COMPMID-1124: Validate CLLSTM

-Enables cell-to-input weights when !cifg and peephole
-Makes projection bias conditional

Change-Id: Iee866db9f5d8479c2dfd95d74a2d42492bf07a8d
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140543
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Les Bell <les.bell@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

4 files changed, 236 insertions, 126 deletions

diff --git a/src/core/CL/cl_kernels/copy_tensor.cl b/src/core/CL/cl_kernels/copy_tensor.cl
index 4b37decf21..930a6762a8 100644
--- a/src/core/CL/cl_kernels/copy_tensor.cl
+++ b/src/core/CL/cl_kernels/copy_tensor.cl
@@ -25,24 +25,35 @@
 
 /** Performs a copy of input tensor to the output tensor.
  *
- * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8.
- * @param[in]  in_stride_x                       Stride of the source image in X dimension (in bytes)
- * @param[in]  in_step_x                         in_stride_x * number of elements along X processed per work item (in bytes)
- * @param[in]  in_offset_first_element_in_bytes  Offset of the first element in the source image
- * @param[out] out_ptr                           Pointer to the destination image. Supported data types: U8.
- * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
- * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per work item (in bytes)
- * @param[in]  out_offset_first_element_in_bytes Offset of the first element in the destination image
+ * @param[in]  in_ptr                            Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  in_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  in_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  in_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  in_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  in_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  in_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out] out_ptr                           Pointer to the destination tensor. Supported data types: same as @p in_ptr
+ * @param[in]  out_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  out_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  out_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  out_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  out_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination tensor
  */
 __kernel void copy_tensor(
-    VECTOR_DECLARATION(in),
-    VECTOR_DECLARATION(out))
+    TENSOR3D_DECLARATION(in),
+    TENSOR3D_DECLARATION(out))
 {
-    Vector in  = CONVERT_TO_VECTOR_STRUCT(in);
-    Vector out = CONVERT_TO_VECTOR_STRUCT(out);
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(in);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
 
-    VEC_DATA_TYPE(DATA_TYPE, 16)
-    data = vload16(0, (__global DATA_TYPE *)in.ptr);
+    // Load data
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr);
 
-    vstore16(data, 0, (__global DATA_TYPE *)out.ptr);
+    // Store result
+    VSTORE(VEC_SIZE)
+    (data, 0, (__global DATA_TYPE *)out.ptr);
 }
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLCopyKernel.cpp b/src/core/CL/kernels/CLCopyKernel.cpp
index 4f00ef9eef..1fc8b5bfbe 100644
--- a/src/core/CL/kernels/CLCopyKernel.cpp
+++ b/src/core/CL/kernels/CLCopyKernel.cpp
@@ -33,10 +33,44 @@
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
 
-#include <algorithm>
-
 using namespace arm_compute;
 
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+
+    // Validate output if initialized
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(input->tensor_shape(), output->tensor_shape());
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, *input);
+
+    // Configure window
+    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
+
+    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+
+    bool window_changed = update_window_and_padding(win, input_access, output_access);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
 CLCopyKernel::CLCopyKernel()
     : _input(nullptr), _output(nullptr)
 {
@@ -44,28 +78,32 @@ CLCopyKernel::CLCopyKernel()
 
 void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(input->info()->tensor_shape(), output->info()->tensor_shape());
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
 
     _input  = input;
     _output = output;
 
+    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+
     // Create kernel
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("copy_tensor", build_opts.options()));
 
-    // Configure window
-    constexpr unsigned int num_elems_processed_per_iteration = 16;
-
-    Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
-
-    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure(win_config.second);
+}
 
-    update_window_and_padding(win, input_access, output_access);
+Status CLCopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
 
-    ICLKernel::configure(win);
+    return Status{};
 }
 
 void CLCopyKernel::run(const Window &window, cl::CommandQueue &queue)
@@ -73,15 +111,15 @@ void CLCopyKernel::run(const Window &window, cl::CommandQueue &queue)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 
-    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimX);
-    Window slice     = collapsed.first_slice_window_1D();
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+    Window slice     = collapsed.first_slice_window_3D();
 
     do
     {
         unsigned int idx = 0;
-        add_1D_tensor_argument(idx, _input, slice);
-        add_1D_tensor_argument(idx, _output, slice);
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice);
         enqueue(queue, *this, slice);
     }
-    while(collapsed.slide_window_slice_1D(slice));
+    while(collapsed.slide_window_slice_3D(slice));
 }
diff --git a/src/runtime/CL/functions/CLCopy.cpp b/src/runtime/CL/functions/CLCopy.cpp
index 3442e3781f..d1b7926812 100644
--- a/src/runtime/CL/functions/CLCopy.cpp
+++ b/src/runtime/CL/functions/CLCopy.cpp
@@ -41,3 +41,8 @@ void CLCopy::configure(ICLTensor *input, ICLTensor *output)
     k->configure(input, output);
     _kernel = std::move(k);
 }
+
+Status CLCopy::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output)
+{
+    return CLCopyKernel::validate(input, output);
+}
diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp
index d384400ed3..3458135799 100644
--- a/src/runtime/CL/functions/CLLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayer.cpp
@@ -45,19 +45,27 @@ CLLSTMLayer::CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)
       _accum_output1(), _accum_output2(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(), _fully_connected_output_state(), _gemm_output_state(), _accum_output_state(),
       _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _input_gate_out1(), _input_gate_out2(), _input_gate_out3(), _input_gate_out4(), _input_gate_out5(),
       _forget_gate_out1(), _forget_gate_out2(), _forget_gate_out3(), _forget_gate_out4(), _forget_gate_out5(), _cell_state_out1(), _cell_state_out2(), _cell_state_out3(), _cell_state_out4(),
-      _cell_state_out5(), _output1(), _output2(), _output3(), _output4(), _output5(), _cell_state_activation(), _output_projection1(), _ones(), _run_peephole_opt(false), _run_cifg_opt(false),
+      _cell_state_out5(), _output1(), _output2(), _output3(), _output4(), _output5(), _cell_state_activation(), _output_state1(), _ones(), _run_peephole_opt(false), _run_cifg_opt(false),
       _perform_cell_clipping(false), _has_projection_weights(false), _perform_projection_clipping(false)
 {
 }
 
-void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
+void CLLSTMLayer::configure(const ICLTensor *input,
+                            const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
                             const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
                             const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
-                            ICLTensor *output_state, ICLTensor *cell_state, ICLTensor *scratch_buffer, ICLTensor *output, const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info,
-                            float cell_threshold, float projection_threshold)
+                            const ICLTensor *output_state_in, const ICLTensor *cell_state_in,
+                            ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
+                            const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
-                                 forget_gate_bias, cell_bias, output_gate_bias, output_state, cell_state);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input,
+                                 input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
+                                 recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
+                                 forget_gate_bias, cell_bias, output_gate_bias,
+                                 output_state_in, cell_state_in,
+                                 scratch_buffer, output_state_out, cell_state_out, output);
+
+    // Set lstm parameters
     LSTMParams<ITensorInfo> lstm_params_info;
     if(lstm_params.has_peephole_opt())
     {
@@ -65,36 +73,41 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
     }
     if(lstm_params.has_projection())
     {
-        lstm_params_info.set_projection_params(lstm_params.projection_weights()->info(), lstm_params.projection_bias()->info());
+        lstm_params_info.set_projection_params(lstm_params.projection_weights()->info(),
+                                               lstm_params.projection_bias() != nullptr ? lstm_params.projection_bias()->info() : nullptr);
     }
     if(!lstm_params.has_cifg_opt())
     {
+        const ITensorInfo *cell_to_input_weights_info = (lstm_params.has_peephole_opt()) ? lstm_params.cell_to_input_weights()->info() : nullptr;
         lstm_params_info.set_cifg_params(lstm_params.input_to_input_weights()->info(), lstm_params.recurrent_to_input_weights()->info(),
-                                         lstm_params.cell_to_input_weights()->info(), lstm_params.input_gate_bias()->info());
+                                         cell_to_input_weights_info, lstm_params.input_gate_bias()->info());
     }
+
+    // Validate
     ARM_COMPUTE_ERROR_THROW_ON(CLLSTMLayer::validate(input->info(), input_to_forget_weights->info(),
                                                      input_to_cell_weights->info(), input_to_output_weights->info(),
                                                      recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(),
                                                      forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(),
-                                                     output_state->info(), cell_state->info(), scratch_buffer->info(), output->info(), lstm_params_info,
-                                                     activation_info, cell_threshold, projection_threshold));
+                                                     output_state_in->info(), cell_state_in->info(),
+                                                     scratch_buffer->info(), output_state_out->info(), cell_state_out->info(), output->info(),
+                                                     lstm_params_info, activation_info, cell_threshold, projection_threshold));
 
-    const TensorShape cell_state_shape = cell_state->info()->tensor_shape();
+    const TensorShape cell_state_shape = cell_state_in->info()->tensor_shape();
 
+    // Configure block that calculates the forget gate
+    // forget_gate = Activation(input * input_to_forget_weights + output_state_in * recurrent_to_forget_weights + PixelWiseMul(cell_state, cell_to_forget_weights) + forget_gate_bias)
     TensorShape forget_gate1_shape = compute_transposed_shape(*recurrent_to_output_weights->info());
     _forget_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
     _forget_gate_out2.allocator()->init(TensorInfo(forget_gate1_shape, 1, input->info()->data_type()));
     _forget_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
     _forget_gate_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
 
-    // Configure block that calculates the forget gate
-    // forget_gate = Activation(input * input_to_forget_weights + output_state * recurrent_to_forget_weights + PixelWiseMul(cell_state, cell_to_forget_weights) + forget_gate_bias)
     _memory_group.manage(&_forget_gate_out1);
     _fully_connected_forget_gate.configure(input, input_to_forget_weights, forget_gate_bias, &_forget_gate_out1);
     _memory_group.manage(&_forget_gate_out2);
     _transpose_forget_gate.configure(recurrent_to_forget_weights, &_forget_gate_out2);
     _memory_group.manage(&_forget_gate_out3);
-    _gemm_forget_gate.configure(output_state, &_forget_gate_out2, nullptr, &_forget_gate_out3, 1.f, 0.f);
+    _gemm_forget_gate.configure(output_state_in, &_forget_gate_out2, nullptr, &_forget_gate_out3, 1.f, 0.f);
     _forget_gate_out2.allocator()->allocate();
     _memory_group.manage(&_forget_gate_out5);
     _accum_forget_gate1.configure(&_forget_gate_out1, &_forget_gate_out3, &_forget_gate_out5, ConvertPolicy::SATURATE);
@@ -106,7 +119,7 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
 
         _run_peephole_opt = true;
         _memory_group.manage(&_forget_gate_out4);
-        _pixelwise_mul_forget_gate.configure(cell_state, lstm_params.cell_to_forget_weights(), &_forget_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+        _pixelwise_mul_forget_gate.configure(cell_state_in, lstm_params.cell_to_forget_weights(), &_forget_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
         _accum_forget_gate2.configure(&_forget_gate_out5, &_forget_gate_out4, &_forget_gate_out3, ConvertPolicy::SATURATE);
         _forget_gate_out4.allocator()->allocate();
         _forget_gate_out5.allocator()->allocate();
@@ -119,11 +132,10 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
     _activation_forget_gate.configure(forget_gate_out, &_forget_gate_out1, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
     forget_gate_out->allocator()->allocate();
 
-    _input_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
-
     // Configure block that calculates the input gate
     // input_gate = Activation(input * input_to_input_weights + output_state * recurrent_to_input_weights + PixelWiseMul(cell_state, cell_to_input_weights) + input_gate_bias), without CIFG
     // input_gate = 1 - forget_gate, with CIFG
+    _input_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
     if(lstm_params.has_cifg_opt())
     {
         _memory_group.manage(&_input_gate_out1);
@@ -146,19 +158,24 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
         _memory_group.manage(&_input_gate_out2);
         _transpose_input_gate.configure(lstm_params.recurrent_to_input_weights(), &_input_gate_out2);
         _memory_group.manage(&_input_gate_out3);
-        _gemm_input_gate.configure(output_state, &_input_gate_out2, nullptr, &_input_gate_out3, 1.f, 0.f);
+        _gemm_input_gate.configure(output_state_in, &_input_gate_out2, nullptr, &_input_gate_out3, 1.f, 0.f);
         _input_gate_out2.allocator()->allocate();
         _memory_group.manage(&_input_gate_out4);
-        _pixelwise_mul_input_gate.configure(cell_state, lstm_params.cell_to_input_weights(), &_input_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
-        _memory_group.manage(&_input_gate_out5);
-        _accum_input_gate1.configure(&_input_gate_out1, &_input_gate_out3, &_input_gate_out5, ConvertPolicy::SATURATE);
+        _accum_input_gate1.configure(&_input_gate_out1, &_input_gate_out3, &_input_gate_out4, ConvertPolicy::SATURATE);
+        if(_run_peephole_opt)
+        {
+            _memory_group.manage(&_input_gate_out5);
+            _pixelwise_mul_input_gate.configure(cell_state_in, lstm_params.cell_to_input_weights(), &_input_gate_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+            _accum_input_gate2.configure(&_input_gate_out4, &_input_gate_out5, &_input_gate_out1, ConvertPolicy::SATURATE);
+            _input_gate_out5.allocator()->allocate();
+        }
         _input_gate_out3.allocator()->allocate();
-        _accum_input_gate2.configure(&_input_gate_out5, &_input_gate_out4, &_input_gate_out1, ConvertPolicy::SATURATE);
         _input_gate_out4.allocator()->allocate();
-        _input_gate_out5.allocator()->allocate();
         _activation_input_gate.configure(&_input_gate_out1, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
     }
 
+    // Configure block that calculates the cell state
+    // cell_state = Clip((PixelwiseMul(input_gate, Activation(input * input_to_cell_weights + output_state_in * recurrent_to_cell_weights + cell_bias)) + PixelwiseMul(forget_gate, cell_state)), cell_threshold)
     TensorShape cell_state1_shape = compute_transposed_shape(*recurrent_to_output_weights->info());
     _cell_state_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
     _cell_state_out2.allocator()->init(TensorInfo(cell_state1_shape, 1, input->info()->data_type()));
@@ -166,14 +183,12 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
     _cell_state_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
     _cell_state_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
 
-    // Configure block that calculates the cell state
-    // cell_state = Clip((PixelwiseMul(input_gate, Activation(input * input_to_cell_weights + output_state * recurrent_to_cell_weights + cell_bias)) + PixelwiseMul(forget_gate, cell_state)), cell_threshold)
     _memory_group.manage(&_cell_state_out1);
     _fully_connected_cell_state.configure(input, input_to_cell_weights, cell_bias, &_cell_state_out1);
     _memory_group.manage(&_cell_state_out2);
     _transpose_cell_state.configure(recurrent_to_cell_weights, &_cell_state_out2);
     _memory_group.manage(&_cell_state_out3);
-    _gemm_cell_state1.configure(output_state, &_cell_state_out2, nullptr, &_cell_state_out3, 1.f, 0.f);
+    _gemm_cell_state1.configure(output_state_in, &_cell_state_out2, nullptr, &_cell_state_out3, 1.f, 0.f);
     _cell_state_out2.allocator()->allocate();
     _memory_group.manage(&_cell_state_out4);
     _accum_cell_state1.configure(&_cell_state_out1, &_cell_state_out3, &_cell_state_out4, ConvertPolicy::SATURATE);
@@ -182,12 +197,11 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
     _pixelwise_mul_cell_state1.configure(&_cell_state_out4, &_input_gate_out1, &_cell_state_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
     _input_gate_out1.allocator()->allocate();
     _cell_state_out4.allocator()->allocate();
-    _pixelwise_mul_cell_state2.configure(&_forget_gate_out1, cell_state, &_cell_state_out3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+    _pixelwise_mul_cell_state2.configure(&_forget_gate_out1, cell_state_in, &_cell_state_out3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
     _forget_gate_out1.allocator()->allocate();
     _accum_cell_state2.configure(&_cell_state_out5, &_cell_state_out3, &_cell_state_out1, ConvertPolicy::SATURATE);
     _cell_state_out3.allocator()->allocate();
     _cell_state_out5.allocator()->allocate();
-
     // Perform clipping
     if(cell_threshold != 0.f)
     {
@@ -195,20 +209,20 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
         _cell_clip.configure(&_cell_state_out1, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, cell_threshold));
     }
 
+    // Configure block that calculates the output
+    // output_state_out = Activation(input * input_to_output_weights + output_state_in * recurrent_to_output_weights + PixelWiseMul(cell_state, cell_to_output_weights) + output_gate_bias)
     TensorShape output1_shape = compute_transposed_shape(*recurrent_to_output_weights->info());
     _output1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
     _output2.allocator()->init(TensorInfo(output1_shape, 1, input->info()->data_type()));
     _output3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
     _output5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
 
-    // Configure block that calculates the output
-    // output_state = Activation(input * input_to_output_weights + output_state * recurrent_to_output_weights + PixelWiseMul(cell_state, cell_to_output_weights) + output_gate_bias)
     _memory_group.manage(&_output1);
     _fully_connected_output.configure(input, input_to_output_weights, output_gate_bias, &_output1);
     _memory_group.manage(&_output2);
     _transpose_output.configure(recurrent_to_output_weights, &_output2);
     _memory_group.manage(&_output3);
-    _gemm_output.configure(output_state, &_output2, nullptr, &_output3, 1.f, 0.f);
+    _gemm_output.configure(output_state_in, &_output2, nullptr, &_output3, 1.f, 0.f);
     _output2.allocator()->allocate();
     _memory_group.manage(&_output5);
     _accum_output1.configure(&_output1, &_output3, &_output5, ConvertPolicy::SATURATE);
@@ -231,11 +245,9 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
     {
         _output1.allocator()->allocate();
     }
-    _activation_output.configure(output_gate_out, output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
+    _activation_output.configure(output_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
     output_gate_out->allocator()->allocate();
 
-    _cell_state_activation.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
-
     // Configure block that calculates the output state
     /** lstm_res = PixelwiseMul(output, Activation(cell_state))
      *
@@ -245,32 +257,32 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
      *                     \
      *                      -- lstm_res , otherwise
      */
+    ICLTensor *output_state_out_tmp = lstm_params.has_projection() ? &_output_state1 : output_state_out;
+    _cell_state_activation.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+    _output_state1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
+
     _memory_group.manage(&_cell_state_activation);
     _activation_output_state.configure(&_cell_state_out1, &_cell_state_activation, activation_info);
-    _pixelwise_mul_output_state2.configure(&_cell_state_activation, output, output_state, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
+    _pixelwise_mul_output_state2.configure(&_cell_state_activation, output_gate_out, output_state_out_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
     _cell_state_activation.allocator()->allocate();
 
     if(lstm_params.has_projection())
     {
         _has_projection_weights = true;
-        _output_projection1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
-        _memory_group.manage(&_output_projection1);
-        _fully_connected_output_state.configure(output_state, lstm_params.projection_weights(), lstm_params.projection_bias(), &_output_projection1);
+        _fully_connected_output_state.configure(output_state_out_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out);
+        _output_state1.allocator()->allocate();
         // Perform clipping
         if(projection_threshold != 0.f)
         {
             _perform_projection_clipping = true;
-            _projection_clip.configure(&_output_projection1, output_state, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold));
+            _projection_clip.configure(output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold));
         }
-
-        // Allocate intermediate buffer
-        _output_projection1.allocator()->allocate();
     }
 
     // Copy cell state and output
-    _copy_cell_state.configure(&_cell_state_out1, cell_state);
+    _copy_cell_state.configure(&_cell_state_out1, cell_state_out);
     _cell_state_out1.allocator()->allocate();
-    _copy_output.configure(output_state, output);
+    _copy_output.configure(output_state_out, output);
 
     // Vector for holding the tensors to store in scratch buffer
     std::vector<ICLTensor *> scratch_inputs;
@@ -284,17 +296,31 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
     _concat_scratch_buffer.configure(scratch_inputs, scratch_buffer);
 }
 
-Status CLLSTMLayer::validate(const ITensorInfo *input, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
+Status CLLSTMLayer::validate(const ITensorInfo *input,
+                             const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
                              const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
                              const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
-                             const ITensorInfo *output_state, const ITensorInfo *cell_state, const ITensorInfo *scratch_buffer, const ITensorInfo *output,
+                             const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in,
+                             const ITensorInfo *scratch_buffer, const ITensorInfo *output_state_out, const ITensorInfo *cell_state_out, const ITensorInfo *output,
                              const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
-                                        forget_gate_bias, cell_bias, output_gate_bias, output_state, cell_state);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input,
+                                        input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
+                                        recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
+                                        forget_gate_bias, cell_bias, output_gate_bias,
+                                        output_state_in, cell_state_in,
+                                        scratch_buffer, output_state_out, cell_state_out, output);
+
+    // Check data types
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights,
-                                                       recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias, output_state, cell_state);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input,
+                                                       input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
+                                                       recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
+                                                       forget_gate_bias, cell_bias, output_gate_bias,
+                                                       output_state_in, cell_state_in,
+                                                       scratch_buffer, output_state_out, cell_state_out, output);
+
+    // Check dimensions
     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 2);
     ARM_COMPUTE_RETURN_ERROR_ON(input_to_forget_weights->num_dimensions() > 2);
     ARM_COMPUTE_RETURN_ERROR_ON(input_to_cell_weights->num_dimensions() > 2);
@@ -305,12 +331,19 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, const ITensorInfo *input_
     ARM_COMPUTE_RETURN_ERROR_ON(forget_gate_bias->num_dimensions() > 1);
     ARM_COMPUTE_RETURN_ERROR_ON(cell_bias->num_dimensions() > 1);
     ARM_COMPUTE_RETURN_ERROR_ON(output_gate_bias->num_dimensions() > 1);
-    ARM_COMPUTE_RETURN_ERROR_ON(output_state->num_dimensions() > 2);
-    ARM_COMPUTE_RETURN_ERROR_ON(cell_state->num_dimensions() > 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() > 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->num_dimensions() > 2);
     ARM_COMPUTE_RETURN_ERROR_ON(scratch_buffer->num_dimensions() > 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(output_state_out->num_dimensions() > 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(cell_state_out->num_dimensions() > 2);
     ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 2);
-    ARM_COMPUTE_RETURN_ERROR_ON(cell_bias->dimension(0) * 4 != scratch_buffer->dimension(0) && cell_bias->dimension(0) * 3 != scratch_buffer->dimension(0));
+    ARM_COMPUTE_RETURN_ERROR_ON(cell_bias->dimension(0) * 4 != scratch_buffer->dimension(0)
+                                && cell_bias->dimension(0) * 3 != scratch_buffer->dimension(0));
+
+    const unsigned int num_batches = input->dimension(1);
+    const unsigned int num_cells   = input_to_output_weights->dimension(1);
 
+    // Check peephole optimization
     if(lstm_params.has_peephole_opt())
     {
         ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_output_weights(), lstm_params.cell_to_forget_weights());
@@ -319,85 +352,105 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, const ITensorInfo *input_
     }
 
     TensorShape      units_out_transposed_shape = compute_transposed_shape(*recurrent_to_output_weights);
-    TensorShape      gemmv_shape{ 1, output_state->dimension(1) };
     TensorShape      num_units_transposed_shape = compute_transposed_shape(*forget_gate_bias);
     const TensorInfo units_out_transposed_info  = TensorInfo(units_out_transposed_shape, 1, input->data_type());
-    const TensorInfo gemmv_shape_info           = TensorInfo(gemmv_shape, 1, input->data_type());
     const TensorInfo num_units_transposed_info  = TensorInfo(num_units_transposed_shape, 1, input->data_type());
 
+    TensorInfo input_gate      = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
+    TensorInfo forget_gate     = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
+    TensorInfo output_gate_tmp = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
+    TensorInfo cell_state_tmp  = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
+
     // Validate forget gate
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, forget_gate_bias, cell_state));
-    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state, &units_out_transposed_info, nullptr, cell_state, 1.f, 0.f, GEMMInfo()));
-    ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAdditionKernel::validate(cell_state, cell_state, cell_state, ConvertPolicy::SATURATE));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, forget_gate_bias, &forget_gate));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &forget_gate, 1.f, 0.f, GEMMInfo()));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAdditionKernel::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE));
     if(lstm_params.has_peephole_opt())
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(cell_state, &num_units_transposed_info, nullptr, &gemmv_shape_info, 1.f, 0.f, GEMMInfo()));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(cell_state, &gemmv_shape_info, cell_state, ConvertPolicy::SATURATE));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE));
     }
-    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, cell_state, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
 
     // Validate input gate
     if(!lstm_params.has_cifg_opt())
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights(), lstm_params.cell_to_input_weights(), lstm_params.input_gate_bias());
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.input_to_input_weights(),
+                                            lstm_params.recurrent_to_input_weights(),
+                                            lstm_params.input_gate_bias());
         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_to_input_weights()->num_dimensions() > 2);
         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.recurrent_to_input_weights()->num_dimensions() > 2);
-        ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_input_weights()->num_dimensions() > 1);
         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_gate_bias()->num_dimensions() > 1);
-        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), lstm_params.input_gate_bias(), cell_state));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(cell_state, &num_units_transposed_info, nullptr, &gemmv_shape_info, 1.f, 0.f, GEMMInfo()));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(cell_state, &gemmv_shape_info, cell_state, ConvertPolicy::SATURATE));
-        ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
+
+        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), lstm_params.input_gate_bias(), &input_gate));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &input_gate, 1.f, 0.f, GEMMInfo()));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE));
+        if(lstm_params.has_peephole_opt())
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_input_weights());
+            ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_input_weights()->num_dimensions() > 1);
+            ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE));
+        }
+        ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
     }
     else
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticSubtractionKernel::validate(cell_state, cell_state, cell_state, ConvertPolicy::SATURATE));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticSubtractionKernel::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE));
     }
 
     // Validate cell state
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, cell_bias, cell_state));
-    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, nullptr, activation_info));
-    ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state, cell_state, cell_state, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
-
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, cell_bias, &cell_state_tmp));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &cell_state_tmp, 1.f, 0.f, GEMMInfo()));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&cell_state_tmp, nullptr, activation_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
     if(cell_threshold != 0.f)
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, cell_threshold)));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&cell_state_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold,
+                                                                                                                    cell_threshold)));
     }
 
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, output_gate_bias, cell_state));
+    // Validate output gate tmp
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, output_gate_bias, &output_gate_tmp));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &output_gate_tmp, 1.f, 0.f, GEMMInfo()));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE));
     if(lstm_params.has_peephole_opt())
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(cell_state, cell_state, cell_state, ConvertPolicy::SATURATE));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, lstm_params.cell_to_output_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE,
+                                                                              RoundingPolicy::TO_NEAREST_EVEN));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE));
     }
-    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
 
     // Validate output state
-    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, cell_state, activation_info));
-    ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state, output, output_state, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&cell_state_tmp, &cell_state_tmp, activation_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &output_gate_tmp, &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
     if(lstm_params.has_projection())
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(output_state, lstm_params.projection_weights(), lstm_params.projection_bias(), cell_state));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(&output_gate_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out));
         if(projection_threshold != 0.f)
         {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, output_state, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold,
-                                                                                                                        projection_threshold)));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(output_state_out, output_state_out,
+                                                                          ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold)));
         }
     }
 
-    std::vector<TensorInfo> inputs_vector_info;
-    if(lstm_params.has_cifg_opt())
-    {
-        inputs_vector_info.emplace_back(*cell_state);
-    }
-    inputs_vector_info.emplace_back(*cell_state);
-    inputs_vector_info.emplace_back(*cell_state);
-    inputs_vector_info.emplace_back(*cell_state);
+    // Validate copy kernel
+    ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(&cell_state_tmp, cell_state_out));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(output_state_out, output));
 
+    // Validate scratch concatenation
     std::vector<ITensorInfo *> inputs_vector_info_raw;
-    for(auto &input : inputs_vector_info)
+    if(lstm_params.has_cifg_opt())
     {
-        inputs_vector_info_raw.emplace_back(&input);
+        inputs_vector_info_raw.push_back(&input_gate);
     }
+    inputs_vector_info_raw.push_back(&cell_state_tmp);
+    inputs_vector_info_raw.push_back(&forget_gate);
+    inputs_vector_info_raw.push_back(&output_gate_tmp);
 
     ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenateLayer::validate(inputs_vector_info_raw, scratch_buffer));
     return Status{};
@@ -438,9 +491,12 @@ void CLLSTMLayer::run()
         _fully_connected_input_gate.run();
         CLScheduler::get().enqueue(_transpose_input_gate);
         _gemm_input_gate.run();
-        CLScheduler::get().enqueue(_pixelwise_mul_input_gate);
         CLScheduler::get().enqueue(_accum_input_gate1);
-        _accum_input_gate2.run();
+        if(_run_peephole_opt)
+        {
+            CLScheduler::get().enqueue(_pixelwise_mul_input_gate);
+            _accum_input_gate2.run();
+        }
         CLScheduler::get().enqueue(_activation_input_gate);
     }