wip

src/plugins/intel_gpu/include/intel_gpu/primitives/lstm.hpp

@@ -282,8 +282,8 @@ struct lstm_seq : public primitive_base<lstm_seq> {
         if (!cell.empty())
             ret.push_back(cell);
         */
-        //ret.push_back(second_output);
-        //ret.push_back(third_output);
+        ret.push_back(out1_prim_id);
+        ret.push_back(out2_prim_id);
         return ret;
     }
 };

src/plugins/intel_gpu/src/graph/include/lstm_seq_inst.h

@@ -65,8 +65,12 @@ class typed_primitive_inst<lstm_seq> : public typed_primitive_inst_base<lstm_seq
         size_t offset = 8;
         return dep_memory_ptr(offset);
     }
+
+    void update_output_memory() override;
+
+private:
+    void on_execute() override;
 };
 
 using lstm_seq_inst = typed_primitive_inst<lstm_seq>;
-
 }  // namespace cldnn

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

@@ -52,6 +52,26 @@ std::vector<layout> lstm_seq_inst::calc_output_layouts(lstm_seq_node const& node
 
 template std::vector<layout> lstm_seq_inst::calc_output_layouts<ov::PartialShape>(lstm_seq_node const& node, const kernel_impl_params& impl_param);
 
+void lstm_seq_inst::on_execute() {
+    update_output_memory();
+}
+
+void lstm_seq_inst::update_output_memory() {
+    if (!can_be_optimized())
+        return;
+
+    for (size_t i = 1; i < 3; i++) {
+        if (node->get_program().is_new_shape_infer() && input_memory_ptr(i+6) == nullptr)
+            return;
+
+        if (output_memory_ptr(i) != nullptr && _network.get_engine().is_the_same_buffer(output_memory(i), input_memory(i+6)))
+            return;
+
+        _outputs[i] = {_network.get_engine().reinterpret_buffer(input_memory(i+6), _impl_params->get_output_layout(i))};
+    }
+}
+
+
 std::string lstm_seq_inst::to_string(lstm_seq_node const& node) {
     auto desc = node.get_primitive();
     auto node_info = node.desc_to_json();

src/plugins/intel_gpu/src/kernel_selector/cl_kernels/lstm_seq_gpu_bfyx_ref.cl

@@ -139,5 +139,5 @@ KERNEL(lstm_seq)(
         printf("hidden_history[OUTPUT_GET_INDEX_SAFE(b, 0, i, hidden_idx)] is %f\n", hidden_history[OUTPUT_GET_INDEX_SAFE(b, 0, i, hidden_idx)]);
     }
     //printf("cell state for %d is %f \n", OUTPUT2_GET_INDEX_SAFE(b, hidden_idx, 0, 0), cell_state[OUTPUT2_GET_INDEX_SAFE(b, hidden_idx, 0, 0)]);
-    //printf("R is %p B is %p ; %p out0 %p add for out1 for out2 %p batch %d\n", &R, &B, &hidden_history, &hidden_state, &cell_state, b);
+    printf("R is %p B is %p ; hidden history %p hidden state %p cell state %p batch %d\n", &R[0], &B[0], &hidden_history[0], &hidden_state[0], &cell_state[0], b);
 }

src/plugins/intel_gpu/src/kernel_selector/kernels/lstm/lstm_seq_kernel_base.h

@@ -22,12 +22,6 @@ struct lstm_seq_params : public base_params {
     };
 
     lstm_seq_params() : base_params(KernelType::LSTM_SEQ) {}
-
-    DataTensor initial_hidden_state;
-    DataTensor initial_cell_state;
-    DataTensor sequence_lengths;
-    DataTensor WR;
-    DataTensor B;
     bool has_cell = false;
     order_type gate_order = offset_iofz;
     float clip = 0;

src/plugins/intel_gpu/src/plugin/ops/rnn.cpp

@@ -265,9 +265,9 @@ static void CreateLSTMSequenceOp(ProgramBuilder& p, const std::shared_ptr<ov::op
                 cldnn::element_type_to_data_type(mutable_precision_first),
                 cldnn::format::bfyx,
                 tensor_from_dims(op->get_output_shape(1)));
-    cldnn::memory::ptr shared_memory1 = p.get_engine().allocate_memory(out1Layout);
+    cldnn::memory::ptr shared_memory0 = p.get_engine().allocate_memory(out1Layout);
     const cldnn::primitive_id mutable_id_0 = layerName + "_md_write0";
-    const cldnn::mutable_data mutable_prim_0{mutable_id_0, shared_memory1};
+    const cldnn::mutable_data mutable_prim_0{mutable_id_0, shared_memory0};
     p.add_primitive(*op, mutable_prim_0);
 
     inputs.push_back(cldnn::input_info(mutable_id_0));
@@ -278,9 +278,9 @@ static void CreateLSTMSequenceOp(ProgramBuilder& p, const std::shared_ptr<ov::op
                 cldnn::element_type_to_data_type(mutable_precision_second),
                 cldnn::format::bfyx,
                 tensor_from_dims(op->get_output_shape(2)));
-    cldnn::memory::ptr shared_memory2 = p.get_engine().allocate_memory(out2Layout);
+    cldnn::memory::ptr shared_memory1 = p.get_engine().allocate_memory(out2Layout);
     const cldnn::primitive_id mutable_id_1 = layerName + "_md_write1";
-    const cldnn::mutable_data mutable_prim_1{mutable_id_1, shared_memory2};
+    const cldnn::mutable_data mutable_prim_1{mutable_id_1, shared_memory1};
     p.add_primitive(*op, mutable_prim_1);
 
     inputs.push_back(cldnn::input_info(mutable_id_1));
@@ -289,8 +289,8 @@ static void CreateLSTMSequenceOp(ProgramBuilder& p, const std::shared_ptr<ov::op
     inputs[2], inputs[3], inputs[4], inputs[5], cldnn::input_info(bias), inputs[7].pid, inputs[8].pid, \
     "", clip, 0, activations, activation_params, cldnn::lstm_weights_order::fizo, 0);
     p.add_primitive(*op, prim);
-    p.add_primitive(*op, cldnn::mutable_data(lstm_seq_id + ".out1", {cldnn::input_info(lstm_seq_id + ".out0")}, shared_memory1));
-    p.add_primitive(*op, cldnn::mutable_data(lstm_seq_id + ".out2", {cldnn::input_info(lstm_seq_id + ".out0")}, shared_memory2));
+    p.add_primitive(*op, cldnn::mutable_data(lstm_seq_id + ".out1", {cldnn::input_info(lstm_seq_id + ".out0")}, shared_memory0));
+    p.add_primitive(*op, cldnn::mutable_data(lstm_seq_id + ".out2", {cldnn::input_info(lstm_seq_id + ".out0")}, shared_memory1));
 }
 
 REGISTER_FACTORY_IMPL(v4, LSTMCell);