enabled use of GEMM3M functions

common_param.h

@@ -435,6 +435,9 @@ BLASLONG (*icamin_k)(BLASLONG, float *, BLASLONG);
   int    (*chemm_outcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
   int    (*chemm_oltcopy)(BLASLONG, BLASLONG, float *, BLASLONG, BLASLONG, BLASLONG, float *);
 
+  int cgemm3m_p, cgemm3m_q, cgemm3m_r;
+  int cgemm3m_unroll_m, cgemm3m_unroll_n, cgemm3m_unroll_mn;
+
   int    (*cgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG);
 
   int    (*cgemm3m_incopyb)(BLASLONG, BLASLONG, float *, BLASLONG, float *);
@@ -595,6 +598,9 @@ BLASLONG (*izamin_k)(BLASLONG, double *, BLASLONG);
   int    (*zhemm_outcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
   int    (*zhemm_oltcopy)(BLASLONG, BLASLONG, double *, BLASLONG, BLASLONG, BLASLONG, double *);
 
+  int zgemm3m_p, zgemm3m_q, zgemm3m_r;
+  int zgemm3m_unroll_m, zgemm3m_unroll_n, zgemm3m_unroll_mn;
+
   int    (*zgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG);
 
   int    (*zgemm3m_incopyb)(BLASLONG, BLASLONG, double *, BLASLONG, double *);
@@ -757,6 +763,9 @@ BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
   int    (*xhemm_outcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
   int    (*xhemm_oltcopy)(BLASLONG, BLASLONG, xdouble *, BLASLONG, BLASLONG, BLASLONG, xdouble *);
 
+  int xgemm3m_p, xgemm3m_q, xgemm3m_r;
+  int xgemm3m_unroll_m, xgemm3m_unroll_n, xgemm3m_unroll_mn;
+
   int    (*xgemm3m_kernel)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, xdouble *, xdouble *, BLASLONG);
 
   int    (*xgemm3m_incopyb)(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
@@ -900,6 +909,27 @@ extern gotoblas_t *gotoblas;
 #define	XGEMM_UNROLL_N	gotoblas -> xgemm_unroll_n
 #define XGEMM_UNROLL_MN	gotoblas -> xgemm_unroll_mn
 
+#define	CGEMM3M_P		gotoblas -> cgemm3m_p
+#define	CGEMM3M_Q		gotoblas -> cgemm3m_q
+#define	CGEMM3M_R		gotoblas -> cgemm3m_r
+#define	CGEMM3M_UNROLL_M	gotoblas -> cgemm3m_unroll_m
+#define	CGEMM3M_UNROLL_N	gotoblas -> cgemm3m_unroll_n
+#define CGEMM3M_UNROLL_MN	gotoblas -> cgemm3m_unroll_mn
+
+#define	ZGEMM3M_P		gotoblas -> zgemm3m_p
+#define	ZGEMM3M_Q		gotoblas -> zgemm3m_q
+#define	ZGEMM3M_R		gotoblas -> zgemm3m_r
+#define	ZGEMM3M_UNROLL_M	gotoblas -> zgemm3m_unroll_m
+#define	ZGEMM3M_UNROLL_N	gotoblas -> zgemm3m_unroll_n
+#define ZGEMM3M_UNROLL_MN	gotoblas -> zgemm3m_unroll_mn
+
+#define	XGEMM3M_P		gotoblas -> xgemm3m_p
+#define	XGEMM3M_Q		gotoblas -> xgemm3m_q
+#define	XGEMM3M_R		gotoblas -> xgemm3m_r
+#define	XGEMM3M_UNROLL_M	gotoblas -> xgemm3m_unroll_m
+#define	XGEMM3M_UNROLL_N	gotoblas -> xgemm3m_unroll_n
+#define XGEMM3M_UNROLL_MN	gotoblas -> xgemm3m_unroll_mn
+
 #else
 
 #define DTB_ENTRIES  DTB_DEFAULT_ENTRIES
@@ -972,6 +1002,55 @@ extern gotoblas_t *gotoblas;
 #define XGEMM_UNROLL_N	XGEMM_DEFAULT_UNROLL_N
 #define XGEMM_UNROLL_MN	MAX((XGEMM_UNROLL_M), (XGEMM_UNROLL_N))
 
+#ifdef CGEMM_DEFAULT_UNROLL_N
+
+#define	CGEMM3M_P		CGEMM3M_DEFAULT_P
+#define	CGEMM3M_Q		CGEMM3M_DEFAULT_Q
+#define	CGEMM3M_R		CGEMM3M_DEFAULT_R
+#define CGEMM3M_UNROLL_M	CGEMM3M_DEFAULT_UNROLL_M
+#define CGEMM3M_UNROLL_N	CGEMM3M_DEFAULT_UNROLL_N
+#define CGEMM3M_UNROLL_MN	MAX((CGEMM3M_UNROLL_M), (CGEMM3M_UNROLL_N))
+
+#else
+
+#define	CGEMM3M_P		SGEMM_DEFAULT_P
+#define	CGEMM3M_Q		SGEMM_DEFAULT_Q
+#define	CGEMM3M_R		SGEMM_DEFAULT_R
+#define CGEMM3M_UNROLL_M	SGEMM_DEFAULT_UNROLL_M
+#define CGEMM3M_UNROLL_N	SGEMM_DEFAULT_UNROLL_N
+#define CGEMM3M_UNROLL_MN	MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
+
+#endif
+
+
+#ifdef ZGEMM_DEFAULT_UNROLL_N
+
+#define	ZGEMM3M_P		ZGEMM3M_DEFAULT_P
+#define	ZGEMM3M_Q		ZGEMM3M_DEFAULT_Q
+#define	ZGEMM3M_R		ZGEMM3M_DEFAULT_R
+#define ZGEMM3M_UNROLL_M	ZGEMM3M_DEFAULT_UNROLL_M
+#define ZGEMM3M_UNROLL_N	ZGEMM3M_DEFAULT_UNROLL_N
+#define ZGEMM3M_UNROLL_MN	MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
+
+#else
+
+#define	ZGEMM3M_P		DGEMM_DEFAULT_P
+#define	ZGEMM3M_Q		DGEMM_DEFAULT_Q
+#define	ZGEMM3M_R		DGEMM_DEFAULT_R
+#define ZGEMM3M_UNROLL_M	DGEMM_DEFAULT_UNROLL_M
+#define ZGEMM3M_UNROLL_N	DGEMM_DEFAULT_UNROLL_N
+#define ZGEMM3M_UNROLL_MN	MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
+
+#endif
+
+#define	XGEMM3M_P		QGEMM_DEFAULT_P
+#define	XGEMM3M_Q		QGEMM_DEFAULT_Q
+#define	XGEMM3M_R		QGEMM_DEFAULT_R
+#define XGEMM3M_UNROLL_M	QGEMM_DEFAULT_UNROLL_M
+#define XGEMM3M_UNROLL_N	QGEMM_DEFAULT_UNROLL_N
+#define XGEMM3M_UNROLL_MN	MAX((QGEMM_UNROLL_M), (QGEMM_UNROLL_N))
+
+
 #endif
 #endif
 
@@ -1054,14 +1133,14 @@ extern gotoblas_t *gotoblas;
 #endif
 
 #ifdef XDOUBLE
-#define GEMM3M_UNROLL_M	QGEMM_UNROLL_M
-#define GEMM3M_UNROLL_N	QGEMM_UNROLL_N
+#define GEMM3M_UNROLL_M	XGEMM3M_UNROLL_M
+#define GEMM3M_UNROLL_N	XGEMM3M_UNROLL_N
 #elif defined(DOUBLE)
-#define GEMM3M_UNROLL_M	DGEMM_UNROLL_M
-#define GEMM3M_UNROLL_N	DGEMM_UNROLL_N
+#define GEMM3M_UNROLL_M	ZGEMM3M_UNROLL_M
+#define GEMM3M_UNROLL_N	ZGEMM3M_UNROLL_N
 #else
-#define GEMM3M_UNROLL_M	SGEMM_UNROLL_M
-#define GEMM3M_UNROLL_N	SGEMM_UNROLL_N
+#define GEMM3M_UNROLL_M	CGEMM3M_UNROLL_M
+#define GEMM3M_UNROLL_N	CGEMM3M_UNROLL_N
 #endif
 
 
@@ -1123,31 +1202,31 @@ extern gotoblas_t *gotoblas;
 
 #ifndef GEMM3M_P
 #ifdef XDOUBLE
-#define GEMM3M_P	QGEMM_P
+#define GEMM3M_P	XGEMM3M_P
 #elif defined(DOUBLE)
-#define GEMM3M_P	DGEMM_P
+#define GEMM3M_P	ZGEMM3M_P
 #else
-#define GEMM3M_P	SGEMM_P
+#define GEMM3M_P	CGEMM3M_P
 #endif
 #endif
 
 #ifndef GEMM3M_Q
 #ifdef XDOUBLE
-#define GEMM3M_Q	QGEMM_Q
+#define GEMM3M_Q	XGEMM3M_Q
 #elif defined(DOUBLE)
-#define GEMM3M_Q	DGEMM_Q
+#define GEMM3M_Q	ZGEMM3M_Q
 #else
-#define GEMM3M_Q	SGEMM_Q
+#define GEMM3M_Q	CGEMM3M_Q
 #endif
 #endif
 
 #ifndef GEMM3M_R
 #ifdef XDOUBLE
-#define GEMM3M_R	QGEMM_R
+#define GEMM3M_R	XGEMM3M_R
 #elif defined(DOUBLE)
-#define GEMM3M_R	DGEMM_R
+#define GEMM3M_R	ZGEMM3M_R
 #else
-#define GEMM3M_R	SGEMM_R
+#define GEMM3M_R	CGEMM3M_R
 #endif
 #endif
 

driver/level3/Makefile

@@ -4,11 +4,11 @@ include ../../Makefile.system
 USE_GEMM3M = 0
 
 ifeq ($(ARCH), x86)
-USE_GEMM3M = 0
+USE_GEMM3M = 1
 endif
 
 ifeq ($(ARCH), x86_64)
-USE_GEMM3M = 0
+USE_GEMM3M = 1
 endif
 
 ifeq ($(ARCH), ia64)

exports/gensymbol

@@ -75,7 +75,9 @@
 	       );
 
 @gemm3mobjs = (
-	       
+	       cgemm3m,zgemm3m,
+	       chemm3m,zhemm3m,
+	       csymm3m,zsymm3m
 	       );
 
 

interface/Makefile

@@ -4,11 +4,11 @@ include $(TOPDIR)/Makefile.system
 SUPPORT_GEMM3M = 0
 
 ifeq ($(ARCH), x86)
-SUPPORT_GEMM3M = 0
+SUPPORT_GEMM3M = 1
 endif
 
 ifeq ($(ARCH), x86_64)
-SUPPORT_GEMM3M = 0
+SUPPORT_GEMM3M = 1
 endif
 
 ifeq ($(ARCH), ia64)

kernel/setparam-ref.c

@@ -293,6 +293,14 @@ gotoblas_t TABLE_NAME = {
 #endif
   chemm_outcopyTS,  chemm_oltcopyTS,
 
+  0, 0, 0,
+#ifdef CGEMM3M_DEFAULT_UNROLL_M
+  CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N, MAX(CGEMM3M_DEFAULT_UNROLL_M, CGEMM3M_DEFAULT_UNROLL_N),
+#else
+  SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N, MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
+#endif
+
+
   cgemm3m_kernelTS,
 
   cgemm3m_incopybTS,  cgemm3m_incopyrTS,
@@ -391,6 +399,14 @@ gotoblas_t TABLE_NAME = {
 #endif
   zhemm_outcopyTS,  zhemm_oltcopyTS,
 
+  0, 0, 0,
+#ifdef ZGEMM3M_DEFAULT_UNROLL_M
+  ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N, MAX(ZGEMM3M_DEFAULT_UNROLL_M, ZGEMM3M_DEFAULT_UNROLL_N),
+#else
+  DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N, MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
+#endif
+
+
   zgemm3m_kernelTS,
 
   zgemm3m_incopybTS,  zgemm3m_incopyrTS,
@@ -486,6 +502,9 @@ gotoblas_t TABLE_NAME = {
 #endif
   xhemm_outcopyTS,  xhemm_oltcopyTS,
 
+  0, 0, 0,
+  QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N, MAX(QGEMM_DEFAULT_UNROLL_M, QGEMM_DEFAULT_UNROLL_N),
+
   xgemm3m_kernelTS,
 
   xgemm3m_incopybTS,  xgemm3m_incopyrTS,
@@ -661,9 +680,23 @@ static void init_parameter(void) {
   TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
   TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
   TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
+
+#ifdef CGEMM3M_DEFAULT_Q
+  TABLE_NAME.cgemm3m_q = CGEMM3M_DEFAULT_Q;
+#else
+  TABLE_NAME.cgemm3m_q = SGEMM_DEFAULT_Q;
+#endif
+
+#ifdef ZGEMM3M_DEFAULT_Q
+  TABLE_NAME.zgemm3m_q = ZGEMM3M_DEFAULT_Q;
+#else
+  TABLE_NAME.zgemm3m_q = DGEMM_DEFAULT_Q;
+#endif
+
 #ifdef EXPRECISION
   TABLE_NAME.qgemm_q = QGEMM_DEFAULT_Q;
   TABLE_NAME.xgemm_q = XGEMM_DEFAULT_Q;
+  TABLE_NAME.xgemm3m_q = QGEMM_DEFAULT_Q;
 #endif
 
 #if defined(CORE_KATMAI)  || defined(CORE_COPPERMINE) || defined(CORE_BANIAS) || defined(CORE_YONAH) || defined(CORE_ATHLON)
@@ -918,20 +951,56 @@ static void init_parameter(void) {
   TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
   TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
   TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
+
+
+
 #ifdef EXPRECISION
   TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
   TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
 #endif
+
+#endif
+
+
+#ifdef CGEMM3M_DEFAULT_P
+  TABLE_NAME.cgemm3m_p = CGEMM3M_DEFAULT_P;
+#else
+  TABLE_NAME.cgemm3m_p = TABLE_NAME.sgemm_p;
+#endif
+
+#ifdef ZGEMM3M_DEFAULT_P
+  TABLE_NAME.zgemm3m_p = ZGEMM3M_DEFAULT_P;
+#else
+  TABLE_NAME.zgemm3m_p = TABLE_NAME.dgemm_p;
+#endif
+
+#ifdef EXPRECISION
+  TABLE_NAME.xgemm3m_p = TABLE_NAME.qgemm_p;
 #endif
 
 
+
   TABLE_NAME.sgemm_p = (TABLE_NAME.sgemm_p + SGEMM_DEFAULT_UNROLL_M - 1) & ~(SGEMM_DEFAULT_UNROLL_M - 1);
   TABLE_NAME.dgemm_p = (TABLE_NAME.dgemm_p + DGEMM_DEFAULT_UNROLL_M - 1) & ~(DGEMM_DEFAULT_UNROLL_M - 1);
   TABLE_NAME.cgemm_p = (TABLE_NAME.cgemm_p + CGEMM_DEFAULT_UNROLL_M - 1) & ~(CGEMM_DEFAULT_UNROLL_M - 1);
   TABLE_NAME.zgemm_p = (TABLE_NAME.zgemm_p + ZGEMM_DEFAULT_UNROLL_M - 1) & ~(ZGEMM_DEFAULT_UNROLL_M - 1);
+
+#ifdef CGEMM3M_DEFAULT_UNROLL_M
+  TABLE_NAME.cgemm3m_p = (TABLE_NAME.cgemm3m_p + CGEMM3M_DEFAULT_UNROLL_M - 1) & ~(CGEMM3M_DEFAULT_UNROLL_M - 1);
+#else
+  TABLE_NAME.cgemm3m_p = (TABLE_NAME.cgemm3m_p + SGEMM_DEFAULT_UNROLL_M - 1) & ~(SGEMM_DEFAULT_UNROLL_M - 1);
+#endif
+
+#ifdef ZGEMM3M_DEFAULT_UNROLL_M
+  TABLE_NAME.zgemm3m_p = (TABLE_NAME.zgemm3m_p + ZGEMM3M_DEFAULT_UNROLL_M - 1) & ~(ZGEMM3M_DEFAULT_UNROLL_M - 1);
+#else
+  TABLE_NAME.zgemm3m_p = (TABLE_NAME.zgemm3m_p + DGEMM_DEFAULT_UNROLL_M - 1) & ~(DGEMM_DEFAULT_UNROLL_M - 1);
+#endif
+
 #ifdef QUAD_PRECISION
   TABLE_NAME.qgemm_p = (TABLE_NAME.qgemm_p + QGEMM_DEFAULT_UNROLL_M - 1) & ~(QGEMM_DEFAULT_UNROLL_M - 1);
   TABLE_NAME.xgemm_p = (TABLE_NAME.xgemm_p + XGEMM_DEFAULT_UNROLL_M - 1) & ~(XGEMM_DEFAULT_UNROLL_M - 1);
+  TABLE_NAME.xgemm3m_p = (TABLE_NAME.xgemm3m_p + QGEMM_DEFAULT_UNROLL_M - 1) & ~(QGEMM_DEFAULT_UNROLL_M - 1);
 #endif
 
 #ifdef DEBUG
@@ -965,11 +1034,32 @@ static void init_parameter(void) {
 				 + TABLE_NAME.align) & ~TABLE_NAME.align)
 			       ) / (TABLE_NAME.zgemm_q * 16) - 15) & ~15);
 
+  TABLE_NAME.cgemm3m_r = (((BUFFER_SIZE -
+			       ((TABLE_NAME.cgemm3m_p * TABLE_NAME.cgemm3m_q *  8 + TABLE_NAME.offsetA
+				 + TABLE_NAME.align) & ~TABLE_NAME.align)
+			       ) / (TABLE_NAME.cgemm3m_q *  8) - 15) & ~15);
+
+  TABLE_NAME.zgemm3m_r = (((BUFFER_SIZE -
+			       ((TABLE_NAME.zgemm3m_p * TABLE_NAME.zgemm3m_q * 16 + TABLE_NAME.offsetA
+				 + TABLE_NAME.align) & ~TABLE_NAME.align)
+			       ) / (TABLE_NAME.zgemm3m_q * 16) - 15) & ~15);
+
+
+
+
 #ifdef EXPRECISION
   TABLE_NAME.xgemm_r = (((BUFFER_SIZE -
 			       ((TABLE_NAME.xgemm_p * TABLE_NAME.xgemm_q * 32 + TABLE_NAME.offsetA
 				 + TABLE_NAME.align) & ~TABLE_NAME.align)
 		       ) / (TABLE_NAME.xgemm_q * 32) - 15) & ~15);
+
+  TABLE_NAME.xgemm3m_r = (((BUFFER_SIZE -
+			       ((TABLE_NAME.xgemm3m_p * TABLE_NAME.xgemm3m_q * 32 + TABLE_NAME.offsetA
+				 + TABLE_NAME.align) & ~TABLE_NAME.align)
+		       ) / (TABLE_NAME.xgemm3m_q * 32) - 15) & ~15);
+
 #endif
 
+
+
 }

param.h

@@ -289,6 +289,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ZGEMM_DEFAULT_Q 224
 #define XGEMM_DEFAULT_Q 224
 
+#define CGEMM3M_DEFAULT_P 448
+#define ZGEMM3M_DEFAULT_P 224
+#define XGEMM3M_DEFAULT_P 112
+#define CGEMM3M_DEFAULT_Q 224
+#define ZGEMM3M_DEFAULT_Q 224
+#define XGEMM3M_DEFAULT_Q 224
+#define CGEMM3M_DEFAULT_R 12288
+#define ZGEMM3M_DEFAULT_R 12288
+#define XGEMM3M_DEFAULT_R 12288
+
 #define SGEMM_DEFAULT_R sgemm_r
 #define QGEMM_DEFAULT_R qgemm_r
 #define DGEMM_DEFAULT_R dgemm_r
@@ -371,6 +381,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define QGEMM_DEFAULT_Q 224
 #define XGEMM_DEFAULT_Q 224
 
+#define CGEMM3M_DEFAULT_P 448
+#define ZGEMM3M_DEFAULT_P 224
+#define XGEMM3M_DEFAULT_P 112
+#define CGEMM3M_DEFAULT_Q 224
+#define ZGEMM3M_DEFAULT_Q 224
+#define XGEMM3M_DEFAULT_Q 224
+#define CGEMM3M_DEFAULT_R 12288
+#define ZGEMM3M_DEFAULT_R 12288
+#define XGEMM3M_DEFAULT_R 12288
+
 #define SGEMM_DEFAULT_R 12288
 #define QGEMM_DEFAULT_R qgemm_r
 #define DGEMM_DEFAULT_R 12288
@@ -1073,10 +1093,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define GETRF_FACTOR 0.72
 
-#define CGEMM3M_DEFAULT_UNROLL_N 4
-#define CGEMM3M_DEFAULT_UNROLL_M 8
-#define ZGEMM3M_DEFAULT_UNROLL_N 2
-#define ZGEMM3M_DEFAULT_UNROLL_M 8
 #endif
 
 
@@ -1157,6 +1173,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ZGEMM3M_DEFAULT_UNROLL_N 2
 #define ZGEMM3M_DEFAULT_UNROLL_M 8
 
+#define CGEMM3M_DEFAULT_P 448
+#define ZGEMM3M_DEFAULT_P 224
+#define XGEMM3M_DEFAULT_P 112
+#define CGEMM3M_DEFAULT_Q 224
+#define ZGEMM3M_DEFAULT_Q 224
+#define XGEMM3M_DEFAULT_Q 224
+#define CGEMM3M_DEFAULT_R 12288
+#define ZGEMM3M_DEFAULT_R 12288
+#define XGEMM3M_DEFAULT_R 12288
+
+
+
 #define GETRF_FACTOR 0.72
 
 #endif
@@ -1263,6 +1291,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CGEMM3M_DEFAULT_UNROLL_M 8
 #define ZGEMM3M_DEFAULT_UNROLL_N 2
 #define ZGEMM3M_DEFAULT_UNROLL_M 8
+
+
+#define CGEMM3M_DEFAULT_P 448
+#define ZGEMM3M_DEFAULT_P 224
+#define XGEMM3M_DEFAULT_P 112
+#define CGEMM3M_DEFAULT_Q 224
+#define ZGEMM3M_DEFAULT_Q 224
+#define XGEMM3M_DEFAULT_Q 224
+#define CGEMM3M_DEFAULT_R 12288
+#define ZGEMM3M_DEFAULT_R 12288
+#define XGEMM3M_DEFAULT_R 12288
+
 #endif