc*2)
+     &             + denom * 3 * (f1tbc*(f1tbc+f3tcb+f4ncb*2)
+     &             + f2tbc*f2ncb+f3tbc*f4nbc))
+
+            emp5i = emp5i
+     &             + (denom * t1v1b * dintx1c
+     &             * (f1tbc+f2nbc+f4ncb
+     &             - (f3tbc+f4nbc+f2ncb+f1nbc+f2tbc+f3ncb)*2
+     &             + (f3nbc+f4tbc+f1ncb)*4)
+     &             + denom * t1v1b * dintc1c
+     &             * (f1nbc+f4nbc+f1tcb -(f2nbc+f3nbc+f2tcb)*2))
+
+           emp5k = emp5k
+     &             + (denom * t1v2b * dintx2c
+     &             * (f1nbc+f2tbc+f4tcb
+     &             - (f3nbc+f4tbc+f2tcb +f1tbc+f2nbc+f3tcb)*2
+     &             + (f3tbc+f4nbc+f1tcb)*4)
+     &             + denom * t1v2b * dintc2c
+     &             * (f1tbc+f4tbc+f1ncb -(f2tbc+f3tbc+f2ncb)*2))
+         enddo
+      enddo
+      !$omp end target teams distribute parallel do
+
+      emp4 = emp4 + emp4i
+      emp5 = emp5 + emp5i
+      if (i.ne.k) then
+         emp4 = emp4 + emp4k
+         emp5 = emp5 + emp5k
+      end if ! (i.ne.k)
+
+      end

src/ccsd/module/GNUmakefile

@@ -0,0 +1,32 @@
+#     $Id$
+
+include ../../config/makefile.h
+
+ifdef USE_IMAX_OPENMP_TRPDRV
+ OBJ_OPTIMIZE = \
+
+  OBJ_OPTIMIZE  += ccsd_trpdrv_mkl_module.o
+
+
+  FOPTIONS += -fiopenmp -fopenmp-targets=spir64="-mllvm -vpo-paropt-enable-64bit-opencl-atomics=true -mllvm -vpo-paropt-opt-data-sharing-for-reduction=false" -qmkl -DMKL_ILP64 -I"${MKLROOT}/include" -fpp -fixed -free
+
+#OBJ_OPTIMIZE += ccsd_trpdrv_mkl_module.o
+#.SECONDARY: ccsd_trpdrv_mkl_module.o
+#ccsd_trpdrv_mkl_module.o: ccsd_trpdrv_mkl_module.F
+#	$(FC) $(MK_MODULE_FLAG) -c $< -o $@
+
+#-include $(ccsd_trpdrv_mkl_module_)
+#.PRECIOUS: ccsd_trpdrv_mkl_module.F
+#	$(FC) $(MK_MODULE_FLAG) -c $< -o $@
+#USES_BLAS += ccsd_trpdrv_mkl_module.F
+#ccsd_trpdrv_mkl_module.o : ccsd_trpdrv_mkl_module.F
+#	$(FC) $(MK_MODULE_FLAG) -c $< -o $@
+
+endif
+
+  LIB_DEFINES +=-DUSE_F90INTERFACE
+
+  LIBRARY = libccsd.a
+
+include ../../config/makelib.h
+

src/ccsd/module/ccsd_trpdrv_mkl_module.F

@@ -0,0 +1,5 @@
+include "mkl_omp_offload.f90"
+
+      module ccsd_mklmodule
+
+      end module ccsd_mklmodule

src/config/makefile.h

@@ -2358,32 +2358,6 @@ ifneq ($(TARGET),LINUX)
             endif
 
             # support for Intel(R) Fortran compiler
-            ifeq ($(_FC),ifxold)
-                DEFINES += -DIFCV8 -DIFCLINUX
-                FOPTIONS += -fpp -align
-                ifneq ($(V),1)
-                    FOPTIONS += -Qoption,fpp,-w0
-		endif
-                FOPTIMIZE = -g -O3 -fimf-arch-consistency=true
-                ifdef USE_I4FLAGS
-                else
-                    FOPTIONS += -i8
-                endif
-                ifdef USE_OPENMP
-                    FOPTIONS += -fopenmp
-                    ifdef USE_OFFLOAD
-                        FOPTIONS += -fopenmp-targets=spirv64
-                    endif
-                endif
-                ifdef OPENBLAS_USES_OPENMP
-                     LDOPTIONS += -fopenmp
-                endif
-                ifdef IFX_DEBUG
-                    # debugging remove at some point
-                    FOPTIONS += -std95 -what
-                endif
-                FDEBUG = $(FOPTIMIZE)
-            endif
 
 
             # support for traditional Intel(R) Fortran compiler
@@ -2466,7 +2440,10 @@ ifneq ($(TARGET),LINUX)
                         ifdef USE_IFX
                             FOPTIONS += -fiopenmp
                             ifdef USE_OFFLOAD
-                                FOPTIONS += -fopenmp-targets=spirv64
+                                FOPTIONS += -fopenmp-targets=spir64
+                                ifdef USE_IMAX_OPENMP_TRPDRV
+                                    DEFINES += -DUSE_IMAX_OPENMP_TRPDRV
+                                endif
                             endif
                         else
                             FOPTIONS += -qopenmp

src/util/util_getenv.F

@@ -111,7 +111,9 @@ c     offloading enabled.
       span = offload_span()
 #if USE_OPENMP
       ndev = omp_get_num_devices()
+#ifdef DEBUG
       write(6,*) ' omp_get_num_devices() ',ndev
+#endif
       offload_device = mod(ga_nodeid() / span, ndev)
       call omp_set_default_device(offload_device)
 #else