[ARM] [WA] Reconfigure ACL Matmul on each inference (#191)

3.5 squash list:
[FORK][FIX] Reconfigure ACL matmul WA

src/cpu/aarch64/matmul/acl_matmul.cpp

@@ -178,6 +178,14 @@ status_t acl_matmul_t::execute_forward(const exec_ctx_t &ctx) const {
     bool use_dst_acc_for_sum = amp.use_dst_acc_for_sum;
 
     const auto scratchpad = ctx.get_scratchpad_grantor();
+    // [WA] ACL Matmul produces wrong results in case it is not reconfigured on each inference
+    if (do_transC) {
+        acl_obj.gemm.configure(&acl_obj.wei_tensor, &acl_obj.src_tensor,
+            nullptr, &acl_obj.dst_acc_tensor, 1.0f, 0.0f, pd()->amp_.gemm_info);
+    } else {
+        acl_obj.gemm.configure(&acl_obj.src_tensor, &acl_obj.wei_tensor,
+            nullptr, &acl_obj.dst_tensor, 1.0f, 0.0f, pd()->amp_.gemm_info);
+    }
 
     arm_compute::Tensor src_tensor;
     arm_compute::Tensor wei_tensor;