loongarch: Optimizing the performance of the GEMM on servers

common_loongarch64.h

@@ -96,6 +96,32 @@ static inline int WhereAmI(void){
 }
 #endif
 
+static inline int get_cpu_model(char *model_name) {
+  FILE *cpuinfo_file = fopen("/proc/cpuinfo", "r");
+  if (!cpuinfo_file) {
+    return 0;
+  }
+  char line[1024];
+  while (fgets(line, sizeof(line), cpuinfo_file)) {
+    if (strstr(line, "model name")) {
+      char *token = strtok(line, ":");
+      token = strtok(NULL, ":");
+      while (*token == ' ')
+        token++;
+      char *end = token + strlen(token) - 1;
+      while (end > token && (*end == '\n' || *end == '\r')) {
+        *end = '\0';
+        end--;
+      }
+      strcpy(model_name, token);
+      fclose(cpuinfo_file);
+      return 1;
+    }
+  }
+  fclose(cpuinfo_file);
+  return 0;
+}
+
 #ifdef DOUBLE
 #define GET_IMAGE(res)  __asm__ __volatile__("fmov.d %0, $f2" : "=f"(res)  : : "memory")
 #else

driver/others/blas_server.c

@@ -484,6 +484,14 @@ blas_queue_t *tscq;
       main_status[cpu] = MAIN_RUNNING1;
 #endif
 
+//For Loongson servers, like the 3C5000 (featuring 16 cores), applying an
+//offset to the buffer is essential for minimizing cache conflicts and optimizing performance.
+#if defined(LOONGSON3R5) && !defined(NO_AFFINITY)
+      char model_name[128];
+      get_cpu_model(model_name);
+      if ((strstr(model_name, "3C5000") != NULL) || (strstr(model_name, "3D5000") != NULL))
+        if (sa == NULL) sa = (void *)((BLASLONG)buffer + (WhereAmI() & 0xf) * GEMM_OFFSET_A);
+#endif
       if (sa == NULL) sa = (void *)((BLASLONG)buffer + GEMM_OFFSET_A);
 
       if (sb == NULL) {
@@ -1006,7 +1014,7 @@ void goto_set_num_threads(int num_threads) {
 
   blas_cpu_number  = num_threads;
 
-#if defined(ARCH_MIPS64)
+#if defined(ARCH_MIPS64) || defined(ARCH_LOONGARCH64)
 #ifndef DYNAMIC_ARCH
   //set parameters for different number of threads.
   blas_set_parameter();

driver/others/blas_server_omp.c

@@ -113,7 +113,7 @@ void goto_set_num_threads(int num_threads) {
   blas_cpu_number  = num_threads;
 
   adjust_thread_buffers();
-#if defined(ARCH_MIPS64)
+#if defined(ARCH_MIPS64) || defined(ARCH_LOONGARCH64)
   //set parameters for different number of threads.
   blas_set_parameter();
 #endif

driver/others/memory.c

@@ -1219,7 +1219,7 @@ UNLOCK_COMMAND(&alloc_lock);
       if (!blas_num_threads) blas_cpu_number = blas_get_cpu_number();
 #endif
 
-#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64)
+#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64) || defined(ARCH_LOONGARCH64)
 #ifndef DYNAMIC_ARCH
       blas_set_parameter();
 #endif
@@ -2814,7 +2814,7 @@ void *blas_memory_alloc(int procpos){
     if (!blas_num_threads) blas_cpu_number = blas_get_cpu_number();
 #endif
 
-#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64)
+#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64) || defined(ARCH_LOONGARCH64)
 #ifndef DYNAMIC_ARCH
     blas_set_parameter();
 #endif

driver/others/parameter.c

@@ -739,6 +739,100 @@ void blas_set_parameter(void){
 }
 #endif
 
+#if defined(ARCH_LOONGARCH64)
+int get_L3_size() {
+  int ret = 0, id = 0x14;
+  __asm__ volatile (
+    "cpucfg %[ret], %[id]"
+    : [ret]"=r"(ret)
+    : [id]"r"(id)
+    : "memory"
+  );
+  return ((ret & 0xffff) + 1) * pow(2, ((ret >> 16) & 0xff)) * pow(2, ((ret >> 24) & 0x7f)) / 1024 / 1024; // MB
+}
+
+void blas_set_parameter(void){
+#if defined(LOONGSON3R5)
+  int L3_size = get_L3_size();
+#ifdef SMP
+  if(blas_num_threads == 1){
+#endif
+    //single thread
+    if (L3_size == 32){ // 3C5000 and 3D5000
+      sgemm_p = 256;
+      sgemm_q = 384;
+      sgemm_r = 8192;
+
+      dgemm_p = 112;
+      dgemm_q = 289;
+      dgemm_r = 4096;
+
+      cgemm_p = 128;
+      cgemm_q = 256;
+      cgemm_r = 4096;
+
+      zgemm_p = 128;
+      zgemm_q = 128;
+      zgemm_r = 2048;
+    } else { // 3A5000 and 3C5000L
+      sgemm_p = 256;
+      sgemm_q = 384;
+      sgemm_r = 4096;
+
+      dgemm_p = 112;
+      dgemm_q = 300;
+      dgemm_r = 3024;
+
+      cgemm_p = 128;
+      cgemm_q = 256;
+      cgemm_r = 2048;
+
+      zgemm_p = 128;
+      zgemm_q = 128;
+      zgemm_r = 1024;
+    }
+#ifdef SMP
+  }else{
+    //multi thread
+    if (L3_size == 32){ // 3C5000 and 3D5000
+      sgemm_p = 256;
+      sgemm_q = 384;
+      sgemm_r = 1024;
+
+      dgemm_p = 112;
+      dgemm_q = 289;
+      dgemm_r = 342;
+
+      cgemm_p = 128;
+      cgemm_q = 256;
+      cgemm_r = 512;
+
+      zgemm_p = 128;
+      zgemm_q = 128;
+      zgemm_r = 512;
+    } else { // 3A5000 and 3C5000L
+      sgemm_p = 256;
+      sgemm_q = 384;
+      sgemm_r = 2048;
+
+      dgemm_p = 112;
+      dgemm_q = 300;
+      dgemm_r = 738;
+
+      cgemm_p = 128;
+      cgemm_q = 256;
+      cgemm_r = 1024;
+
+      zgemm_p = 128;
+      zgemm_q = 128;
+      zgemm_r = 1024;
+    }
+  }
+#endif
+#endif
+}
+#endif
+
 #if defined(ARCH_ARM64)
 
 void blas_set_parameter(void)

interface/gemm.c

@@ -521,7 +521,18 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 
   buffer = (XFLOAT *)blas_memory_alloc(0);
 
+//For Loongson servers, like the 3C5000 (featuring 16 cores), applying an
+//offset to the buffer is essential for minimizing cache conflicts and optimizing performance.
+#if defined(LOONGSON3R5) && !defined(NO_AFFINITY)
+  char model_name[128];
+  get_cpu_model(model_name);
+  if ((strstr(model_name, "3C5000") != NULL) || (strstr(model_name, "3D5000") != NULL))
+    sa = (XFLOAT *)((BLASLONG)buffer + (WhereAmI() & 0xf) * GEMM_OFFSET_A);
+  else
+    sa = (XFLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
+#else
   sa = (XFLOAT *)((BLASLONG)buffer +GEMM_OFFSET_A);
+#endif
   sb = (XFLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
 
 #ifdef SMP

kernel/setparam-ref.c

@@ -1066,31 +1066,123 @@ static void init_parameter(void) {
 }
 #else // (ARCH_MIPS64)
 #if (ARCH_LOONGARCH64)
+static int get_L3_size() {
+  int ret = 0, id = 0x14;
+  __asm__ volatile (
+    "cpucfg %[ret], %[id]"
+    : [ret]"=r"(ret)
+    : [id]"r"(id)
+    : "memory"
+  );
+  return ((ret & 0xffff) + 1) * pow(2, ((ret >> 16) & 0xff)) * pow(2, ((ret >> 24) & 0x7f)) / 1024 / 1024; // MB
+}
 static void init_parameter(void) {
 
 #ifdef BUILD_BFLOAT16
   TABLE_NAME.sbgemm_p = SBGEMM_DEFAULT_P;
 #endif
+
+#ifdef BUILD_BFLOAT16
+  TABLE_NAME.sbgemm_r = SBGEMM_DEFAULT_R;
+#endif
+
+#if defined(LOONGSON3R5)
+  int L3_size = get_L3_size();
+#ifdef SMP
+  if(blas_num_threads == 1){
+#endif
+    //single thread
+    if (L3_size == 32){ // 3C5000 and 3D5000
+      TABLE_NAME.sgemm_p = 256;
+      TABLE_NAME.sgemm_q = 384;
+      TABLE_NAME.sgemm_r = 8192;
+
+      TABLE_NAME.dgemm_p = 112;
+      TABLE_NAME.dgemm_q = 289;
+      TABLE_NAME.dgemm_r = 4096;
+
+      TABLE_NAME.cgemm_p = 128;
+      TABLE_NAME.cgemm_q = 256;
+      TABLE_NAME.cgemm_r = 4096;
+
+      TABLE_NAME.zgemm_p = 128;
+      TABLE_NAME.zgemm_q = 128;
+      TABLE_NAME.zgemm_r = 2048;
+    } else { // 3A5000 and 3C5000L
+      TABLE_NAME.sgemm_p = 256;
+      TABLE_NAME.sgemm_q = 384;
+      TABLE_NAME.sgemm_r = 4096;
+
+      TABLE_NAME.dgemm_p = 112;
+      TABLE_NAME.dgemm_q = 300;
+      TABLE_NAME.dgemm_r = 3024;
+
+      TABLE_NAME.cgemm_p = 128;
+      TABLE_NAME.cgemm_q = 256;
+      TABLE_NAME.cgemm_r = 2048;
+
+      TABLE_NAME.zgemm_p = 128;
+      TABLE_NAME.zgemm_q = 128;
+      TABLE_NAME.zgemm_r = 1024;
+    }
+#ifdef SMP
+  }else{
+    //multi thread
+    if (L3_size == 32){ // 3C5000 and 3D5000
+      TABLE_NAME.sgemm_p = 256;
+      TABLE_NAME.sgemm_q = 384;
+      TABLE_NAME.sgemm_r = 1024;
+
+      TABLE_NAME.dgemm_p = 112;
+      TABLE_NAME.dgemm_q = 289;
+      TABLE_NAME.dgemm_r = 342;
+
+      TABLE_NAME.cgemm_p = 128;
+      TABLE_NAME.cgemm_q = 256;
+      TABLE_NAME.cgemm_r = 512;
+
+      TABLE_NAME.zgemm_p = 128;
+      TABLE_NAME.zgemm_q = 128;
+      TABLE_NAME.zgemm_r = 512;
+    } else { // 3A5000 and 3C5000L
+      TABLE_NAME.sgemm_p = 256;
+      TABLE_NAME.sgemm_q = 384;
+      TABLE_NAME.sgemm_r = 2048;
+
+      TABLE_NAME.dgemm_p = 112;
+      TABLE_NAME.dgemm_q = 300;
+      TABLE_NAME.dgemm_r = 738;
+
+      TABLE_NAME.cgemm_p = 128;
+      TABLE_NAME.cgemm_q = 256;
+      TABLE_NAME.cgemm_r = 1024;
+
+      TABLE_NAME.zgemm_p = 128;
+      TABLE_NAME.zgemm_q = 128;
+      TABLE_NAME.zgemm_r = 1024;
+    }
+  }
+#endif
+#else
   TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
   TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
   TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
   TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
 
-#ifdef BUILD_BFLOAT16
-  TABLE_NAME.sbgemm_r = SBGEMM_DEFAULT_R;
-#endif
+  TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
+  TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
+  TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
+  TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
+
   TABLE_NAME.sgemm_r = SGEMM_DEFAULT_R;
   TABLE_NAME.dgemm_r = DGEMM_DEFAULT_R;
   TABLE_NAME.cgemm_r = CGEMM_DEFAULT_R;
   TABLE_NAME.zgemm_r = ZGEMM_DEFAULT_R;
+#endif
 
 #ifdef BUILD_BFLOAT16
   TABLE_NAME.sbgemm_q = SBGEMM_DEFAULT_Q;
 #endif
-  TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
-  TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
-  TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
-  TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
 }
 #else // (ARCH_LOONGARCH64)
 #if (ARCH_POWER)

param.h

@@ -2836,7 +2836,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SNUMOPT         2
 #define DNUMOPT         2
 
-#define GEMM_DEFAULT_OFFSET_A 0
+#define GEMM_DEFAULT_OFFSET_A 0x20000
 #define GEMM_DEFAULT_OFFSET_B 0
 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
 
@@ -2866,20 +2866,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define QGEMM_DEFAULT_UNROLL_M 2
 #define XGEMM_DEFAULT_UNROLL_M 1
 
-#define SGEMM_DEFAULT_P 256
-#define DGEMM_DEFAULT_P 32
+#define SGEMM_DEFAULT_P sgemm_p
+#define DGEMM_DEFAULT_P dgemm_p
 #define CGEMM_DEFAULT_P 128
-#define ZGEMM_DEFAULT_P 128
+#define ZGEMM_DEFAULT_P zgemm_p
 
-#define SGEMM_DEFAULT_R 1024
-#define DGEMM_DEFAULT_R 858
+#define SGEMM_DEFAULT_R sgemm_r
+#define DGEMM_DEFAULT_R dgemm_r
 #define CGEMM_DEFAULT_R 4096
-#define ZGEMM_DEFAULT_R 4096
+#define ZGEMM_DEFAULT_R zgemm_r
 
-#define SGEMM_DEFAULT_Q 256
-#define DGEMM_DEFAULT_Q 152
+#define SGEMM_DEFAULT_Q sgemm_q
+#define DGEMM_DEFAULT_Q dgemm_q
 #define CGEMM_DEFAULT_Q 128
-#define ZGEMM_DEFAULT_Q 128
+#define ZGEMM_DEFAULT_Q zgemm_q
 
 #define SYMV_P  16
 #endif