CCSD_T2_8 DGEMM w/ CUBLAS

src/tce/ccsd/GNUmakefile

@@ -1,25 +1,34 @@
 #$Id$
 
+include ../../config/makefile.h
+
 OBJ_OPTIMIZE = ccsd_e.o ccsd_t1.o ccsd_t2.o cc2_t1.o cc2_t2.o \
                ccsd_1prdm.o ccsd_1prdm_hh.o ccsd_1prdm_hp.o \
                ccsd_1prdm_ph.o ccsd_1prdm_pp.o \
                icsd_t1.o icsd_t2.o \
-	       ccsd_kernels.o ccsd_t2_8.o tce_1b_dens_print.o
+	       ccsd_kernels.o ccsd_t2_7.o ccsd_t2_8.o tce_1b_dens_print.o
 
 LIB_INCLUDES = -I../include
 
 LIBRARY = libtce.a
 
 USES_BLAS = ccsd_e.F ccsd_t1.F ccsd_t2.F cc2_t1.F cc2_t2.F \
             ccsd_1prdm_hh.F ccsd_1prdm_hp.F ccsd_1prdm_ph.F \
-            ccsd_1prdm_pp.F ccsd_1prdm.F \
-            icsd_t1.F icsd_t2.F ccsd_t2_8.F ccsd_kernels.F sd_t2_8_loops.F
-
+            ccsd_1prdm_pp.F ccsd_1prdm.F ccsd_kernels.F sd_t2_8_loops.F \
+            icsd_t1.F icsd_t2.F ccsd_t2_7.F ccsd_t2_8.F
 
 LIB_DEFINES = -DDEBUG_PRINT
 
-# This replaces 3*TCE_SORT4+DGEMM with 6D loops (ccsd_kernels.F).
-#LIB_DEFINES += -DUSE_LOOPS_NOT_DGEMM
+# replace this with something better later
+ifdef USE_OPENACC_TRPDRV
+  FOPTIONS += -DUSE_TCE_CUBLAS
+  ifeq ($(_FC),pgf90)
+      FOPTIONS += -Mextend -acc -cuda -cudalib=cublas
+  endif
+  ifeq ($(_FC),gfortran)
+      FOPTIONS += -ffree-form -fopenacc -lcublas
+  endif
+endif
 
 #
 # Possible #defines
@@ -32,6 +41,5 @@ LIB_DEFINES = -DDEBUG_PRINT
 
 HEADERS = 
 
-include ../../config/makefile.h
 include ../../config/makelib.h