Merge pull request #3316 from xianyi/develop

Merge develop for bugfix release 0.3.17

Changelog.txt

@@ -1,5 +1,20 @@
 OpenBLAS ChangeLog
 ====================================================================
+Version 0.3.17
+ 15-Jul-2021
+
+common:
+ - reverted the optimization of SGEMV_N/DGEMV_N for small input sizes
+   and consecutive arguments as it led to stack overflows on x86_64
+   with some operating systems (notably OSX and Windows)
+ 
+ x86_64:
+  - reverted the performance patch for SGEMV_T on AVX512 as it caused
+    wrong results in some applications
+    
+ SPARC:
+  - fixed compilation with compilers other than gcc
+====================================================================
 Version 0.3.16
  11-Jul-2021
 

Makefile.rule

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.3.16
+VERSION = 0.3.16.dev
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library

cpuid.h

@@ -54,6 +54,7 @@
 #define VENDOR_TRANSMETA  9
 #define VENDOR_NSC	 10
 #define VENDOR_HYGON	 11
+#define VENDOR_ZHAOXIN   12
 #define VENDOR_UNKNOWN   99
 
 #define BITMASK(a, b, c) ((((a) >> (b)) & (c)))

cpuid_x86.c

@@ -283,7 +283,7 @@ int get_vendor(void){
   if (!strcmp(vendor, "CyrixInstead")) return VENDOR_CYRIX;
   if (!strcmp(vendor, "NexGenDriven")) return VENDOR_NEXGEN;
   if (!strcmp(vendor, "CentaurHauls")) return VENDOR_CENTAUR;
-  if (!strcmp(vendor, "  Shanghai  ")) return VENDOR_CENTAUR;
+  if (!strcmp(vendor, "  Shanghai  ")) return VENDOR_ZHAOXIN;
   if (!strcmp(vendor, "RiseRiseRise")) return VENDOR_RISE;
   if (!strcmp(vendor, " SiS SiS SiS")) return VENDOR_SIS;
   if (!strcmp(vendor, "GenuineTMx86")) return VENDOR_TRANSMETA;
@@ -1067,7 +1067,8 @@ int get_cacheinfo(int type, cache_info_t *cacheinfo){
 
   if ((get_vendor() == VENDOR_AMD) ||
       (get_vendor() == VENDOR_HYGON) ||
-      (get_vendor() == VENDOR_CENTAUR)) {
+      (get_vendor() == VENDOR_CENTAUR) ||
+      (get_vendor() == VENDOR_ZHAOXIN)) {
     cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
 
     LDTB.size        = 4096;
@@ -1190,7 +1191,7 @@ int get_cacheinfo(int type, cache_info_t *cacheinfo){
 
 int get_cpuname(void){
 
-  int family, exfamily, model, vendor, exmodel;
+  int family, exfamily, model, vendor, exmodel, stepping;
 
   if (!have_cpuid()) return CPUTYPE_80386;
 
@@ -1198,6 +1199,7 @@ int get_cpuname(void){
   exfamily = get_cputype(GET_EXFAMILY);
   model    = get_cputype(GET_MODEL);
   exmodel  = get_cputype(GET_EXMODEL);
+  stepping = get_cputype(GET_STEPPING);
 
   vendor = get_vendor();
 
@@ -1628,15 +1630,20 @@ int get_cpuname(void){
     switch (family) {
     case 0x5:
       return CPUTYPE_CENTAURC6;
-      break;
     case 0x6:
-      return CPUTYPE_NANO;
+      if (model == 0xf && stepping < 0xe)
-      break;
+        return CPUTYPE_NANO;
-    case 0x7:
       return CPUTYPE_NEHALEM;
-      break;
+    default:
+      if (family >= 0x7)
+        return CPUTYPE_NEHALEM;
+      else
+        return CPUTYPE_VIAC3;
     }
-    return CPUTYPE_VIAC3;
+  }
+
+  if (vendor == VENDOR_ZHAOXIN){
+    return CPUTYPE_NEHALEM;
   }
 
   if (vendor == VENDOR_RISE){
@@ -1869,7 +1876,7 @@ char *get_lower_cpunamechar(void){
 
 int get_coretype(void){
 
-  int family, exfamily, model, exmodel, vendor;
+  int family, exfamily, model, exmodel, vendor, stepping;
 
   if (!have_cpuid()) return CORE_80486;
 
@@ -1877,6 +1884,7 @@ int get_coretype(void){
   exfamily = get_cputype(GET_EXFAMILY);
   model    = get_cputype(GET_MODEL);
   exmodel  = get_cputype(GET_EXMODEL);
+  stepping = get_cputype(GET_STEPPING);
 
   vendor = get_vendor();
 
@@ -2286,13 +2294,19 @@ int get_coretype(void){
   if (vendor == VENDOR_CENTAUR) {
     switch (family) {
     case 0x6:
-      return CORE_NANO;
+      if (model == 0xf && stepping < 0xe)
-      break;
+        return CORE_NANO;
-    case 0x7:
       return CORE_NEHALEM;
-      break;
+    default:
+      if (family >= 0x7)
+        return CORE_NEHALEM;
+      else
+        return CORE_VIAC3;
     }
-    return CORE_VIAC3;
+  }
+
+  if (vendor == VENDOR_ZHAOXIN) {
+     return CORE_NEHALEM;
   }
 
   return CORE_UNKNOWN;

driver/others/dynamic.c

@@ -292,6 +292,7 @@ extern gotoblas_t  gotoblas_COOPERLAKE;
 #define VENDOR_AMD        2
 #define VENDOR_CENTAUR    3
 #define VENDOR_HYGON	  4
+#define VENDOR_ZHAOXIN    5
 #define VENDOR_UNKNOWN   99
 
 #define BITMASK(a, b, c) ((((a) >> (b)) & (c)))
@@ -404,7 +405,7 @@ static int get_vendor(void){
   if (!strcmp(vendor.vchar, "GenuineIntel")) return VENDOR_INTEL;
   if (!strcmp(vendor.vchar, "AuthenticAMD")) return VENDOR_AMD;
   if (!strcmp(vendor.vchar, "CentaurHauls")) return VENDOR_CENTAUR;
-  if (!strcmp(vendor.vchar, "  Shanghai  ")) return VENDOR_CENTAUR;
+  if (!strcmp(vendor.vchar, "  Shanghai  ")) return VENDOR_ZHAOXIN;
   if (!strcmp(vendor.vchar, "HygonGenuine")) return VENDOR_HYGON;
 
   if ((eax == 0) || ((eax & 0x500) != 0)) return VENDOR_INTEL;
@@ -415,7 +416,7 @@ static int get_vendor(void){
 static gotoblas_t *get_coretype(void){
 
   int eax, ebx, ecx, edx;
-  int family, exfamily, model, vendor, exmodel;
+  int family, exfamily, model, vendor, exmodel, stepping;
 
   cpuid(1, &eax, &ebx, &ecx, &edx);
 
@@ -423,6 +424,7 @@ static gotoblas_t *get_coretype(void){
   exfamily = BITMASK(eax, 20, 0xff);
   model    = BITMASK(eax,  4, 0x0f);
   exmodel  = BITMASK(eax, 16, 0x0f);
+  stepping = BITMASK(eax,  0, 0x0f);
 
   vendor = get_vendor();
 
@@ -824,13 +826,19 @@ static gotoblas_t *get_coretype(void){
   if (vendor == VENDOR_CENTAUR) {
     switch (family) {
     case 0x6:
-      return &gotoblas_NANO;
+      if (model == 0xf && stepping < 0xe)
-      break;
+        return &gotoblas_NANO;
-    case 0x7:
       return &gotoblas_NEHALEM;
+    default:
+      if (family >= 0x7)
+        return &gotoblas_NEHALEM;
     }
   }
 
+  if (vendor == VENDOR_ZHAOXIN) {
+      return &gotoblas_NEHALEM;
+  }
+
   return NULL;
 }
 

interface/gemv.c

@@ -201,12 +201,14 @@ void CNAME(enum CBLAS_ORDER order,
   if (beta != ONE) SCAL_K(leny, 0, 0, beta, y, blasabs(incy), NULL, 0, NULL, 0);
 
   if (alpha == ZERO) return;
-
+	
+#if 0
+/* this optimization causes stack corruption on x86_64 under OSX, Windows and FreeBSD */	
   if (trans == 0 && incx == 1 && incy == 1 && m*n < 2304 *GEMM_MULTITHREAD_THRESHOLD) {
     GEMV_N(m, n, 0, alpha, a, lda, x, incx, y, incy, NULL);
     return;
   }    
-
+#endif
   IDEBUG_START;
 
   FUNCTION_PROFILE_START();

kernel/x86_64/sgemv_t_4.c

@@ -38,7 +38,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "sgemv_t_microk_haswell-4.c"
 #elif defined (SKYLAKEX) || defined (COOPERLAKE)
 #include "sgemv_t_microk_haswell-4.c"
-#include "sgemv_t_microk_skylakex.c"
+/*#include "sgemv_t_microk_skylakex.c"*/
 #endif
 
 #if defined(STEAMROLLER) || defined(EXCAVATOR)

openblas_config_template.h

@@ -99,6 +99,8 @@ typedef int blasint;
 
 /* Inclusion of Linux-specific header is needed for definition of cpu_set_t. */
 #ifdef OPENBLAS_OS_LINUX
-#define _GNU_SOURCE
+#ifndef _GNU_SOURCE
+ #define _GNU_SOURCE
+#endif
 #include <sched.h>
 #endif

param.h

@@ -2502,7 +2502,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define GEMM_DEFAULT_OFFSET_A 0
 #define GEMM_DEFAULT_OFFSET_B 2048
-#define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
+#define GEMM_DEFAULT_ALIGN 0x03fffUL
 
 #define SGEMM_DEFAULT_UNROLL_M 2
 #define SGEMM_DEFAULT_UNROLL_N 8
@@ -2534,7 +2534,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define GEMM_DEFAULT_OFFSET_A 0
 #define GEMM_DEFAULT_OFFSET_B 2048
-#define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
+#define GEMM_DEFAULT_ALIGN 0x03fffUL
 
 #define SGEMM_DEFAULT_UNROLL_M 4
 #define SGEMM_DEFAULT_UNROLL_N 4