one less primitive in post optimize weights, deleted legacy output

src/plugins/intel_gpu/src/graph/graph_optimizer/post_optimize_weights.cpp

@@ -11,7 +11,6 @@
 #include "fully_connected_inst.h"
 #include "lstm_seq_inst.h"
 #include "intel_gpu/runtime/format.hpp"
-#include "intel_gpu/primitives/mutable_data.hpp"
 #include "permute_inst.h"
 #include "crop_inst.h"
 #ifdef ENABLE_ONEDNN_FOR_GPU
@@ -180,20 +179,20 @@ void post_optimize_weights::add_lstm_weights_reorder(primitive_id input_id, std:
 
     auto& crop0_node = get_crop_node(0);
     auto& crop1_node = get_crop_node(1);
-    auto& crop2_node = get_crop_node(2);
-    auto& crop3_node = get_crop_node(3);
-    std::vector<input_info> con_input{input_info(crop_id_b + "1"), input_info(crop_id_b + "0"), input_info(crop_id_b + "2"), input_info(crop_id_b + "3")};
+    auto crop2_id = primitive_id(crop_id_b + std::to_string(2));
+    auto crop2_prim = std::make_shared<cldnn::crop>(crop2_id, reorder_id,  cldnn::tensor{dir_num, static_cast<int>(2*hiddenSize), 1, size_third},
+        cldnn::tensor{0, static_cast<int>(2*hiddenSize), 0, 0});
+    auto& crop2_node = p.get_or_create(crop2_prim);
+    std::vector<input_info> con_input{input_info(crop_id_b + "1"), input_info(crop_id_b + "0"), input_info(crop_id_b + "2")};
     cldnn::primitive_id concat_id{input_id + "cont"};
     auto con = std::make_shared<cldnn::concatenation>(concat_id, con_input, 1);
     auto& con_node = p.get_or_create(con);
     p.add_intermediate(con_node, node, prev, true);
     p.add_intermediate(crop1_node, con_node, prev, true);
     p.add_connection(prev, crop0_node, 0);
     p.add_connection(prev, crop2_node, 0);
-    p.add_connection(prev, crop3_node, 0);
     p.add_connection(crop0_node, con_node, 0);
     p.add_connection(crop2_node, con_node, 0);
-    p.add_connection(crop3_node, con_node, 0);
     std::string permute_id = input_id + "_perx";
     std::vector<uint16_t> ord{0, 2, 1};
     auto permute = std::make_shared<cldnn::permute>(permute_id, input_info{concat_id}, ord);
@@ -208,7 +207,6 @@ void post_optimize_weights::add_lstm_weights_reorder(primitive_id input_id, std:
     set_implementation_and_output(crop1_node);
     set_implementation_and_output(crop0_node);
     set_implementation_and_output(crop2_node);
-    set_implementation_and_output(crop3_node);
     set_implementation_and_output(con_node);
     set_implementation_and_output(permute_node);
 }
@@ -227,20 +225,20 @@ void post_optimize_weights::add_lstm_bias_reorder(primitive_id input_id, std::sh
     };
     auto& crop0_node = get_crop_node(0);
     auto& crop1_node = get_crop_node(1);
-    auto& crop2_node = get_crop_node(2);
-    auto& crop3_node = get_crop_node(3);
-    std::vector<input_info> con_input{input_info(crop1_node.id()), input_info(crop0_node.id()), input_info(crop2_node.id()), input_info(crop3_node.id())};
+    auto crop2_id = primitive_id(crop_id_b + std::to_string(2));
+    auto crop2_prim = std::make_shared<cldnn::crop>(crop2_id, input_id,  cldnn::tensor{dir_num, static_cast<int>(2*hiddenSize), 1, 1},
+        cldnn::tensor{0, static_cast<int>(2*hiddenSize), 0, 0});
+    auto& crop2_node = p.get_or_create(crop2_prim);
+    std::vector<input_info> con_input{input_info(crop1_node.id()), input_info(crop0_node.id()), input_info(crop2_node.id())};
     cldnn::primitive_id concat_id{input_id + "concat"};
     auto con = std::make_shared<cldnn::concatenation>(concat_id, con_input, 1);
     auto& con_node = p.get_or_create(con);
     p.add_intermediate(con_node, node, prev, true);
     p.add_intermediate(crop1_node, con_node, prev, true);
     p.add_connection(prev, crop0_node, 0);
     p.add_connection(prev, crop2_node, 0);
-    p.add_connection(prev, crop3_node, 0);
     p.add_connection(crop0_node, con_node, 0);
     p.add_connection(crop2_node, con_node, 0);
-    p.add_connection(crop3_node, con_node, 0);
     auto set_implementation_and_output = [this, &p](program_node& node) {
         node.get_output_layout(false);
         select_implementation(p, node);
@@ -250,7 +248,6 @@ void post_optimize_weights::add_lstm_bias_reorder(primitive_id input_id, std::sh
     set_implementation_and_output(crop0_node);
     set_implementation_and_output(crop1_node);
     set_implementation_and_output(crop2_node);
-    set_implementation_and_output(crop3_node);
     set_implementation_and_output(con_node);
 }
 

src/plugins/intel_gpu/src/graph/impls/ocl/lstm_cell.cpp

@@ -28,17 +28,13 @@ struct lstm_cell_impl : typed_primitive_impl_ocl<lstm_cell> {
 protected:
     kernel_arguments_data get_arguments(const typed_primitive_inst<lstm_cell>& instance) const override {
         kernel_arguments_data args;
-        size_t op_input_size = 6;
-        for (size_t i = 0; i < op_input_size; i++) {
+        for (size_t i = 0; i < instance.inputs_memory_count(); i++) {
             args.inputs.push_back(instance.input_memory_ptr(i));
         }
 
         for (size_t i = 0; i < instance.outputs_memory_count(); i++) {
             args.outputs.push_back(instance.output_memory_ptr(i));
         }
-        for (size_t i = op_input_size; i < instance.inputs_memory_count(); i++) {
-            args.outputs.push_back(instance.dep_memory_ptr(i));
-        }
         return args;
     }
 
@@ -68,8 +64,6 @@ struct lstm_cell_impl : typed_primitive_impl_ocl<lstm_cell> {
         params.SetOffsetOrder(static_cast<int32_t>(primitive->offset_order));
         params.clip = primitive->clip;
         params.direction = primitive->direction;
-        //Legacy multi-output
-        params.outputs.push_back(convert_data_tensor(impl_param.input_layouts[1]));
 
         return params;
     }

src/plugins/intel_gpu/src/graph/impls/ocl/rnn_seq.cpp

@@ -29,17 +29,13 @@ struct rnn_seq_impl : typed_primitive_impl_ocl<lstm_seq> {
 protected:
     kernel_arguments_data get_arguments(const typed_primitive_inst<lstm_seq>& instance) const override {
         kernel_arguments_data args;
-        size_t op_input_size = 6 + (instance.has_cell() ? 1 : 0);
-        for (size_t i = 0; i < op_input_size; i++) {
+        for (size_t i = 0; i < instance.inputs_memory_count(); i++) {
             args.inputs.push_back(instance.input_memory_ptr(i));
         }
 
         for (size_t i = 0; i < instance.outputs_memory_count(); i++) {
             args.outputs.push_back(instance.output_memory_ptr(i));
         }
-        for (size_t i = op_input_size; i < instance.inputs_memory_count(); i++) {
-            args.outputs.push_back(instance.dep_memory_ptr(i));
-        }
         return args;
     }
 
@@ -70,11 +66,6 @@ struct rnn_seq_impl : typed_primitive_impl_ocl<lstm_seq> {
         params.SetOffsetOrder(static_cast<int32_t>(primitive->offset_order));
         params.clip = primitive->clip;
         params.direction = primitive->direction;
-        //Legacy multi-output
-        params.outputs.push_back(convert_data_tensor(impl_param.input_layouts[1]));
-        if (!primitive->initial_cell_state.pid.empty()) {
-            params.outputs.push_back(convert_data_tensor(impl_param.input_layouts[1]));
-        }
         return params;
     }
 

src/plugins/intel_gpu/src/graph/impls/onednn/lstm_seq_onednn.hpp

@@ -25,8 +25,6 @@ struct LSTMSeqImplementationManager : public ImplementationManager {
         if (info.arch == gpu_arch::unknown)
             return false;
 
-        const auto& lstm_seq_node = node.as<lstm_seq>();
-        const auto& lstm_seq_prim = lstm_seq_node.get_primitive();
         auto in0_dt = node.get_input_layout(0).data_type;
         auto in1_dt = node.get_input_layout(1).data_type;
         auto in2_dt = node.get_input_layout(2).data_type;

src/plugins/intel_gpu/src/graph/lstm_seq.cpp

@@ -1,4 +1,4 @@
-// Copyright (C) 2018-2024 Intel Corporation
+// Copyright (C) 2024 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 //
 #include "lstm_seq_inst.h"