hypre-multiprecision build

src/blas/Makefile

@@ -45,18 +45,67 @@ OBJS = ${BLAS_FILES:.c=.o}
 # Targets
 ##################################################################
 
-all: ${OBJS}
+#all: ${OBJS}
+
+all: blas_OBJS
 
 install: all
 	cp -fR $(srcdir)/_hypre_blas.h $(HYPRE_INC_INSTALL)
 
-clean:
+clean: clean_mup
 	rm -rf *.o
 
 distclean: clean
 
+ifeq (${MP_BUILD}, 1) 
+clean_mup:
+	rm -rf ${MuP_OBJDIR}	
+else
+clean_mup:
+endif
+
 ##################################################################
-# Rules
+# Build Rules
 ##################################################################
 
+ifeq (${MP_BUILD}, 1) 
+
+MuP_OBJDIR = MuP_obj.dir
+MuP_OBJDIR_single = ${MuP_OBJDIR}/${join ${shell basename ${CURDIR}},_flt}
+MuP_OBJDIR_double = ${MuP_OBJDIR}/${join ${shell basename ${CURDIR}},_dbl}
+MuP_OBJDIR_longdouble = ${MuP_OBJDIR}/${join ${shell basename ${CURDIR}},_ldbl}
+#
+#MKDIR_P = mkdir -p
+#mkmpdir: 
+#	${MKDIR_P} ${MuP_OBJDIR}
+#
+ADD_OBJS_single = ${addprefix ${MuP_OBJDIR_single}/, ${notdir ${OBJS}}}
+${ADD_OBJS_single}: CFLAGS += -DMP_BUILD_SINGLE=1
+${MuP_OBJDIR_single}/%.o: %.c ${BLAS_HEADERS}
+#	@echo  "MuP_OBJDIR is ${ADD_OBJS_single} ... $%"
+	@echo  "processing C object $@ ..."
+	@mkdir -p ${@D}
+	${CC} ${CFLAGS} -c -o $@ $<
+#
+ADD_OBJS_double = ${addprefix ${MuP_OBJDIR_double}/, ${notdir ${OBJS}}}
+${ADD_OBJS_double}: CFLAGS += -DMP_BUILD_DOUBLE=1
+${MuP_OBJDIR_double}/%.o: %.c ${BLAS_HEADERS}
+	@echo  "processing C object $@ ..."
+	@mkdir -p ${@D}
+	${CC} ${CFLAGS} -c -o $@ $<
+#
+ADD_OBJS_longdouble = ${addprefix ${MuP_OBJDIR_longdouble}/, ${notdir ${OBJS}}}
+${ADD_OBJS_longdouble}: CFLAGS += -DMP_BUILD_LONGDOUBLE=1
+${MuP_OBJDIR_longdouble}/%.o: %.c ${BLAS_HEADERS}
+	@echo  "processing C object $@ ..."
+	@mkdir -p ${@D}
+	${CC} ${CFLAGS} -c -o $@ $<
+#
+# multiprecision object files
+blas_OBJS: ${ADD_OBJS_single} ${ADD_OBJS_double} ${ADD_OBJS_longdouble}
+#
+else
 ${OBJS}: ${BLAS_HEADERS}
+blas_OBJS: ${OBJS}
+endif
+

src/blas/_hypre_blas.h

@@ -19,6 +19,10 @@
 
 #include <HYPRE_config.h>
 
+#ifdef HYPRE_MIXED_PRECISION
+#include "blas_mup_func.h"
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif

src/blas/blas_func_header

@@ -0,0 +1,41 @@
+#!/bin/sh
+# Copyright (c) 1998 Lawrence Livermore National Security, LLC and other
+# HYPRE Project Developers. See the top-level COPYRIGHT file for details.
+#
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+# generate header file for transforming multiprecision function names
+
+MUP_HEADER=blas_mup_func.h
+
+cat > $MUP_HEADER <<@
+/*** DO NOT EDIT THIS FILE DIRECTLY (use 'blas_func_header' to generate) ***/
+@
+# Generate copyright header
+../config/writeHeader.sh $MUP_HEADER
+
+cat >> $MUP_HEADER <<@
+
+/******************************************************************************
+ * Header file for transforming multiprecision functions names
+ *****************************************************************************/
+
+#ifndef BLAS_MUP_FUNC_HEADER
+#define BLAS_MUP_FUNC_HEADER
+
+#include "multiprecision.h"
+
+@
+
+FIN=blas_functions.saved
+cat>> $MUP_HEADER <<@
+$(
+cat "$FIN" | while read -r func_name
+do
+# print dressed-up function 
+echo "#define $func_name HYPRE_MULTIPRECISION_FUNC ( $func_name )"
+done
+)
+
+#endif
+@

src/blas/blas_functions.saved

@@ -0,0 +1,30 @@
+ hypre_dasum
+ hypre_daxpy
+ hypre_dcopy
+ hypre_ddot
+ hypre_dgemm
+ hypre_dgemv
+ hypre_dger
+ hypre_dnrm2
+ hypre_drot
+ hypre_dscal
+ hypre_dswap
+ hypre_dsymm
+ hypre_dsymv
+ hypre_dsyr2
+ hypre_dsyr2k
+ hypre_dsyrk
+ hypre_dtrmm
+ hypre_dtrmv
+ hypre_dtrsm
+ hypre_dtrsv
+ hypre_d_lg10
+ hypre_d_sign
+ hypre_pow_dd
+ hypre_pow_di
+ hypre_s_cat
+ hypre_s_cmp
+ hypre_s_copy
+ hypre_idamax
+ hypre_blas_lsame
+ hypre_blas_xerbla

src/blas/blas_mup_func.h

@@ -0,0 +1,50 @@
+/*** DO NOT EDIT THIS FILE DIRECTLY (use 'blas_func_header' to generate) ***/
+
+/******************************************************************************
+ * Copyright (c) 1998 Lawrence Livermore National Security, LLC and other
+ * HYPRE Project Developers. See the top-level COPYRIGHT file for details.
+ *
+ * SPDX-License-Identifier: (Apache-2.0 OR MIT)
+ ******************************************************************************/
+
+/******************************************************************************
+ * Header file for transforming multiprecision functions names
+ *****************************************************************************/
+
+#ifndef BLAS_MUP_FUNC_HEADER
+#define BLAS_MUP_FUNC_HEADER
+
+#include "multiprecision.h"
+
+#define hypre_dasum HYPRE_MULTIPRECISION_FUNC ( hypre_dasum )
+#define hypre_daxpy HYPRE_MULTIPRECISION_FUNC ( hypre_daxpy )
+#define hypre_dcopy HYPRE_MULTIPRECISION_FUNC ( hypre_dcopy )
+#define hypre_ddot HYPRE_MULTIPRECISION_FUNC ( hypre_ddot )
+#define hypre_dgemm HYPRE_MULTIPRECISION_FUNC ( hypre_dgemm )
+#define hypre_dgemv HYPRE_MULTIPRECISION_FUNC ( hypre_dgemv )
+#define hypre_dger HYPRE_MULTIPRECISION_FUNC ( hypre_dger )
+#define hypre_dnrm2 HYPRE_MULTIPRECISION_FUNC ( hypre_dnrm2 )
+#define hypre_drot HYPRE_MULTIPRECISION_FUNC ( hypre_drot )
+#define hypre_dscal HYPRE_MULTIPRECISION_FUNC ( hypre_dscal )
+#define hypre_dswap HYPRE_MULTIPRECISION_FUNC ( hypre_dswap )
+#define hypre_dsymm HYPRE_MULTIPRECISION_FUNC ( hypre_dsymm )
+#define hypre_dsymv HYPRE_MULTIPRECISION_FUNC ( hypre_dsymv )
+#define hypre_dsyr2 HYPRE_MULTIPRECISION_FUNC ( hypre_dsyr2 )
+#define hypre_dsyr2k HYPRE_MULTIPRECISION_FUNC ( hypre_dsyr2k )
+#define hypre_dsyrk HYPRE_MULTIPRECISION_FUNC ( hypre_dsyrk )
+#define hypre_dtrmm HYPRE_MULTIPRECISION_FUNC ( hypre_dtrmm )
+#define hypre_dtrmv HYPRE_MULTIPRECISION_FUNC ( hypre_dtrmv )
+#define hypre_dtrsm HYPRE_MULTIPRECISION_FUNC ( hypre_dtrsm )
+#define hypre_dtrsv HYPRE_MULTIPRECISION_FUNC ( hypre_dtrsv )
+#define hypre_d_lg10 HYPRE_MULTIPRECISION_FUNC ( hypre_d_lg10 )
+#define hypre_d_sign HYPRE_MULTIPRECISION_FUNC ( hypre_d_sign )
+#define hypre_pow_dd HYPRE_MULTIPRECISION_FUNC ( hypre_pow_dd )
+#define hypre_pow_di HYPRE_MULTIPRECISION_FUNC ( hypre_pow_di )
+#define hypre_s_cat HYPRE_MULTIPRECISION_FUNC ( hypre_s_cat )
+#define hypre_s_cmp HYPRE_MULTIPRECISION_FUNC ( hypre_s_cmp )
+#define hypre_s_copy HYPRE_MULTIPRECISION_FUNC ( hypre_s_copy )
+#define hypre_idamax HYPRE_MULTIPRECISION_FUNC ( hypre_idamax )
+#define hypre_blas_lsame HYPRE_MULTIPRECISION_FUNC ( hypre_blas_lsame )
+#define hypre_blas_xerbla HYPRE_MULTIPRECISION_FUNC ( hypre_blas_xerbla )
+
+#endif

src/blas/hypre_blas.h

@@ -11,6 +11,11 @@
  * This header renames the functions in BLAS to avoid conflicts
  *--------------------------------------------------------------------------*/
 
+#ifdef HYPRE_MIXED_PRECISION
+#include "blas_mup_func.h"
+#endif
+
+
 /* blas */
 #define dasum_   hypre_dasum
 #define daxpy_   hypre_daxpy

src/config/HYPRE_config.h.in

@@ -121,6 +121,9 @@
 /* Define to 1 if using long long int for HYPRE_BigInt */
 #undef HYPRE_MIXEDINT
 
+/* Define to 1 if using mixed-precision for HYPRE_Real type */
+#undef HYPRE_MIXED_PRECISION
+
 /* Print HYPRE errors */
 #undef HYPRE_PRINT_ERRORS
 

src/config/Makefile.config.in

@@ -78,6 +78,8 @@ SHARED_BUILD_FLAG   = @SHARED_BUILD_FLAG@
 SHARED_SET_SONAME   = @SHARED_SET_SONAME@
 SHARED_OPTIONS      = @SHARED_OPTIONS@
 
+MP_BUILD     = @MP_BUILD@
+
 BUILD_PYTHON = @BUILD_PYTHON@
 PYTHON       = @PYTHON@
 

src/config/configure.in

@@ -352,6 +352,25 @@ then
    AC_DEFINE(HYPRE_COMPLEX, 1, [Define to 1 if using complex values])
 fi
 
+dnl *** Mixed-precision support build option
+AC_ARG_ENABLE(mixed-precision,
+AS_HELP_STRING([--enable-mixed-precision],
+               [Enable mixed-precision build (default is NO).]),
+[case "${enableval}" in
+    yes) hypre_using_fei=no
+         hypre_using_mixed_precision=yes ;;
+    no)  hypre_using_mixed_precision=no ;;
+    *)   AC_MSG_ERROR([Bad value ${enableval} for --enable-mixed-precision]) ;;
+ esac],
+[hypre_using_mixed_precision=no]
+)
+MP_BUILD=0
+if test "$hypre_using_mixed_precision" = "yes"
+then
+   AC_DEFINE(HYPRE_MIXED_PRECISION, 1, [Define to 1 if using mixed-precision for HYPRE_Real type])
+   MP_BUILD=1
+fi
+
 AC_ARG_ENABLE(maxdim,
 AS_HELP_STRING([--enable-maxdim=MAXDIM],
                [Change max dimension size to MAXDIM (default is 3).
@@ -2915,6 +2934,11 @@ dnl *********************************************************************
 AC_SUBST(DSUPERLU_INCLUDE)
 AC_SUBST(DSUPERLU_LIBS)
 
+dnl *********************************************************************
+dnl * Mixed precision build
+dnl *********************************************************************
+AC_SUBST(MP_BUILD)
+
 dnl *********************************************************************
 dnl * ar & ranlib substitution
 dnl *********************************************************************

src/config/generate_function_list.sh

@@ -0,0 +1,21 @@
+#!/bin/sh
+# Copyright (c) 1998 Lawrence Livermore National Security, LLC and other
+# HYPRE Project Developers. See the top-level COPYRIGHT file for details.
+#
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+# generate file containing function names in current folder.
+# This will be used to generate a header file for transforming 
+# multiprecision function names. It will also be useful for 
+# regression testing for the multiprecision build. 
+#
+# NOTE: It must be run on symbols generated from the 
+# non-multiprecision build of hypre.
+
+rootdir=$PWD
+rootdir="${rootdir%/}"
+rootname="${rootdir##*/}"
+
+## extract function names and remove leading and trailing underscores, if any
+## To include local functions (static functions) test on $2=="t" as well
+nm -A --defined-only *.o* | awk -F'[ ]' '$2=="T"{print $3}' | sed 's/_*//;s/_*$//' > ${rootname}_functions.out

src/config/writeHeader.sh

@@ -0,0 +1,17 @@
+#!/bin/sh
+# Copyright (c) 1998 Lawrence Livermore National Security, LLC and other
+# HYPRE Project Developers. See the top-level COPYRIGHT file for details.
+#
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+FNAME=$1
+cat >> $FNAME <<@
+
+/******************************************************************************
+ * Copyright (c) 1998 Lawrence Livermore National Security, LLC and other
+ * HYPRE Project Developers. See the top-level COPYRIGHT file for details.
+ *
+ * SPDX-License-Identifier: (Apache-2.0 OR MIT)
+ ******************************************************************************/
+@
+

src/configure

@@ -654,6 +654,7 @@ ac_header_c_list=
 ac_subst_vars='LTLIBOBJS
 LIBOBJS
 AR
+MP_BUILD
 DSUPERLU_LIBS
 DSUPERLU_INCLUDE
 SUPERLU_LIBS
@@ -805,6 +806,7 @@ enable_bigint
 enable_single
 enable_longdouble
 enable_complex
+enable_mixed_precision
 enable_maxdim
 enable_persistent
 enable_hopscotch
@@ -1548,6 +1550,8 @@ Optional Features:
   --enable-single         Use single precision values (default is NO).
   --enable-longdouble     Use long double precision values (default is NO).
   --enable-complex        Use complex values (default is NO).
+  --enable-mixed-precision
+                          Enable mixed-precision build (default is NO).
   --enable-maxdim=MAXDIM  Change max dimension size to MAXDIM (default is 3).
                           Currently must be at least 3.
   --enable-persistent     Uses persistent communication (default is NO).
@@ -3588,6 +3592,29 @@ printf "%s\n" "#define HYPRE_COMPLEX 1" >>confdefs.h
 
 fi
 
+# Check whether --enable-mixed-precision was given.
+if test ${enable_mixed_precision+y}
+then :
+  enableval=$enable_mixed_precision; case "${enableval}" in
+    yes) hypre_using_fei=no
+         hypre_using_mixed_precision=yes ;;
+    no)  hypre_using_mixed_precision=no ;;
+    *)   as_fn_error $? "Bad value ${enableval} for --enable-mixed-precision" "$LINENO" 5 ;;
+ esac
+else $as_nop
+  hypre_using_mixed_precision=no
+
+fi
+
+MP_BUILD=0
+if test "$hypre_using_mixed_precision" = "yes"
+then
+
+printf "%s\n" "#define HYPRE_MIXED_PRECISION 1" >>confdefs.h
+
+   MP_BUILD=1
+fi
+
 # Check whether --enable-maxdim was given.
 if test ${enable_maxdim+y}
 then :
@@ -11164,6 +11191,8 @@ printf "%s\n" "#define HYPRE_LINUX 1" >>confdefs.h
 
 
 
+
+