COMPMID-2418: CLDequantizationLayer support for QASYMM8_PER_CHANNEL

Add support for QASYMM8_PER_CHANNEL in CLDequantiazationLayer.
Added tests for NHWC and also updated NEON code to work with NHWC
data layout.
Cleaned up the reference implementation.

Change-Id: Ic1d51f16f7f625503fffdbbb66f6487aa588f08c
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1828
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

4 files changed, 73 insertions, 39 deletions

diff --git a/tests/validation/CL/DequantizationLayer.cpp b/tests/validation/CL/DequantizationLayer.cpp
index 2ef8c60998..acc0022d3e 100644
--- a/tests/validation/CL/DequantizationLayer.cpp
+++ b/tests/validation/CL/DequantizationLayer.cpp
@@ -41,6 +41,33 @@ namespace test
 {
 namespace validation
 {
+namespace
+{
+const auto dataset_quant_f32 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
+                                               framework::dataset::make("DataType", DataType::F32)),
+                                       framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_f16 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
+                                               framework::dataset::make("DataType", DataType::F16)),
+                                       framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_per_channel_f32 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedPerChannelTypes()),
+                                                           framework::dataset::make("DataType", DataType::F32)),
+                                                   framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+const auto dataset_quant_per_channel_f16 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedPerChannelTypes()),
+                                                           framework::dataset::make("DataType", DataType::F16)),
+                                                   framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+const auto dataset_quant_nightly_f32 = combine(combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
+                                                       framework::dataset::make("DataType", DataType::F32)),
+                                               framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_nightly_f16 = combine(combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
+                                                       framework::dataset::make("DataType", DataType::F16)),
+                                               framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_per_channel_nightly_f32 = combine(combine(combine(datasets::LargeShapes(), datasets::QuantizedPerChannelTypes()),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                           framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+const auto dataset_quant_per_channel_nightly_f16 = combine(combine(combine(datasets::LargeShapes(), datasets::QuantizedPerChannelTypes()),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                           framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+} // namespace
 TEST_SUITE(CL)
 TEST_SUITE(DequantizationLayer)
 
@@ -97,14 +124,12 @@ template <typename T>
 using CLDequantizationLayerFixture = DequantizationValidationFixture<CLTensor, CLAccessor, CLDequantizationLayer, T>;
 
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
-                                                                                                                framework::dataset::make("DataType", DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, concat(dataset_quant_f16, dataset_quant_per_channel_f16))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, CLDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
-                                                                                                              framework::dataset::make("DataType", DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, concat(dataset_quant_nightly_f16, dataset_quant_per_channel_nightly_f16))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -112,14 +137,12 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLDequantizationLayerFixture<half>, framework::
 TEST_SUITE_END() // FP16
 
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
-                                                                                                                 framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, concat(dataset_quant_f32, dataset_quant_per_channel_f32))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, CLDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
-                                                                                                               framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, concat(dataset_quant_nightly_f32, dataset_quant_per_channel_nightly_f32))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
diff --git a/tests/validation/NEON/DequantizationLayer.cpp b/tests/validation/NEON/DequantizationLayer.cpp
index 005ed6900c..0dce76a933 100644
--- a/tests/validation/NEON/DequantizationLayer.cpp
+++ b/tests/validation/NEON/DequantizationLayer.cpp
@@ -48,6 +48,31 @@ const auto data_types = framework::dataset::make("DataType", { DataType::F16, Da
 #else  /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 const auto data_types = framework::dataset::make("DataType", { DataType::F32 });
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
+const auto dataset_quant_f32 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
+                                               framework::dataset::make("DataType", DataType::F32)),
+                                       framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_f16 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
+                                               framework::dataset::make("DataType", DataType::F16)),
+                                       framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_per_channel_f32 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedPerChannelTypes()),
+                                                           framework::dataset::make("DataType", DataType::F32)),
+                                                   framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+const auto dataset_quant_per_channel_f16 = combine(combine(combine(datasets::SmallShapes(), datasets::QuantizedPerChannelTypes()),
+                                                           framework::dataset::make("DataType", DataType::F16)),
+                                                   framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+const auto dataset_quant_nightly_f32 = combine(combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
+                                                       framework::dataset::make("DataType", DataType::F32)),
+                                               framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_nightly_f16 = combine(combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
+                                                       framework::dataset::make("DataType", DataType::F16)),
+                                               framework::dataset::make("DataLayout", { DataLayout::NCHW }));
+const auto dataset_quant_per_channel_nightly_f32 = combine(combine(combine(datasets::LargeShapes(), datasets::QuantizedPerChannelTypes()),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                           framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
+const auto dataset_quant_per_channel_nightly_f16 = combine(combine(combine(datasets::LargeShapes(), datasets::QuantizedPerChannelTypes()),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                           framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }));
 } // namespace
 
 TEST_SUITE(NEON)
@@ -107,14 +132,12 @@ using NEDequantizationLayerFixture = DequantizationValidationFixture<Tensor, Acc
 
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
-                                                                                                                framework::dataset::make("DataType", DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, concat(dataset_quant_f16, dataset_quant_per_channel_f16))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
-                                                                                                              framework::dataset::make("DataType", DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, concat(dataset_quant_nightly_f16, dataset_quant_per_channel_nightly_f16))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -123,16 +146,12 @@ TEST_SUITE_END() // FP16
 #endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), concat(datasets::QuantizedTypes(),
-                                                                                                                 datasets::QuantizedPerChannelTypes())),
-                                                                                                                 framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, concat(dataset_quant_f32, dataset_quant_per_channel_f32))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), concat(datasets::QuantizedTypes(),
-                                                                                                                       datasets::QuantizedPerChannelTypes())),
-                                                                                                               framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, concat(dataset_quant_nightly_f32, dataset_quant_per_channel_nightly_f32))
 {
     // Validate output
     validate(Accessor(_target), _reference);
diff --git a/tests/validation/fixtures/DequantizationLayerFixture.h b/tests/validation/fixtures/DequantizationLayerFixture.h
index 4842ee1c59..c7a818fcc7 100644
--- a/tests/validation/fixtures/DequantizationLayerFixture.h
+++ b/tests/validation/fixtures/DequantizationLayerFixture.h
@@ -32,6 +32,7 @@
 #include "tests/IAccessor.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Fixture.h"
+#include "tests/validation/Helpers.h"
 #include "tests/validation/reference/DequantizationLayer.h"
 
 #include <random>
@@ -47,10 +48,10 @@ class DequantizationValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(TensorShape shape, DataType src_data_type, DataType dst_datatype)
+    void setup(TensorShape shape, DataType src_data_type, DataType dst_datatype, DataLayout data_layout)
     {
         _quantization_info = generate_quantization_info(src_data_type, shape.z());
-        _target            = compute_target(shape, src_data_type, dst_datatype);
+        _target            = compute_target(shape, src_data_type, dst_datatype, data_layout);
         _reference         = compute_reference(shape, src_data_type);
     }
 
@@ -61,11 +62,16 @@ protected:
         library->fill_tensor_uniform(tensor, 0);
     }
 
-    TensorType compute_target(const TensorShape &shape, DataType src_data_type, DataType dst_datatype)
+    TensorType compute_target(TensorShape shape, DataType src_data_type, DataType dst_datatype, DataLayout data_layout)
     {
+        if(data_layout == DataLayout::NHWC)
+        {
+            permute(shape, PermutationVector(2U, 0U, 1U));
+        }
+
         // Create tensors
-        TensorType src = create_tensor<TensorType>(shape, src_data_type, 1, _quantization_info);
-        TensorType dst = create_tensor<TensorType>(shape, dst_datatype);
+        TensorType src = create_tensor<TensorType>(shape, src_data_type, 1, _quantization_info, data_layout);
+        TensorType dst = create_tensor<TensorType>(shape, dst_datatype, 1, QuantizationInfo(), data_layout);
 
         // Create and configure function
         FunctionType dequantization_layer;
diff --git a/tests/validation/reference/DequantizationLayer.cpp b/tests/validation/reference/DequantizationLayer.cpp
index 74686bdaaf..69a49a3d6d 100644
--- a/tests/validation/reference/DequantizationLayer.cpp
+++ b/tests/validation/reference/DequantizationLayer.cpp
@@ -50,9 +50,9 @@ TOut dequantize(int16_t val, const UniformQuantizationInfo qinfo)
 {
     return static_cast<TOut>(dequantize_qsymm16(val, qinfo));
 }
-
+} // namespace
 template <typename TOut, typename TIn>
-SimpleTensor<TOut> dequantization_layer_nchw(const SimpleTensor<TIn> &src)
+SimpleTensor<TOut> dequantization_layer(const SimpleTensor<TIn> &src)
 {
     const DataType src_data_type = src.data_type();
     const DataType dst_data_type = std::is_same<TOut, float>::value ? DataType::F32 : DataType::F16;
@@ -97,20 +97,6 @@ SimpleTensor<TOut> dequantization_layer_nchw(const SimpleTensor<TIn> &src)
 
     return dst;
 }
-} // namespace
-template <typename TOut, typename TIn>
-SimpleTensor<TOut> dequantization_layer(const SimpleTensor<TIn> &src)
-{
-    if(src.data_layout() == DataLayout::NHWC && src.data_type() == DataType::QSYMM8_PER_CHANNEL)
-    {
-        SimpleTensor<TIn> src_nchw = reference::permute<TIn>(src, PermutationVector(1U, 2U, 0U));
-        return reference::permute<TOut>(dequantization_layer_nchw<TOut>(src_nchw), PermutationVector(2U, 0U, 1U));
-    }
-    else
-    {
-        return dequantization_layer_nchw<TOut>(src);
-    }
-}
 
 template SimpleTensor<half> dequantization_layer(const SimpleTensor<uint8_t> &src);
 template SimpleTensor<float> dequantization_layer(const SimpleTensor<uint8_t> &src);