Merge pull request #1589 from fenrus75/skylakex

Initial support for SkylakeX / AVX512

Author: martin-frbg

Makefile.system

@@ -62,6 +62,9 @@ ifeq ($(BINARY), 32)
 ifeq ($(TARGET), HASWELL)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
+ifeq ($(TARGET), SKYLAKEX)
+GETARCH_FLAGS := -DFORCE_NEHALEM
+endif
 ifeq ($(TARGET), SANDYBRIDGE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
@@ -95,6 +98,9 @@ ifeq ($(BINARY), 32)
 ifeq ($(TARGET_CORE), HASWELL)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
+ifeq ($(TARGET_CORE), SKYLAKEX)
+GETARCH_FLAGS := -DFORCE_NEHALEM
+endif
 ifeq ($(TARGET_CORE), SANDYBRIDGE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
@@ -141,6 +147,10 @@ ifeq ($(NO_AVX2), 1)
 GETARCH_FLAGS	+= -DNO_AVX2
 endif
 
+ifeq ($(NO_AVX512), 1)
+GETARCH_FLAGS	+= -DNO_AVX512
+endif
+
 ifeq ($(DEBUG), 1)
 GETARCH_FLAGS	+= -g
 endif
@@ -469,6 +479,11 @@ endif
 ifneq ($(NO_AVX2), 1)
 DYNAMIC_CORE += HASWELL ZEN
 endif
+ifneq ($(NO_AVX512), 1)
+ifneq ($(NO_AVX2), 1)
+DYNAMIC_CORE += SKYLAKEX
+endif
+endif
 endif
 
 # If DYNAMIC_CORE is not set, DYNAMIC_ARCH cannot do anything, so force it to empty
@@ -924,6 +939,10 @@ ifeq ($(NO_AVX2), 1)
 CCOMMON_OPT	+= -DNO_AVX2
 endif
 
+ifeq ($(NO_AVX512), 1)
+CCOMMON_OPT	+= -DNO_AVX512
+endif
+
 ifdef SMP
 CCOMMON_OPT	+= -DSMP_SERVER
 
@@ -1230,6 +1249,7 @@ export MSA_FLAGS
 export KERNELDIR
 export FUNCTION_PROFILE
 export TARGET_CORE
+export NO_AVX512
 
 export SGEMM_UNROLL_M
 export SGEMM_UNROLL_N

TargetList.txt

@@ -20,6 +20,7 @@ DUNNINGTON
 NEHALEM
 SANDYBRIDGE
 HASWELL
+SKYLAKEX
 ATOM
 
 b)AMD CPU:

cmake/arch.cmake

@@ -56,6 +56,9 @@ if (DYNAMIC_ARCH)
     if (NOT NO_AVX2)
       set(DYNAMIC_CORE ${DYNAMIC_CORE} HASWELL ZEN)
     endif ()
+    if (NOT NO_AVX512)
+      set(DYNAMIC_CORE ${DYNAMIC_CORE} SKYLAKEX)
+    endif ()
   endif ()
 
   if (NOT DYNAMIC_CORE)

cmake/system.cmake

@@ -33,7 +33,7 @@ endif ()
 if (DEFINED BINARY AND DEFINED TARGET AND BINARY EQUAL 32)
   message(STATUS "Compiling a ${BINARY}-bit binary.")
   set(NO_AVX 1)
-  if (${TARGET} STREQUAL "HASWELL" OR ${TARGET} STREQUAL "SANDYBRIDGE")
+  if (${TARGET} STREQUAL "HASWELL" OR ${TARGET} STREQUAL "SANDYBRIDGE" OR ${TARGET} STREQUAL "SKYLAKEX")
     set(TARGET "NEHALEM")
   endif ()
   if (${TARGET} STREQUAL "BULLDOZER" OR ${TARGET} STREQUAL "PILEDRIVER" OR ${TARGET} STREQUAL "ZEN")

cpuid.h

@@ -115,6 +115,7 @@
 #define CORE_STEAMROLLER 25
 #define CORE_EXCAVATOR   26
 #define CORE_ZEN         27
+#define CORE_SKYLAKEX    28
 
 #define HAVE_SSE      (1 <<  0)
 #define HAVE_SSE2     (1 <<  1)
@@ -137,6 +138,7 @@
 #define HAVE_AVX      (1 <<  18)
 #define HAVE_FMA4     (1 <<  19)
 #define HAVE_FMA3     (1 <<  20)
+#define HAVE_AVX512VL (1 <<  21)
 
 #define CACHE_INFO_L1_I     1
 #define CACHE_INFO_L1_D     2
@@ -211,5 +213,6 @@ typedef struct {
 #define CPUTYPE_STEAMROLLER 		49
 #define CPUTYPE_EXCAVATOR 		50
 #define CPUTYPE_ZEN 			51
+#define CPUTYPE_SKYLAKEX		52
 
 #endif

cpuid_x86.c

@@ -50,6 +50,8 @@
 #ifdef NO_AVX
 #define CPUTYPE_HASWELL CPUTYPE_NEHALEM
 #define CORE_HASWELL CORE_NEHALEM
+#define CPUTYPE_SKYLAKEX CPUTYPE_NEHALEM
+#define CORE_SKYLAKEX CORE_NEHALEM
 #define CPUTYPE_SANDYBRIDGE CPUTYPE_NEHALEM
 #define CORE_SANDYBRIDGE CORE_NEHALEM
 #define CPUTYPE_BULLDOZER CPUTYPE_BARCELONA
@@ -1299,6 +1301,19 @@ int get_cpuname(void){
           else
 	    return CPUTYPE_NEHALEM;
 	case 5:
+	  // Skylake X
+#ifndef NO_AVX512
+	  return CPUTYPE_SKYLAKEX;
+#else
+	  if(support_avx())
+#ifndef NO_AVX2
+	  return CPUTYPE_HASWELL;
+#else
+	  return CPUTYPE_SANDYBRIDGE;
+#endif
+	  else
+	  return CPUTYPE_NEHALEM;
+#endif			
         case 14:
 	  // Skylake
           if(support_avx())
@@ -1556,6 +1571,7 @@ static char *cpuname[] = {
   "STEAMROLLER",
   "EXCAVATOR",
   "ZEN",
+  "SKYLAKEX"	
 };
 
 static char *lowercpuname[] = {
@@ -1610,6 +1626,7 @@ static char *lowercpuname[] = {
   "steamroller",
   "excavator",
   "zen",
+  "skylakex"
 };
 
 static char *corename[] = {
@@ -1641,6 +1658,7 @@ static char *corename[] = {
   "STEAMROLLER",
   "EXCAVATOR",
   "ZEN",
+  "SKYLAKEX"	
 };
 
 static char *corename_lower[] = {
@@ -1672,6 +1690,7 @@ static char *corename_lower[] = {
   "steamroller",
   "excavator",
   "zen",
+  "skylakex"	
 };
 
 
@@ -1860,6 +1879,19 @@ int get_coretype(void){
           else
 	    return CORE_NEHALEM;
 	case 5:
+	 // Skylake X
+#ifndef NO_AVX512
+	    return CORE_SKYLAKEX;
+#else
+	  if(support_avx())
+#ifndef NO_AVX2
+	    return CORE_HASWELL;
+#else
+	    return CORE_SANDYBRIDGE;
+#endif
+	  else
+	    return CORE_NEHALEM;
+#endif			
 	case 14:
 	  // Skylake
           if(support_avx())

driver/others/dynamic.c

@@ -74,15 +74,22 @@ extern gotoblas_t  gotoblas_STEAMROLLER;
 extern gotoblas_t  gotoblas_EXCAVATOR;
 #ifdef NO_AVX2
 #define gotoblas_HASWELL gotoblas_SANDYBRIDGE
+#define gotoblas_SKYLAKEX gotoblas_SANDYBRIDGE
 #define gotoblas_ZEN gotoblas_SANDYBRIDGE
 #else
 extern gotoblas_t  gotoblas_HASWELL;
 extern gotoblas_t  gotoblas_ZEN;
+#ifndef NO_AVX512
+extern gotoblas_t  gotoblas_SKYLAKEX;
+#else
+#define gotoblas_SKYLAKEX gotoblas_HASWELL
+#endif
 #endif
 #else
 //Use NEHALEM kernels for sandy bridge
 #define gotoblas_SANDYBRIDGE gotoblas_NEHALEM
 #define gotoblas_HASWELL gotoblas_NEHALEM
+#define gotoblas_SKYLAKEX gotoblas_NEHALEM
 #define gotoblas_BULLDOZER gotoblas_BARCELONA
 #define gotoblas_PILEDRIVER gotoblas_BARCELONA
 #define gotoblas_STEAMROLLER gotoblas_BARCELONA
@@ -284,8 +291,21 @@ static gotoblas_t *get_coretype(void){
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
+	if (model == 5) {	
+	// Intel Skylake X
+#ifndef NO_AVX512
+	  return &gotoblas_SKYLAKEX;
+#else		
+	  if(support_avx())
+	    return &gotoblas_HASWELL;
+	  else {
+	    openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
+	    return &gotoblas_NEHALEM;
+	  }
+#endif		
+	}
 	//Intel Skylake
-	if (model == 14 || model == 5) {
+	if (model == 14) {
 	  if(support_avx())
 	    return &gotoblas_HASWELL;
 	  else{
@@ -445,7 +465,8 @@ static char *corename[] = {
     "Haswell",
     "Steamroller",
     "Excavator",
-    "Zen"
+    "Zen",
+    "SkylakeX"	
 };
 
 char *gotoblas_corename(void) {
@@ -473,7 +494,7 @@ char *gotoblas_corename(void) {
   if (gotoblas == &gotoblas_STEAMROLLER)  return corename[21];
   if (gotoblas == &gotoblas_EXCAVATOR)    return corename[22];
   if (gotoblas == &gotoblas_ZEN)          return corename[23];
-
+  if (gotoblas == &gotoblas_SKYLAKEX)     return corename[24];
   return corename[0];
 }
 
@@ -485,7 +506,7 @@ static gotoblas_t *force_coretype(char *coretype){
 	char message[128];
 	//char mname[20];
 
-	for ( i=1 ; i <= 23; i++)
+	for ( i=1 ; i <= 24; i++)
 	{
 		if (!strncasecmp(coretype,corename[i],20))
 		{
@@ -503,6 +524,7 @@ static gotoblas_t *force_coretype(char *coretype){
 
 	switch (found)
 	{
+		case 24: return (&gotoblas_SKYLAKEX);	
 		case 23: return (&gotoblas_ZEN);
 		case 22: return (&gotoblas_EXCAVATOR);
 		case 21: return (&gotoblas_STEAMROLLER);

driver/others/parameter.c

@@ -167,7 +167,7 @@ int get_L2_size(void){
 #if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER) || \
     defined(CORE_PRESCOTT) || defined(CORE_CORE2) || defined(PENRYN) || defined(DUNNINGTON) || \
     defined(CORE_NEHALEM) || defined(CORE_SANDYBRIDGE) || defined(ATOM) || defined(GENERIC) || \
-    defined(PILEDRIVER) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined(ZEN)
+    defined(PILEDRIVER) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined(ZEN) || defined(SKYLAKEX)
 
   cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
 
@@ -251,7 +251,7 @@ int get_L2_size(void){
 void blas_set_parameter(void){
 
   int factor;
-#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(SANDYBRIDGE) || defined(NEHALEM) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined(ZEN)
+#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(SANDYBRIDGE) || defined(NEHALEM) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR) || defined(ZEN) || defined(SKYLAKEX)
   int size = 16;
 #else
   int size = get_L2_size();

getarch.c

@@ -326,6 +326,21 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "HASWELL"
 #endif
 
+#ifdef FORCE_SKYLAKEX
+#define FORCE
+#define FORCE_INTEL
+#define ARCHITECTURE    "X86"
+#define SUBARCHITECTURE "SKYLAKEX"
+#define ARCHCONFIG   "-DSKYLAKEX " \
+		     "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
+		     "-DL2_SIZE=262144 -DL2_LINESIZE=64 " \
+		     "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+		     "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 -DHAVE_SSE4_1 -DHAVE_SSE4_2 -DHAVE_AVX " \
+                     "-DFMA3 -DHAVE_AVX512VL -march=skylake-avx512"
+#define LIBNAME   "skylakex"
+#define CORENAME  "SKYLAKEX"
+#endif
+
 #ifdef FORCE_ATOM
 #define FORCE
 #define FORCE_INTEL

kernel/CMakeLists.txt

@@ -121,7 +121,7 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
     # Makefile.L3
     set(USE_TRMM false)
 
-    if (ARM OR ARM64 OR "${TARGET_CORE}" STREQUAL "LONGSOON3B" OR "${TARGET_CORE}" STREQUAL "GENERIC" OR "${CORE}" STREQUAL "generic" OR "${TARGET_CORE}" STREQUAL "HASWELL" OR "${CORE}" STREQUAL "haswell" OR "${CORE}" STREQUAL "zen")
+    if (ARM OR ARM64 OR "${TARGET_CORE}" STREQUAL "LONGSOON3B" OR "${TARGET_CORE}" STREQUAL "GENERIC" OR "${CORE}" STREQUAL "generic" OR "${TARGET_CORE}" STREQUAL "HASWELL" OR "${CORE}" STREQUAL "haswell" OR "${CORE}" STREQUAL "zen" OR "${TARGET_CORE}" STREQUAL "SKYLAKEX" OR "${CORE}" STREQUAL "skylakex")
       set(USE_TRMM true)
     endif ()