From 52526ec567f73dcd30c6f6073a3c068132fc066d Mon Sep 17 00:00:00 2001
From: Steve Yoo <steve.yoo@intel.com>
Date: Sun, 2 Feb 2025 21:26:01 +0000
Subject: [PATCH 1/3] [GPU] Insert scratchpad memory in concat when it's null

---
 .../src/graph/impls/onednn/concatenation_onednn.cpp         | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp b/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp
index 4e18e07f699174..2c439d98669952 100644
--- a/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp
@@ -27,6 +27,12 @@ struct concatenation_onednn : typed_primitive_onednn_impl<concatenation, dnnl::c
     std::unordered_map<int, dnnl::memory> get_arguments(concatenation_inst& instance) const override {
         std::unordered_map<int, dnnl::memory> args;
 
+        if (_scratchpad_md.get_size() != 0) {
+            // onednn primitive can have only 1 scratchpad memory.
+            auto scratchpad = instance.get_intermediates_memories()[0];
+            args.insert({DNNL_ARG_SCRATCHPAD, scratchpad->get_onednn_memory(_scratchpad_md, 0)});
+        }
+
         int input_idx = DNNL_ARG_MULTIPLE_SRC;
         for (size_t i = 0; i < instance.inputs_memory_count(); i++) {
             auto& input = instance.input_memory(i);

From 80269e89e324389866ad5b3d1f71ec85fa62dddb Mon Sep 17 00:00:00 2001
From: Steve Yoo <steve.yoo@intel.com>
Date: Mon, 3 Feb 2025 14:04:07 +0000
Subject: [PATCH 2/3] [GPU] Change patch code location

---
 .../src/graph/impls/onednn/concatenation_onednn.cpp  | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp b/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp
index 2c439d98669952..77d4aaed0ffc7f 100644
--- a/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp
@@ -27,12 +27,6 @@ struct concatenation_onednn : typed_primitive_onednn_impl<concatenation, dnnl::c
     std::unordered_map<int, dnnl::memory> get_arguments(concatenation_inst& instance) const override {
         std::unordered_map<int, dnnl::memory> args;
 
-        if (_scratchpad_md.get_size() != 0) {
-            // onednn primitive can have only 1 scratchpad memory.
-            auto scratchpad = instance.get_intermediates_memories()[0];
-            args.insert({DNNL_ARG_SCRATCHPAD, scratchpad->get_onednn_memory(_scratchpad_md, 0)});
-        }
-
         int input_idx = DNNL_ARG_MULTIPLE_SRC;
         for (size_t i = 0; i < instance.inputs_memory_count(); i++) {
             auto& input = instance.input_memory(i);
@@ -46,6 +40,12 @@ struct concatenation_onednn : typed_primitive_onednn_impl<concatenation, dnnl::c
             args.insert({DNNL_ARG_DST, output.get_onednn_memory(_pd.dnnl::primitive_desc_base::dst_desc(0), offset)});
         }
 
+        if (_scratchpad_md.get_size() != 0) {
+            // onednn primitive can have only 1 scratchpad memory.
+            auto scratchpad = instance.get_intermediates_memories()[0];
+            args.insert({DNNL_ARG_SCRATCHPAD, scratchpad->get_onednn_memory(_scratchpad_md, 0)});
+        }
+
         configure_post_ops_arguments(instance, args);
 
         return args;

From 35bb66edc1ed500458594837339d6e2c52df79da Mon Sep 17 00:00:00 2001
From: Steve Yoo <steve.yoo@intel.com>
Date: Tue, 4 Feb 2025 12:37:53 +0000
Subject: [PATCH 3/3] [GPU] Add comments about the patch

---
 .../intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp   | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp b/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp
index 77d4aaed0ffc7f..ce55cb4ebc7d1f 100644
--- a/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp
+++ b/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp
@@ -40,6 +40,8 @@ struct concatenation_onednn : typed_primitive_onednn_impl<concatenation, dnnl::c
             args.insert({DNNL_ARG_DST, output.get_onednn_memory(_pd.dnnl::primitive_desc_base::dst_desc(0), offset)});
         }
 
+        // When scratchpad is available, it is used;
+        // for example, internal reorder is used.
         if (_scratchpad_md.get_size() != 0) {
             // onednn primitive can have only 1 scratchpad memory.
             auto scratchpad = instance.get_intermediates_memories()[0];