Merge pull request #2780 from Guobing-Chen/CPL_build_support

Enable COOPERLAKE build target

Author: martin-frbg

Makefile.system

@@ -88,6 +88,9 @@ endif
 ifeq ($(TARGET), SKYLAKEX)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
+ifeq ($(TARGET), COOPERLAKE)
+GETARCH_FLAGS := -DFORCE_NEHALEM
+endif
 ifeq ($(TARGET), SANDYBRIDGE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
@@ -130,6 +133,9 @@ endif
 ifeq ($(TARGET_CORE), SKYLAKEX)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
+ifeq ($(TARGET_CORE), COOPERLAKE)
+GETARCH_FLAGS := -DFORCE_NEHALEM
+endif
 ifeq ($(TARGET_CORE), SANDYBRIDGE)
 GETARCH_FLAGS := -DFORCE_NEHALEM
 endif
@@ -553,7 +559,7 @@ DYNAMIC_CORE += HASWELL ZEN
 endif
 ifneq ($(NO_AVX512), 1)
 ifneq ($(NO_AVX2), 1)
-DYNAMIC_CORE += SKYLAKEX
+DYNAMIC_CORE += SKYLAKEX COOPERLAKE
 endif
 endif
 endif

Makefile.x86_64

@@ -27,6 +27,32 @@ endif
 endif
 endif
 
+ifeq ($(CORE), COOPERLAKE)
+ifndef DYNAMIC_ARCH
+ifndef NO_AVX512
+ifeq ($(C_COMPILER), GCC)
+# cooperlake support was added in 10.1
+GCCVERSIONGTEQ10 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 10)
+GCCMINORVERSIONGTEQ1 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 1)
+ifeq ($(GCCVERSIONGTEQ10)$(GCCMINORVERSIONGTEQ1), 11)
+CCOMMON_OPT += -march=cooperlake
+FCOMMON_OPT += -march=cooperlake
+endif
+endif
+ifeq ($(OSNAME), CYGWIN_NT)
+CCOMMON_OPT += -fno-asynchronous-unwind-tables
+FCOMMON_OPT += -fno-asynchronous-unwind-tables
+endif
+ifeq ($(OSNAME), WINNT)
+ifeq ($(C_COMPILER), GCC)
+CCOMMON_OPT += -fno-asynchronous-unwind-tables
+FCOMMON_OPT += -fno-asynchronous-unwind-tables
+endif
+endif
+endif
+endif
+endif
+
 ifeq ($(CORE), HASWELL)
 ifndef DYNAMIC_ARCH
 ifndef NO_AVX2

TargetList.txt

@@ -22,6 +22,7 @@ SANDYBRIDGE
 HASWELL
 SKYLAKEX
 ATOM
+COOPERLAKE
 
 b)AMD CPU:
 ATHLON

cmake/arch.cmake

@@ -76,9 +76,9 @@ if (DYNAMIC_ARCH)
       set(DYNAMIC_CORE ${DYNAMIC_CORE} HASWELL ZEN)
     endif ()
     if (NOT NO_AVX512)
-      set(DYNAMIC_CORE ${DYNAMIC_CORE} SKYLAKEX)
+      set(DYNAMIC_CORE ${DYNAMIC_CORE} SKYLAKEX COOPERLAKE)
       string(REGEX REPLACE "-march=native" "" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
-   endif ()
+    endif ()
     if (DYNAMIC_LIST)
 	set(DYNAMIC_CORE PRESCOTT ${DYNAMIC_LIST})
     endif ()

cmake/cc.cmake

@@ -103,3 +103,16 @@ if (${CORE} STREQUAL "SKYLAKEX")
     endif ()
   endif ()
 endif ()
+
+if (${CORE} STREQUAL "COOPERLAKE")
+  if (NOT DYNAMIC_ARCH)
+    if (NOT NO_AVX512)
+      execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
+      if (${GCC_VERSION} VERSION_GREATER 10.1 OR ${GCC_VERSION} VERSION_EQUAL 10.1)
+        set (CCOMMON_OPT = "${CCOMMON_OPT} -march=cooperlake")
+      else ()
+        set (CCOMMON_OPT "${CCOMMON_OPT} -march=skylake-avx512")
+      endif()  
+    endif ()
+  endif ()
+endif ()

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" OR ${TARGET} STREQUAL "SKYLAKEX")
+  if (${TARGET} STREQUAL "HASWELL" OR ${TARGET} STREQUAL "SANDYBRIDGE" OR ${TARGET} STREQUAL "SKYLAKEX" OR ${TARGET} STREQUAL "COOPERLAKE")
     set(TARGET "NEHALEM")
   endif ()
   if (${TARGET} STREQUAL "BULLDOZER" OR ${TARGET} STREQUAL "PILEDRIVER" OR ${TARGET} STREQUAL "ZEN")
@@ -45,6 +45,18 @@ if (DEFINED BINARY AND DEFINED TARGET AND BINARY EQUAL 32)
 endif ()
 
 if (DEFINED TARGET)
+  if (${TARGET} STREQUAL "COOPERLAKE" AND NOT NO_AVX512)
+#    if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
+      execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
+        if (${GCC_VERSION} VERSION_GREATER 10.1 OR ${GCC_VERSION} VERSION_EQUAL 10.1)
+          set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=cooperlake")
+        else()
+          set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
+        endif()
+#    elseif (${CMAKE_C_COMPILER_ID} STREQUAL "CLANG")
+#      set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
+#    endif()    
+  endif()
   if (${TARGET} STREQUAL "SKYLAKEX" AND NOT NO_AVX512)
     set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -march=skylake-avx512")
   endif()

cpuid.h

@@ -118,6 +118,7 @@
 #define CORE_ZEN         27
 #define CORE_SKYLAKEX    28
 #define CORE_DHYANA	 29
+#define CORE_COOPERLAKE  30
 
 #define HAVE_SSE      (1 <<  0)
 #define HAVE_SSE2     (1 <<  1)
@@ -137,11 +138,12 @@
 #define HAVE_MISALIGNSSE (1 << 15)
 #define HAVE_128BITFPU   (1 << 16)
 #define HAVE_FASTMOVU    (1 << 17)
-#define HAVE_AVX      (1 <<  18)
-#define HAVE_FMA4     (1 <<  19)
-#define HAVE_FMA3     (1 <<  20)
-#define HAVE_AVX512VL (1 <<  21)
-#define HAVE_AVX2     (1 <<  22)
+#define HAVE_AVX      (1 << 18)
+#define HAVE_FMA4     (1 << 19)
+#define HAVE_FMA3     (1 << 20)
+#define HAVE_AVX512VL (1 << 21)
+#define HAVE_AVX2     (1 << 22)
+#define HAVE_AVX512BF16  (1 << 23)
 
 #define CACHE_INFO_L1_I     1
 #define CACHE_INFO_L1_D     2
@@ -218,7 +220,8 @@ typedef struct {
 #define CPUTYPE_ZEN 			51
 #define CPUTYPE_SKYLAKEX		52
 #define CPUTYPE_DHYANA			53
+#define CPUTYPE_COOPERLAKE		54
 
-#define CPUTYPE_HYGON_UNKNOWN		54
+#define CPUTYPE_HYGON_UNKNOWN		99
 
 #endif

cpuid_x86.c

@@ -249,6 +249,22 @@ int support_avx512(){
 #endif
 }
 
+int support_avx512_bf16(){
+#if !defined(NO_AVX) && !defined(NO_AVX512)
+  int eax, ebx, ecx, edx;
+  int ret=0;
+
+  if (!support_avx512())
+    return 0;
+  cpuid_count(7, 1, &eax, &ebx, &ecx, &edx);
+  if((eax & 32) == 32){
+      ret=1;  // CPUID.7.1:EAX[bit 5] indicates whether avx512_bf16 supported or not
+  }
+  return ret;
+#else
+  return 0;
+#endif
+}
 
 int get_vendor(void){
   int eax, ebx, ecx, edx;
@@ -335,6 +351,7 @@ int get_cputype(int gettype){
     if (support_avx()) feature |= HAVE_AVX;
     if (support_avx2()) feature |= HAVE_AVX2;
     if (support_avx512()) feature |= HAVE_AVX512VL;
+    if (support_avx512_bf16()) feature |= HAVE_AVX512BF16;
     if ((ecx & (1 << 12)) != 0) feature |= HAVE_FMA3;
 #endif
 
@@ -1337,6 +1354,8 @@ int get_cpuname(void){
 	    return CPUTYPE_NEHALEM;
 	case 5:
 	  // Skylake X
+          if(support_avx512_bf16())
+            return CPUTYPE_COOPERLAKE;
           if(support_avx512())
             return CPUTYPE_SKYLAKEX;
           if(support_avx2())
@@ -1677,7 +1696,8 @@ static char *cpuname[] = {
   "EXCAVATOR",
   "ZEN",
   "SKYLAKEX",
-  "DHYANA"
+  "DHYANA",
+  "COOPERLAKE"
 };
 
 static char *lowercpuname[] = {
@@ -1733,7 +1753,8 @@ static char *lowercpuname[] = {
   "excavator",
   "zen",
   "skylakex",
-  "dhyana"
+  "dhyana",
+  "cooperlake"
 };
 
 static char *corename[] = {
@@ -1766,7 +1787,8 @@ static char *corename[] = {
   "EXCAVATOR",
   "ZEN",
   "SKYLAKEX",
-  "DHYANA"
+  "DHYANA",
+  "COOPERLAKE"
 };
 
 static char *corename_lower[] = {
@@ -1799,7 +1821,8 @@ static char *corename_lower[] = {
   "excavator",
   "zen",
   "skylakex",
-  "dhyana"
+  "dhyana",
+  "cooperlake"
 };
 
 
@@ -2007,7 +2030,9 @@ int get_coretype(void){
 	case 5:
 	 // Skylake X
 #ifndef NO_AVX512
-	    return CORE_SKYLAKEX;
+          if(support_avx512_bf16())
+            return CORE_COOPERLAKE;
+	  return CORE_SKYLAKEX;
 #else
 	  if(support_avx())
 #ifndef NO_AVX2
@@ -2276,6 +2301,7 @@ void get_cpuconfig(void){
     if (features & HAVE_AVX )    printf("#define HAVE_AVX\n");
     if (features & HAVE_AVX2 )    printf("#define HAVE_AVX2\n");
     if (features & HAVE_AVX512VL )    printf("#define HAVE_AVX512VL\n");
+    if (features & HAVE_AVX512BF16 )    printf("#define HAVE_AVX512BF16\n");
     if (features & HAVE_3DNOWEX) printf("#define HAVE_3DNOWEX\n");
     if (features & HAVE_3DNOW)   printf("#define HAVE_3DNOW\n");
     if (features & HAVE_FMA4 )    printf("#define HAVE_FMA4\n");
@@ -2346,6 +2372,7 @@ void get_sse(void){
   if (features & HAVE_AVX )    printf("HAVE_AVX=1\n");
   if (features & HAVE_AVX2 )    printf("HAVE_AVX2=1\n");
   if (features & HAVE_AVX512VL )    printf("HAVE_AVX512VL=1\n");
+  if (features & HAVE_AVX512BF16 )    printf("HAVE_AVX512BF16=1\n");
   if (features & HAVE_3DNOWEX) printf("HAVE_3DNOWEX=1\n");
   if (features & HAVE_3DNOW)   printf("HAVE_3DNOW=1\n");
   if (features & HAVE_FMA4 )    printf("HAVE_FMA4=1\n");

driver/level3/level3.c

@@ -333,7 +333,7 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 #else
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-#ifdef SKYLAKEX
+#if defined(SKYLAKEX) || defined(COOPERLAKE)
 	/* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve best performance */
 	if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N;
 #else

driver/level3/level3_thread.c

@@ -367,7 +367,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
       /* Split local region of B into parts */
       for(jjs = js; jjs < MIN(n_to, js + div_n); jjs += min_jj){
 	min_jj = MIN(n_to, js + div_n) - jjs;
-#ifdef SKYLAKEX
+#if defined(SKYLAKEX) || defined(COOPERLAKE)
 	/* the current AVX512 s/d/c/z GEMM kernel requires n>=6*GEMM_UNROLL_N to achieve the best performance */
 	if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N;
 #else