Merge pull request #3802 from martin-frbg/relafix

Fix cmake compilation of ReLAPACK and expose its INCLUDE_ALL option

CMakeLists.txt

@@ -212,10 +212,10 @@ if(NOT NO_LAPACKE)
   add_library(LAPACKE OBJECT ${LAPACKE_SOURCES})
   list(APPEND TARGET_OBJS "$<TARGET_OBJECTS:LAPACKE>")
 endif()
-if(BUILD_RELAPACK)
-  add_library(RELAPACK OBJECT ${RELA_SOURCES})
-  list(APPEND TARGET_OBJS "$<TARGET_OBJECTS:RELAPACK>")
-endif()
+#if(BUILD_RELAPACK)
+#  add_library(RELAPACK OBJECT ${RELA_SOURCES})
+#  list(APPEND TARGET_OBJS "$<TARGET_OBJECTS:RELAPACK>")
+#endif()
 set(OpenBLAS_LIBS "")
 if(BUILD_STATIC_LIBS)
   add_library(${OpenBLAS_LIBNAME}_static STATIC ${TARGET_OBJS} ${OpenBLAS_DEF_FILE})

Makefile.rule

@@ -131,6 +131,9 @@ BUILD_LAPACK_DEPRECATED = 1
 
 # Build RecursiveLAPACK on top of LAPACK
 # BUILD_RELAPACK = 1
+# Have RecursiveLAPACK actually replace standard LAPACK routines instead of 
+# just adding its equivalents with a RELAPACK_ prefix
+# RELAPACK_REPLACE = 1
 
 # If you want to use the legacy threaded Level 3 implementation.
 # USE_SIMPLE_THREADED_LEVEL3 = 1

Makefile.system

@@ -9,6 +9,10 @@ ifndef TOPDIR
 TOPDIR = . 
 endif
 
+ifndef RELAPACK_REPLACE
+RELAPACK_REPLACE=0
+endif
+
 # we need to use the host system's architecture for getarch compile options even especially when cross-compiling
 HOSTARCH := $(shell uname -m)
 ifeq ($(HOSTARCH), amd64)

interface/CMakeLists.txt

@@ -189,7 +189,16 @@ if (NOT DEFINED NO_LAPACK)
   )
 
   GenerateNamedObjects("${LAPACK_SOURCES}")
+  if (NOT RELAPACK_REPLACE)
   GenerateNamedObjects("${LAPACK_MANGLED_SOURCES}" "" "" 0 "" "" 0 3)
+  else ()
+  GenerateNamedObjects("lapack/getrs.c" "" "" 0 "" "" 0 3)
+  GenerateNamedObjects("lapack/getf2.c" "" "" 0 "" "" 0 3)
+  GenerateNamedObjects("lapack/potf2.c" "" "" 0 "" "" 0 3)
+  GenerateNamedObjects("lapack/laswp.c" "" "" 0 "" "" 0 3)
+  GenerateNamedObjects("lapack/lauu2.c" "" "" 0 "" "" 0 3)
+  GenerateNamedObjects("lapack/trti2.c" "" "" 0 "" "" 0 3)
+  endif()
 endif ()
 
 if ( BUILD_COMPLEX AND NOT  BUILD_SINGLE)

lapack/CMakeLists.txt

@@ -39,8 +39,12 @@ set(UNIT_SOURCES2
   trti2/trti2_L.c
 )
 
+if (NOT RELAPACK_REPLACE)
 GenerateNamedObjects("${LAPACK_SOURCES}")
 GenerateNamedObjects("${LAPACK_MANGLED_SOURCES}" "" "" false "" "" false 3)
+else()
+GenerateNamedObjects("${LAPACK_MANGLED_SOURCES}" "" "" false "" "" false 3)
+endif()
 
 GenerateNamedObjects("laswp/generic/laswp_k_4.c" "" "laswp_plus" false "" ""  false 3)
 GenerateNamedObjects("laswp/generic/laswp_k_4.c" "MINUS" "laswp_minus" false "" ""  false 3)
@@ -113,4 +117,3 @@ GenerateCombinationObjects("${UNIT_SOURCES}" "UNIT" "N" "" 4)
 GenerateCombinationObjects("${UNIT_SOURCES2}" "UNIT" "N" "" 0 "" "" 3)
 
 add_library(lapack OBJECT ${OPENBLAS_SRC})
-

relapack/Makefile

@@ -1,53 +1,61 @@
 TOPDIR  = ..
 include $(TOPDIR)/Makefile.system
 
-
+ifeq ($(RELAPACK_REPLACE),0)
+RELAPREFIX=RELAPACK_
+INCLALL=-DINCLUDE_ALL=0
+else
+INCLALL=-DINCLUDE_ALL=1
+endif
 
 SRC = $(wildcard src/*.c)
 
 SRC1 = \
-	src/slauum.c src/clauum.c src/dlauum.c src/zlauum.c \
-	src/strtri.c src/dtrtri.c src/ctrtri.c src/ztrtri.c \
-	src/spotrf.c src/dpotrf.c src/cpotrf.c src/zpotrf.c \
-	src/sgetrf.c src/dgetrf.c src/cgetrf.c src/zgetrf.c
+	slauum.c clauum.c dlauum.c zlauum.c \
+	strtri.c dtrtri.c ctrtri.c ztrtri.c \
+	spotrf.c dpotrf.c cpotrf.c zpotrf.c \
+	sgetrf.c dgetrf.c cgetrf.c zgetrf.c
 
 SRC2 = \
-	src/cgbtrf.c src/cpbtrf.c src/dsytrf_rec2.c src/sgbtrf.c src/ssytrf_rook.c src/zhegst.c src/zsytrf_rec2.c \
-	src/cgemmt.c src/dgbtrf.c src/dsytrf_rook.c src/sgemmt.c src/ssytrf_rook_rec2.c src/zhetrf.c src/zsytrf_rook.c \
-	src/csytrf.c src/dgemmt.c src/dsytrf_rook_rec2.c src/stgsyl.c src/zhetrf_rec2.c src/zsytrf_rook_rec2.c \
-	src/chegst.c src/csytrf_rec2.c src/dtgsyl.c src/strsyl.c src/zhetrf_rook.c src/ztgsyl.c \
-	src/chetrf.c src/csytrf_rook.c src/dtrsyl.c src/spbtrf.c src/strsyl_rec2.c src/zhetrf_rook_rec2.c src/ztrsyl.c \
-	src/chetrf_rec2.c src/csytrf_rook_rec2.c src/dpbtrf.c  src/dtrsyl_rec2.c  src/ztrsyl_rec2.c \
-	src/chetrf_rook.c src/ctgsyl.c src/ssygst.c src/zgbtrf.c src/zpbtrf.c  \
-	src/chetrf_rook_rec2.c src/ctrsyl.c src/dsygst.c src/f2c.c src/ssytrf.c src/zgemmt.c \
-	src/ctrsyl_rec2.c src/dsytrf.c src/lapack_wrappers.c src/ssytrf_rec2.c src/zsytrf.c
+	cgbtrf.c cpbtrf.c dsytrf_rec2.c sgbtrf.c ssytrf_rook.c zhegst.c zsytrf_rec2.c \
+	cgemmt.c dgbtrf.c dsytrf_rook.c sgemmt.c ssytrf_rook_rec2.c zhetrf.c zsytrf_rook.c \
+	csytrf.c dgemmt.c dsytrf_rook_rec2.c stgsyl.c zhetrf_rec2.c zsytrf_rook_rec2.c \
+	chegst.c csytrf_rec2.c dtgsyl.c strsyl.c zhetrf_rook.c ztgsyl.c \
+	chetrf.c csytrf_rook.c dtrsyl.c spbtrf.c strsyl_rec2.c zhetrf_rook_rec2.c ztrsyl.c \
+	chetrf_rec2.c csytrf_rook_rec2.c dpbtrf.c dtrsyl_rec2.c ztrsyl_rec2.c \
+	chetrf_rook.c ctgsyl.c ssygst.c zgbtrf.c zpbtrf.c  \
+	chetrf_rook_rec2.c ctrsyl.c dsygst.c f2c.c ssytrf.c zgemmt.c \
+	ctrsyl_rec2.c dsytrf.c lapack_wrappers.c ssytrf_rec2.c zsytrf.c
 
 SRCX = \
-	src/cgbtrf.c src/cpbtrf.c src/ctrtri.c src/dsytrf_rec2.c src/sgbtrf.c src/ssytrf_rook.c src/zhegst.c src/zsytrf_rec2.c \
-	src/cgemmt.c src/cpotrf.c src/dgbtrf.c src/dsytrf_rook.c src/sgemmt.c src/ssytrf_rook_rec2.c src/zhetrf.c src/zsytrf_rook.c \
-	src/cgetrf.c src/csytrf.c src/dgemmt.c src/dsytrf_rook_rec2.c src/sgetrf.c src/stgsyl.c src/zhetrf_rec2.c src/zsytrf_rook_rec2.c \
-	src/chegst.c src/csytrf_rec2.c src/dgetrf.c src/dtgsyl.c src/slauum.c src/strsyl.c src/zhetrf_rook.c src/ztgsyl.c \
-	src/chetrf.c src/csytrf_rook.c src/dlauum.c src/dtrsyl.c src/spbtrf.c src/strsyl_rec2.c src/zhetrf_rook_rec2.c src/ztrsyl.c \
-	src/chetrf_rec2.c src/csytrf_rook_rec2.c src/dpbtrf.c  src/dtrsyl_rec2.c src/spotrf.c src/strtri.c src/zlauum.c src/ztrsyl_rec2.c \
-	src/chetrf_rook.c src/ctgsyl.c src/dpotrf.c src/dtrtri.c src/ssygst.c src/zgbtrf.c src/zpbtrf.c src/ztrtri.c \
-	src/chetrf_rook_rec2.c src/ctrsyl.c src/dsygst.c src/f2c.c src/ssytrf.c src/zgemmt.c src/zpotrf.c \
-	src/clauum.c src/ctrsyl_rec2.c src/dsytrf.c src/lapack_wrappers.c src/ssytrf_rec2.c src/zgetrf.c src/zsytrf.c
-
-OBJS1 = $(SRC1:%.c=%.$(SUFFIX))
-OBJS2 = $(SRC2:%.c=%.o)
+	cgbtrf.c cpbtrf.c ctrtri.c dsytrf_rec2.c sgbtrf.c ssytrf_rook.c zhegst.c zsytrf_rec2.c \
+	cgemmt.c cpotrf.c dgbtrf.c dsytrf_rook.c sgemmt.c ssytrf_rook_rec2.c zhetrf.c zsytrf_rook.c \
+	cgetrf.c csytrf.c dgemmt.c dsytrf_rook_rec2.c sgetrf.c stgsyl.c zhetrf_rec2.c zsytrf_rook_rec2.c \
+	chegst.c csytrf_rec2.c dgetrf.c dtgsyl.c slauum.c strsyl.c zhetrf_rook.c ztgsyl.c \
+	chetrf.c csytrf_rook.c dlauum.c dtrsyl.c spbtrf.c strsyl_rec2.c zhetrf_rook_rec2.c ztrsyl.c \
+	chetrf_rec2.c csytrf_rook_rec2.c dpbtrf.c dtrsyl_rec2.c spotrf.c strtri.c zlauum.c ztrsyl_rec2.c \
+	chetrf_rook.c ctgsyl.c dpotrf.c dtrtri.c ssygst.c zgbtrf.c zpbtrf.c ztrtri.c \
+	chetrf_rook_rec2.c ctrsyl.c dsygst.c f2c.c ssytrf.c zgemmt.c zpotrf.c \
+	clauum.c ctrsyl_rec2.c dsytrf.c lapack_wrappers.c ssytrf_rec2.c zgetrf.c zsytrf.c
+
+
+OBJS1 = $(SRC1:%.c=src/$(RELAPREFIX)%.$(SUFFIX))
+OBJS2 = $(SRC2:%.c=src/$(RELAPREFIX)%.o)
 OBJS = $(OBJS1) $(OBJS2)
 
 TEST_SUITS = \
-	slauum dlauum clauum zlauum \
-	spotrf dpotrf cpotrf zpotrf \
-	spbtrf dpbtrf cpbtrf zpbtrf \
-	ssygst dsygst chegst zhegst \
-	ssytrf dsytrf csytrf chetrf zsytrf zhetrf \
-	sgetrf dgetrf cgetrf zgetrf \
-	sgbtrf dgbtrf cgbtrf zgbtrf \
-	strsyl dtrsyl ctrsyl ztrsyl \
-	stgsyl dtgsyl ctgsyl ztgsyl \
 	sgemmt dgemmt cgemmt zgemmt
+
+	#	slauum dlauum clauum zlauum \
+	#	spotrf dpotrf cpotrf zpotrf \
+	#	spbtrf dpbtrf cpbtrf zpbtrf \
+	#	ssygst dsygst chegst zhegst \
+	#	ssytrf dsytrf csytrf chetrf zsytrf zhetrf \
+	#	sgetrf dgetrf cgetrf zgetrf \
+	#	sgbtrf dgbtrf cgbtrf zgbtrf \
+	#	strsyl dtrsyl ctrsyl ztrsyl \
+	#	stgsyl dtgsyl ctgsyl ztgsyl \
+
 TESTS = $(TEST_SUITS:%=test/%.pass)  # dummies
 TEST_EXES = $(TEST_SUITS:%=test/%.x)
 
@@ -63,11 +71,11 @@ libs:	$(OBJS)
 	$(AR) -r  $(TOPDIR)/$(LIBNAME) $(OBJS)
 	$(RANLIB) $(TOPDIR)/$(LIBNAME)
 
-%.$(SUFFIX): %.c config.h
-	$(CC) $(CFLAGS) -c $< -o $@
+src/$(RELAPREFIX)%.$(SUFFIX): src/%.c relapack_config.h
+	$(CC) -v $(CFLAGS) -I. $(INCLALL) -c $< -o $@
 
-%.o: %.c config.h
-	$(CC) $(CFLAGS) -c $< -o $@
+src/$(RELAPREFIX)%.o: src/%.c relapack_config.h
+	$(CC) -v $(CFLAGS) -I. $(INCLALL) -c $< -o $@
 
 
 # ReLAPACK testing

relapack/config.h → relapack/relapack_config.h

@@ -45,7 +45,7 @@
 // The following macros specify which routines are included in the library under
 // LAPACK's symbol names: 1 included, 0 not included
 
-#define INCLUDE_ALL 1
+// #define INCLUDE_ALL 1
 
 #define INCLUDE_XLAUUM INCLUDE_ALL
 #define INCLUDE_SLAUUM INCLUDE_XLAUUM
@@ -115,7 +115,7 @@
 #define INCLUDE_CTGSYL INCLUDE_XTGSYL
 #define INCLUDE_ZTGSYL INCLUDE_XTGSYL
 
-#define INCLUDE_XGEMMT 1
+#define INCLUDE_XGEMMT INCLUDE_ALL
 #define INCLUDE_SGEMMT INCLUDE_XGEMMT
 #define INCLUDE_DGEMMT INCLUDE_XGEMMT
 #define INCLUDE_CGEMMT INCLUDE_XGEMMT

relapack/src/CMakeLists.txt

@@ -1,5 +1,6 @@
 include_directories(${PROJECT_SOURCE_DIR})
 include_directories(${PROJECT_BINARY_DIR})
+include_directories(${PROJECT_SOURCE_DIR}/relapack)
 
 set(RELAFILES
 clauum.c

relapack/src/ctrsyl_rec2.c

@@ -10,7 +10,7 @@
 		http://www.netlib.org/f2c/libf2c.zip
 */
 
-#include "../config.h"
+#include "relapack_config.h"
 #include "f2c.h"
 
 #if BLAS_COMPLEX_FUNCTIONS_AS_ROUTINES

relapack/src/relapack.h

@@ -1,7 +1,7 @@
 #ifndef RELAPACK_INT_H
 #define RELAPACK_INT_H
 #include <string.h>
-#include "../../config.h"
+#include "config.h"
 #if defined(OS_WINDOWS) && defined(__64BIT__)
 typedef long long BLASLONG;
 typedef unsigned long long BLASULONG;
@@ -9,7 +9,7 @@ typedef unsigned long long BLASULONG;
 typedef long BLASLONG;
 typedef unsigned long BLASULONG;
 #endif
-#include "../config.h"
+#include "relapack_config.h"
 
 #include "../inc/relapack.h"
 

relapack/src/ztrsyl_rec2.c

@@ -10,7 +10,7 @@
 		http://www.netlib.org/f2c/libf2c.zip
 */
 
-#include "../config.h"
+#include "relapack_config.h"
 #include "f2c.h"
 
 #if BLAS_COMPLEX_FUNCTIONS_AS_ROUTINES