Define SBGEMM_ALIGN_K for DYNAMIC_ARCH build

common_param.h

@@ -50,6 +50,7 @@ typedef struct {
 #ifdef BUILD_BFLOAT16
   int sbgemm_p, sbgemm_q, sbgemm_r;
   int sbgemm_unroll_m, sbgemm_unroll_n, sbgemm_unroll_mn;
+  int sbgemm_align_k;
 
   void   (*sbstobf16_k) (BLASLONG, float    *, BLASLONG, bfloat16 *, BLASLONG);
   void   (*sbdtobf16_k) (BLASLONG, double   *, BLASLONG, bfloat16 *, BLASLONG);
@@ -1193,7 +1194,6 @@ BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
 #ifdef BUILD_COMPLEX16
   int    (*zgeadd_k) (BLASLONG, BLASLONG, double, double, double *, BLASLONG, double, double, double *, BLASLONG); 
 #endif
-  int align_k;  // must be 2^n
 } gotoblas_t;
 
 extern gotoblas_t *gotoblas;

driver/level3/level3.c

@@ -305,13 +305,12 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
       }
 
       BLASLONG pad_min_l = min_l;
-      
-#if defined(HALF) && defined(DYNAMIC_ARCH)
-      pad_min_l = (min_l + gotoblas->align_k - 1) & ~(gotoblas->align_k-1);
+#if defined(HALF)
+#if defined(DYNAMIC_ARCH)
+      pad_min_l = (min_l + gotoblas->sbgemm_align_k - 1) & ~(gotoblas->sbgemm_align_k-1);
+#else
+      pad_min_l = (min_l + SBGEMM_ALIGN_K - 1) & ~(SBGEMM_ALIGN_K - 1);;
 #endif
-
-#if defined(HALF) && !defined(DYNAMIC_ARCH) && defined(NEOVERSEN2)
-      pad_min_l = (min_l + 3) & ~3;
 #endif
 
       /* First, we have to move data A to L2 cache */

driver/level3/level3_thread.c

@@ -327,12 +327,12 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
     
     BLASLONG pad_min_l = min_l;
 
-#if defined(HALF) && defined(DYNAMIC_ARCH)
-    pad_min_l = (min_l + gotoblas->align_k - 1) & ~(gotoblas->align_k-1);
+#if defined(HALF)
+#if defined(DYNAMIC_ARCH)
+    pad_min_l = (min_l + gotoblas->sbgemm_align_k - 1) & ~(gotoblas->sbgemm_align_k-1);
+#else
+    pad_min_l = (min_l + SBGEMM_ALIGN_K - 1) & ~(SBGEMM_ALIGN_K - 1);;
 #endif
-
-#if defined(HALF) && !defined(DYNAMIC_ARCH) && defined(NEOVERSEN2)
-    pad_min_l = (min_l + 3) & ~3;
 #endif
 
     /* Determine step size in m

kernel/setparam-ref.c

@@ -62,6 +62,8 @@ gotoblas_t TABLE_NAME = {
  MAX(SBGEMM_DEFAULT_UNROLL_M, SBGEMM_DEFAULT_UNROLL_N),
 #endif
 
+  SBGEMM_ALIGN_K,
+
   sbstobf16_kTS, sbdtobf16_kTS, sbf16tos_kTS, dbf16tod_kTS,
 
   samax_kTS,  samin_kTS,  smax_kTS,  smin_kTS,
@@ -973,12 +975,6 @@ static void init_parameter(void) {
   TABLE_NAME.xgemm3m_r = TABLE_NAME.qgemm_r;
 #endif
 #endif
-  
-#if defined(NEOVERSEN2) && BUILD_BFLOAT16 == 1
-  TABLE_NAME.align_k = 4;
-#else
-  TABLE_NAME.align_k = 1;
-#endif
 
 }
 #else // (ARCH_ARM64)

param.h

@@ -79,6 +79,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SBGEMM_DEFAULT_P 256
 #define SBGEMM_DEFAULT_R 256
 #define SBGEMM_DEFAULT_Q 256
+#define SBGEMM_ALIGN_K 1  // must be 2^x
+
 #ifdef OPTERON
 
 #define SNUMOPT		4
@@ -3394,6 +3396,9 @@ is a big desktop or server with abundant cache rather than a phone or embedded d
 
 #elif defined(NEOVERSEN2)
 
+#undef SBGEMM_ALIGN_K
+#define SBGEMM_ALIGN_K 4
+
 #undef SBGEMM_DEFAULT_UNROLL_M
 #undef SBGEMM_DEFAULT_UNROLL_N
 #define SBGEMM_DEFAULT_UNROLL_M 8