modified param.h

param.h

@@ -304,9 +304,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #endif
 
 #ifdef PILEDRIVER
-
+#define SNUMOPT         8
-#define SNUMOPT		8
+#define DNUMOPT         4
-#define DNUMOPT		4
 
 #define GEMM_DEFAULT_OFFSET_A  64
 #define GEMM_DEFAULT_OFFSET_B 832
@@ -344,39 +343,42 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define GEMV_UNROLL 8
 #endif
 
-
 #if defined(ARCH_X86_64)
 #define SGEMM_DEFAULT_P 768
-#define DGEMM_DEFAULT_P 384
+#define DGEMM_DEFAULT_P 768
+#define ZGEMM_DEFAULT_P 384
+#define CGEMM_DEFAULT_P 768
 #else
 #define SGEMM_DEFAULT_P 448
-#define DGEMM_DEFAULT_P 224
+#define DGEMM_DEFAULT_P 480
+#define ZGEMM_DEFAULT_P 112 
+#define CGEMM_DEFAULT_P 224
 #endif
 #define QGEMM_DEFAULT_P 112
-#define CGEMM_DEFAULT_P 224
-#define ZGEMM_DEFAULT_P 112 
 #define XGEMM_DEFAULT_P  56
 
 #if defined(ARCH_X86_64)
-#define SGEMM_DEFAULT_Q 168
+#define SGEMM_DEFAULT_Q 192
 #define DGEMM_DEFAULT_Q 168
+#define ZGEMM_DEFAULT_Q 168
+#define CGEMM_DEFAULT_Q 168
 #else
 #define SGEMM_DEFAULT_Q 224
 #define DGEMM_DEFAULT_Q 224
+#define ZGEMM_DEFAULT_Q 224
+#define CGEMM_DEFAULT_Q 224
 #endif
 #define QGEMM_DEFAULT_Q 224
-#define CGEMM_DEFAULT_Q 224
-#define ZGEMM_DEFAULT_Q 224
 #define XGEMM_DEFAULT_Q 224
 
-#define SGEMM_DEFAULT_R sgemm_r
+#define SGEMM_DEFAULT_R 12288
 #define QGEMM_DEFAULT_R qgemm_r
-#define DGEMM_DEFAULT_R dgemm_r
+#define DGEMM_DEFAULT_R 12288
 #define CGEMM_DEFAULT_R cgemm_r
 #define ZGEMM_DEFAULT_R zgemm_r
 #define XGEMM_DEFAULT_R xgemm_r
 
-#define SYMV_P	16
+#define SYMV_P  16
 #define HAVE_EXCLUSIVE_CACHE
 
 #define GEMM_THREAD gemm_thread_mn
@@ -1150,6 +1152,107 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #endif
 
+#ifdef HASWELL
+
+#define SNUMOPT         8
+#define DNUMOPT         4
+
+#define GEMM_DEFAULT_OFFSET_A     0
+#define GEMM_DEFAULT_OFFSET_B     0
+#define GEMM_DEFAULT_ALIGN 0x03fffUL
+
+#define SYMV_P  8
+
+#define SWITCH_RATIO    4
+
+#ifdef ARCH_X86
+
+#define SGEMM_DEFAULT_UNROLL_M 4
+#define DGEMM_DEFAULT_UNROLL_M 2
+#define QGEMM_DEFAULT_UNROLL_M 2
+#define CGEMM_DEFAULT_UNROLL_M 2
+#define ZGEMM_DEFAULT_UNROLL_M 1
+#define XGEMM_DEFAULT_UNROLL_M 1
+
+#define SGEMM_DEFAULT_UNROLL_N 4
+#define DGEMM_DEFAULT_UNROLL_N 4
+#define QGEMM_DEFAULT_UNROLL_N 2
+#define CGEMM_DEFAULT_UNROLL_N 2
+#define ZGEMM_DEFAULT_UNROLL_N 2
+#define XGEMM_DEFAULT_UNROLL_N 1
+
+#else
+
+#define SGEMM_DEFAULT_UNROLL_M 16
+#define DGEMM_DEFAULT_UNROLL_M 4
+#define QGEMM_DEFAULT_UNROLL_M 2
+#define CGEMM_DEFAULT_UNROLL_M 8
+#define ZGEMM_DEFAULT_UNROLL_M 4
+#define XGEMM_DEFAULT_UNROLL_M 1
+
+#define SGEMM_DEFAULT_UNROLL_N 4
+#define DGEMM_DEFAULT_UNROLL_N 4
+#define QGEMM_DEFAULT_UNROLL_N 2
+#define CGEMM_DEFAULT_UNROLL_N 2
+#define ZGEMM_DEFAULT_UNROLL_N 2
+#define XGEMM_DEFAULT_UNROLL_N 1
+
+#endif
+
+#ifdef ARCH_X86
+
+#define SGEMM_DEFAULT_P 512
+#define SGEMM_DEFAULT_R sgemm_r
+#define DGEMM_DEFAULT_P 512
+#define DGEMM_DEFAULT_R dgemm_r
+#define QGEMM_DEFAULT_P 504
+#define QGEMM_DEFAULT_R qgemm_r
+#define CGEMM_DEFAULT_P 128
+#define CGEMM_DEFAULT_R 1024
+#define ZGEMM_DEFAULT_P 512
+#define ZGEMM_DEFAULT_R zgemm_r
+#define XGEMM_DEFAULT_P 252
+#define XGEMM_DEFAULT_R xgemm_r
+#define SGEMM_DEFAULT_Q 256
+#define DGEMM_DEFAULT_Q 256
+#define QGEMM_DEFAULT_Q 128
+#define CGEMM_DEFAULT_Q 256 
+#define ZGEMM_DEFAULT_Q 192
+#define XGEMM_DEFAULT_Q 128
+
+#else
+
+#define SGEMM_DEFAULT_P 768
+#define DGEMM_DEFAULT_P 512
+#define CGEMM_DEFAULT_P 384
+#define ZGEMM_DEFAULT_P 256
+
+#define SGEMM_DEFAULT_Q 384
+#define DGEMM_DEFAULT_Q 256
+#define CGEMM_DEFAULT_Q 192 
+#define ZGEMM_DEFAULT_Q 128
+
+#define SGEMM_DEFAULT_R sgemm_r
+#define DGEMM_DEFAULT_R 13824
+#define CGEMM_DEFAULT_R cgemm_r
+#define ZGEMM_DEFAULT_R zgemm_r
+
+#define QGEMM_DEFAULT_Q 128
+#define QGEMM_DEFAULT_P 504
+#define QGEMM_DEFAULT_R qgemm_r
+#define XGEMM_DEFAULT_P 252
+#define XGEMM_DEFAULT_R xgemm_r
+#define XGEMM_DEFAULT_Q 128
+
+#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
+
+
+#endif
+
 
 
 #ifdef ATOM
@@ -1793,6 +1896,129 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SYMV_P	16
 #endif
 
+
+#ifdef ARMV7
+#define SNUMOPT		2
+#define DNUMOPT		2
+
+#define GEMM_DEFAULT_OFFSET_A 0
+#define GEMM_DEFAULT_OFFSET_B 0
+#define GEMM_DEFAULT_ALIGN 0x03fffUL
+
+#define SGEMM_DEFAULT_UNROLL_M  4
+#define SGEMM_DEFAULT_UNROLL_N  4
+
+#define DGEMM_DEFAULT_UNROLL_M  4
+#define DGEMM_DEFAULT_UNROLL_N  4
+
+#define CGEMM_DEFAULT_UNROLL_M  2
+#define CGEMM_DEFAULT_UNROLL_N  2
+
+#define ZGEMM_DEFAULT_UNROLL_M  2
+#define ZGEMM_DEFAULT_UNROLL_N  2
+
+#define SGEMM_DEFAULT_P	128
+#define DGEMM_DEFAULT_P	128
+#define CGEMM_DEFAULT_P 96
+#define ZGEMM_DEFAULT_P 64
+
+#define SGEMM_DEFAULT_Q 240
+#define DGEMM_DEFAULT_Q 120
+#define CGEMM_DEFAULT_Q 120
+#define ZGEMM_DEFAULT_Q 120
+
+#define SGEMM_DEFAULT_R 12288
+#define DGEMM_DEFAULT_R 8192
+#define CGEMM_DEFAULT_R 4096
+#define ZGEMM_DEFAULT_R 4096
+
+
+
+#define SYMV_P	16
+#endif
+
+
+#if defined(ARMV6)
+#define SNUMOPT		2
+#define DNUMOPT		2
+
+#define GEMM_DEFAULT_OFFSET_A 0
+#define GEMM_DEFAULT_OFFSET_B 0
+#define GEMM_DEFAULT_ALIGN 0x03fffUL
+
+#define SGEMM_DEFAULT_UNROLL_M  4
+#define SGEMM_DEFAULT_UNROLL_N  2
+
+#define DGEMM_DEFAULT_UNROLL_M  4
+#define DGEMM_DEFAULT_UNROLL_N  2
+
+#define CGEMM_DEFAULT_UNROLL_M  2
+#define CGEMM_DEFAULT_UNROLL_N  2
+
+#define ZGEMM_DEFAULT_UNROLL_M  2
+#define ZGEMM_DEFAULT_UNROLL_N  2
+
+#define SGEMM_DEFAULT_P	128
+#define DGEMM_DEFAULT_P	128
+#define CGEMM_DEFAULT_P 96
+#define ZGEMM_DEFAULT_P 64
+
+#define SGEMM_DEFAULT_Q 240
+#define DGEMM_DEFAULT_Q 120
+#define CGEMM_DEFAULT_Q 120
+#define ZGEMM_DEFAULT_Q 120
+
+#define SGEMM_DEFAULT_R 12288
+#define DGEMM_DEFAULT_R 8192
+#define CGEMM_DEFAULT_R 4096
+#define ZGEMM_DEFAULT_R 4096
+
+
+#define SYMV_P	16
+#endif
+
+#if defined(ARMV8)
+#define SNUMOPT		2
+#define DNUMOPT		2
+
+#define GEMM_DEFAULT_OFFSET_A 0
+#define GEMM_DEFAULT_OFFSET_B 0
+#define GEMM_DEFAULT_ALIGN 0x03fffUL
+
+#define SGEMM_DEFAULT_UNROLL_M  2
+#define SGEMM_DEFAULT_UNROLL_N  2
+
+#define DGEMM_DEFAULT_UNROLL_M  2
+#define DGEMM_DEFAULT_UNROLL_N  2
+
+#define CGEMM_DEFAULT_UNROLL_M  2
+#define CGEMM_DEFAULT_UNROLL_N  2
+
+#define ZGEMM_DEFAULT_UNROLL_M  2
+#define ZGEMM_DEFAULT_UNROLL_N  2
+
+#define SGEMM_DEFAULT_P	128
+#define DGEMM_DEFAULT_P	128
+#define CGEMM_DEFAULT_P 96
+#define ZGEMM_DEFAULT_P 64
+
+#define SGEMM_DEFAULT_Q 240
+#define DGEMM_DEFAULT_Q 120
+#define CGEMM_DEFAULT_Q 120
+#define ZGEMM_DEFAULT_Q 120
+
+#define SGEMM_DEFAULT_R 12288
+#define DGEMM_DEFAULT_R 8192
+#define CGEMM_DEFAULT_R 4096
+#define ZGEMM_DEFAULT_R 4096
+
+
+#define SYMV_P	16
+#endif
+
+
+
+
 #ifdef GENERIC
 
 #define SNUMOPT		2