Merge pull request #58 from xianyi/develop

rebase

Author: martin-frbg

CONTRIBUTORS.md

@@ -183,4 +183,7 @@ In chronological order:
 
 * Rajalakshmi Srinivasaraghavan <https://github.com/RajalakshmiSR>
   * [2020-04-15] Half-precision GEMM for bfloat16
-  
+
+* 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

Makefile.system

@@ -563,8 +563,27 @@ DYNAMIC_CORE += EMAG8180
 endif
 
 ifeq ($(ARCH), zarch)
-DYNAMIC_CORE = Z13
+DYNAMIC_CORE = ZARCH_GENERIC
+
+# Z13 is supported since gcc-5.2, gcc-6, and in RHEL 7.3 and newer
+GCC_GE_52 := $(subst 0,,$(shell expr `$(CC) -dumpversion` \>= "5.2"))
+
+ifeq ($(wildcard /etc/redhat-release), /etc/redhat-release)
+RHEL_WITH_Z13 := $(subst 0,,$(shell source /etc/os-release ; expr $$VERSION_ID \>= "7.3"))
+endif
+
+ifeq ($(or $(GCC_GE_52),$(RHEL_WITH_Z13)), 1)
+DYNAMIC_CORE += Z13
+else
+$(info OpenBLAS: Not building Z13 kernels because gcc is older than 5.2 or 6.x)
+endif
+
+GCC_MAJOR_GE_7 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 7)
+ifeq ($(GCC_MAJOR_GE_7), 1)
 DYNAMIC_CORE += Z14
+else
+$(info OpenBLAS: Not building Z14 kernels because gcc is older than 7.x)
+endif
 endif
 
 ifeq ($(ARCH), power)
@@ -855,7 +874,7 @@ ifneq ($(INTERFACE64), 0)
 FCOMMON_OPT += -i8
 endif
 endif
-FCOMMON_OPT += -recursive
+FCOMMON_OPT += -recursive -fp-model strict -assume protect-parens
 ifeq ($(USE_OPENMP), 1)
 FCOMMON_OPT += -fopenmp
 endif

Makefile.zarch

@@ -5,6 +5,6 @@ FCOMMON_OPT += -march=z13 -mzvector
 endif
 
 ifeq ($(CORE), Z14)
-CCOMMON_OPT += -march=z14 -mzvector
+CCOMMON_OPT += -march=z14 -mzvector -O3
 FCOMMON_OPT += -march=z14 -mzvector
 endif

benchmark/Makefile

@@ -49,6 +49,12 @@ else
 GOTO_LAPACK_TARGETS=
 endif
 
+ifeq ($(BUILD_HALF),1)
+GOTO_HALF_TARGETS=shgemm.goto
+else
+GOTO_HALF_TARGETS=
+endif
+
 ifeq ($(OSNAME), WINNT)
 
 goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
@@ -91,7 +97,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
        sgetri.goto dgetri.goto cgetri.goto zgetri.goto \
        spotrf.goto dpotrf.goto cpotrf.goto zpotrf.goto \
        ssymm.goto dsymm.goto csymm.goto zsymm.goto \
-       saxpby.goto daxpby.goto caxpby.goto zaxpby.goto
+       saxpby.goto daxpby.goto caxpby.goto zaxpby.goto $(GOTO_HALF_TARGETS)
 
 acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        scholesky.acml dcholesky.acml ccholesky.acml zcholesky.acml \
@@ -264,7 +270,7 @@ goto :: sgemm.goto dgemm.goto cgemm.goto zgemm.goto \
        samin.goto damin.goto camin.goto zamin.goto \
        smin.goto dmin.goto \
        saxpby.goto daxpby.goto caxpby.goto zaxpby.goto \
-       snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto $(GOTO_LAPACK_TARGETS)
+       snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto $(GOTO_LAPACK_TARGETS) $(GOTO_HALF_TARGETS)
 
 acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        scholesky.acml dcholesky.acml ccholesky.acml zcholesky.acml \
@@ -614,6 +620,11 @@ zcholesky.essl : zcholesky.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBESSL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 ##################################### Sgemm ####################################################
+ifeq ($(BUILD_HALF),1)
+shgemm.goto : shgemm.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+endif
+
 sgemm.goto : sgemm.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
@@ -2916,6 +2927,11 @@ ccholesky.$(SUFFIX) : cholesky.c
 zcholesky.$(SUFFIX) : cholesky.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
+ifeq ($(BUILD_HALF),1)
+shgemm.$(SUFFIX) : gemm.c
+	$(CC) $(CFLAGS) -c -DHALF -UCOMPLEX -UDOUBLE -o $(@F) $^
+endif
+
 sgemm.$(SUFFIX) : gemm.c
 	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
 

benchmark/gemm.c

@@ -39,6 +39,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifdef DOUBLE
 #define GEMM   BLASFUNC(dgemm)
+#elif defined(HALF)
+#define GEMM   BLASFUNC(shgemm)
 #else
 #define GEMM   BLASFUNC(sgemm)
 #endif
@@ -120,7 +122,8 @@ static void *huge_malloc(BLASLONG size){
 
 int main(int argc, char *argv[]){
 
-  FLOAT *a, *b, *c;
+  IFLOAT *a, *b;
+  FLOAT *c;
   FLOAT alpha[] = {1.0, 0.0};
   FLOAT beta [] = {0.0, 0.0};
   char transa = 'N';
@@ -184,10 +187,10 @@ int main(int argc, char *argv[]){
     k = to;
   }
 
-  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * m * k * COMPSIZE)) == NULL) {
+  if (( a = (IFLOAT *)malloc(sizeof(IFLOAT) * m * k * COMPSIZE)) == NULL) {
     fprintf(stderr,"Out of Memory!!\n");exit(1);
   }
-  if (( b = (FLOAT *)malloc(sizeof(FLOAT) * k * n * COMPSIZE)) == NULL) {
+  if (( b = (IFLOAT *)malloc(sizeof(IFLOAT) * k * n * COMPSIZE)) == NULL) {
     fprintf(stderr,"Out of Memory!!\n");exit(1);
   }
   if (( c = (FLOAT *)malloc(sizeof(FLOAT) * m * n * COMPSIZE)) == NULL) {
@@ -199,10 +202,10 @@ int main(int argc, char *argv[]){
 #endif
 
   for (i = 0; i < m * k * COMPSIZE; i++) {
-    a[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+    a[i] = ((IFLOAT) rand() / (IFLOAT) RAND_MAX) - 0.5;
   }
   for (i = 0; i < k * n * COMPSIZE; i++) {
-    b[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+    b[i] = ((IFLOAT) rand() / (IFLOAT) RAND_MAX) - 0.5;
   }
   for (i = 0; i < m * n * COMPSIZE; i++) {
     c[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;

cmake/utils.cmake

@@ -43,7 +43,8 @@ macro(ParseMakefileVars MAKEFILE_IN)
     if (NOT "${line_match}" STREQUAL "")
 #message(STATUS "match on ${line_match}")
       set(var_name ${CMAKE_MATCH_1})
-      set(var_value ${CMAKE_MATCH_2})
+#      set(var_value ${CMAKE_MATCH_2})
+      string(STRIP ${CMAKE_MATCH_2} var_value)
       # check for Makefile variables in the string, e.g. $(TSUFFIX)
       string(REGEX MATCHALL "\\$\\(([0-9_a-zA-Z]+)\\)" make_var_matches ${var_value})
       foreach (make_var ${make_var_matches})
@@ -63,7 +64,7 @@ macro(ParseMakefileVars MAKEFILE_IN)
         string(REGEX MATCH "ifeq \\(\\$\\(([_A-Z]+)\\),[ \t]*([0-9_A-Z]+)\\)" line_match "${makefile_line}")
         if (NOT "${line_match}" STREQUAL "")
 #          message(STATUS "IFEQ: ${line_match} first: ${CMAKE_MATCH_1} second: ${CMAKE_MATCH_2}")
-          if (${${CMAKE_MATCH_1}} STREQUAL ${CMAKE_MATCH_2})
+          if (DEFINED ${${CMAKE_MATCH_1}} AND ${${CMAKE_MATCH_1}} STREQUAL ${CMAKE_MATCH_2})
 #            message (STATUS "condition is true")
             set (IfElse 1)
           else ()

driver/others/dynamic_zarch.c

@@ -1,12 +1,58 @@
-
 #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>
+#if __GLIBC_PREREQ(2, 16)
+#include <sys/auxv.h>
+#define HAVE_GETAUXVAL 1
+
+static unsigned long get_hwcap(void)
+{
+	unsigned long hwcap = getauxval(AT_HWCAP);
+	char *maskenv;
+
+	// honor requests for not using specific CPU features in LD_HWCAP_MASK
+	maskenv = getenv("LD_HWCAP_MASK");
+	if (maskenv)
+		hwcap &= strtoul(maskenv, NULL, 0);
+
+	return hwcap;
+	// note that a missing auxval is interpreted as no capabilities
+	// available, which is safe.
+}
 
+#else // __GLIBC_PREREQ(2, 16)
+#warn "Cannot detect SIMD support in Z13 or newer architectures since glibc is older than 2.16"
+
+static unsigned long get_hwcap(void) {
+	// treat missing support for getauxval() as no capabilities available,
+	// which is safe.
+	return 0;
+}
+#endif // __GLIBC_PREREQ(2, 16)
+#endif // __GLIBC
+
+extern gotoblas_t gotoblas_ZARCH_GENERIC;
+#ifdef HAVE_Z13_SUPPORT
 extern gotoblas_t gotoblas_Z13;
+#endif
+#ifdef HAVE_Z14_SUPPORT
 extern gotoblas_t gotoblas_Z14;
-//extern gotoblas_t gotoblas_Z15;
-//#if (!defined C_GCC) || (GCC_VERSION >= 60000)
-//extern gotoblas_t gotoblas_Z14;
-//#endif
+#endif
 
 #define NUM_CORETYPES 4
 
@@ -16,47 +62,50 @@ static char* corename[] = {
 	"unknown",
 	"Z13",
 	"Z14",
-//	"Z15",
 	"ZARCH_GENERIC",
 };
 
 char* gotoblas_corename(void) {
+#ifdef HAVE_Z13_SUPPORT
 	if (gotoblas == &gotoblas_Z13)	return corename[1];
+#endif
+#ifdef HAVE_Z14_SUPPORT
 	if (gotoblas == &gotoblas_Z14)	return corename[2];
-//	if (gotoblas == &gotoblas_Z15)	return corename[3];
-//#if (!defined C_GCC) || (GCC_VERSION >= 60000)
-//	if (gotoblas == &gotoblas_POWER9)	return corename[3];
-//#endif
-	return corename[0]; // try generic?
+#endif
+	if (gotoblas == &gotoblas_ZARCH_GENERIC) return corename[3];
+
+	return corename[0];
 }
 
-// __builtin_cpu_is is not supported by zarch
+/**
+ * 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
+ * ("coretype") that exploits most of the features and can be compiled with the
+ * available gcc version.
+ * Note that we cannot use vector registers on a z13 or newer unless supported
+ * by the OS kernel (which needs to handle them properly during context switch).
+ */
 static gotoblas_t* get_coretype(void) {
-	FILE* infile;
-	char buffer[512], * p;
-
-	p = (char*)NULL;
-	infile = fopen("/proc/sysinfo", "r");
-	while (fgets(buffer, sizeof(buffer), infile)) {
-		if (!strncmp("Type", buffer, 4)) {
-			p = strchr(buffer, ':') + 2;
-#if 0
-			fprintf(stderr, "%s\n", p);
-#endif
-			break;
-		}
-	}
 
-	fclose(infile);
+	unsigned long hwcap __attribute__((unused)) = get_hwcap();
 
-	if (strstr(p, "2964")) return &gotoblas_Z13;
-	if (strstr(p, "2965")) return &gotoblas_Z13;
-	if (strstr(p, "3906")) return &gotoblas_Z14;
-	if (strstr(p, "3907")) return &gotoblas_Z14;
-	if (strstr(p, "8561")) return &gotoblas_Z14;        // fallback z15 to z14
-	if (strstr(p, "8562")) return &gotoblas_Z14;        // fallback z15 to 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
+
+	// z13: Vector Facility (SIMD for double)
+#ifdef HAVE_Z13_SUPPORT
+	if (hwcap & HWCAP_S390_VX)
+		return &gotoblas_Z13;
+#endif
 
-	return NULL; // should be ZARCH_GENERIC
+	// fallback in case of missing compiler support, systems before z13, or
+	// when the OS does not advertise support for the Vector Facility (e.g.,
+	// missing support in the OS kernel)
+	return &gotoblas_ZARCH_GENERIC;
 }
 
 static gotoblas_t* force_coretype(char* coretype) {
@@ -76,12 +125,13 @@ static gotoblas_t* force_coretype(char* coretype) {
 
 	switch (found)
 	{
+#ifdef HAVE_Z13_SUPPORT
 	case  1: return (&gotoblas_Z13);
+#endif
+#ifdef HAVE_Z14_SUPPORT
 	case  2: return (&gotoblas_Z14);
-//	case  3: return (&gotoblas_Z15);
-//#if (!defined C_GCC) || (GCC_VERSION >= 60000)
-//	case  3: return (&gotoblas_POWER9);
-//#endif
+#endif
+	case  3: return (&gotoblas_ZARCH_GENERIC);
 	default: return NULL;
 	}
 	snprintf(message, 128, "Core not found: %s\n", coretype);
@@ -109,9 +159,9 @@ void gotoblas_dynamic_init(void) {
 
 	if (gotoblas == NULL)
 	{
-		snprintf(coremsg, 128, "Falling back to Z14 core\n");
+		snprintf(coremsg, 128, "Failed to detect system, falling back to generic z support.\n");
 		openblas_warning(1, coremsg);
-		gotoblas = &gotoblas_Z14;
+		gotoblas = &gotoblas_ZARCH_GENERIC;
 	}
 
 	if (gotoblas && gotoblas->init) {

kernel/x86_64/KERNEL.SKYLAKEX

@@ -24,3 +24,6 @@ DGEMM_BETA = dgemm_beta_skylakex.c
 
 CGEMMKERNEL    =  cgemm_kernel_8x2_skylakex.c
 ZGEMMKERNEL    =  zgemm_kernel_4x2_skylakex.c
+
+CSCALKERNEL    = ../arm/zscal.c
+ZSCALKERNEL    = ../arm/zscal.c

kernel/zarch/KERNEL.Z14

@@ -86,23 +86,23 @@ DGEMVTKERNEL = dgemv_t_4.c
 CGEMVTKERNEL = cgemv_t_4.c
 ZGEMVTKERNEL = zgemv_t_4.c
 
-STRMMKERNEL	= strmm8x4V.S
+STRMMKERNEL	= gemm_vec.c
 DTRMMKERNEL	= trmm8x4V.S
 CTRMMKERNEL	= ctrmm4x4V.S
 ZTRMMKERNEL	= ztrmm4x4V.S
 
-SGEMMKERNEL    =  strmm8x4V.S
-SGEMMINCOPY    = ../generic/gemm_ncopy_8.c
-SGEMMITCOPY    = ../generic/gemm_tcopy_8.c
-SGEMMONCOPY    = ../generic/gemm_ncopy_4.c
-SGEMMOTCOPY    = ../generic/gemm_tcopy_4.c
+SGEMMKERNEL    = gemm_vec.c
+ifneq ($(SGEMM_UNROLL_M),$(SGEMM_UNROLL_N))
+SGEMMINCOPY    = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
+SGEMMITCOPY    = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
 SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
 SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+SGEMMONCOPY    = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
+SGEMMOTCOPY    = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
 SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
 SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
-
- 
 DGEMMKERNEL    =  gemm8x4V.S
 DGEMMINCOPY    = ../generic/gemm_ncopy_8.c
 DGEMMITCOPY    = ../generic/gemm_tcopy_8.c
@@ -145,7 +145,3 @@ ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
 ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
 ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
 
-
-
-
-