Merge pull request #83 from xianyi/develop

rebase

.travis.yml

@@ -43,6 +43,18 @@ matrix:
         - TARGET_BOX=IBMZ_LINUX
         - BTYPE="BINARY=64 USE_OPENMP=1"
 
+    - <<: *test-ubuntu
+      os: linux
+      dist: focal
+      arch: s390x
+      compiler: clang
+      before_script:
+        - COMMON_FLAGS="DYNAMIC_ARCH=1 TARGET=Z13 NUM_THREADS=32"
+      env:
+        # for matrix annotation only
+        - TARGET_BOX=IBMZ_LINUX
+        - BTYPE="BINARY=64 USE_OPENMP=0 CC=clang"
+
     - <<: *test-ubuntu
       env:
         - TARGET_BOX=LINUX64

CONTRIBUTORS.md

@@ -187,6 +187,7 @@ In chronological order:
 * Marius Hillenbrand <https://github.com/mhillenibm>
   * [2020-05-12] Revise dynamic architecture detection for IBM z
   * [2020-05-12] Add new sgemm and strmm kernel for IBM z14
+  * [2020-09-07] Fix builds with clang on IBM z, including dynamic architecture support
 
 * Danfeng Zhang <https://github.com/craft-zhang>
   * [2020-05-20] Improve performance of SGEMM and STRMM on Arm Cortex-A53

Makefile.system

@@ -295,7 +295,6 @@ endif
 ifeq ($(C_COMPILER), GCC)
 GCCVERSIONGTEQ4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 4)
 GCCVERSIONGT4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 4)
-GCCVERSIONEQ5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` = 5)
 GCCVERSIONGT5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 5)
 GCCVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 7)
 GCCVERSIONGTEQ9 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 9)
@@ -594,34 +593,36 @@ endif
 ifeq ($(ARCH), zarch)
 DYNAMIC_CORE = ZARCH_GENERIC
 
-# Z13 is supported since gcc-5.2, gcc-6, and in RHEL 7.3 and newer
-ifeq ($(GCCVERSIONGT5), 1)
-	ZARCH_SUPPORT_Z13 := 1
-else ifeq ($(GCCVERSIONEQ5), 1)
-ifeq ($(GCCMINORVERSIONGTEQ2), 1)
-	ZARCH_SUPPORT_Z13 := 1
-endif
-endif
-
-ifeq ($(wildcard /etc/redhat-release), /etc/redhat-release)
-ifeq ($(shell source /etc/os-release ; expr $$VERSION_ID \>= "7.3"), 1)
-	ZARCH_SUPPORT_Z13 := 1
-endif
-endif
-
-ifeq ($(ZARCH_SUPPORT_Z13), 1)
+# if the compiler accepts -march=arch11 or -march=z13 and can compile a file
+# with z13-specific inline assembly, then we can include support for Z13.
+# note: -march=z13 is equivalent to -march=arch11 yet some compiler releases
+# only support one or the other.
+# note: LLVM version 6.x supported -march=z13 yet could not handle vector
+# registers in inline assembly, so the check for supporting the -march flag is
+# not enough.
+ZARCH_TEST_COMPILE=-c $(TOPDIR)/kernel/zarch/damin_z13.c -I$(TOPDIR) -o /dev/null > /dev/null 2> /dev/null
+ZARCH_CC_SUPPORTS_ARCH11=$(shell $(CC) -march=arch11 $(ZARCH_TEST_COMPILE) && echo 1)
+ZARCH_CC_SUPPORTS_Z13=$(shell $(CC) -march=z13 $(ZARCH_TEST_COMPILE) && echo 1)
+
+ifeq ($(or $(ZARCH_CC_SUPPORTS_ARCH11), $(ZARCH_CC_SUPPORTS_Z13)), 1)
 DYNAMIC_CORE += Z13
+CCOMMON_OPT += -DDYN_Z13
 else
-$(info OpenBLAS: Not building Z13 kernels because gcc is older than 5.2 or 6.x)
+$(info OpenBLAS: Not building Z13 kernels because the compiler $(CC) does not support it)
 endif
 
-ifeq ($(GCCVERSIONGTEQ7), 1)
+# as above for z13, check for -march=arch12 and z14 support in the compiler.
+ZARCH_CC_SUPPORTS_ARCH12=$(shell $(CC) -march=arch12 $(ZARCH_TEST_COMPILE) && echo 1)
+ZARCH_CC_SUPPORTS_Z14=$(shell $(CC) -march=z14 $(ZARCH_TEST_COMPILE) && echo 1)
+ifeq ($(or $(ZARCH_CC_SUPPORTS_ARCH12), $(ZARCH_CC_SUPPORTS_Z14)), 1)
 DYNAMIC_CORE += Z14
+CCOMMON_OPT += -DDYN_Z14
 else
-$(info OpenBLAS: Not building Z14 kernels because gcc is older than 7.x)
-endif
+$(info OpenBLAS: Not building Z14 kernels because the compiler $(CC) does not support it)
 endif
 
+endif # ARCH zarch
+
 ifeq ($(ARCH), power)
 DYNAMIC_CORE = POWER6
 DYNAMIC_CORE += POWER8

driver/others/dynamic_zarch.c

@@ -1,18 +1,6 @@
 #include "common.h"
 #include <stdbool.h>
 
-// Gate kernels for z13 and z14 on gcc version
-#if (__GNUC__ == 5 && __GNUC_MINOR__ >= 2) || __GNUC__ >= 6 ||           \
-    /* RHEL 7 since 7.3: */                                              \
-    (__GNUC__ == 4 && __GNUC_MINOR__ == 8 && __GNUC_PATCHLEVEL__ == 5 && \
-     __GNUC_RH_RELEASE__ >= 11)
-#define HAVE_Z13_SUPPORT
-#endif
-
-#if __GNUC__ >= 7
-#define HAVE_Z14_SUPPORT
-#endif
-
 // Guard the use of getauxval() on glibc version >= 2.16
 #ifdef __GLIBC__
 #include <features.h>
@@ -47,10 +35,10 @@ static unsigned long get_hwcap(void) {
 #endif // __GLIBC
 
 extern gotoblas_t gotoblas_ZARCH_GENERIC;
-#ifdef HAVE_Z13_SUPPORT
+#ifdef DYN_Z13
 extern gotoblas_t gotoblas_Z13;
 #endif
-#ifdef HAVE_Z14_SUPPORT
+#ifdef DYN_Z14
 extern gotoblas_t gotoblas_Z14;
 #endif
 
@@ -66,10 +54,10 @@ static char* corename[] = {
 };
 
 char* gotoblas_corename(void) {
-#ifdef HAVE_Z13_SUPPORT
+#ifdef DYN_Z13
 	if (gotoblas == &gotoblas_Z13)	return corename[1];
 #endif
-#ifdef HAVE_Z14_SUPPORT
+#ifdef DYN_Z14
 	if (gotoblas == &gotoblas_Z14)	return corename[2];
 #endif
 	if (gotoblas == &gotoblas_ZARCH_GENERIC) return corename[3];
@@ -77,6 +65,10 @@ char* gotoblas_corename(void) {
 	return corename[0];
 }
 
+#ifndef HWCAP_S390_VXE
+#define HWCAP_S390_VXE 8192
+#endif
+
 /**
  * Detect the fitting set of kernels by retrieving the CPU features supported by
  * OS from the auxiliary value AT_HWCAP and choosing the set of kernels
@@ -89,15 +81,15 @@ static gotoblas_t* get_coretype(void) {
 
 	unsigned long hwcap __attribute__((unused)) = get_hwcap();
 
+#ifdef DYN_Z14
 	// z14 and z15 systems: exploit Vector Facility (SIMD) and
 	// Vector-Enhancements Facility 1 (float SIMD instructions), if present.
-#ifdef HAVE_Z14_SUPPORT
 	if ((hwcap & HWCAP_S390_VX) && (hwcap & HWCAP_S390_VXE))
 		return &gotoblas_Z14;
 #endif
 
+#ifdef DYN_Z13
 	// z13: Vector Facility (SIMD for double)
-#ifdef HAVE_Z13_SUPPORT
 	if (hwcap & HWCAP_S390_VX)
 		return &gotoblas_Z13;
 #endif
@@ -123,19 +115,27 @@ static gotoblas_t* force_coretype(char* coretype) {
 		}
 	}
 
-	switch (found)
-	{
-#ifdef HAVE_Z13_SUPPORT
-	case  1: return (&gotoblas_Z13);
+	if (found == 1) {
+#ifdef DYN_Z13
+		return &gotoblas_Z13;
+#else
+		openblas_warning(1, "Z13 support not compiled in");
+		return NULL;
 #endif
-#ifdef HAVE_Z14_SUPPORT
-	case  2: return (&gotoblas_Z14);
+	} else if (found == 2) {
+#ifdef DYN_Z14
+		return &gotoblas_Z14;
+#else
+		openblas_warning(1, "Z14 support not compiled in");
+		return NULL;
 #endif
-	case  3: return (&gotoblas_ZARCH_GENERIC);
-	default: return NULL;
+	} else if (found == 3) {
+		return &gotoblas_ZARCH_GENERIC;
 	}
+
 	snprintf(message, 128, "Core not found: %s\n", coretype);
 	openblas_warning(1, message);
+	return NULL;
 }
 
 void gotoblas_dynamic_init(void) {

lapack-netlib/TESTING/EIG/cchkst2stg.f

@@ -1014,8 +1014,8 @@
 *           the one from above. Compare it with D1 computed 
 *           using the 1-stage.
 *
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
             CALL CLACPY( 'U', N, N, A, LDA, V, LDU )
             LH = MAX(1, 4*N)
             LW = LWORK - LH
@@ -1048,8 +1048,8 @@
 *           the one from above. Compare it with D1 computed 
 *           using the 1-stage. 
 *
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
             CALL CLACPY( 'L', N, N, A, LDA, V, LDU )
             CALL CHETRD_2STAGE( 'N', "L", N, V, LDU, SD, SE, TAU, 
      $                   WORK, LH, WORK( LH+1 ), LW, IINFO )

lapack-netlib/TESTING/EIG/dchksb2stg.f

@@ -670,8 +670,8 @@
 *              the one from above. Compare it with D1 computed 
 *              using the DSBTRD.
 *            
-               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
                CALL DLACPY( ' ', K+1, N, A, LDA, U, LDU )
                LH = MAX(1, 4*N)
                LW = LWORK - LH
@@ -743,8 +743,8 @@
 *              the one from above. Compare it with D1 computed 
 *              using the DSBTRD. 
 *           
-               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
                CALL DLACPY( ' ', K+1, N, A, LDA, U, LDU )
                LH = MAX(1, 4*N)
                LW = LWORK - LH

lapack-netlib/TESTING/EIG/dchkst2stg.f

@@ -999,8 +999,8 @@
 *           the one from above. Compare it with D1 computed 
 *           using the 1-stage.
 *
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
             CALL DLACPY( "U", N, N, A, LDA, V, LDU )
             LH = MAX(1, 4*N)
             LW = LWORK - LH
@@ -1032,8 +1032,8 @@
 *           the one from above. Compare it with D1 computed 
 *           using the 1-stage. 
 *
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
             CALL DLACPY( "L", N, N, A, LDA, V, LDU )
             CALL DSYTRD_2STAGE( 'N', "L", N, V, LDU, SD, SE, TAU, 
      $                   WORK, LH, WORK( LH+1 ), LW, IINFO )

lapack-netlib/TESTING/EIG/zchkhb2stg.f

@@ -680,8 +680,8 @@
 *              the one from above. Compare it with D1 computed 
 *              using the DSBTRD.
 *            
-               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
                CALL ZLACPY( ' ', K+1, N, A, LDA, U, LDU )
                LH = MAX(1, 4*N)
                LW = LWORK - LH
@@ -753,8 +753,8 @@
 *              the one from above. Compare it with D1 computed 
 *              using the DSBTRD. 
 *           
-               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+               CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
                CALL ZLACPY( ' ', K+1, N, A, LDA, U, LDU )
                LH = MAX(1, 4*N)
                LW = LWORK - LH

lapack-netlib/TESTING/EIG/zchkst2stg.f

@@ -1014,8 +1014,8 @@
 *           the one from above. Compare it with D1 computed 
 *           using the 1-stage.
 *
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
             CALL ZLACPY( 'U', N, N, A, LDA, V, LDU )
             LH = MAX(1, 4*N)
             LW = LWORK - LH
@@ -1048,8 +1048,8 @@
 *           the one from above. Compare it with D1 computed 
 *           using the 1-stage. 
 *
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, 1 )
-            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, 1 )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SD, N )
+            CALL DLASET( 'Full', N, 1, ZERO, ZERO, SE, N )
             CALL ZLACPY( 'L', N, N, A, LDA, V, LDU )
             CALL ZHETRD_2STAGE( 'N', "L", N, V, LDU, SD, SE, TAU, 
      $                   WORK, LH, WORK( LH+1 ), LW, IINFO )