Optimize Quantized/Integer Bilinear Scale for Neon™

This patch introduces several performance optimizations regarding the Bilinear Scale operator with REPLICATE Border mode. Changes apply only to NHWC.

This patch
   - Reduces the memory footprint by disabling precomputation of indices and weights when they're not used
   - Rewrites the kernels for QASYMM8/QASYMM8_SIGNED/U8(Uint8)
   - Adds S8(Int8) Bilinear Scale for Border mode REPLICATE
   - Removes Bilinear Scale SVE kernels for Quantized and Integer types and adjust the heuristics to choose the Neon™ implementation
   - Adds new test cases where the input and output of the Bilinear Scale operator have different quantization scale and offset

Resolves: COMPMID-5453, COMPMID-5454

Change-Id: I3d251e76e0c6978fd5a0a1795ec62ab536bec93c
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8250
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 110 insertions, 49 deletions

diff --git a/tests/validation/NEON/Scale.cpp b/tests/validation/NEON/Scale.cpp
index e386d804ca..9927e21490 100644
--- a/tests/validation/NEON/Scale.cpp
+++ b/tests/validation/NEON/Scale.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,16 +22,10 @@
  * SOFTWARE.
  */
 #include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/NEON/functions/NEScale.h"
-#include "arm_compute/runtime/Tensor.h"
-#include "arm_compute/runtime/TensorAllocator.h"
 #include "tests/NEON/Accessor.h"
-#include "tests/PaddingCalculator.h"
 #include "tests/datasets/ScaleValidationDataset.h"
-#include "tests/framework/Asserts.h"
 #include "tests/framework/Macros.h"
-#include "tests/validation/Helpers.h"
 #include "tests/validation/Validation.h"
 #include "tests/validation/fixtures/ScaleFixture.h"
 
@@ -51,7 +45,7 @@ using datasets::ScaleAlignCornersSamplingPolicySet;
 
 /** We consider vector size in byte 64 since the maximum size of
  * a vector used by the kernel is currently 64-byte (float32x4x4).
- * There are possibility to reduce test time further by using
+ * There is possibility to reduce test time further by using
  * smaller vector sizes for different data types where applicable.
  */
 constexpr uint32_t vector_byte = 64;
@@ -62,26 +56,31 @@ constexpr uint32_t num_elements_per_vector()
     return vector_byte / sizeof(T);
 }
 
-/** Scale data types */
-const auto ScaleDataTypes = framework::dataset::make("DataType",
+/** Quantization information data set */
+const auto QuantizationInfoSet = framework::dataset::make("QuantizationInfo",
 {
-    DataType::U8,
-    DataType::S16,
-    DataType::F32,
+    QuantizationInfo(0.5f, -10),
 });
 
 /** Quantization information data set */
-const auto QuantizationInfoSet = framework::dataset::make("QuantizationInfo",
+const auto InputQuantizationInfoSet = framework::dataset::make("InputQuantizationInfo",
 {
     QuantizationInfo(0.5f, -10),
 });
 
+/** Quantization information data set */
+const auto OutputQuantizationInfoSet = framework::dataset::make("OutputQuantizationInfo",
+{
+    QuantizationInfo(0.2f, 20),
+});
+
 /** Tolerance */
 constexpr AbsoluteTolerance<uint8_t> tolerance_u8(1);
+constexpr AbsoluteTolerance<int8_t>  tolerance_s8(1);
 constexpr AbsoluteTolerance<int16_t> tolerance_s16(1);
 RelativeTolerance<float>             tolerance_f32(0.05);
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-constexpr float          abs_tolerance_f16(0.01f);
+constexpr float         abs_tolerance_f16(0.01f);
 RelativeTolerance<half> tolerance_f16(half(0.1));
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
@@ -321,6 +320,8 @@ using NEScaleMixedDataLayoutFixture = ScaleValidationFixture<Tensor, Accessor, N
 template <typename T>
 using NEScaleQuantizedFixture = ScaleValidationQuantizedFixture<Tensor, Accessor, NEScale, T>;
 template <typename T>
+using NEScaleDifferentOutputQuantizedFixture = ScaleValidationDifferentOutputQuantizedFixture<Tensor, Accessor, NEScale, T>;
+template <typename T>
 using NEScaleQuantizedMixedDataLayoutFixture = ScaleValidationQuantizedFixture<Tensor, Accessor, NEScale, T, true>;
 
 TEST_SUITE(Float)
@@ -457,6 +458,27 @@ FIXTURE_DATA_TEST_CASE(RunSmallAlignCorners, NEScaleFixture<uint8_t>, framework:
     validate(Accessor(_target), _reference, valid_region, tolerance_u8);
 }
 TEST_SUITE_END() // U8
+TEST_SUITE(S8)
+const auto s8_shape = combine((SCALE_SHAPE_DATASET(num_elements_per_vector<int8_t>())), framework::dataset::make("DataType", DataType::S8));
+FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleFixture<int8_t>, framework::DatasetMode::ALL, ASSEMBLE_S8_DATASET(s8_shape, ScaleSamplingPolicySet))
+{
+    //Create valid region
+    TensorInfo  src_info(_shape, 1, _data_type);
+    ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED));
+
+    // Validate output
+    validate(Accessor(_target), _reference, valid_region, tolerance_s8);
+}
+FIXTURE_DATA_TEST_CASE(RunSmallAlignCorners, NEScaleFixture<int8_t>, framework::DatasetMode::ALL, ASSEMBLE_S8_DATASET(s8_shape, ScaleAlignCornersSamplingPolicySet))
+{
+    //Create valid region
+    TensorInfo  src_info(_shape, 1, _data_type);
+    ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED));
+
+    // Validate output
+    validate(Accessor(_target), _reference, valid_region, tolerance_s8);
+}
+TEST_SUITE_END() // S8
 TEST_SUITE(S16)
 const auto s16_shape = combine((SCALE_SHAPE_DATASET(num_elements_per_vector<int16_t>())), framework::dataset::make("DataType", DataType::S16));
 FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleFixture<int16_t>, framework::DatasetMode::ALL, ASSEMBLE_DATASET(s16_shape, ScaleSamplingPolicySet))
@@ -492,6 +514,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleQuantizedFixture<uint8_t>, framework::Da
     // Validate output
     validate(Accessor(_target), _reference, valid_region, tolerance_u8);
 }
+FIXTURE_DATA_TEST_CASE(RunSmallDifferentOutputQuantization, NEScaleDifferentOutputQuantizedFixture<uint8_t>, framework::DatasetMode::ALL,
+                       ASSEMBLE_DIFFERENTLY_QUANTIZED_DATASET(qasymm8_shape, ScaleSamplingPolicySet, InputQuantizationInfoSet, OutputQuantizationInfoSet))
+{
+    //Create valid region
+    TensorInfo  src_info(_shape, 1, _data_type);
+    ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED));
+
+    // Validate output
+    validate(Accessor(_target), _reference, valid_region, tolerance_u8);
+}
 FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEScaleQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::ALL, ASSEMBLE_QUANTIZED_DATASET(qasymm8_shape, ScaleSamplingPolicySet,
                        QuantizationInfoSet))
 {
@@ -525,6 +557,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleQuantizedFixture<int8_t>, framework::Dat
     // Validate output
     validate(Accessor(_target), _reference, valid_region, tolerance_qasymm8_signed);
 }
+FIXTURE_DATA_TEST_CASE(RunSmallDifferentOutputQuantization, NEScaleDifferentOutputQuantizedFixture<int8_t>, framework::DatasetMode::ALL,
+                       ASSEMBLE_DIFFERENTLY_QUANTIZED_DATASET(qasymm8_signed_shape, ScaleSamplingPolicySet, InputQuantizationInfoSet, OutputQuantizationInfoSet))
+{
+    //Create valid region
+    TensorInfo  src_info(_shape, 1, _data_type);
+    ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED));
+
+    // Validate output
+    validate(Accessor(_target), _reference, valid_region, tolerance_qasymm8_signed);
+}
 FIXTURE_DATA_TEST_CASE(RunSmallAlignCorners, NEScaleQuantizedFixture<int8_t>, framework::DatasetMode::ALL, ASSEMBLE_QUANTIZED_DATASET(qasymm8_signed_shape, ScaleAlignCornersSamplingPolicySet,
                        QuantizationInfoSet))
 {
diff --git a/tests/validation/fixtures/ScaleFixture.h b/tests/validation/fixtures/ScaleFixture.h
index c0b44bcb5f..72feb62016 100644
--- a/tests/validation/fixtures/ScaleFixture.h
+++ b/tests/validation/fixtures/ScaleFixture.h
@@ -24,12 +24,6 @@
 #ifndef ARM_COMPUTE_TEST_SCALE_FIXTURE
 #define ARM_COMPUTE_TEST_SCALE_FIXTURE
 
-#include "arm_compute/core/TensorShape.h"
-#include "arm_compute/core/Types.h"
-#include "tests/AssetsLibrary.h"
-#include "tests/Globals.h"
-#include "tests/IAccessor.h"
-#include "tests/framework/Asserts.h"
 #include "tests/framework/Fixture.h"
 #include "tests/validation/reference/Permute.h"
 #include "tests/validation/reference/Scale.h"
@@ -46,16 +40,17 @@ class ScaleValidationGenericFixture : public framework::Fixture
 public:
     template <typename...>
     void setup(TensorShape shape, DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout, InterpolationPolicy policy, BorderMode border_mode, SamplingPolicy sampling_policy,
-               bool align_corners, bool mixed_layout)
+               bool align_corners, bool mixed_layout, QuantizationInfo output_quantization_info)
     {
-        _shape             = shape;
-        _policy            = policy;
-        _border_mode       = border_mode;
-        _sampling_policy   = sampling_policy;
-        _data_type         = data_type;
-        _quantization_info = quantization_info;
-        _align_corners     = align_corners;
-        _mixed_layout      = mixed_layout;
+        _shape                    = shape;
+        _policy                   = policy;
+        _border_mode              = border_mode;
+        _sampling_policy          = sampling_policy;
+        _data_type                = data_type;
+        _input_quantization_info  = quantization_info;
+        _output_quantization_info = output_quantization_info;
+        _align_corners            = align_corners;
+        _mixed_layout             = mixed_layout;
 
         generate_scale(shape);
 
@@ -144,7 +139,7 @@ protected:
         }
 
         // Create tensors
-        TensorType src = create_tensor<TensorType>(shape, _data_type, 1, _quantization_info, data_layout);
+        TensorType src = create_tensor<TensorType>(shape, _data_type, 1, _input_quantization_info, data_layout);
 
         const int idx_width  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
         const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
@@ -152,7 +147,7 @@ protected:
         TensorShape shape_scaled(shape);
         shape_scaled.set(idx_width, shape[idx_width] * _scale_x, /* apply_dim_correction = */ false);
         shape_scaled.set(idx_height, shape[idx_height] * _scale_y, /* apply_dim_correction = */ false);
-        TensorType dst = create_tensor<TensorType>(shape_scaled, _data_type, 1, _quantization_info, data_layout);
+        TensorType dst = create_tensor<TensorType>(shape_scaled, _data_type, 1, _output_quantization_info, data_layout);
 
         // Create and configure function
         FunctionType scale;
@@ -188,12 +183,12 @@ protected:
     SimpleTensor<T> compute_reference(const TensorShape &shape)
     {
         // Create reference
-        SimpleTensor<T> src{ shape, _data_type, 1, _quantization_info };
+        SimpleTensor<T> src{ shape, _data_type, 1, _input_quantization_info };
 
         // Fill reference
         fill(src);
 
-        return reference::scale<T>(src, _scale_x, _scale_y, _policy, _border_mode, _constant_border_value, _sampling_policy, /* ceil_policy_scale */ false, _align_corners);
+        return reference::scale<T>(src, _scale_x, _scale_y, _policy, _border_mode, _constant_border_value, _sampling_policy, /* ceil_policy_scale */ false, _align_corners, _output_quantization_info);
     }
 
     TensorType          _target{};
@@ -204,7 +199,8 @@ protected:
     T                   _constant_border_value{};
     SamplingPolicy      _sampling_policy{};
     DataType            _data_type{};
-    QuantizationInfo    _quantization_info{};
+    QuantizationInfo    _input_quantization_info{};
+    QuantizationInfo    _output_quantization_info{};
     bool                _align_corners{ false };
     bool                _mixed_layout{ false };
     float               _scale_x{ 1.f };
@@ -227,7 +223,29 @@ public:
                                                                                         border_mode,
                                                                                         sampling_policy,
                                                                                         align_corners,
-                                                                                        mixed_layout);
+                                                                                        mixed_layout,
+                                                                                        quantization_info);
+    }
+};
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false>
+class ScaleValidationDifferentOutputQuantizedFixture : public ScaleValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape, DataType data_type, QuantizationInfo input_quantization_info, QuantizationInfo output_quantization_info, DataLayout data_layout, InterpolationPolicy policy,
+               BorderMode border_mode, SamplingPolicy sampling_policy,
+               bool align_corners)
+    {
+        ScaleValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape,
+                                                                                        data_type,
+                                                                                        input_quantization_info,
+                                                                                        data_layout,
+                                                                                        policy,
+                                                                                        border_mode,
+                                                                                        sampling_policy,
+                                                                                        align_corners,
+                                                                                        mixed_layout,
+                                                                                        output_quantization_info);
     }
 };
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false>
@@ -245,7 +263,8 @@ public:
                                                                                         border_mode,
                                                                                         sampling_policy,
                                                                                         align_corners,
-                                                                                        mixed_layout);
+                                                                                        mixed_layout,
+                                                                                        QuantizationInfo());
     }
 };
 } // namespace validation
diff --git a/tests/validation/reference/Scale.cpp b/tests/validation/reference/Scale.cpp
index 71e98fd776..2f429cb29b 100644
--- a/tests/validation/reference/Scale.cpp
+++ b/tests/validation/reference/Scale.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2020, 2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,7 +25,6 @@
 #include "Scale.h"
 
 #include "Utils.h"
-#include "arm_compute/core/utils/misc/Utility.h"
 #include "src/core/utils/ScaleUtils.h"
 #include "support/Rounding.h"
 
@@ -183,14 +182,15 @@ SimpleTensor<T> scale_core(const SimpleTensor<T> &in, float scale_x, float scale
 
 template <typename T>
 SimpleTensor<T> scale(const SimpleTensor<T> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, T constant_border_value,
-                      SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners)
+                      SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info)
 {
+    ARM_COMPUTE_UNUSED(output_quantization_info);
     return scale_core<T>(src, scale_x, scale_y, policy, border_mode, constant_border_value, sampling_policy, ceil_policy_scale, align_corners);
 }
 
 template <>
 SimpleTensor<uint8_t> scale(const SimpleTensor<uint8_t> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, uint8_t constant_border_value,
-                            SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners)
+                            SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info)
 {
     SimpleTensor<uint8_t> dst;
     if(src.quantization_info().uniform().scale != 0.f)
@@ -198,7 +198,7 @@ SimpleTensor<uint8_t> scale(const SimpleTensor<uint8_t> &src, float scale_x, flo
         SimpleTensor<float> src_tmp                 = convert_from_asymmetric(src);
         float               constant_border_value_f = dequantize_qasymm8(constant_border_value, src.quantization_info());
         SimpleTensor<float> dst_tmp                 = scale_core<float>(src_tmp, scale_x, scale_y, policy, border_mode, constant_border_value_f, sampling_policy, ceil_policy_scale, align_corners);
-        dst                                         = convert_to_asymmetric<uint8_t>(dst_tmp, src.quantization_info());
+        dst                                         = convert_to_asymmetric<uint8_t>(dst_tmp, output_quantization_info);
     }
     else
     {
@@ -209,7 +209,7 @@ SimpleTensor<uint8_t> scale(const SimpleTensor<uint8_t> &src, float scale_x, flo
 
 template <>
 SimpleTensor<int8_t> scale(const SimpleTensor<int8_t> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, int8_t constant_border_value,
-                           SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners)
+                           SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info)
 {
     SimpleTensor<int8_t> dst;
     if(src.quantization_info().uniform().scale != 0.f)
@@ -217,7 +217,7 @@ SimpleTensor<int8_t> scale(const SimpleTensor<int8_t> &src, float scale_x, float
         SimpleTensor<float> src_tmp                 = convert_from_asymmetric(src);
         float               constant_border_value_f = dequantize_qasymm8_signed(constant_border_value, src.quantization_info());
         SimpleTensor<float> dst_tmp                 = scale_core<float>(src_tmp, scale_x, scale_y, policy, border_mode, constant_border_value_f, sampling_policy, ceil_policy_scale, align_corners);
-        dst                                         = convert_to_asymmetric<int8_t>(dst_tmp, src.quantization_info());
+        dst                                         = convert_to_asymmetric<int8_t>(dst_tmp, output_quantization_info);
     }
     else
     {
@@ -227,11 +227,11 @@ SimpleTensor<int8_t> scale(const SimpleTensor<int8_t> &src, float scale_x, float
 }
 
 template SimpleTensor<int16_t> scale(const SimpleTensor<int16_t> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, int16_t constant_border_value,
-                                     SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners);
+                                     SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info);
 template SimpleTensor<half> scale(const SimpleTensor<half> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, half constant_border_value,
-                                  SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners);
+                                  SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info);
 template SimpleTensor<float> scale(const SimpleTensor<float> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, float constant_border_value,
-                                   SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners);
+                                   SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/reference/Scale.h b/tests/validation/reference/Scale.h
index c66af8d94e..c32c07d1c0 100644
--- a/tests/validation/reference/Scale.h
+++ b/tests/validation/reference/Scale.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2020, 2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,7 +37,7 @@ namespace reference
 {
 template <typename T>
 SimpleTensor<T> scale(const SimpleTensor<T> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, T constant_border_value = 0,
-                      SamplingPolicy sampling_policy = SamplingPolicy::CENTER, bool ceil_policy_scale = false, bool align_corners = false);
+                      SamplingPolicy sampling_policy = SamplingPolicy::CENTER, bool ceil_policy_scale = false, bool align_corners = false, QuantizationInfo output_quantization_info = QuantizationInfo());
 } // namespace reference
 } // namespace validation
 } // namespace test