COMPMID-2759 add support for QASYMM8_SIGNED to CLFullyConnectedLayer

Change-Id: I7092390b01a56065a442be0d14e2f9bfce2cdc9c
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2583
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 27 insertions, 14 deletions

diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index 39c3c2b85b..cbd28603fc 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -26,14 +26,12 @@
 
 #include "arm_compute/runtime/CL/ICLSimpleFunction.h"
 
-#include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h"
 #include "arm_compute/core/CL/kernels/CLTransposeKernel.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h"
 #include "arm_compute/runtime/CL/functions/CLFlattenLayer.h"
 #include "arm_compute/runtime/CL/functions/CLGEMM.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
-#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
 #include "arm_compute/runtime/IWeightsManager.h"
 #include "arm_compute/runtime/MemoryGroup.h"
 
@@ -117,7 +115,6 @@ private:
  *  -# @ref CLIm2ColKernel (called when the input comes from a convolutional layer)
  *  -# @ref CLFullyConnectedLayerReshapeWeights (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
  *  -# @ref CLGEMMMatrixMultiplyKernel or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
- *  -# @ref CLGEMMMatrixAccumulateBiasesKernel or @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if quantized asymmetric) (if @p biases is not equal to nullptr)
  *
  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
  */
@@ -136,7 +133,7 @@ public:
     CLFullyConnectedLayer &operator=(CLFullyConnectedLayer &&) = default;
     /** Set the input and output tensors.
      *
-     * @param[in]  input   Source tensor. Data type supported: QASYMM8/F16/F32.
+     * @param[in]  input   Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in]  weights Weights tensor. The weights must be 2 dimensional.
      *                     If this function is called after a Convolution Layer, the (transposed) weights will have as many rows as the product of the first 3 input's dimensions.
      *                     If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension.
@@ -152,7 +149,7 @@ public:
                    FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
     /** Static function to check if given info will lead to a valid configuration of @ref CLFullyConnectedLayer
      *
-     * @param[in]  input   Source tensor info. Data type supported: QASYMM8/F16/F32.
+     * @param[in]  input   Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in]  weights Weights tensor info. The weights must be 2 dimensional.
      *                     If this function is called after a Convolution Layer, the (transposed) weights will have as many rows as the product of the first 3 input's dimensions.
      *                     If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension.
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index ad0714ed15..dcaa12645e 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -48,8 +48,10 @@ Status construct_gemmlowp_output_stage(const ITensorInfo &input, const ITensorIn
     gemmlowp_output_stage.gemmlowp_multiplier = 0;
     gemmlowp_output_stage.gemmlowp_shift      = 0;
 
+    const auto data_type = input.data_type();
+
     // Configure output stage for quantized case
-    if(is_data_type_quantized_asymmetric(input.data_type()))
+    if(is_data_type_quantized_asymmetric(data_type))
     {
         const UniformQuantizationInfo iq_info = input.quantization_info().uniform();
         const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
@@ -62,14 +64,18 @@ Status construct_gemmlowp_output_stage(const ITensorInfo &input, const ITensorIn
         int         output_shift      = 0;
         ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
 
+        PixelValue type_min{};
+        PixelValue type_max{};
+        std::tie(type_min, type_max) = get_min_max(data_type);
+
         // Set the GEMMLowp output stage info
         gemmlowp_output_stage.gemmlowp_offset     = output_quant_info.offset;
         gemmlowp_output_stage.gemmlowp_multiplier = output_multiplier;
         gemmlowp_output_stage.gemmlowp_shift      = output_shift;
-        gemmlowp_output_stage.gemmlowp_min_bound  = 0;
-        gemmlowp_output_stage.gemmlowp_max_bound  = 255;
         gemmlowp_output_stage.gemmlowp_multipliers.push_back(output_multiplier);
         gemmlowp_output_stage.gemmlowp_shifts.push_back(output_shift);
+        type_min.get(gemmlowp_output_stage.gemmlowp_min_bound);
+        type_max.get(gemmlowp_output_stage.gemmlowp_max_bound);
     }
 
     return Status{};
@@ -304,7 +310,7 @@ Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorIn
                                        FullyConnectedLayerInfo fc_info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
 
diff --git a/tests/validation/CL/FullyConnectedLayer.cpp b/tests/validation/CL/FullyConnectedLayer.cpp
index 091d9411b7..e57dd4e7b1 100644
--- a/tests/validation/CL/FullyConnectedLayer.cpp
+++ b/tests/validation/CL/FullyConnectedLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -57,6 +57,7 @@ const auto CNNDataTypes = framework::dataset::make("DataType",
     DataType::F16,
     DataType::F32,
     DataType::QASYMM8,
+    DataType::QASYMM8_SIGNED,
 });
 
 const auto FullyConnectedParameters = combine(framework::dataset::make("TransposeWeights", { false, true }), framework::dataset::make("ReshapeWeights", { false, true }));
@@ -222,8 +223,16 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerQuantizedFixture<uint8_t>,
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_qasymm8);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() /* QASYMM8 */
+TEST_SUITE(QASYMM8_SIGNED)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), QuantizationData))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END() /* QASYMM8_SIGNED */
+TEST_SUITE_END() /* Quantized */
 
 TEST_SUITE_END()
 TEST_SUITE_END()
diff --git a/tests/validation/fixtures/FullyConnectedLayerFixture.h b/tests/validation/fixtures/FullyConnectedLayerFixture.h
index ff6ac17744..7f0ceadea1 100644
--- a/tests/validation/fixtures/FullyConnectedLayerFixture.h
+++ b/tests/validation/fixtures/FullyConnectedLayerFixture.h
@@ -49,7 +49,8 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ
 class FullyConnectedLayerValidationGenericFixture : public framework::Fixture
 {
 public:
-    using TBias = typename std::conditional < std::is_same<typename std::decay<T>::type, uint8_t>::value || std::is_same<typename std::decay<T>::type, int8_t>::value, int32_t, T >::type;
+    using TDecay = typename std::decay<T>::type;
+    using TBias  = typename std::conditional < (std::is_same<TDecay, uint8_t>::value || std::is_same<TDecay, int8_t>::value), int32_t, T >::type;
 
 public:
     template <typename...>