Fix mkldnn_matmul error on AArch64 (pytorch#114851)

Fixes pytorch#110149

Cherry pick pytorch#110150. This is a bug fix against 2.1 release

Author: imzhuhl

aten/src/ATen/native/LinearAlgebra.cpp

@@ -1483,12 +1483,14 @@ static void addmm_impl_cpu_(
   // it is faster to call oneDNN matrix multiplication primitive with RHS*LHS
   // that will call then into Arm® Compute Library (ACL) GEMM kernel and also
   // additionally have support for running kernel with BF16 instructions
-  bool apply_heur = apply_mkldnn_matmul_heur(b.sizes()[0], b.sizes()[1], a.sizes()[1]);
-  if (apply_heur && transpose_a && !transpose_b && result.scalar_type() == at::ScalarType::Float) {
-      mkldnn_matmul(b, a, c, beta.to<float>(), alpha.to<float>());
-      // We have dispatched to ACL GEMM for single precision float
-      // so do not need to dispatch to BLAS GEMM below
-      dispatched = true;
+  if (transpose_c) {
+    bool apply_heur = apply_mkldnn_matmul_heur(b.sizes()[0], b.sizes()[1], a.sizes()[1]);
+    if (apply_heur && transpose_a && !transpose_b && result.scalar_type() == at::ScalarType::Float) {
+        mkldnn_matmul(b, a, c, beta.to<float>(), alpha.to<float>());
+        // We have dispatched to ACL GEMM for single precision float
+        // so do not need to dispatch to BLAS GEMM below
+        dispatched = true;
+    }
   }
 #endif