xgemm_p = ((xgemm_p + XGEMM_UNROLL_M - 1)/XGEMM_UNROLL_M) * XGEMM_UNROLL_M;
#endif
+ sbgemm_r = (((BUFFER_SIZE - ((SBGEMM_P * SBGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SBGEMM_Q * 4)) - 15) & ~15;
sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q * 4)) - 15) & ~15;
dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q * 8)) - 15) & ~15;
cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q * 8)) - 15) & ~15;
#endif
#define USE_SGEMM_KERNEL_DIRECT 1
+#undef SBGEMM_DEFAULT_UNROLL_N
+#undef SBGEMM_DEFAULT_UNROLL_M
+#undef SBGEMM_DEFAULT_P
+#undef SBGEMM_DEFAULT_R
+#undef SBGEMM_DEFAULT_Q
+#define SBGEMM_DEFAULT_UNROLL_N 4
+#define SBGEMM_DEFAULT_UNROLL_M 16
+#define SBGEMM_DEFAULT_P 384
+#define SBGEMM_DEFAULT_Q 768
+#define SBGEMM_DEFAULT_R sbgemm_r
+
#ifdef ARCH_X86
#define SGEMM_DEFAULT_UNROLL_M 4