Merge pull request #13 from xianyi/develop

rebase

cmake/lapacke.cmake

@@ -2499,6 +2499,5 @@ foreach (Utils_FILE ${Utils_SRC})
 endforeach ()
 
 set(lapacke_include_dir "${NETLIB_LAPACK_DIR}/LAPACKE/include")
-configure_file("${lapacke_include_dir}/lapacke_mangling_with_flags.h.in" "${lapacke_include_dir}/lapacke_mangling.h" COPYONLY)
 include_directories(${lapacke_include_dir})
 set_source_files_properties(${LAPACKE_SOURCES} PROPERTIES COMPILE_FLAGS "${LAPACK_CFLAGS}")

cpuid_x86.c

@@ -1418,6 +1418,15 @@ int get_cpuname(void){
       case 9:
       case 8:      
         switch (model) {
+        case 12: // Tiger Lake
+          if(support_avx512())
+            return CPUTYPE_SKYLAKEX;
+          if(support_avx2())
+            return CPUTYPE_HASWELL;
+          if(support_avx())
+            return CPUTYPE_SANDYBRIDGE;
+          else
+          return CPUTYPE_NEHALEM;
 	case 14: // Kaby Lake and refreshes
           if(support_avx2())
             return CPUTYPE_HASWELL;
@@ -2124,6 +2133,16 @@ int get_coretype(void){
         break;    	
       case 9:
       case 8:
+       if (model == 12) { // Tiger Lake
+          if(support_avx512())
+            return CPUTYPE_SKYLAKEX;
+          if(support_avx2())
+            return CPUTYPE_HASWELL;
+          if(support_avx())
+            return CPUTYPE_SANDYBRIDGE;
+          else
+          return CPUTYPE_NEHALEM;
+        }
         if (model == 14) { // Kaby Lake 
 	  if(support_avx())
 #ifndef NO_AVX2

driver/others/blas_server.c

@@ -1024,38 +1024,39 @@ int BLASFUNC(blas_thread_shutdown)(void){
 
   int i;
 
-  if (!blas_server_avail) return 0;
-
   LOCK_COMMAND(&server_lock);
 
-  for (i = 0; i < blas_num_threads - 1; i++) {
+  if (blas_server_avail) {
+
+    for (i = 0; i < blas_num_threads - 1; i++) {
 
 
-    pthread_mutex_lock (&thread_status[i].lock);
+      pthread_mutex_lock (&thread_status[i].lock);
 
-    atomic_store_queue(&thread_status[i].queue, (blas_queue_t *)-1);
+      atomic_store_queue(&thread_status[i].queue, (blas_queue_t *)-1);
-    thread_status[i].status = THREAD_STATUS_WAKEUP;
+      thread_status[i].status = THREAD_STATUS_WAKEUP;
-    pthread_cond_signal (&thread_status[i].wakeup);
+      pthread_cond_signal (&thread_status[i].wakeup);
 
-    pthread_mutex_unlock(&thread_status[i].lock);
+      pthread_mutex_unlock(&thread_status[i].lock);
 
-  }
+    }
 
-  for(i = 0; i < blas_num_threads - 1; i++){
+    for(i = 0; i < blas_num_threads - 1; i++){
-    pthread_join(blas_threads[i], NULL);
+      pthread_join(blas_threads[i], NULL);
-  }
+    }
 
-  for(i = 0; i < blas_num_threads - 1; i++){
+    for(i = 0; i < blas_num_threads - 1; i++){
-    pthread_mutex_destroy(&thread_status[i].lock);
+      pthread_mutex_destroy(&thread_status[i].lock);
-    pthread_cond_destroy (&thread_status[i].wakeup);
+      pthread_cond_destroy (&thread_status[i].wakeup);
-  }
+    }
 
 #ifdef NEED_STACKATTR
-  pthread_attr_destory(&attr);
+    pthread_attr_destroy(&attr);
 #endif
 
-  blas_server_avail = 0;
+    blas_server_avail = 0;
 
+  }
   UNLOCK_COMMAND(&server_lock);
 
   return 0;

driver/others/dynamic.c

@@ -644,6 +644,21 @@ static gotoblas_t *get_coretype(void){
         return NULL;  
       case 9:
       case 8:
+        if (model == 12) { // Tiger Lake
+          if (support_avx512()) 
+            return &gotoblas_SKYLAKEX;
+          if(support_avx2()){
+            openblas_warning(FALLBACK_VERBOSE, HASWELL_FALLBACK);
+            return &gotoblas_HASWELL;
+          }
+          if(support_avx()) {
+            openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
+            return &gotoblas_SANDYBRIDGE;
+          } else {
+          openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
+          return &gotoblas_NEHALEM;
+          }
+        }
 	if (model == 14 ) { // Kaby Lake, Coffee Lake
 	  if(support_avx2())
 	    return &gotoblas_HASWELL;

driver/others/memory.c

@@ -222,11 +222,11 @@ int get_num_procs(void);
 #else
 int get_num_procs(void) {
   static int nums = 0;
+
+#if defined(__GLIBC_PREREQ)
   cpu_set_t cpuset,*cpusetp;
   size_t size;
   int ret;
-
-#if defined(__GLIBC_PREREQ)
 #if !__GLIBC_PREREQ(2, 7)
   int i;
 #if !__GLIBC_PREREQ(2, 6)

f_check

@@ -75,6 +75,7 @@ if ($compiler eq "") {
 
 	} elsif ($data =~ /GNU/ || $data =~ /GCC/ ) {
 
+            $data =~ s/\(+.*?\)+//g;
 	    $data =~ /(\d+)\.(\d+).(\d+)/;
 	    $major = $1;
 	    $minor = $2;

kernel/x86_64/KERNEL.ZEN

@@ -97,3 +97,5 @@ ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
 CGEMM3MKERNEL    =  cgemm3m_kernel_8x4_haswell.c
 ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_haswell.c
 
+SROTKERNEL = srot.c
+DROTKERNEL = drot.c

kernel/x86_64/dasum.c

@@ -6,7 +6,7 @@
 
 #if defined(SKYLAKEX)
 #include "dasum_microk_skylakex-2.c"
-#elif defined(HASWELL)
+#elif defined(HASWELL) || defined(ZEN)
 #include "dasum_microk_haswell-2.c"
 #endif
 

kernel/x86_64/drot.c

@@ -2,7 +2,7 @@
 
 #if defined(SKYLAKEX)
 #include "drot_microk_skylakex-2.c"
-#elif defined(HASWELL)
+#elif defined(HASWELL) || defined(ZEN)
 #include "drot_microk_haswell-2.c"
 #endif
 

kernel/x86_64/sasum.c

@@ -11,7 +11,7 @@
 
 #if defined(SKYLAKEX)
 #include "sasum_microk_skylakex-2.c"
-#elif defined(HASWELL)
+#elif defined(HASWELL) || defined(ZEN)
 #include "sasum_microk_haswell-2.c"
 #endif
 

kernel/x86_64/srot.c

@@ -2,7 +2,7 @@
 
 #if defined(SKYLAKEX)
 #include "srot_microk_skylakex-2.c"
-#elif defined(HASWELL)
+#elif defined(HASWELL) || defined(ZEN)
 #include "srot_microk_haswell-2.c"
 #endif
 

param.h

@@ -2443,8 +2443,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define SGEMM_DEFAULT_UNROLL_M 16
 #define SGEMM_DEFAULT_UNROLL_N 8
+#if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#define DGEMM_DEFAULT_UNROLL_M 16
+#define DGEMM_DEFAULT_UNROLL_N 4
+#else
 #define DGEMM_DEFAULT_UNROLL_M 8
 #define DGEMM_DEFAULT_UNROLL_N 8
+#endif
 #define CGEMM_DEFAULT_UNROLL_M 8
 #define CGEMM_DEFAULT_UNROLL_N 4
 #define ZGEMM_DEFAULT_UNROLL_M 8