COMPMID-3239: Fix projection and peephole in NEQLSTMLayer

- Peephole and projection has been fixed to be working
- Small internal kernel copying data between tensors to
  cover the case where num_units and output_size is different
  is added.

Below is strictly outside of this patch's scope but are changes
helping this patch working (directly or indirectly) or making
NEQLSTM more complete.

- Consideration for layer normalization is added to InfoHelpers
- QSYMM8 data type is added to helper function to
  print out tensors.
- NE/CLLSTMLayer::validate() logic has been modified to use correct
  value for shape validation.

Change-Id: I40b4e71dfdbe8432caa2fe4a9af60a725362cc33
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3157
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 62 insertions, 2 deletions

diff --git a/arm_compute/core/utils/misc/InfoHelpers.h b/arm_compute/core/utils/misc/InfoHelpers.h
index 8cf701c124..6ecda7a0dd 100644
--- a/arm_compute/core/utils/misc/InfoHelpers.h
+++ b/arm_compute/core/utils/misc/InfoHelpers.h
@@ -90,6 +90,23 @@ inline void build_lstm_params_tensor_info(const LSTMParams<T>     &lstm_params,
         lstm_params_info->set_cifg_params(lstm_params.input_to_input_weights()->info(), lstm_params.recurrent_to_input_weights()->info(),
                                           cell_to_input_weights_info, lstm_params.input_gate_bias()->info());
     }
+    if(lstm_params.use_layer_norm())
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(lstm_params.forget_layer_norm_weights(),
+                                     lstm_params.output_layer_norm_weights(),
+                                     lstm_params.cell_layer_norm_weights());
+        if(!lstm_params.has_cifg_opt())
+        {
+            ARM_COMPUTE_ERROR_ON_NULLPTR(lstm_params.input_layer_norm_weights());
+        }
+
+        const ITensorInfo *forget_info = lstm_params.forget_layer_norm_weights()->info();
+        const ITensorInfo *cell_info   = lstm_params.cell_layer_norm_weights()->info();
+        const ITensorInfo *output_info = lstm_params.output_layer_norm_weights()->info();
+        const ITensorInfo *input_info  = lstm_params.has_cifg_opt() ? nullptr : lstm_params.input_layer_norm_weights()->info();
+
+        lstm_params_info->set_layer_normalization_params(input_info, forget_info, cell_info, output_info);
+    }
 }
 } // namespace info_helpers
 } // namespace utils
diff --git a/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h b/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h
index 0553e4f266..9eb0654cfe 100644
--- a/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEQLSTMLayer.h
@@ -178,7 +178,8 @@ private:
         Output,
         Count
     };
-    static constexpr uint8_t _layer_norm_count = static_cast<uint8_t>(LayerNormGate::Count);
+    static constexpr uint8_t  _layer_norm_count                    = static_cast<uint8_t>(LayerNormGate::Count);
+    static constexpr uint32_t _out_state_output_size_dimension_idx = 0;
 
     /** Internal method to configure matrix multiplication plus output stage of each gate.
      *
@@ -201,6 +202,35 @@ private:
 
     MemoryGroup _memory_group{};
 
+    /** A small internel kernel do the copy between two tensors */
+    class TensorCopyKernel
+    {
+        static constexpr uint32_t max_dimension_supported = 2;
+
+        ITensor *_src{ nullptr };
+        ITensor *_dst{ nullptr };
+        size_t   _row_size{};
+        Window   _window{};
+
+    public:
+        /** Static function to check if given info will lead to a valid configuration of @ref NEQLSTMLayer::TensorCopyKernel
+         *
+         * @param[in] src Source tensor info.
+         * @param[in] dst Destination tensor info
+         *
+         * @return a status
+         */
+        static Status validate(const ITensorInfo &src, const ITensorInfo &dst);
+        /** Set the input and output tensors.
+         *
+         * @param[in]  src Source tensor
+         * @param[out] dst Destination tensor
+         */
+        void configure(ITensor &src, ITensor &dst);
+        /** run the kernel */
+        void run();
+    };
+
     // Functions used
     NETranspose                      _transpose_input_to_forget_weights{};
     NETranspose                      _transpose_input_to_cell_weights{};
@@ -245,7 +275,7 @@ private:
     NEPixelWiseMultiplicationKernel  _pixelwise_mul_cell_to_input{};
     NEGEMMLowpOutputStage            _cell_to_input_outstage{};
     NEArithmeticAdditionKernel       _accumulate_cell_input{};
-    NEActivationLayer                _input_gate_tanh{};
+    NEActivationLayer                _input_gate_sigmoid{};
     NEPixelWiseMultiplicationKernel  _pixelwise_mul_forget_cell{};
     NEPixelWiseMultiplicationKernel  _pixelwise_mul_input_cell{};
     NEArithmeticAdditionKernel       _add_forget_cell{};
@@ -256,6 +286,7 @@ private:
     NEGEMMLowpOutputStage            _recurrent_to_output_outstage{};
     NEArithmeticAdditionKernel       _accumulate_input_recurrent_output{};
     NEPixelWiseMultiplicationKernel  _pixelwise_mul_cell_to_output{};
+    NEGEMMLowpOutputStage            _cell_to_output_outstage{};
     NEArithmeticAdditionKernel       _accumulate_cell_to_output{};
     NEActivationLayer                _output_gate_sigmoid{};
     NEActivationLayer                _hidden_tanh{};
@@ -265,6 +296,12 @@ private:
     NEGEMMLowpOutputStage            _projection_outstage{};
     NEArithmeticAdditionKernel       _accumulate_projection{};
     NEActivationLayer                _projection_clip{};
+
+    TensorCopyKernel _projection_bias_copy{};
+    TensorCopyKernel _projection_output_to_accumulate_copy{};
+    TensorCopyKernel _projection_accumulate_to_output_copy{};
+    TensorCopyKernel _hidden_to_output_copy{};
+
     std::array<NEQLSTMLayerNormalizationKernel, _layer_norm_count> _layer_norms{ {} };
 
     // Tensor pointers
@@ -375,11 +412,16 @@ private:
     Tensor _input_to_output_outstage_res{ nullptr };
     Tensor _mm_recurrent_to_output_res{ nullptr };
     Tensor _mul_cell_to_output_res{ nullptr };
+    Tensor _cell_to_output_outstage_res{ nullptr };
     Tensor _recurrent_to_output_outstage_res{ nullptr };
     Tensor _output_gate{ nullptr };
     Tensor _hidden_mul_res{ nullptr };
+    Tensor _hidden_gate{ nullptr };
     Tensor _mm_projection_res{ nullptr };
     Tensor _projection_outstage_res{ nullptr };
+    Tensor _projection_out_res{ nullptr };
+    Tensor _projection_eff_bias_adjusted{ nullptr };
+    Tensor _projection_accumulate_res{ nullptr };
     Tensor _ones{ nullptr };
     std::array<Tensor, _layer_norm_count> _layer_norm_output{ {} };
 
@@ -395,6 +437,7 @@ private:
     bool _has_projection_clipping{ false };
     bool _has_peephole{ false };
     bool _has_layer_norm{ false };
+    bool _projection_tensor_copy_required{ false };
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_NEQLSTMLAYER_H */