-int sgemm_small_kernel_a1b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
-int sgemm_small_kernel_a1b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
-int sgemm_small_kernel_a1b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
-int sgemm_small_kernel_a1b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
-
-int dgemm_small_kernel_a1b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
-int dgemm_small_kernel_a1b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
-int dgemm_small_kernel_a1b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
-int dgemm_small_kernel_a1b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+int sgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+int sgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+int sgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+int sgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+
+int dgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+int dgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+int dgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+int dgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);