From eadb2c7371f4eaeace2463771ba7584530480590 Mon Sep 17 00:00:00 2001
From: ESCOBAR Juan <escj@nuwa>
Date: Mon, 14 Sep 2020 15:14:47 +0200
Subject: [PATCH] Juan 14/09/2020: Rules.LXnvhpc2005.mk , add support of arch
CPU type = TP & MANAGEDO2/OPENACCO2 optlevel <=> without "-g"
---
src/Rules.LXnvhpc2005.mk | 63 +++++++++++++++++++++++++++++-----------
1 file changed, 46 insertions(+), 17 deletions(-)
diff --git a/src/Rules.LXnvhpc2005.mk b/src/Rules.LXnvhpc2005.mk
index 9aafd7a10..8f9f2ab8f 100644
--- a/src/Rules.LXnvhpc2005.mk
+++ b/src/Rules.LXnvhpc2005.mk
@@ -9,13 +9,18 @@
##########################################################
#OBJDIR_PATH=${WORKDIR}
#
+# Processor type
+#TP= -tp=sandybridge
+#TP= -tp=px
+TP= -tp=p7-64
+#
#Version of CUDA
#(8.0 at least if compute capability >= 6.0)
-CUDALEVEL=cuda10.2
+CUDALEVEL=cuda10.1
#
#Compute capability of GPU
#
-OPT_CPTCAP=cc35,cc50
+OPT_CPTCAP=cc35,cc50,cc70
#Aeropc45: cc50
#Nuwa: cc35
#Ouessant Firestone K80: cc35
@@ -30,7 +35,7 @@ OPT_INFO = -Minfo=accel
#
#PW: if -Ktrap=fp: nvprof/pgprof do not work OPT_BASE = -Ktrap=fp ...
#PW: -g: big impact on performance
-OPT_BASE = -Mbackslash -Mextend -Kieee -nofma
+OPT_BASE = $(TP) -Mbackslash -Mextend -Kieee -nofma -Mallocatable=95
#
OPT_PERF0 = -O0
OPT_PERF1 = -O1
@@ -39,7 +44,7 @@ OPT_PERF2 = -O2
OPT_MANAGED = -Mframe -Mnostack_arrays -Mallocatable=95 -acc=host,gpu -gpu=nofma,$(OPT_CPTCAP),$(CUDALEVEL),managed $(OPT_INFO) $(OPT_PROF)
OPT_MULTICORE = -acc=multicore $(OPT_INFO) $(OPT_PROF)
OPT_NOOPENACC = -acc=host $(OPT_INFO) $(OPT_PROF)
-OPT_OPENACC = -Mnostack_arrays -Mallocatable=95 -acc=host,gpu -gpu=nofma,$(OPT_CPTCAP),$(CUDALEVEL) $(OPT_INFO) $(OPT_PROF)
+OPT_OPENACC = -Mnostack_arrays -Mnostack_arrays -Mallocatable=95 -acc=host,gpu -gpu=nofma,$(OPT_CPTCAP),$(CUDALEVEL) $(OPT_INFO) $(OPT_PROF)
#
OPT_CHECK = -C #-Mchkfpstk -Mchkptr
OPT_I8 = -i8
@@ -78,8 +83,6 @@ ifeq "$(OPTLEVEL)" "DEBUG"
OPT = -g $(OPT_BASE) $(OPT_PERF0) $(OPT_CHECK)
OPT0 = -g $(OPT_BASE) $(OPT_PERF0) $(OPT_CHECK)
OPT_NOCB = -g $(OPT_BASE) $(OPT_PERF0)
-OBJS_DWARF3 = spll_model_n.o
-$(OBJS_DWARF3) : OPT = -g -Mdwarf2 $(OPT_BASE) $(OPT_PERF0) $(OPT_CHECK)
endif
#
ifeq "$(OPTLEVEL)" "MANAGED"
@@ -92,6 +95,16 @@ OBJS_REPROD= spll_mode_sum_ll.o
$(OBJS_REPROD) : OPT = $(OPT_BASE) $(OPT_MANAGED) $(OPT_PERF2) -Mvect=nosimd -Minfo=all -g
endif
#
+ifeq "$(OPTLEVEL)" "MANAGEDO2"
+CPPFLAGS += -DMNH_OPENACC
+OPT = $(OPT_BASE) $(OPT_MANAGED) $(OPT_PERF2)
+OPT0 = $(OPT_BASE) $(OPT_MANAGED) $(OPT_PERF0)
+OPT_NOCB = $(OPT_BASE) $(OPT_MANAGED) $(OPT_PERF2)
+CXXFLAGS = -acc -Kieee -Mnofma $(OPT_MANAGED) -gpu=nofma
+OBJS_REPROD= spll_mode_sum_ll.o
+$(OBJS_REPROD) : OPT = $(OPT_BASE) $(OPT_MANAGED) $(OPT_PERF2) -Mvect=nosimd -Minfo=all
+endif
+#
ifeq "$(OPTLEVEL)" "MULTICORE"
CPPFLAGS += -DMNH_OPENACC
OPT = $(OPT_BASE) $(OPT_MULTICORE)
@@ -112,12 +125,12 @@ $(OBJS_REPROD) : OPT = $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF2) -Mvect=nosimd -Mi
endif
ifeq "$(OPTLEVEL)" "OPENACCO2"
CPPFLAGS += -DMNH_OPENACC
-OPT = -g $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF2)
-OPT0 = -g $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF0)
-OPT_NOCB = -g $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF2)
-CXXFLAGS = -g -acc -Kieee -Mnofma $(OPT_OPENACC)
+OPT = $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF2)
+OPT0 = $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF0)
+OPT_NOCB = $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF2)
+CXXFLAGS = -acc -Kieee -Mnofma $(OPT_OPENACC)
OBJS_REPROD= spll_mode_sum_ll.o
-$(OBJS_REPROD) : OPT = $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF2) -Mvect=nosimd -Minfo=all -g
+$(OBJS_REPROD) : OPT = $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF2) -Mvect=nosimd -Minfo=all
endif
#
ifeq "$(OPTLEVEL)" "OPENACCDEFONLY"
@@ -125,9 +138,9 @@ CPPFLAGS += -DMNH_OPENACC -D_FAKEOPENACC
OPT = $(OPT_BASE) $(OPT_NOOPENACC) $(OPT_PERF2)
OPT0 = $(OPT_BASE) $(OPT_NOOPENACC) $(OPT_PERF0)
OPT_NOCB = $(OPT_BASE) $(OPT_NOOPENACC) $(OPT_PERF2)
-CXXFLAGS = -Kieee -Mnofma $(OPT_NOOPENACC)
+CXXFLAGS = -Kieee -Mnofma $(OPT_NOOPENACC)
OBJS_REPROD= spll_mode_sum_ll.o
-$(OBJS_REPROD) : OPT = $(OPT_BASE) $(OPT_NOOPENACC) $(OPT_PERF2) -Mvect=nosimd -Minfo=all -g
+$(OBJS_REPROD) : OPT = $(OPT_BASE) $(OPT_NOOPENACC) $(OPT_PERF2) -Mvect=nosimd -Minfo=all
endif
#
ifeq "$(OPTLEVEL)" "NOOPENACC"
@@ -141,6 +154,8 @@ $(OBJS_REPROD) : OPT = $(OPT_BASE) $(OPT_NOOPENACC) $(OPT_PERF2) -Mvect=nosimd -
endif
#
CC = pgcc
+CXXFLAGS += $(TP)
+CFLAGS += $(TP)
FC = pgf90
ifeq "$(VER_MPI)" "MPIAUTO"
F90 = mpif90
@@ -176,7 +191,14 @@ CPPFLAGS_MNH += -Uvector -Upixel
TARGET_GRIBEX=linux
CNAME_GRIBEX=_pgf77
GRIBAPI_CONF="CPP=cpp"
+#
+# Netcdf/HDF5 flags
+#
+#HDF_CONF= CXXFLAGS=$(TP)
+HDF_OPT ?= $(TP)
+NETCDF_OPT ?= $(TP)
CDF_CONF="CPP=cpp"
+HDF_CONF="CPP=cpp"
#
# BITREP flags
#
@@ -220,11 +242,14 @@ include Makefile.MESONH.mk
# etc ... #
# #
##########################################################
-OBJS_O1= spll_modd_isba_n.o spll_mode_construct_ll.o \
+OBJS_O1 += spll_modd_isba_n.o spll_mode_construct_ll.o \
spll_init_surf_atm_n.o spll_mode_scatter_ll.o spll_convert_patch_teb.o \
spll_define_mask_n.o spll_del1dfield_ll.o spll_mode_fm.o spll_mode_gather_ll.o \
spll_convect_updraft.o spll_convect_updraft_shal.o \
- spll_mode_dustopt.o spll_mode_saltopt.o
+ spll_mode_dustopt.o spll_mode_saltopt.o \
+ spll_aeroopt_get.o spll_write_lfifm1_for_diag_supp.o spll_write_lfifm1_for_diag.o spll_write_lfifm_n.o \
+
+#spll_unpack_1d_2d_from2d.o
#spll_pack_isba_patch_n.o
#spll_phys_param_n.o
$(OBJS_O1) : OPT = $(OPT_BASE) $(OPT_PERF1)
@@ -232,9 +257,13 @@ $(OBJS_O1) : OPT = $(OPT_BASE) $(OPT_PERF1)
OBJS_O0= spll_mode_mppdb.o \
spll_fft55.o spll_fft.o spll_flat_invz.o \
spll_mode_repro_sum.o \
-# spll_fast_terms.o
+ spll_modd_les_n.o
+
+# spll_fast_terms.o
+# spll_modd_ch_solver_n.o \
+# spll_modd_dummy_gr_field_n.o spll_modd_dyn_n.o
# spll_mode_sum_ll.o
-$(OBJS_O0) : OPT = $(OPT_BASE) $(OPT_PERF0)
+$(OBJS_O0) : OPT = -g $(OPT_BASE) $(OPT_PERF0)
OBJS_O2= spll_mode_device.o
$(OBJS_O2) : OPT = $(OPT_BASE) $(OPT_OPENACC) $(OPT_PERF2)
--
GitLab