COMPMID-1401 Implement NEFullyConnectedLayer for QASYMM8

Change-Id: I0404df6d369855e2f458f2db8f26e81c80a1ee87
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140148
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

4 files changed, 65 insertions, 54 deletions

diff --git a/tests/validation/CL/FullyConnectedLayer.cpp b/tests/validation/CL/FullyConnectedLayer.cpp
index cd050e378e..91f698cb01 100644
--- a/tests/validation/CL/FullyConnectedLayer.cpp
+++ b/tests/validation/CL/FullyConnectedLayer.cpp
@@ -162,7 +162,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(
 // *INDENT-ON*
 
 template <typename T>
-using CLFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T, false>;
+using CLFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T>;
 
 TEST_SUITE(Float)
 TEST_SUITE(FP16)
@@ -199,7 +199,7 @@ TEST_SUITE_END()
 TEST_SUITE_END()
 
 template <typename T>
-using CLFullyConnectedLayerQuantizedFixture = FullyConnectedLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T, false>;
+using CLFullyConnectedLayerQuantizedFixture = FullyConnectedLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T>;
 
 TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
diff --git a/tests/validation/GLES_COMPUTE/FullyConnectedLayer.cpp b/tests/validation/GLES_COMPUTE/FullyConnectedLayer.cpp
index c82a8a1a43..53de8b9d10 100644
--- a/tests/validation/GLES_COMPUTE/FullyConnectedLayer.cpp
+++ b/tests/validation/GLES_COMPUTE/FullyConnectedLayer.cpp
@@ -102,7 +102,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame
 }
 
 template <typename T>
-using GCFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<GCTensor, GCAccessor, GCFullyConnectedLayer, T, false>;
+using GCFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<GCTensor, GCAccessor, GCFullyConnectedLayer, T>;
 
 TEST_SUITE(Float)
 TEST_SUITE(FP16)
diff --git a/tests/validation/NEON/FullyConnectedLayer.cpp b/tests/validation/NEON/FullyConnectedLayer.cpp
index 80fdf1784e..3aeba7a969 100644
--- a/tests/validation/NEON/FullyConnectedLayer.cpp
+++ b/tests/validation/NEON/FullyConnectedLayer.cpp
@@ -48,6 +48,9 @@ constexpr RelativeTolerance<float> tolerance_f32(0.01f);
 constexpr RelativeTolerance<float> tolerance_f16(0.01f);
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC*/
 
+/** Tolerance for quantized asymmetric operations */
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1);
+
 /** CNN data types */
 const auto CNNDataTypes = framework::dataset::make("DataType",
 {
@@ -68,6 +71,9 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame
                                                                    CNNDataTypes),
                src_shape, weights_shape, bias_shape, dst_shape, transpose_weights, reshape_weights, data_type)
 {
+    const DataType         bias_data_type    = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type;
+    const QuantizationInfo quantization_info = is_data_type_quantized_asymmetric(data_type) ? QuantizationInfo(2.f / 255.f, 127) : QuantizationInfo();
+
     TensorShape ws(weights_shape);
 
     // Transpose weights if not done in the function
@@ -76,23 +82,13 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame
         const size_t shape_x = ws.x();
         ws.set(0, ws.y());
         ws.set(1, shape_x);
-
-        // Weights have to be passed reshaped
-        // Transpose 1xW for batched version
-        if(!reshape_weights && dst_shape.y() > 1)
-        {
-            const float  transpose_width = 16.0f / data_size_from_type(data_type);
-            const size_t shape_x         = ws.x();
-            ws.set(0, ws.y() * static_cast<unsigned int>(transpose_width));
-            ws.set(1, static_cast<unsigned int>(std::ceil(shape_x / transpose_width)));
-        }
     }
 
     // Create tensors
-    Tensor src     = create_tensor<Tensor>(src_shape, data_type, 1);
-    Tensor weights = create_tensor<Tensor>(ws, data_type, 1);
-    Tensor bias    = create_tensor<Tensor>(bias_shape, data_type, 1);
-    Tensor dst     = create_tensor<Tensor>(dst_shape, data_type, 1);
+    Tensor src     = create_tensor<Tensor>(src_shape, data_type, 1, quantization_info);
+    Tensor weights = create_tensor<Tensor>(ws, data_type, 1, quantization_info);
+    Tensor bias    = create_tensor<Tensor>(bias_shape, bias_data_type, 1, quantization_info);
+    Tensor dst     = create_tensor<Tensor>(dst_shape, data_type, 1, quantization_info);
 
     ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -104,6 +100,9 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame
     fc_info.transpose_weights    = transpose_weights;
     fc_info.are_weights_reshaped = !reshape_weights;
 
+    const QuantizationInfo src_quantization_info     = src.info()->quantization_info();
+    const QuantizationInfo weights_quantization_info = weights.info()->quantization_info();
+
     // Create and configure function.
     NEFullyConnectedLayer fc;
     fc.configure(&src, &weights, &bias, &dst, fc_info);
@@ -111,6 +110,10 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame
     // Validate valid region
     const ValidRegion dst_valid_region = shape_to_valid_region(dst_shape);
     validate(dst.info()->valid_region(), dst_valid_region);
+
+    // Validate QuantizationInfo
+    ARM_COMPUTE_EXPECT(src.info()->quantization_info() == src_quantization_info, framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(weights.info()->quantization_info() == weights_quantization_info, framework::LogLevel::ERRORS);
 }
 
 // *INDENT-OFF*
@@ -161,7 +164,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(
 // *INDENT-ON*
 
 template <typename T>
-using NEFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<Tensor, Accessor, NEFullyConnectedLayer, T, true>;
+using NEFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<Tensor, Accessor, NEFullyConnectedLayer, T>;
 
 TEST_SUITE(Float)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
@@ -199,6 +202,32 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixture<float>, framework:
 TEST_SUITE_END()
 TEST_SUITE_END()
 
+template <typename T>
+using NEFullyConnectedLayerQuantizedFixture = FullyConnectedLayerValidationQuantizedFixture<Tensor, Accessor, NEFullyConnectedLayer, T>;
+
+TEST_SUITE(Quantized)
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(
+                           combine(datasets::SmallFullyConnectedLayerDataset(),
+                                   FullyConnectedParameters),
+                           framework::dataset::make("DataType", DataType::QASYMM8)),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 10) })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(
+                           combine(datasets::LargeFullyConnectedLayerDataset(),
+                                   FullyConnectedParameters),
+                           framework::dataset::make("DataType", DataType::QASYMM8)),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 256.f, 10) })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
 TEST_SUITE_END()
 TEST_SUITE_END()
 } // namespace validation
diff --git a/tests/validation/fixtures/FullyConnectedLayerFixture.h b/tests/validation/fixtures/FullyConnectedLayerFixture.h
index 18321480f8..49c3be0c2e 100644
--- a/tests/validation/fixtures/FullyConnectedLayerFixture.h
+++ b/tests/validation/fixtures/FullyConnectedLayerFixture.h
@@ -45,7 +45,7 @@ namespace test
 {
 namespace validation
 {
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool run_interleave>
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
 class FullyConnectedLayerValidationGenericFixture : public framework::Fixture
 {
 public:
@@ -103,8 +103,8 @@ protected:
         // -----------+-----------+---------------------------
         //  transpose |           | ***
         // -----------+-----------+---------------------------
-        // !transpose | transpose | transpose &
-        //            |           | transpose1xW (if required)
+        // !transpose | transpose | transpose
+        //            |           |
         //
         // ***: That combination is invalid. But we can ignore the transpose flag and handle all !reshape the same
         if(!reshape_weights || !transpose_weights)
@@ -112,16 +112,6 @@ protected:
             const size_t shape_x = reshaped_weights_shape.x();
             reshaped_weights_shape.set(0, reshaped_weights_shape.y());
             reshaped_weights_shape.set(1, shape_x);
-
-            // Weights have to be passed reshaped
-            // Transpose 1xW for batched version
-            if(!reshape_weights && output_shape.y() > 1 && run_interleave)
-            {
-                const int   transpose_width = 16 / data_size_from_type(_data_type);
-                const float shape_x         = reshaped_weights_shape.x();
-                reshaped_weights_shape.set(0, reshaped_weights_shape.y() * transpose_width);
-                reshaped_weights_shape.set(1, static_cast<unsigned int>(std::ceil(shape_x / transpose_width)));
-            }
         }
 
         // Create tensors
@@ -170,14 +160,6 @@ protected:
             // Transpose elementwise
             tmp = transpose(tmp);
 
-            // Reshape weights for batched runs
-            if(!reshape_weights && output_shape.y() > 1 && run_interleave)
-            {
-                // Transpose with interleave
-                const int interleave_size = 16 / tmp.element_size();
-                tmp                       = transpose(tmp, interleave_size);
-            }
-
             AccessorType weights_accessor(weights);
 
             for(int i = 0; i < tmp.num_elements(); ++i)
@@ -222,43 +204,43 @@ protected:
     QuantizationInfo _quantization_info{};
 };
 
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool run_interleave>
-class FullyConnectedLayerValidationFixture : public FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, run_interleave>
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class FullyConnectedLayerValidationFixture : public FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
     void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, bool transpose_weights, bool reshape_weights, DataType data_type)
     {
-        FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, run_interleave>::setup(input_shape, weights_shape, bias_shape, output_shape, transpose_weights,
-                                                                                                                      reshape_weights, data_type,
-                                                                                                                      QuantizationInfo());
+        FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, transpose_weights,
+                                                                                                      reshape_weights, data_type,
+                                                                                                      QuantizationInfo());
     }
 };
 
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool run_interleave>
-class FullyConnectedLayerValidationFixedPointFixture : public FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, run_interleave>
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class FullyConnectedLayerValidationFixedPointFixture : public FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
     void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, bool transpose_weights, bool reshape_weights, DataType data_type)
     {
-        FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, run_interleave>::setup(input_shape, weights_shape, bias_shape, output_shape, transpose_weights,
-                                                                                                                      reshape_weights, data_type,
-                                                                                                                      QuantizationInfo());
+        FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, transpose_weights,
+                                                                                                      reshape_weights, data_type,
+                                                                                                      QuantizationInfo());
     }
 };
 
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool run_interleave>
-class FullyConnectedLayerValidationQuantizedFixture : public FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, run_interleave>
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class FullyConnectedLayerValidationQuantizedFixture : public FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
     void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, bool transpose_weights, bool reshape_weights, DataType data_type,
                QuantizationInfo quantization_info)
     {
-        FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, run_interleave>::setup(input_shape, weights_shape, bias_shape, output_shape, transpose_weights,
-                                                                                                                      reshape_weights, data_type,
-                                                                                                                      quantization_info);
+        FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, transpose_weights,
+                                                                                                      reshape_weights, data_type,
+                                                                                                      quantization_info);
     }
 };
 } // namespace validation