COMPMID-2236: QUANTIZED_16BIT_LSTM operator for NEON

Change-Id: I554023508e09b790ecc1bbdada529697d6c7b616
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1551
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

6 files changed, 214 insertions, 8 deletions

diff --git a/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h b/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
index 3320ba6889..f0a2a57d1a 100644
--- a/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
@@ -52,13 +52,13 @@ public:
     ~NEDequantizationLayerKernel() = default;
     /** Set input, output tensors.
      *
-     * @param[in]  input  Source tensor. Data type supported: QASYMM8/QSYMM8.
+     * @param[in]  input  Source tensor. Data type supported: QASYMM8/QSYMM8/QSYMM16.
      * @param[out] output Destination tensor with the same dimensions of input. Data type supported: F16/F32.
      */
     void configure(const ITensor *input, ITensor *output);
     /** Static function to check if given info will lead to a valid configuration of @ref NEDequantizationLayerKernel
      *
-     * @param[in] input  Input tensor info. Data types supported: QASYMM8/QSYMM8.
+     * @param[in] input  Input tensor info. Data types supported: QASYMM8/QSYMM8/QSYMM16.
      * @param[in] output Output tensor info. Data types supported: F16/F32.
      *
      * @return a status
diff --git a/arm_compute/core/Validate.h b/arm_compute/core/Validate.h
index dab4221a3b..37c7b50ec7 100644
--- a/arm_compute/core/Validate.h
+++ b/arm_compute/core/Validate.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -565,7 +565,7 @@ inline arm_compute::Status error_on_mismatching_quantization_info(const char *fu
     DataType             &&first_data_type         = tensor_info_1->data_type();
     const QuantizationInfo first_quantization_info = tensor_info_1->quantization_info();
 
-    if(!is_data_type_quantized_asymmetric(first_data_type))
+    if(!is_data_type_quantized(first_data_type))
     {
         return arm_compute::Status{};
     }
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index d44afcbb0f..b59f24eed5 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -95,6 +95,7 @@
 #include "arm_compute/runtime/NEON/functions/NEIntegralImage.h"
 #include "arm_compute/runtime/NEON/functions/NEL2NormalizeLayer.h"
 #include "arm_compute/runtime/NEON/functions/NELSTMLayer.h"
+#include "arm_compute/runtime/NEON/functions/NELSTMLayerQuantized.h"
 #include "arm_compute/runtime/NEON/functions/NELaplacianPyramid.h"
 #include "arm_compute/runtime/NEON/functions/NELaplacianReconstruct.h"
 #include "arm_compute/runtime/NEON/functions/NELocallyConnectedLayer.h"
diff --git a/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h b/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h
index 8c24b38cee..c08366e5a7 100644
--- a/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h
@@ -39,13 +39,13 @@ class NEDequantizationLayer : public INESimpleFunctionNoBorder
 public:
     /** Configure the kernel.
      *
-     * @param[in]  input  Source tensor. Data types supported: QASYMM8/QSYMM8.
+     * @param[in]  input  Source tensor. Data types supported: QASYMM8/QSYMM8/QSYMM16.
      * @param[out] output Destination tensor with the same dimensions of input. Data type supported: F16/F32.
      */
     void configure(const ITensor *input, ITensor *output);
     /** Static function to check if given info will lead to a valid configuration of @ref NEDequantizationLayer
      *
-     * @param[in] input  Input tensor info. Data types supported: QASYMM8/QSYMM8.
+     * @param[in] input  Input tensor info. Data types supported: QASYMM8/QSYMM8/QSYMM16.
      * @param[in] output Output tensor info. Data type supported: F16/F32.
      *
      * @return a status
diff --git a/arm_compute/runtime/NEON/functions/NELSTMLayerQuantized.h b/arm_compute/runtime/NEON/functions/NELSTMLayerQuantized.h
new file mode 100644
index 0000000000..b45d714990
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NELSTMLayerQuantized.h
@@ -0,0 +1,205 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_NELSTMLAYERQUANTIZED_H__
+#define __ARM_COMPUTE_NELSTMLAYERQUANTIZED_H__
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEActivationLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEArithmeticAddition.h"
+#include "arm_compute/runtime/NEON/functions/NEConcatenateLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEDequantizationLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEElementwiseOperations.h"
+#include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
+#include "arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h"
+#include "arm_compute/runtime/NEON/functions/NEQuantizationLayer.h"
+#include "arm_compute/runtime/NEON/functions/NESlice.h"
+#include "arm_compute/runtime/NEON/functions/NETranspose.h"
+
+#include "arm_compute/runtime/common/LSTMParams.h"
+
+namespace arm_compute
+{
+// Forward declarations
+class ITensor;
+
+/** Basic function to run @ref NELSTMLayerQuantized
+ *
+ * This function calls the following NEON functions/kernels:
+ *
+ * -# @ref NEGEMMLowpMatrixMultiplyCore                          Quantized matrix multiplication core. Accumulators are 32-bit integers
+ * -# @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint   Convert 32-bit integers into QSYMM16
+ * -# @ref NETranspose                                           Matrix transpose
+ * -# @ref NEConcatenateLayer                                    Tensor concatenation
+ * -# @ref NEActivationLayer                                     Activation functions (tanh and logistig)
+ * -# @ref NEArithmeticAddition                                  Elementwise addition
+ * -# @ref NEPixelWiseMultiplication                             Elementwise multiplication
+ * -# @ref NESlice                                               Tensor slicing
+ * -# @ref NEDequantizationLayer                                 Dequantize into float
+ * -# @ref NEQuantizationLayer                                   Quantize from float
+ * */
+class NELSTMLayerQuantized : public IFunction
+{
+public:
+    /** Default constructor */
+    NELSTMLayerQuantized(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NELSTMLayerQuantized(const NELSTMLayerQuantized &) = delete;
+    /** Default move constructor */
+    NELSTMLayerQuantized(NELSTMLayerQuantized &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NELSTMLayerQuantized &operator=(const NELSTMLayerQuantized &) = delete;
+    /** Default move assignment operator */
+    NELSTMLayerQuantized &operator=(NELSTMLayerQuantized &&) = default;
+    /** Initialize function's tensors.
+     *
+     * @param[in]  input                       Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: QASYMM8.
+     * @param[in]  input_to_input_weights      2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  input_to_forget_weights     2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  input_to_cell_weights       2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  input_to_output_weights     2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  recurrent_to_input_weights  2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  recurrent_to_cell_weights   2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  recurrent_to_output_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  input_gate_bias             1D weights tensor with dimensions [output_size]. Data type supported: S32.
+     * @param[in]  forget_gate_bias            1D weights tensor with dimensions [output_size]. Data type supported: S32.
+     * @param[in]  cell_bias                   1D weights tensor with dimensions [output_size]. Data type supported: S32.
+     * @param[in]  output_gate_bias            1D weights tensor with dimensions [output_size]. Data type supported: S32.
+     * @param[in]  cell_state_in               2D tensor with dimensions [output_size, batch_size]. Data type supported:  QSYMM16.
+     * @param[in]  output_state_in             2D tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
+     * @param[out] cell_state_out              Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size]. Data type supported:  QSYMM16.
+     * @param[out] output_state_out            Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].Data types supported: Same as @p input.
+     */
+    void configure(const ITensor *input,
+                   const ITensor *input_to_input_weights, const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights,
+                   const ITensor *recurrent_to_input_weights, const ITensor *recurrent_to_forget_weights, const ITensor *recurrent_to_cell_weights, const ITensor *recurrent_to_output_weights,
+                   const ITensor *input_gate_bias, const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias,
+                   ITensor *cell_state_in, const ITensor *output_state_in,
+                   ITensor *cell_state_out, ITensor *output_state_out);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref NELSTMLayer
+     *
+     * @param[in]  input                       Source tensor info. Input is a 2D tensor info with dimensions [input_size, batch_size]. Data types supported: QASYMM8.
+     * @param[in]  input_to_input_weights      2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  input_to_forget_weights     2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  input_to_cell_weights       2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  input_to_output_weights     2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  recurrent_to_input_weights  2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  recurrent_to_cell_weights   2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
+     * @param[in]  input_gate_bias             1D weights tensor info with dimensions [output_size]. Data type supported: S32.
+     * @param[in]  forget_gate_bias            1D weights tensor info with dimensions [output_size]. Data type supported: S32.
+     * @param[in]  cell_bias                   1D weights tensor info with dimensions [output_size]. Data type supported: S32.
+     * @param[in]  output_gate_bias            1D weights tensor info with dimensions [output_size]. Data type supported: S32.
+     * @param[in]  cell_state_in               2D tensor info with dimensions [output_size, batch_size]. Data type supported:  QSYMM16.
+     * @param[in]  output_state_in             2D tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
+     * @param[out] cell_state_out              Destination tensor info. Output is a 2D tensor info with dimensions [output_size, batch_size]. Data type supported:  QSYMM16.
+     * @param[out] output_state_out            Destination tensor info. Output is a 2D tensor info with dimensions [output_size, batch_size].Data types supported: Same as @p input.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input,
+                           const ITensorInfo *input_to_input_weights, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
+                           const ITensorInfo *recurrent_to_input_weights, const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
+                           const ITensorInfo *input_gate_bias, const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
+                           const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in,
+                           const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out);
+
+    // Inherited methods overridden:
+    void run() override;
+    void prepare() override;
+
+private:
+    MemoryGroup _memory_group;
+
+    // Functions used
+    NEGEMMLowpMatrixMultiplyCore                        _gemmlowp;
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint _output_stage;
+    NETranspose                                         _transpose_weights;
+    NEConcatenateLayer                                  _concat_input_weights;
+    NEConcatenateLayer                                  _concat_recurrent_weights;
+    NEConcatenateLayer                                  _concat_weights;
+    NEConcatenateLayer                                  _concat_inputs;
+    NEConcatenateLayer                                  _concat_bias;
+    NEActivationLayer                                   _sigmoid_forget_gate;
+    NEActivationLayer                                   _sigmoid_input_gate;
+    NEActivationLayer                                   _sigmoid_output_gate;
+    NEActivationLayer                                   _tanh_modulation_gate;
+    NEActivationLayer                                   _tanh_output_state;
+    NEArithmeticAddition                                _add1;
+    NEArithmeticAddition                                _add2;
+    NEPixelWiseMultiplication                           _mul1;
+    NEPixelWiseMultiplication                           _mul2;
+    NEPixelWiseMultiplication                           _mul3;
+    NESlice                                             _slice_input_tensor;
+    NESlice                                             _slice_forget_tensor;
+    NESlice                                             _slice_cell_tensor;
+    NESlice                                             _slice_output_tensor;
+    NEDequantizationLayer                               _dequantize;
+    NEQuantizationLayer                                 _quantize;
+
+    // Tensor pointers
+    const ITensor *_input_to_input_weights;
+    const ITensor *_input_to_forget_weights;
+    const ITensor *_input_to_cell_weights;
+    const ITensor *_input_to_output_weights;
+    const ITensor *_recurrent_to_input_weights;
+    const ITensor *_recurrent_to_forget_weights;
+    const ITensor *_recurrent_to_cell_weights;
+    const ITensor *_recurrent_to_output_weights;
+    const ITensor *_input_gate_bias;
+    const ITensor *_forget_gate_bias;
+    const ITensor *_cell_bias;
+    const ITensor *_output_gate_bias;
+
+    // Temporary tensors
+    Tensor _recurrent_weights;
+    Tensor _input_weights;
+    Tensor _weights;
+    Tensor _input;
+    Tensor _weights_transposed;
+    Tensor _output_highp;
+    Tensor _output_lowp;
+    Tensor _bias;
+    Tensor _forget_gate_input;
+    Tensor _input_gate_input;
+    Tensor _output_gate_input;
+    Tensor _input_modulation_gate_input;
+    Tensor _forget_gate_output;
+    Tensor _input_gate_output;
+    Tensor _output_gate_output;
+    Tensor _input_modulation_gate_output;
+    Tensor _cell_state1;
+    Tensor _cell_state2;
+    Tensor _output_state_tmp;
+    Tensor _output_state_out_symm;
+    Tensor _output_state_out_f32;
+
+    bool _is_prepared;
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NELSTMLAYERQUANTIZED_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h b/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h
index 5e4b4f754c..46a62bd903 100644
--- a/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h
@@ -49,13 +49,13 @@ public:
     /** Set the input and output tensors.
      *
      * @param[in]  input  Source tensor. The dimensions over the third will be interpreted as batches. Data types supported: F32/F16.
-     * @param[out] output Destination tensor with the same dimensions of input. Data types supported: QASYMM8
+     * @param[out] output Destination tensor with the same dimensions of input. Data types supported: QASYMM8/QSYMM16
      */
     void configure(const ITensor *input, ITensor *output);
     /** Static function to check if given info will lead to a valid configuration of @ref NEQuantizationLayer
      *
      * @param[in] input  Input tensor info. The dimensions over the third will be interpreted as batches. Data types supported: F32/F16.
-     * @param[in] output Output tensor info. Data types supported: QASYMM8
+     * @param[in] output Output tensor info. Data types supported: QASYMM8/QSYMM16
      *
      * @return a status
      */