Merge pull request #3061 from martin-frbg/arm64-pgi

Support NVIDIA HPC SDK on ARM64

Author: martin-frbg

Makefile.arm64

@@ -1,4 +1,4 @@
-
+ifneq ($(C_COMPILER), PGI)
 ifeq ($(CORE), ARMV8)
 CCOMMON_OPT += -march=armv8-a
 FCOMMON_OPT += -march=armv8-a
@@ -77,4 +77,4 @@ CCOMMON_OPT += -march=armv8.2-a -mtune=tsv110
 FCOMMON_OPT += -march=armv8.2-a -mtune=tsv110
 endif
 endif
-
+endif

cmake/utils.cmake

@@ -74,6 +74,9 @@ macro(ParseMakefileVars MAKEFILE_IN)
           string(REGEX MATCH "ifneq \\(\\$\\(([_A-Z]+)\\),[ \t]*([0-9_A-Z]+)\\)" line_match "${makefile_line}")
           if (NOT "${line_match}" STREQUAL "")
 #            message(STATUS "IFNEQ: ${line_match} first: ${CMAKE_MATCH_1} second: ${CMAKE_MATCH_2}")
+            if ( ${CMAKE_MATCH_1} STREQUAL C_COMPILER)
+                 set (CMAKE_MATCH_1 CMAKE_C_COMPILER)
+            endif ()
             if (NOT ( ${${CMAKE_MATCH_1}} STREQUAL ${CMAKE_MATCH_2}))
 #              message (STATUS "condition is true")
               set (IfElse 1)

common_arm64.h

@@ -39,7 +39,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define INLINE inline
 
-#ifdef F_INTERFACE_FLANG
+#if defined( F_INTERFACE_FLANG) || defined(F_INTERFACE_PGI)
 #define RETURN_BY_STACK
 #else
 #define RETURN_BY_COMPLEX

kernel/arm/zdot.c

@@ -48,7 +48,7 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
 
 	dot[0]=0.0;
 	dot[1]=0.0;
-#if !defined(__PPC__) && !defined(__SunOS)
+#if !defined(__PPC__) && !defined(__SunOS) && !defined(__PGI)
 	CREAL(result) = 0.0 ;
 	CIMAG(result) = 0.0 ;
 #else
@@ -73,7 +73,7 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
 		i++ ;
 
 	}
-#if !defined(__PPC__)	&& !defined(__SunOS)
+#if !defined(__PPC__)	&& !defined(__SunOS) && !defined(__PGI)
         CREAL(result) = dot[0];
 	CIMAG(result) = dot[1];
 #else

kernel/arm64/KERNEL.ARMV8

@@ -97,9 +97,18 @@ CNRM2KERNEL    = znrm2.S
 ZNRM2KERNEL    = znrm2.S
 
 DDOTKERNEL     = dot.S
+ifneq ($(C_COMPILER), PGI)
 SDOTKERNEL     = ../generic/dot.c
+else
+SDOTKERNEL = dot.S
+endif
+ifneq ($(C_COMPILER), PGI)
 CDOTKERNEL     = zdot.S
 ZDOTKERNEL     = zdot.S
+else
+CDOTKERNEL = ../arm/zdot.c
+ZDOTKERNEL = ../arm/zdot.c
+endif
 DSDOTKERNEL    = dot.S
 
 DGEMM_BETA     = dgemm_beta.S

kernel/arm64/KERNEL.CORTEXA53

@@ -96,11 +96,20 @@ DNRM2KERNEL    = nrm2.S
 CNRM2KERNEL    = znrm2.S
 ZNRM2KERNEL    = znrm2.S
 
-DDOTKERNEL     = dot.S
-SDOTKERNEL     = ../generic/dot.c
-CDOTKERNEL     = zdot.S
-ZDOTKERNEL     = zdot.S
-DSDOTKERNEL    = dot.S
+ifneq ($(C_COMPILER), PGI)
+SDOTKERNEL   = ../generic/dot.c
+else
+SDOTKERNEL   = dot.S
+endif
+DDOTKERNEL   = dot.S
+ifneq ($(C_COMPILER), PGI)
+CDOTKERNEL   = zdot.S
+ZDOTKERNEL   = zdot.S
+else
+CDOTKERNEL = ../arm/zdot.c
+ZDOTKERNEL = ../arm/zdot.c
+endif
+DSDOTKERNEL  = dot.S
 
 DGEMM_BETA     = dgemm_beta.S
 SGEMM_BETA     = sgemm_beta.S

kernel/arm64/KERNEL.CORTEXA57

@@ -70,10 +70,19 @@ DCOPYKERNEL  = copy.S
 CCOPYKERNEL  = copy.S
 ZCOPYKERNEL  = copy.S
 
+ifneq ($(C_COMPILER), PGI)
 SDOTKERNEL   = ../generic/dot.c
+else
+SDOTKERNEL   = dot.S
+endif
 DDOTKERNEL   = dot.S
+ifneq ($(C_COMPILER), PGI)
 CDOTKERNEL   = zdot.S
 ZDOTKERNEL   = zdot.S
+else
+CDOTKERNEL = ../arm/zdot.c
+ZDOTKERNEL = ../arm/zdot.c
+endif
 DSDOTKERNEL  = dot.S
 
 SNRM2KERNEL  = nrm2.S

kernel/arm64/KERNEL.THUNDERX

@@ -47,8 +47,13 @@ ZCOPYKERNEL  = copy.S
 
 SDOTKERNEL   = dot_thunderx.c
 DDOTKERNEL   = ddot_thunderx.c
+ifneq ($(C_COMPILER), PGI)
 CDOTKERNEL   = zdot.S
 ZDOTKERNEL   = zdot.S
+else
+CDOTKERNEL = ../arm/zdot.c
+ZDOTKERNEL = ../arm/zdot.c
+endif
 DSDOTKERNEL  = dot.S
 
 SNRM2KERNEL  = nrm2.S

kernel/arm64/KERNEL.TSV110

@@ -72,8 +72,13 @@ ZCOPYKERNEL  = copy.S
 
 SDOTKERNEL   = dot.S
 DDOTKERNEL   = dot.S
+ifneq ($(C_COMPILER), PGI)
 CDOTKERNEL   = zdot.S
 ZDOTKERNEL   = zdot.S
+else
+CDOTKERNEL = ../arm/zdot.c
+ZDOTKERNEL = ../arm/zdot.c
+endif
 DSDOTKERNEL  = dot.S
 
 SNRM2KERNEL  = nrm2.S