optimize AVX2 SGEMM

Author: wjc404

param.h

@@ -625,7 +625,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #else
 
-#define SGEMM_DEFAULT_UNROLL_M 16
+#define SGEMM_DEFAULT_UNROLL_M 8
 #define DGEMM_DEFAULT_UNROLL_M 4
 #define QGEMM_DEFAULT_UNROLL_M 2
 #define CGEMM_DEFAULT_UNROLL_M 8
@@ -666,7 +666,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #else
 
-#define SGEMM_DEFAULT_P 768
+#define SGEMM_DEFAULT_P 320
 #define DGEMM_DEFAULT_P 512
 #define CGEMM_DEFAULT_P 256
 #define ZGEMM_DEFAULT_P 192
@@ -675,7 +675,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SGEMM_DEFAULT_Q 320
 #define DGEMM_DEFAULT_Q 128
 #else
-#define SGEMM_DEFAULT_Q 384
+#define SGEMM_DEFAULT_Q 320
 #define DGEMM_DEFAULT_Q 256
 #endif
 #define CGEMM_DEFAULT_Q 256
@@ -1528,7 +1528,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #else
 
-#define SGEMM_DEFAULT_UNROLL_M 16
+#define SGEMM_DEFAULT_UNROLL_M 8
 #define DGEMM_DEFAULT_UNROLL_M 4
 #define QGEMM_DEFAULT_UNROLL_M 2
 #define CGEMM_DEFAULT_UNROLL_M 8
@@ -1569,7 +1569,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #else
 
-#define SGEMM_DEFAULT_P 768
+#define SGEMM_DEFAULT_P 320
 #define DGEMM_DEFAULT_P 512
 #define CGEMM_DEFAULT_P 256
 #define ZGEMM_DEFAULT_P 192
@@ -1578,7 +1578,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SGEMM_DEFAULT_Q 320
 #define DGEMM_DEFAULT_Q 128
 #else
-#define SGEMM_DEFAULT_Q 384
+#define SGEMM_DEFAULT_Q 320
 #define DGEMM_DEFAULT_Q 256
 #endif
 #define CGEMM_DEFAULT_Q 256