Merge pull request #3720 from FlyGoat/mips64

Make it work on general MIPS64 processors

Makefile.system

@@ -677,7 +677,12 @@ endif
 endif
 
 ifeq ($(ARCH), mips64)
-DYNAMIC_CORE = LOONGSON3R3 LOONGSON3R4
+DYNAMIC_CORE = LOONGSON3R3 LOONGSON3R4 MIPS64_GENERIC
+ifdef DYNAMIC_LIST
+override DYNAMIC_CORE = MIPS64_GENERIC $(DYNAMIC_LIST)
+XCCOMMON_OPT = -DDYNAMIC_LIST -DDYN_MIPS64_GENERIC
+XCCOMMON_OPT += $(foreach dcore,$(DYNAMIC_LIST),-DDYN_$(dcore))
+endif
 endif
 
 ifeq ($(ARCH), loongarch64)
@@ -856,6 +861,11 @@ CCOMMON_OPT += -mabi=32
 BINARY_DEFINED = 1
 endif
 
+ifneq (, $(filter $(CORE), MIPS64_GENERIC))
+CCOMMON_OPT += -DNO_MSA
+FCOMMON_OPT += -DNO_MSA
+endif
+
 ifneq (, $(filter $(CORE),LOONGSON3R3 LOONGSON3R4))
 CCOMMON_OPT += -march=loongson3a
 FCOMMON_OPT += -march=loongson3a

TargetList.txt

@@ -65,6 +65,7 @@ MIPS1004K
 MIPS24K
 
 4.MIPS64 CPU:
+MIPS64_GENERIC
 SICORTEX
 LOONGSON3A
 LOONGSON3B

common_mips64.h

@@ -86,7 +86,9 @@ static inline unsigned int rpcc(void){
   //__asm__ __volatile__("dmfc0 %0, $25, 1": "=r"(tmp):: "memory");
   //ret=tmp;
   __asm__ __volatile__(".set push \n"
+#if !defined(__mips_isa_rev) || __mips_isa_rev < 2
                        ".set mips32r2\n"
+#endif
                        "rdhwr %0, $2\n"
                        ".set pop": "=r"(ret):: "memory");
 
@@ -99,7 +101,9 @@ static inline unsigned int rpcc(void){
 static inline int WhereAmI(void){
   int ret=0;
   __asm__ __volatile__(".set push \n"
+#if !defined(__mips_isa_rev) || __mips_isa_rev < 2
                        ".set mips32r2\n"
+#endif
                        "rdhwr %0, $0\n"
                        ".set pop": "=r"(ret):: "memory");
   return ret;
@@ -197,9 +201,15 @@ static inline int blas_quickdivide(blasint x, blasint y){
 
 #if defined(ASSEMBLER) && !defined(NEEDPARAM)
 
+#if defined(__mips_isa_rev) && __mips_isa_rev >= 6
+#define ASSEMBLER_ARCH mips64r6
+#else
+#define ASSEMBLER_ARCH mips64
+#endif
+
 #define PROLOGUE \
 	.text ;\
-	.set	mips64 ;\
+	.set	ASSEMBLER_ARCH ;\
 	.align 5 ;\
 	.globl	REALNAME ;\
 	.ent	REALNAME ;\

cpuid_mips64.c

@@ -70,16 +70,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /* or implied, of The University of Texas at Austin.                 */
 /*********************************************************************/
 
-#define CPU_UNKNOWN      0
+#define CPU_UNKNOWN         0
-#define CPU_SICORTEX     1
+#define CPU_MIPS64_GENERIC  1
-#define CPU_LOONGSON3R3  2
+#define CPU_SICORTEX        2
-#define CPU_LOONGSON3R4  3
+#define CPU_LOONGSON3R3     3
-#define CPU_I6400        4
+#define CPU_LOONGSON3R4     4
-#define CPU_P6600        5
+#define CPU_I6400           5
-#define CPU_I6500        6
+#define CPU_P6600           6
+#define CPU_I6500           7
 
 static char *cpuname[] = {
   "UNKNOWN",
+  "MIPS64_GENERIC"
   "SICORTEX",
   "LOONGSON3R3",
   "LOONGSON3R4",
@@ -113,8 +115,11 @@ int detect(void){
       return CPU_SICORTEX;
     }
   }
+
+  return CPU_MIPS64_GENERIC;
+#else
+  return CPU_UNKNOWN;
 #endif
-    return CPU_UNKNOWN;
 }
 
 char *get_corename(void){
@@ -136,9 +141,11 @@ void get_subarchitecture(void){
     printf("P6600");
   }else if(detect()==CPU_I6500){
     printf("I6500");
-  }else{
+  }else if(detect()==CPU_SICORTEX){
     printf("SICORTEX");
-  }
+  }else{
+    printf("MIPS64_GENERIC");
+  } 
 }
 
 void get_subdirname(void){
@@ -215,8 +222,8 @@ void get_libname(void){
     printf("p6600\n");
   }else if(detect()==CPU_I6500) {
     printf("i6500\n");
-  }else{
+  }else {
-    printf("mips64\n");
+    printf("mips64_generic\n");
   }
 }
 

driver/others/dynamic_mips64.c

@@ -38,22 +38,48 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <sys/resource.h>
 #include "common.h"
 
+#if (defined OS_LINUX || defined OS_ANDROID)
+#include <asm/hwcap.h>
+#include <sys/auxv.h>
+
+#ifndef HWCAP_LOONGSON_CPUCFG
+#define HWCAP_LOONGSON_CPUCFG (1 << 14)
+#endif
+#endif
+
+#ifdef DYNAMIC_LIST
+extern gotoblas_t  gotoblas_MIPS64_GENERIC;
+#ifdef DYN_LOONGSON3R3
+extern gotoblas_t  gotoblas_LOONGSON3R3;
+#else
+#define  gotoblas_LOONGSON3R3 gotoblas_MIPS64_GENERIC
+#endif
+#ifdef DYN_LOONGSON3R4
+extern gotoblas_t  gotoblas_LOONGSON3R4;
+#else
+#define  gotoblas_LOONGSON3R4 gotoblas_MIPS64_GENERIC
+#endif
+#else
 extern gotoblas_t  gotoblas_LOONGSON3R3;
 extern gotoblas_t  gotoblas_LOONGSON3R4;
+extern gotoblas_t  gotoblas_MIPS64_GENERIC;
+#endif
 
 extern void openblas_warning(int verbose, const char * msg);
 
-#define NUM_CORETYPES    2
+#define NUM_CORETYPES    3
 
 static char *corename[] = {
+  "MIPS64_GENERIC"
   "loongson3r3",
   "loongson3r4",
   "UNKNOWN"
 };
 
 char *gotoblas_corename(void) {
-  if (gotoblas == &gotoblas_LOONGSON3R3)    return corename[0];
+  if (gotoblas == &gotoblas_MIPS64_GENERIC) return corename[0];
-  if (gotoblas == &gotoblas_LOONGSON3R4)    return corename[1];
+  if (gotoblas == &gotoblas_LOONGSON3R3)    return corename[1];
+  if (gotoblas == &gotoblas_LOONGSON3R4)    return corename[2];
   return corename[NUM_CORETYPES];
 }
 
@@ -73,77 +99,32 @@ static gotoblas_t *force_coretype(char *coretype) {
 
   switch (found)
   {
-    case  0: return (&gotoblas_LOONGSON3R3);
+    case  0: return (&gotoblas_MIPS64_GENERIC);
-    case  1: return (&gotoblas_LOONGSON3R4);
+    case  1: return (&gotoblas_LOONGSON3R3);
+    case  2: return (&gotoblas_LOONGSON3R4);
   }
   snprintf(message, 128, "Core not found: %s\n", coretype);
   openblas_warning(1, message);
   return NULL;
 }
 
+#if (defined OS_LINUX || defined OS_ANDROID)
 #define MMI_MASK    0x00000010
 #define MSA_MASK    0x00000020
 
-int fd[2];
-int support_cpucfg;
-
-static void handler(int signum)
-{
-    close(fd[1]);
-    exit(1);
-}
-
-/* Brief :  Function to check if cpucfg supported on loongson
- * Return:  1   supported
- *          0   not supported
- */
-static int cpucfg_test(void) {
-    pid_t pid;
-    int status = 0;
-
-    support_cpucfg = 0;
-    pipe(fd);
-    pid = fork();
-    if (pid == 0) { /* Subprocess */
-        struct sigaction act;
-        close(fd[0]);
-        /* Set signal action for SIGILL. */
-        act.sa_handler = handler;
-        sigaction(SIGILL,&act,NULL);
-
-        /* Execute cpucfg in subprocess. */
-        __asm__ volatile(
-            ".insn              \n\t"
-            ".word (0xc8080118) \n\t"
-            :::
-        );
-        support_cpucfg = 1;
-        write(fd[1],&support_cpucfg,sizeof(support_cpucfg));
-        close(fd[1]);
-        exit(0);
-    } else if (pid > 0){ /* Parent process*/
-        close(fd[1]);
-        if ((waitpid(pid,&status,0) <= 0) ||
-            (read(fd[0],&support_cpucfg,sizeof(support_cpucfg)) <= 0))
-            support_cpucfg = 0;
-        close(fd[0]);
-    } else {
-        support_cpucfg = 0;
-    }
-
-    return support_cpucfg;
-}
-
 static gotoblas_t *get_coretype_from_cpucfg(void) {
     int flag = 0;
     __asm__ volatile(
+        ".set push                 \n\t"
+        ".set noat                 \n\t"
         ".insn                     \n\t"
-        "dli    $8,    0x01        \n\t"
+        "dli    $1,    0x01        \n\t"
-        ".word (0xc9084918)        \n\t"
+        ".word (0xc8080118)        \n\t"
-        "usw    $9,    0x00(%0)    \n\t"
+        "move   %0,    $1          \n\t"
+        ".set pop                  \n\t"
+        : "=r"(flag)
+        :
         :
-        : "r"(&flag)
-        : "memory"
     );
     if (flag & MSA_MASK)
         return (&gotoblas_LOONGSON3R4);
@@ -153,7 +134,7 @@ static gotoblas_t *get_coretype_from_cpucfg(void) {
 }
 
 static gotoblas_t *get_coretype_from_cpuinfo(void) {
-#ifdef linux
+#ifdef __linux
   FILE *infile;
   char buffer[512], *p;
 
@@ -176,17 +157,19 @@ static gotoblas_t *get_coretype_from_cpuinfo(void) {
      return NULL;
   }
 #endif
-    return NULL;
+  return NULL;
 }
+#endif
 
 static gotoblas_t *get_coretype(void) {
-    int ret = 0;
+#if (!defined OS_LINUX && !defined OS_ANDROID)
-
+  return NULL;
-    ret = cpucfg_test();
+#else
-    if (ret == 1)
+  if (!(getauxval(AT_HWCAP) & HWCAP_LOONGSON_CPUCFG))
-        return get_coretype_from_cpucfg();
+    return get_coretype_from_cpucfg();
-    else
+  else
-        return get_coretype_from_cpuinfo();
+    return get_coretype_from_cpuinfo();
+#endif
 }
 
 void gotoblas_dynamic_init(void) {
@@ -208,9 +191,9 @@ void gotoblas_dynamic_init(void) {
 
   if (gotoblas == NULL)
   {
-    snprintf(coremsg, 128, "Falling back to loongson3r3 core\n");
+    snprintf(coremsg, 128, "Falling back to MIPS64_GENEIRC\n");
     openblas_warning(1, coremsg);
-    gotoblas = &gotoblas_LOONGSON3R3;
+    gotoblas = &gotoblas_MIPS64_GENERIC;
   }
 
   if (gotoblas && gotoblas->init) {

getarch.c

@@ -131,6 +131,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /* #define FORCE_PPC440		*/
 /* #define FORCE_PPC440FP2	*/
 /* #define FORCE_CELL		*/
+/* #define FORCE_MIPS64_GENERIC	*/
 /* #define FORCE_SICORTEX	*/
 /* #define FORCE_LOONGSON3R3     */
 /* #define FORCE_LOONGSON3R4     */
@@ -918,6 +919,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "CELL"
 #endif
 
+#ifdef FORCE_MIPS64_GENERIC
+#define FORCE
+#define ARCHITECTURE    "MIPS"
+#define SUBARCHITECTURE "MIPS64_GENERIC"
+#define SUBDIRNAME      "mips64"
+#define ARCHCONFIG   "-DMIPS64_GENERIC " \
+       "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=32 " \
+       "-DL2_SIZE=1048576 -DL2_LINESIZE=32 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
+#define LIBNAME   "mips64_generic"
+#define CORENAME  "MIPS64_GENERIC"
+#else
+#endif
+
 #ifdef FORCE_SICORTEX
 #define FORCE
 #define ARCHITECTURE    "MIPS"

kernel/mips64/KERNEL

@@ -42,50 +42,58 @@ endif
 
 ifndef SGEMMKERNEL
 SGEMMKERNEL    =  gemm_kernel.S
+ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
 SGEMMINCOPY    = ../generic/gemm_ncopy_2.c
 SGEMMITCOPY    = ../generic/gemm_tcopy_2.c
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
 SGEMMONCOPY    = ../generic/gemm_ncopy_8.c
 SGEMMOTCOPY    = ../generic/gemm_tcopy_8.c
-SGEMMINCOPYOBJ =  sgemm_incopy.o
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
-SGEMMITCOPYOBJ =  sgemm_itcopy.o
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
-SGEMMONCOPYOBJ =  sgemm_oncopy.o
-SGEMMOTCOPYOBJ =  sgemm_otcopy.o
 endif
 
 ifndef DGEMMKERNEL
 DGEMMKERNEL    =  gemm_kernel.S
+ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
 DGEMMINCOPY    = ../generic/gemm_ncopy_2.c
 DGEMMITCOPY    = ../generic/gemm_tcopy_2.c
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
 DGEMMONCOPY    = ../generic/gemm_ncopy_8.c
 DGEMMOTCOPY    = ../generic/gemm_tcopy_8.c
-DGEMMINCOPYOBJ =  dgemm_incopy.o
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
-DGEMMITCOPYOBJ =  dgemm_itcopy.o
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
-DGEMMONCOPYOBJ =  dgemm_oncopy.o
-DGEMMOTCOPYOBJ =  dgemm_otcopy.o
 endif
 
 ifndef CGEMMKERNEL
 CGEMMKERNEL    =  zgemm_kernel.S
+ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N))
 CGEMMINCOPY    = ../generic/zgemm_ncopy_1.c
 CGEMMITCOPY    = ../generic/zgemm_tcopy_1.c
+CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
 CGEMMONCOPY    = ../generic/zgemm_ncopy_4.c
 CGEMMOTCOPY    = ../generic/zgemm_tcopy_4.c
-CGEMMINCOPYOBJ =  cgemm_incopy.o
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
-CGEMMITCOPYOBJ =  cgemm_itcopy.o
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
-CGEMMONCOPYOBJ =  cgemm_oncopy.o
-CGEMMOTCOPYOBJ =  cgemm_otcopy.o
 endif
 
 ifndef ZGEMMKERNEL
 ZGEMMKERNEL    =  zgemm_kernel.S
+ifneq ($(ZGEMM_UNROLL_M), $(ZGEMM_UNROLL_N))
 ZGEMMINCOPY    = ../generic/zgemm_ncopy_1.c
 ZGEMMITCOPY    = ../generic/zgemm_tcopy_1.c
+ZGEMMINCOPYOBJ =  zgemm_incopy$(TSUFFIX).$(SUFFIX)
+ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
 ZGEMMONCOPY    = ../generic/zgemm_ncopy_4.c
 ZGEMMOTCOPY    = ../generic/zgemm_tcopy_4.c
-ZGEMMINCOPYOBJ =  zgemm_incopy.o
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
-ZGEMMITCOPYOBJ =  zgemm_itcopy.o
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
-ZGEMMONCOPYOBJ =  zgemm_oncopy.o
-ZGEMMOTCOPYOBJ =  zgemm_otcopy.o
 endif
 
 ifndef SGEMM_BETA

param.h

@@ -2945,7 +2945,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SYMV_P	16
 #endif
 
-#if defined(P5600) || defined(MIPS1004K) || defined(MIPS24K) || defined(I6400) || defined(P6600) || defined(I6500)
+#if defined(MIPS64_GENERIC) || defined(P5600) || defined(MIPS1004K) || defined(MIPS24K) || defined(I6400) || defined(P6600) || defined(I6500)
 #define SNUMOPT  2
 #define DNUMOPT  2