Merge pull request #2 from xianyi/develop

merge develop

Makefile.install

@@ -48,6 +48,7 @@ ifndef NO_CBLAS
 	@sed 's/common/openblas_config/g' cblas.h > "$(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/cblas.h"
 endif
 
+ifneq ($(OSNAME), AIX)
 ifndef NO_LAPACKE
 	@echo Copying LAPACKE header files to $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)
 	@-install -pm644 $(NETLIB_LAPACK_DIR)/LAPACKE/include/lapacke.h "$(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke.h"
@@ -72,6 +73,7 @@ ifeq ($(OSNAME), $(filter $(OSNAME),Linux SunOS Android Haiku))
 	ln -fs $(LIBSONAME) $(LIBPREFIX).so ; \
 	ln -fs $(LIBSONAME) $(LIBPREFIX).so.$(MAJOR_VERSION)
 endif
+
 ifeq ($(OSNAME), $(filter $(OSNAME),FreeBSD OpenBSD NetBSD DragonFly))
 	@cp $(LIBSONAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
 	@cd "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" ; \
@@ -93,6 +95,33 @@ ifeq ($(OSNAME), CYGWIN_NT)
 endif
 endif
 
+else
+#install on AIX has different options syntax
+ifndef NO_LAPACKE
+	@echo Copying LAPACKE header files to $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)
+	@-installbsd -c -m 644 $(NETLIB_LAPACK_DIR)/LAPACKE/include/lapacke.h "$(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke.h"
+	@-installbsd -c -m 644 $(NETLIB_LAPACK_DIR)/LAPACKE/include/lapacke_config.h "$(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke_config.h"
+	@-installbsd -c -m 644 $(NETLIB_LAPACK_DIR)/LAPACKE/include/lapacke_mangling_with_flags.h.in "$(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke_mangling.h"
+	@-installbsd -c -m 644 $(NETLIB_LAPACK_DIR)/LAPACKE/include/lapacke_utils.h "$(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke_utils.h"
+endif
+
+#for install static library
+ifndef NO_STATIC
+	@echo Copying the static library to $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)
+	@installbsd -c -m 644 $(LIBNAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
+	@cd "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" ; \
+	ln -fs $(LIBNAME) $(LIBPREFIX).$(LIBSUFFIX)
+endif
+#for install shared library
+ifndef NO_SHARED
+	@echo Copying the shared library to $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)
+	@installbsd -c -m 755 $(LIBSONAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
+	@cd "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" ; \
+	ln -fs $(LIBSONAME) $(LIBPREFIX).so ; \
+	ln -fs $(LIBSONAME) $(LIBPREFIX).so.$(MAJOR_VERSION)
+endif
+
+endif
 
 #Generating openblas.pc
 	@echo Generating openblas.pc in "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)"

Makefile.rule

@@ -192,8 +192,8 @@ NO_AFFINITY = 1
 # Flags for POWER8 are defined in Makefile.power. Don't modify COMMON_OPT
 # COMMON_OPT = -O2
 
-# gfortran option for LAPACK
+# gfortran option for LAPACK to improve thread-safety
-# enable this flag only on 64bit Linux and if you need a thread safe lapack library
+# It is enabled by default in Makefile.system for gfortran
 # Flags for POWER8 are defined in Makefile.power. Don't modify FCOMMON_OPT
 # FCOMMON_OPT = -frecursive
 

Makefile.system

@@ -725,6 +725,8 @@ endif
 ifeq ($(F_COMPILER), GFORTRAN)
 CCOMMON_OPT += -DF_INTERFACE_GFORT
 FCOMMON_OPT += -Wall
+# make single-threaded LAPACK calls thread-safe #1847
+FCOMMON_OPT += -frecursive
 #Don't include -lgfortran, when NO_LAPACK=1 or lsbcc
 ifneq ($(NO_LAPACK), 1)
 EXTRALIB += -lgfortran
@@ -1211,7 +1213,11 @@ endif
 
 LIBDLLNAME   = $(LIBPREFIX).dll
 IMPLIBNAME   = lib$(LIBNAMEBASE).dll.a
+ifneq ($(OSNAME), AIX)
 LIBSONAME    = $(LIBNAME:.$(LIBSUFFIX)=.so)
+else
+LIBSONAME    = $(LIBNAME:.$(LIBSUFFIX)=.a)
+endif
 LIBDYNNAME   = $(LIBNAME:.$(LIBSUFFIX)=.dylib)
 LIBDEFNAME   = $(LIBNAME:.$(LIBSUFFIX)=.def)
 LIBEXPNAME   = $(LIBNAME:.$(LIBSUFFIX)=.exp)

cmake/fc.cmake

@@ -44,7 +44,7 @@ endif ()
 
 if (${F_COMPILER} STREQUAL "GFORTRAN")
   set(CCOMMON_OPT "${CCOMMON_OPT} -DF_INTERFACE_GFORT")
-  set(FCOMMON_OPT "${FCOMMON_OPT} -Wall")
+  set(FCOMMON_OPT "${FCOMMON_OPT} -Wall -frecursive")
   #Don't include -lgfortran, when NO_LAPACK=1 or lsbcc
   if (NOT NO_LAPACK)
     set(EXTRALIB "{EXTRALIB} -lgfortran")

common.h

@@ -183,7 +183,7 @@ extern "C" {
 
 #define ALLOCA_ALIGN 63UL
 
-#define NUM_BUFFERS (MAX_CPU_NUMBER * 2 * MAX_PARALLEL_NUMBER)
+#define NUM_BUFFERS MAX(50,(MAX_CPU_NUMBER * 2 * MAX_PARALLEL_NUMBER))
 
 #ifdef NEEDBUNDERSCORE
 #define BLASFUNC(FUNC) FUNC##_

cpuid_arm64.c

@@ -115,8 +115,8 @@ int detect(void)
 	fclose(infile);
 	if(cpu_part != NULL && cpu_implementer != NULL) {
 		if (strstr(cpu_implementer, "0x41") && 
-		(strstr(cpu_part, "0xd07") || strstr(cpu_part,"0xd08") || strstr(cpu_part,"0xd03") ))
+		(strstr(cpu_part, "0xd07") || strstr(cpu_part,"0xd08")))
-			return CPU_CORTEXA57; //or compatible A53, A72
+			return CPU_CORTEXA57; //or compatible, ex. A72
 		else if (strstr(cpu_part, "0x516") && strstr(cpu_implementer, "0x42"))
 			return CPU_VULCAN;
 		else if (strstr(cpu_part, "0x0a1") && strstr(cpu_implementer, "0x43"))

cpuid_x86.c

@@ -2009,6 +2009,8 @@ int get_coretype(void){
 	switch (model) {
 	case 1:
 	  // AMD Ryzen
+	case 8:
+	  // Ryzen 2		
 	  if(support_avx())
 #ifndef NO_AVX2
 	    return CORE_ZEN;

driver/level3/level3_thread.c

@@ -48,6 +48,10 @@
 #define SWITCH_RATIO 2
 #endif
 
+#ifndef GEMM_PREFERED_SIZE
+#define GEMM_PREFERED_SIZE 1
+#endif
+
 //The array of job_t may overflow the stack.
 //Instead, use malloc to alloc job_t.
 #if MAX_CPU_NUMBER > BLAS3_MEM_ALLOC_THRESHOLD
@@ -510,6 +514,16 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
   return 0;
 }
 
+static int round_up(int remainder, int width, int multiple)
+{
+	if (multiple > remainder || width <= multiple)
+		return width;
+	width = (width + multiple - 1) / multiple;
+	width = width * multiple;
+	return width;
+}
+
+
 static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
 		       *range_n, FLOAT *sa, FLOAT *sb,
                        BLASLONG nthreads_m, BLASLONG nthreads_n) {
@@ -601,9 +615,14 @@ static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
   num_parts = 0;
   while (m > 0){
     width = blas_quickdivide(m + nthreads_m - num_parts - 1, nthreads_m - num_parts);
+
+    width = round_up(m, width, GEMM_PREFERED_SIZE);
+
     m -= width;
+
     if (m < 0) width = width + m;
     range_M[num_parts + 1] = range_M[num_parts] + width;
+
     num_parts ++;
   }
   for (i = num_parts; i < MAX_CPU_NUMBER; i++) {
@@ -645,9 +664,12 @@ static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
       if (width < SWITCH_RATIO) {
         width = SWITCH_RATIO;
       }
+      width = round_up(n, width, GEMM_PREFERED_SIZE);
+
       n -= width;
       if (n < 0) width = width + n;
       range_N[num_parts + 1] = range_N[num_parts] + width;
+
       num_parts ++;
     }
     for (j = num_parts; j < MAX_CPU_NUMBER; j++) {

driver/others/blas_server.c

@@ -850,6 +850,11 @@ void goto_set_num_threads(int num_threads) {
 
   long i;
 
+#ifdef SMP_SERVER
+  // Handle lazy re-init of the thread-pool after a POSIX fork
+  if (unlikely(blas_server_avail == 0)) blas_thread_init();
+#endif
+
   if (num_threads < 1) num_threads = blas_num_threads;
 
 #ifndef NO_AFFINITY

driver/others/blas_server_win32.c

@@ -478,7 +478,12 @@ int BLASFUNC(blas_thread_shutdown)(void){
 
 void goto_set_num_threads(int num_threads)
 {
-	 long i;
+	long i;
+
+#if defined(SMP_SERVER) && defined(OS_CYGWIN_NT)
+	// Handle lazy re-init of the thread-pool after a POSIX fork
+	if (unlikely(blas_server_avail == 0)) blas_thread_init();
+#endif
 
 	if (num_threads < 1) num_threads = blas_cpu_number;
 

driver/others/memory.c

@@ -259,6 +259,16 @@ int get_num_procs(void) {
 }
 #endif
 
+#ifdef OS_AIX
+int get_num_procs(void) {
+  static int nums = 0;
+  if (!nums) nums = sysconf(_SC_NPROCESSORS_CONF);
+  return nums;
+}
+#endif
+
+
+
 #ifdef OS_WINDOWS
 
 int get_num_procs(void) {
@@ -1738,6 +1748,22 @@ int get_num_procs(void) {
   return nums;
 }
 #endif
+	
+#ifdef OS_HAIKU
+int get_num_procs(void) {
+  static int nums = 0;
+  if (!nums) nums = sysconf(_SC_NPROCESSORS_CONF);
+  return nums;
+}
+#endif
+
+#ifdef OS_AIX
+int get_num_procs(void) {
+  static int nums = 0;
+  if (!nums) nums = sysconf(_SC_NPROCESSORS_CONF);
+  return nums;
+}
+#endif
 
 #ifdef OS_WINDOWS
 

interface/zswap.c

@@ -42,6 +42,14 @@
 #include "functable.h"
 #endif
 
+#if defined(THUNDERX2T99) || defined(VULCAN) || defined(ARMV8)
+// Multithreaded swap gives performance benefits in ThunderX2T99
+#else
+// Disable multi-threading as it does not show any performance
+// benefits. Keep the multi-threading code for the record.
+#undef SMP
+#endif
+
 #ifndef CBLAS
 
 void NAME(blasint *N, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY){
@@ -81,7 +89,7 @@ FLOAT *y = (FLOAT*)vy;
 #ifdef SMP
   //disable multi-thread when incx==0 or incy==0
   //In that case, the threads would be dependent.
-  if (incx == 0 || incy == 0)
+  if (incx == 0 || incy == 0 || n < 1048576 * GEMM_MULTITHREAD_THRESHOLD / sizeof(FLOAT))
 	  nthreads = 1;
   else
 	  nthreads = num_cpu_avail(1);

kernel/x86_64/dgemm_beta_skylakex.c

@@ -55,6 +55,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
 	return 0;
   }
 
+  if (m == 0 || n == 0)
+	return 0;
 
   c_offset = c;
 
@@ -69,7 +71,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
 
       i = m;
 
-      while (i > 32) {
+      while (i >= 32) {
 	  _mm512_storeu_pd(c_offset1, z_zero);
 	  _mm512_storeu_pd(c_offset1 + 8, z_zero);
 	  _mm512_storeu_pd(c_offset1 + 16, z_zero);
@@ -77,7 +79,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
 	  c_offset1 += 32;
 	  i -= 32;
       }
-      while (i > 8) {
+      while (i >= 8) {
 	  _mm512_storeu_pd(c_offset1, z_zero);
 	  c_offset1 += 8;
 	  i -= 8;

kernel/x86_64/sgemm_beta_skylakex.c

@@ -55,6 +55,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
 	return 0;
   }
 
+  if (n == 0 || m == 0)
+	return;
 
   c_offset = c;
 
@@ -71,13 +73,13 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
 
       i = m;
 
-      while (i > 32) {
+      while (i >= 32) {
 	  _mm512_storeu_ps(c_offset1, z_zero);
 	  _mm512_storeu_ps(c_offset1 + 16, z_zero);
 	  c_offset1 += 32;
 	  i -= 32;
       }
-      while (i > 8) {
+      while (i >= 8) {
 	  _mm256_storeu_ps(c_offset1, y_zero);
 	  c_offset1 += 8;
 	  i -= 8;

lapack-netlib/LAPACKE/include/lapacke_config.h

@@ -34,6 +34,13 @@
 #ifndef _LAPACKE_CONFIG_H_
 #define _LAPACKE_CONFIG_H_
 
+// For Android prior to API 21 (no <complex> include)
+#if defined(__ANDROID__)
+#if __ANDROID_API__ < 21
+#define LAPACK_COMPLEX_STRUCTURE
+#endif
+#endif
+
 #ifdef __cplusplus
 #if defined(LAPACK_COMPLEX_CPP)
 #include <complex>

param.h

@@ -1627,6 +1627,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SYMV_P  8
 
 #define SWITCH_RATIO	32
+#define GEMM_PREFERED_SIZE	32
 
 #ifdef ARCH_X86