Merge pull request #6 from xianyi/develop

rebase

Makefile

@@ -59,6 +59,9 @@ endif
 	@$(CC) --version > /dev/null 2>&1;\
 	if [ $$? -eq 0 ]; then \
 	   cverinfo=`$(CC) --version | sed -n '1p'`; \
+	   if [ -z "$${cverinfo}" ]; then \
+	   cverinfo=`$(CC) --version | sed -n '2p'`; \
+	   fi; \
 	   echo "  C compiler       ... $(C_COMPILER)  (cmd & version : $${cverinfo})";\
 	else  \
 	   echo "  C compiler       ... $(C_COMPILER)  (command line : $(CC))";\
@@ -67,6 +70,9 @@ ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 	@$(FC) --version > /dev/null 2>&1;\
 	if [ $$? -eq 0 ]; then \
 	   fverinfo=`$(FC) --version | sed -n '1p'`; \
+	   if [ -z "$${fverinfo}" ]; then \
+	   fverinfo=`$(FC) --version | sed -n '2p'`; \
+	   fi; \
 	   echo "  Fortran compiler ... $(F_COMPILER)  (cmd & version : $${fverinfo})";\
 	else \
 	   echo "  Fortran compiler ... $(F_COMPILER)  (command line : $(FC))";\

Makefile.power

@@ -10,9 +10,11 @@ USE_OPENMP = 1
 endif
 
 ifeq ($(CORE), POWER10)
+ifneq ($(C_COMPILER), PGI)
 CCOMMON_OPT += -Ofast -mcpu=power10 -mtune=power10 -mvsx -fno-fast-math
 FCOMMON_OPT += -O2 -frecursive -mcpu=power10 -mtune=power10  -fno-fast-math
 endif
+endif
 
 ifeq ($(CORE), POWER9)
 ifneq ($(C_COMPILER), PGI)

Makefile.system

@@ -181,7 +181,7 @@ endif
 
 # On x86_64 build getarch with march=native unless the compiler is PGI. This is required to detect AVX512 support in getarch.
 ifeq ($(HOSTARCH), x86_64)
-ifeq ($(findstring pgcc,$(HOSTCC)),)
+ifeq ($(findstring pgcc,$(HOSTCC))$(findstring nvc,$(HOSTCC)),)
 GETARCH_FLAGS += -march=native
 endif
 endif
@@ -663,6 +663,7 @@ endif
 endif # ARCH zarch
 
 ifeq ($(ARCH), power)
+ifneq ($(C_COMPILER), PGI)
 DYNAMIC_CORE = POWER6
 DYNAMIC_CORE += POWER8
 ifneq ($(C_COMPILER), GCC)
@@ -689,6 +690,10 @@ else
 $(info, OpenBLAS: Your gcc version is too old to build the POWER10 kernels.)
 endif
 endif
+else
+DYNAMIC_CORE = POWER8
+DYNAMIC_CORE += POWER9
+endif
 endif
 
 # If DYNAMIC_CORE is not set, DYNAMIC_ARCH cannot do anything, so force it to empty
@@ -847,9 +852,19 @@ endif
 endif
 
 ifeq ($(C_COMPILER), PGI)
+PGCVERSIONGT20 := $(shell expr `$(CC) --version|sed -n "2p" |sed -e "s/[^0-9.]//g" |cut -d "." -f 1` \> 20)
+PGCVERSIONGTEQ20 := $(shell expr `$(CC) --version|sed -n "2p" |sed -e "s/[^0-9.]//g" |cut -d "." -f 1` \>= 20)
+PGCMINORVERSIONGE11 := $(shell expr `$(CC) --version|sed -n "2p" |sed -e "s/[^0-9.]//g" |cut -c 4-5` == 11)
+PGCVERSIONCHECK := $(PGCVERSIONGT20)$(PGCVERSIONEQ20)$(PGCMINORVERSIONGE11)
+ifeq ($(PGCVERSIONCHECK), $(filter $(PGCVERSIONCHECK), 110 111 011))
+NEWPGI := 1
+endif
 ifdef BINARY64
 ifeq ($(ARCH), x86_64)
-CCOMMON_OPT += -tp p7-64 -D__MMX__ -Mnollvm
+CCOMMON_OPT += -tp p7-64
+ifneq ($(NEWPGI),1)
+CCOMMON_OPT +=  -D__MMX__ -Mnollvm
+endif
 else
 ifeq ($(ARCH), power)
 ifeq ($(CORE), POWER8)
@@ -1029,18 +1044,24 @@ ifeq ($(ARCH), x86_64)
 FCOMMON_OPT += -tp p7-64
 else
 ifeq ($(ARCH), power)
+ifeq ($(CORE), POWER6)
+$(warning NVIDIA HPC compilers do not support POWER6.)
+endif
 ifeq ($(CORE), POWER8)
 FCOMMON_OPT += -tp pwr8
 endif
 ifeq ($(CORE), POWER9)
 FCOMMON_OPT += -tp pwr9
 endif
+ifeq ($(CORE), POWER10)
+$(warning NVIDIA HPC compilers do not support POWER10.)
+endif
 endif
 endif
 else
 FCOMMON_OPT += -tp p7
 endif
-FCOMMON_OPT += -Mrecursive
+FCOMMON_OPT += -Mrecursive -Kieee
 ifeq ($(USE_OPENMP), 1)
 FCOMMON_OPT += -mp
 endif

README.md

@@ -13,7 +13,7 @@ Drone CI: [![Build Status](https://cloud.drone.io/api/badges/xianyi/OpenBLAS/sta
 
 ## Introduction
 
-OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version.
+OpenBLAS is an optimized Basic Linear Algebra Subprograms library based on GotoBLAS2 1.13 BSD version.
 
 Please read the documentation on the OpenBLAS wiki pages: <https://github.com/xianyi/OpenBLAS/wiki>.
 

driver/others/dynamic_power.c

@@ -27,7 +27,9 @@ static char *corename[] = {
 #define NUM_CORETYPES 4
 
 char *gotoblas_corename(void) {
+#ifndef C_PGI
 	if (gotoblas == &gotoblas_POWER6)	return corename[1];
+#endif
 	if (gotoblas == &gotoblas_POWER8)	return corename[2];
 #if (!defined __GNUC__) || ( __GNUC__ >= 6)
 	if (gotoblas == &gotoblas_POWER9)	return corename[3];
@@ -38,10 +40,157 @@ char *gotoblas_corename(void) {
 	return corename[0];
 }
 
+#ifdef C_PGI
+/*
+ * NV HPC compilers do not yet implement __builtin_cpu_is().
+ * Fake a version here for use in the CPU detection code below.
+ *
+ * Strategy here is to first check the CPU to see what it actually is,
+ * and then test the input to see if what the CPU actually is matches
+ * what was requested.
+ */
+
+#include <string.h>
+
+/*
+ *  Define POWER processor version table.
+ *
+ *  NOTE NV HPC SDK compilers only support POWER8 and POWER9 at this time
+ */
+
+#define CPU_UNKNOWN 0
+#define CPU_POWER5 5
+#define CPU_POWER6 6
+#define CPU_POWER8 8
+#define CPU_POWER9 9
+#define CPU_POWER10 10
+
+static  struct {
+    uint32_t    pvr_mask;
+    uint32_t    pvr_value;
+    const char* cpu_name;
+    uint32_t    cpu_type;
+} pvrPOWER [] = {
+
+    {   /* POWER6 in P5+ mode; 2.04-compliant processor */
+        .pvr_mask       = 0xffffffff,
+        .pvr_value      = 0x0f000001,
+        .cpu_name       = "POWER5+",
+        .cpu_type       = CPU_POWER5,
+    },
+
+    {   /* Power6 aka POWER6X*/
+        .pvr_mask       = 0xffff0000,
+        .pvr_value      = 0x003e0000,
+        .cpu_name       = "POWER6 (raw)",
+        .cpu_type       = CPU_POWER6,
+    },
+
+    {   /* Power7 */
+        .pvr_mask       = 0xffff0000,
+        .pvr_value      = 0x003f0000,
+        .cpu_name       = "POWER7 (raw)",
+        .cpu_type       = CPU_POWER6,
+    },
+
+    {   /* Power7+ */
+        .pvr_mask       = 0xffff0000,
+        .pvr_value      = 0x004A0000,
+        .cpu_name       = "POWER7+ (raw)",
+        .cpu_type       = CPU_POWER6,
+    },
+
+    {   /* Power8E */
+        .pvr_mask       = 0xffff0000,
+        .pvr_value      = 0x004b0000,
+        .cpu_name       = "POWER8E (raw)",
+        .cpu_type       = CPU_POWER8,
+    },
+
+    {   /* Power8NVL */
+        .pvr_mask       = 0xffff0000,
+        .pvr_value      = 0x004c0000,
+        .cpu_name       = "POWER8NVL (raw)",
+        .cpu_type       = CPU_POWER8,
+    },
+
+    {   /* Power8 */
+        .pvr_mask       = 0xffff0000,
+        .pvr_value      = 0x004d0000,
+        .cpu_name       = "POWER8 (raw)",
+        .cpu_type       = CPU_POWER8,
+    },
+
+    {   /* Power9 DD2.0 */
+        .pvr_mask       = 0xffffefff,
+        .pvr_value      = 0x004e0200,
+        .cpu_name       = "POWER9 (raw)",
+        .cpu_type       = CPU_POWER9,
+    },
+
+    {   /* Power9 DD 2.1 */
+        .pvr_mask       = 0xffffefff,
+        .pvr_value      = 0x004e0201,
+        .cpu_name       = "POWER9 (raw)",
+        .cpu_type       = CPU_POWER9,
+    },
+
+    {   /* Power9 DD2.2 or later */
+        .pvr_mask       = 0xffff0000,
+        .pvr_value      = 0x004e0000,
+        .cpu_name       = "POWER9 (raw)",
+        .cpu_type       = CPU_POWER9,
+    },
+
+    {   /* Power10 */
+        .pvr_mask       = 0xffff0000,
+        .pvr_value      = 0x00800000,
+        .cpu_name       = "POWER10 (raw)",
+        .cpu_type       = CPU_POWER10,
+    },
+
+    {   /* End of table, pvr_mask and pvr_value must be zero */
+        .pvr_mask       = 0x0,
+        .pvr_value      = 0x0,
+        .cpu_name       = "Unknown",
+        .cpu_type       = CPU_UNKNOWN,
+    },
+};
+
+static int __builtin_cpu_is(const char *cpu) {
+	int i;
+	uint32_t pvr;
+	uint32_t cpu_type;
+
+	asm("mfpvr    %0" : "=r"(pvr));
+
+	for (i = 0 ; i < sizeof pvrPOWER / sizeof *pvrPOWER ; ++i) {
+		if ((pvr & pvrPOWER[i].pvr_mask) == pvrPOWER[i].pvr_value) {
+			break;
+		}
+	}
+
+#if defined(DEBUG)
+	printf("%s: returning CPU=%s, cpu_type=%p\n", __func__,
+		pvrPOWER[i].cpu_name, pvrPOWER[i].cpu_type);
+#endif
+	cpu_type = pvrPOWER[i].cpu_type;
+
+	if (!strcmp(cpu, "power8"))
+		return cpu_type == CPU_POWER8;
+	if (!strcmp(cpu, "power9"))
+		return cpu_type == CPU_POWER9;
+	return 0;
+}
+
+#endif  /* C_PGI */
+
 static gotoblas_t *get_coretype(void) {
 
+#ifndef C_PGI
 	if (__builtin_cpu_is("power6") || __builtin_cpu_is("power6x"))
 		return &gotoblas_POWER6;
+#endif
 	if (__builtin_cpu_is("power8"))
 		return &gotoblas_POWER8;
 #if (!defined __GNUC__) || ( __GNUC__ >= 6)
@@ -77,7 +226,9 @@ static gotoblas_t *force_coretype(char * coretype) {
 
 	switch (found)
 	{
+#ifndef C_PGI
 	case  1: return (&gotoblas_POWER6);
+#endif
 	case  2: return (&gotoblas_POWER8);
 #if (!defined __GNUC__) || ( __GNUC__ >= 6)
 	case  3: return (&gotoblas_POWER9);

f_check

@@ -32,7 +32,7 @@ if ($compiler eq "") {
               "xlf95", "xlf90", "xlf",
               "ppuf77", "ppuf95", "ppuf90", "ppuxlf",
 	      "pathf90", "pathf95",
-	      "pgf95", "pgf90", "pgf77",
+	      "pgf95", "pgf90", "pgf77", "pgfortran", "nvfortran",
 	      "flang", "egfortran",
               "ifort");
 
@@ -64,7 +64,6 @@ if ($compiler eq "") {
     if (!$?) {
 
 	$data = `$compiler -O2 -S ftest.f > /dev/null 2>&1 && cat ftest.s && rm -f ftest.s`;
-
 	if ($data =~ /zhoge_/) {
 	    $bu       = "_";
 	}
@@ -87,7 +86,7 @@ if ($compiler eq "") {
 		if ($compiler =~ /flang/) {
 		    $vendor = FLANG;
 		    $openmp = "-fopenmp";
-	    } elsif ($compiler =~ /pgf/) {
+	    } elsif ($compiler =~ /pgf/ || $compiler =~ /nvf/) {
 		    $vendor = PGI;
 		    $openmp = "-mp";
 		} else {
@@ -123,7 +122,7 @@ if ($compiler eq "") {
 	    $openmp = "-mp";
 	}
 
-	if ($data =~ /PGF/) {
+	if ($data =~ /PGF/ || $data =~ /NVF/) {
 	    $vendor = PGI;
 	    $openmp = "-mp";
 	}
@@ -177,7 +176,7 @@ if ($compiler eq "") {
 	    $openmp = "-mp";
 	}
 
-	if ($compiler =~ /pgf/) {
+	if ($compiler =~ /pgf/ || $compiler =~ /nvf/) {
 	    $vendor = PGI;
 	    $bu       = "_";
 	    $openmp = "-mp";
@@ -330,7 +329,7 @@ if ($link ne "") {
 	    $flags =~ s/\@/\,/g;
 	    $linker_L .= "-Wl,". $flags . " " ;
 	}
-	if ($flags =~ /-lgomp/ && $ENV{"CC"} =~ /clang/) {
+	if ($flags =~ /-lgomp/ && $CC =~ /clang/) {
 	    $flags = "-lomp";
 	}
 

kernel/Makefile

@@ -36,7 +36,7 @@ ifeq ($(TARGET_CORE), COOPERLAKE)
  ifeq ($(GCCVERSIONGTEQ10), 1) 
   override CFLAGS += -march=cooperlake
  else 
-  override CFLAGS += -march=skylake-avx512
+  override CFLAGS += -march=skylake-avx512 -mavx512f
  endif 
  ifeq ($(OSNAME), CYGWIN_NT)
   override CFLAGS += -fno-asynchronous-unwind-tables
@@ -47,7 +47,7 @@ ifeq ($(TARGET_CORE), COOPERLAKE)
   endif
  endif
 else ifeq ($(TARGET_CORE), SKYLAKEX)
- override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE) -march=skylake-avx512
+ override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE) -march=skylake-avx512 -mavx512f
  ifeq ($(OSNAME), CYGWIN_NT)
   override CFLAGS += -fno-asynchronous-unwind-tables
  endif

kernel/arm64/KERNEL.NEOVERSEN1

@@ -91,10 +91,10 @@ IDAMAXKERNEL   = iamax_thunderx2t99.c
 ICAMAXKERNEL   = izamax_thunderx2t99.c
 IZAMAXKERNEL   = izamax_thunderx2t99.c
 
-SNRM2KERNEL    = scnrm2_thunderx2t99.c
+SNRM2KERNEL    = nrm2.S
-DNRM2KERNEL    = dznrm2_thunderx2t99.c
+DNRM2KERNEL    = nrm2.S
-CNRM2KERNEL    = scnrm2_thunderx2t99.c
+CNRM2KERNEL    = znrm2.S
-ZNRM2KERNEL    = dznrm2_thunderx2t99.c
+ZNRM2KERNEL    = znrm2.S
 
 DDOTKERNEL     = dot_thunderx2t99.c
 SDOTKERNEL     = dot_thunderx2t99.c

kernel/arm64/KERNEL.THUNDERX2T99

@@ -153,12 +153,12 @@ IDAMAXKERNEL   = iamax_thunderx2t99.c
 ICAMAXKERNEL   = izamax_thunderx2t99.c
 IZAMAXKERNEL   = izamax_thunderx2t99.c
 
-SNRM2KERNEL    = scnrm2_thunderx2t99.c
+SNRM2KERNEL    = nrm2.S
-CNRM2KERNEL    = scnrm2_thunderx2t99.c
+CNRM2KERNEL    = nrm2.S
 #DNRM2KERNEL    = dznrm2_thunderx2t99_fast.c
 #ZNRM2KERNEL    = dznrm2_thunderx2t99_fast.c
-DNRM2KERNEL    = dznrm2_thunderx2t99.c
+DNRM2KERNEL    = znrm2.S
-ZNRM2KERNEL    = dznrm2_thunderx2t99.c
+ZNRM2KERNEL    = znrm2.S
 
 
 DDOTKERNEL     = dot_thunderx2t99.c

kernel/arm64/KERNEL.THUNDERX3T110

@@ -153,13 +153,16 @@ IDAMAXKERNEL   = iamax_thunderx2t99.c
 ICAMAXKERNEL   = izamax_thunderx2t99.c
 IZAMAXKERNEL   = izamax_thunderx2t99.c
 
-SNRM2KERNEL    = scnrm2_thunderx2t99.c
+#SNRM2KERNEL    = scnrm2_thunderx2t99.c
-CNRM2KERNEL    = scnrm2_thunderx2t99.c
+#CNRM2KERNEL    = scnrm2_thunderx2t99.c
-#DNRM2KERNEL    = dznrm2_thunderx2t99_fast.c
+##DNRM2KERNEL    = dznrm2_thunderx2t99_fast.c
-#ZNRM2KERNEL    = dznrm2_thunderx2t99_fast.c
+##ZNRM2KERNEL    = dznrm2_thunderx2t99_fast.c
-DNRM2KERNEL    = dznrm2_thunderx2t99.c
+#DNRM2KERNEL    = dznrm2_thunderx2t99.c
-ZNRM2KERNEL    = dznrm2_thunderx2t99.c
+#ZNRM2KERNEL    = dznrm2_thunderx2t99.c
-
+SNRM2KERNEL = nrm2.S
+DNRM2KERNEL = nrm2.S
+CNRM2KERNEL = znrm2.S
+ZNRM2KERNEL = znrm2.S
 
 DDOTKERNEL     = dot_thunderx2t99.c
 SDOTKERNEL     = dot_thunderx2t99.c