graph: backend: dnnl: fix the dst pointer setting

Author: gyhintel

src/graph/backend/dnnl/kernels/mqa_decomp.cpp

@@ -220,6 +220,9 @@ status_t mqa_decomp_kernel_t<quantized, dt>::execute_impl(
         //reorder3
         auto &sub_dst_user_tid = res->mem_map[mqa_cfg_.sub_dst_user.get()][tid];
 
+        // matmul2
+        auto &sub_mm2_dst_tid = res->mem_map[mqa_cfg_.sub_mm2_dst.get()][tid];
+
         const size_t sub_src1_offset
                 = bo * M1 * K1 * get_mem_dt_size(sub_src1_tid);
         const size_t sub_wei1_offset = (bo * MBI * K1 * N1 + bi * N1)
@@ -239,6 +242,14 @@ status_t mqa_decomp_kernel_t<quantized, dt>::execute_impl(
         sub_dst_user_tid.set_data_handle(
                 dst2_user_pointer + sub_dst_user_offset);
 
+        // If the last reorder is inplace, it means we don't have to do
+        // extra reorder, thus we should set matmul's output to the user's
+        // output directly.
+        if (mqa_cfg_.sub_reorder3.get_inplace()) {
+            sub_mm2_dst_tid.set_data_handle(
+                    dst2_user_pointer + sub_dst_user_offset);
+        }
+
         // in parallel region - these primitives should use single thread.
         mqa_cfg_.sub_reorder0.execute(strm, res->sub_reorder0_args[tid]);
         mqa_cfg_.sub_reorder1.execute(strm, res->sub_reorder1_args[tid]);

src/graph/backend/dnnl/kernels/mqa_decomp_config.hpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2024 Intel Corporation
+* Copyright 2024-2025 Intel Corporation
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -49,6 +49,8 @@ struct mqa_reorder_t {
         return status::success;
     }
 
+    bool get_inplace() const { return is_inplace_; }
+
     status_t execute(const dnnl::stream &astream,
             const std::unordered_map<int, dnnl::memory> &args) const {
         // If the src and dst are the same, we just set the src arg to dst