Merge pull request #1702 from xianyi/develop

Merge develop for 0.3.2

Author: martin-frbg

CMakeLists.txt

@@ -6,7 +6,7 @@ cmake_minimum_required(VERSION 2.8.5)
 project(OpenBLAS C ASM)
 set(OpenBLAS_MAJOR_VERSION 0)
 set(OpenBLAS_MINOR_VERSION 3)
-set(OpenBLAS_PATCH_VERSION 1)
+set(OpenBLAS_PATCH_VERSION 2)
 set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}")
 
 # Adhere to GNU filesystem layout conventions

Makefile

@@ -21,6 +21,17 @@ ifeq ($(BUILD_RELAPACK), 1)
 RELA = re_lapack
 endif
 
+ifeq ($(NO_FORTRAN), 1)
+define NOFORTRAN
+1
+endef
+define NO_LAPACK
+1
+endef
+export NOFORTRAN
+export NO_LAPACK
+endif
+
 LAPACK_NOOPT := $(filter-out -O0 -O1 -O2 -O3 -Ofast,$(LAPACK_FFLAGS))
 
 SUBDIRS_ALL = $(SUBDIRS) test ctest utest exports benchmark ../laswp ../bench
@@ -47,7 +58,7 @@ endif
 endif
 
 	@echo "  C compiler       ... $(C_COMPILER)  (command line : $(CC))"
-ifndef NOFORTRAN
+ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 	@echo "  Fortran compiler ... $(F_COMPILER)  (command line : $(FC))"
 endif
 ifneq ($(OSNAME), AIX)
@@ -108,7 +119,7 @@ endif
 endif
 
 tests :
-ifndef NOFORTRAN
+ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 	touch $(LIBNAME)
 ifndef NO_FBLAS
 	$(MAKE) -C test all
@@ -210,7 +221,7 @@ netlib :
 
 else
 netlib : lapack_prebuild
-ifndef NOFORTRAN
+ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 	@$(MAKE) -C $(NETLIB_LAPACK_DIR) lapacklib
 	@$(MAKE) -C $(NETLIB_LAPACK_DIR) tmglib
 endif
@@ -231,7 +242,7 @@ prof_lapack : lapack_prebuild
 	@$(MAKE) -C $(NETLIB_LAPACK_DIR) lapack_prof
 
 lapack_prebuild :
-ifndef NOFORTRAN
+ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 	-@echo "FORTRAN     = $(FC)" > $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "OPTS        = $(LAPACK_FFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "POPTS       = $(LAPACK_FPFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
@@ -274,21 +285,21 @@ endif
 endif
 
 large.tgz :
-ifndef NOFORTRAN
+ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 	if [ ! -a $< ]; then
 	-wget http://www.netlib.org/lapack/timing/large.tgz;
 	fi
 endif
 
 timing.tgz :
-ifndef NOFORTRAN
+ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 	if [ ! -a $< ]; then
 	-wget http://www.netlib.org/lapack/timing/timing.tgz;
 	fi
 endif
 
 lapack-timing : large.tgz timing.tgz
-ifndef NOFORTRAN
+ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 	(cd $(NETLIB_LAPACK_DIR); $(TAR) zxf ../timing.tgz TIMING)
 	(cd $(NETLIB_LAPACK_DIR)/TIMING; $(TAR) zxf ../../large.tgz )
 	$(MAKE) -C $(NETLIB_LAPACK_DIR)/TIMING

Makefile.rule

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.3.1
+VERSION = 0.3.2
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library

cmake/prebuild.cmake

@@ -85,7 +85,7 @@ if (NOT NOFORTRAN)
 endif ()
 
 # Cannot run getarch on target if we are cross-compiling
-if (DEFINED CORE AND CMAKE_CROSSCOMPILING)
+if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSSTORE"))
   # Write to config as getarch would
 
   # TODO: Set up defines that getarch sets up based on every other target

cmake/system_check.cmake

@@ -68,7 +68,7 @@ endif()
 
 if (X86_64 OR X86)
   file(WRITE ${PROJECT_BINARY_DIR}/avx512.tmp "int main(void){ __asm__ volatile(\"vbroadcastss -4 * 4(%rsi), %zmm2\"); }")
-execute_process(COMMAND ${CMAKE_C_COMPILER} -v -o ${PROJECT_BINARY_DIR}/avx512.o -x c ${PROJECT_BINARY_DIR}/avx512.tmp RESULT_VARIABLE NO_AVX512)
+execute_process(COMMAND ${CMAKE_C_COMPILER} -march=skylake-avx512 -v -o ${PROJECT_BINARY_DIR}/avx512.o -x c ${PROJECT_BINARY_DIR}/avx512.tmp OUTPUT_QUIET ERROR_QUIET RESULT_VARIABLE NO_AVX512)
 if (NO_AVX512 EQUAL 1)
 set (CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX512")
 endif()

cpuid_power.c

@@ -142,6 +142,52 @@ int detect(void){
 
   return  CPUTYPE_PPC970;
 #endif
+
+#if defined(__FreeBSD__) || defined(__OpenBSD__) || defined(__DragonFly__)
+int id;
+id = __asm __volatile("mfpvr %0" : "=r"(id));
+switch ( id >> 16 ) {
+  case 0x4e: // POWER9
+    return  return CPUTYPE_POWER8;
+    break;
+  case 0x4d:
+  case 0x4b: // POWER8/8E 
+    return CPUTYPE_POWER8;
+    break;
+  case 0x4a:
+  case 0x3f:  // POWER7/7E
+    return CPUTYPE_POWER6; 
+    break;
+  case 0x3e:
+    return CPUTYPE_POWER6;
+    break;
+  case 0x3a:
+    return CPUTYPE_POWER5;
+    break;
+  case 0x35:
+  case 0x38: // POWER4 /4+ 
+    return CPUTYPE_POWER4;
+    break;
+  case 0x40:
+  case 0x41: // POWER3 /3+ 
+    return CPUTYPE_POWER3;
+    break;
+  case 0x39:
+  case 0x3c:
+  case 0x44:
+  case 0x45:
+    return CPUTYPE_PPC970;
+    break;
+  case 0x70: 
+    return CPUTYPE_CELL;
+    break;
+  case 0x8003: 
+    return CPUTYPE_PPCG4;
+    break;
+  default:  
+    return  CPUTYPE_UNKNOWN;
+  }
+#endif
 }
 
 void get_architecture(void){

cpuid_x86.c

@@ -1452,6 +1452,8 @@ int get_cpuname(void){
 	switch (model) {
 	case 1:
 	  // AMD Ryzen
+	case 8:
+	  // AMD Ryzen2
 	  if(support_avx())
 #ifndef NO_AVX2
 	    return CPUTYPE_ZEN;

driver/level3/level3_thread.c

@@ -344,6 +344,12 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
     div_n = (n_to - n_from + DIVIDE_RATE - 1) / DIVIDE_RATE;
     for (js = n_from, bufferside = 0; js < n_to; js += div_n, bufferside ++) {
 
+      /* Make sure if no one is using workspace */
+      START_RPCC();
+      for (i = 0; i < args -> nthreads; i++)
+	while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {YIELDING;MB;};
+      STOP_RPCC(waiting1);
+
 #if defined(FUSED_GEMM) && !defined(TIMING)
 
       /* Fused operation to copy region of B into workspace and apply kernel */
@@ -381,15 +387,10 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
       }
 #endif
 
-      for (i = mypos_n * nthreads_m; i < (mypos_n + 1) * nthreads_m; i++) {
-        /* Make sure if no one is using workspace */
-        START_RPCC();
-        while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {YIELDING;MB;};
-        STOP_RPCC(waiting1);
-        /* Set flag so other threads can access local region of B */
+      /* Set flag so other threads can access local region of B */
+      for (i = mypos_n * nthreads_m; i < (mypos_n + 1) * nthreads_m; i++)
         job[mypos].working[i][CACHE_LINE_SIZE * bufferside] = (BLASLONG)buffer[bufferside];
-        WMB;
-      }
+      WMB;
     }
 
     /* Get regions of B from other threads and apply kernel */
@@ -425,13 +426,13 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 
         /* Clear synchronization flag if this thread is done with other region of B */
 	if (m_to - m_from == min_i) {
-	  job[current].working[mypos][CACHE_LINE_SIZE * bufferside] = 0;
+	  job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
 	  WMB;
 	}
       }
     } while (current != mypos);
 
-    /* Iterate through steps of m
+    /* Iterate through steps of m 
      * Note: First step has already been finished */
     for(is = m_from + min_i; is < m_to; is += min_i){
       min_i = m_to - is;
@@ -461,14 +462,14 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 			   sa, (FLOAT *)job[current].working[mypos][CACHE_LINE_SIZE * bufferside],
 			   c, ldc, is, js);
           STOP_RPCC(kernel);
-
+          
 #ifdef TIMING
           ops += 2 * min_i * MIN(range_n[current + 1]  - js, div_n) * min_l;
 #endif
-
+          
           /* Clear synchronization flag if this thread is done with region of B */
           if (is + min_i >= m_to) {
-            job[current].working[mypos][CACHE_LINE_SIZE * bufferside] = 0;
+            job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
             WMB;
           }
 	}

driver/others/dynamic.c

@@ -510,7 +510,7 @@ static gotoblas_t *get_coretype(void){
 #ifndef NO_AVX2
 	  return &gotoblas_HASWELL;
 #else
-	  return &gotblas_SANDYBRIDGE;
+	  return &gotoblas_SANDYBRIDGE;
 #endif
 	  else
 	  return &gotoblas_NEHALEM;
@@ -607,7 +607,7 @@ static gotoblas_t *get_coretype(void){
 	  }
 	}
       } else if (exfamily == 8) {
-	if (model == 1) {
+	if (model == 1 || model == 8) {
 	  if(support_avx())
 	    return &gotoblas_ZEN;
 	  else{

driver/others/memory.c

@@ -140,7 +140,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #endif
 
 #ifndef BUFFERS_PER_THREAD
-#ifdef USE_OPENMP
+#ifdef USE_OPENMP_UNUSED
 #define BUFFERS_PER_THREAD (MAX_CPU_NUMBER * 2 * MAX_PARALLEL_NUMBER)
 #else
 #define BUFFERS_PER_THREAD NUM_BUFFERS
@@ -363,7 +363,7 @@ int blas_get_cpu_number(void){
 #endif
 
   // blas_goto_num = 0;
-#ifndef USE_OPENMP
+#ifndef USE_OPENMP_UNUSED
   blas_goto_num=openblas_num_threads_env();
   if (blas_goto_num < 0) blas_goto_num = 0;
 
@@ -494,10 +494,10 @@ static const int allocation_block_size = BUFFER_SIZE + sizeof(struct alloc_t);
 #endif
 
 /* Holds pointers to allocated memory */
-#if defined(SMP) && !defined(USE_OPENMP)
+#if defined(SMP) && !defined(USE_OPENMP_UNUSED)
 /* This is the number of threads than can be spawned by the server, which is the
    server plus the number of threads in the thread pool */
-#  define MAX_ALLOCATING_THREADS MAX_CPU_NUMBER * 2 * MAX_PARALLEL_NUMBER +1
+#  define MAX_ALLOCATING_THREADS MAX_CPU_NUMBER * 2 * MAX_PARALLEL_NUMBER * 2
 static int next_memory_table_pos = 0;
 #  if defined(HAS_COMPILER_TLS)
 /* Use compiler generated thread-local-storage */
@@ -532,7 +532,7 @@ static BLASULONG  alloc_lock = 0UL;
 
 /* Returns a pointer to the start of the per-thread memory allocation data */
 static __inline struct alloc_t ** get_memory_table() {
-#if defined(SMP) && !defined(USE_OPENMP)
+#if defined(SMP) && !defined(USE_OPENMP_UNUSED)
 #  if !defined(HAS_COMPILER_TLS)
 #    if defined(OS_WINDOWS)
   int local_memory_table_pos = (int)::TlsGetValue(local_storage_key);
@@ -1057,7 +1057,7 @@ static volatile int memory_initialized = 0;
 /*                2 : Thread                 */
 
 static void blas_memory_init(){
-#if defined(SMP) && !defined(USE_OPENMP)
+#if defined(SMP) && !defined(USE_OPENMP_UNUSED)
   next_memory_table_pos = 0;
 #  if !defined(HAS_COMPILER_TLS)
 #    if defined(OS_WINDOWS)
@@ -1279,7 +1279,7 @@ void blas_shutdown(void){
       struct alloc_t *alloc_info = local_memory_table[thread][pos];
       if (alloc_info) {
         alloc_info->release_func(alloc_info);
-        alloc_info = (void *)0;
+        local_memory_table[thread][pos] = (void *)0;
       }
     }
   }