COMPMID-3212 Investigate CLReduceMean failure with QASYMM8v20.02

Signed-off-by: Luca Foschiani <luca.foschiani@arm.com>
Change-Id: I96d8cbef2e167e09484b0e9b5d8a12ffee0f6e67
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2755
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 30 insertions, 14 deletions

diff --git a/tests/validation/CL/ReduceMean.cpp b/tests/validation/CL/ReduceMean.cpp
index 5069711296..55594eda4a 100644
--- a/tests/validation/CL/ReduceMean.cpp
+++ b/tests/validation/CL/ReduceMean.cpp
@@ -44,7 +44,7 @@ namespace
 {
 constexpr AbsoluteTolerance<float>   tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for 32-bit floating-point type */
 constexpr AbsoluteTolerance<float>   tolerance_f16(0.03f);  /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */
-constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(2);  /**< Tolerance value for comparing reference's output against implementation's output for 8-bit asymmetric quantized type */
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1);  /**< Tolerance value for comparing reference's output against implementation's output for 8-bit asymmetric quantized type */
 
 const auto axis_keep = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1, 0), Coordinates(1, 2), Coordinates(0, 2), Coordinates(1, 3), Coordinates(0, 1, 2, 3) }),
                                framework::dataset::make("KeepDims", { true }));
@@ -133,7 +133,7 @@ TEST_SUITE(QASYMM8)
 FIXTURE_DATA_TEST_CASE(RunSmall,
                        CLReduceMeanQuantizedFixture<uint8_t>,
                        framework::DatasetMode::PRECOMMIT,
-                       combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 0) })))
+                       combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 5) })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_qasymm8);
@@ -142,7 +142,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall,
 FIXTURE_DATA_TEST_CASE(RunLarge,
                        CLReduceMeanQuantizedFixture<uint8_t>,
                        framework::DatasetMode::NIGHTLY,
-                       combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 0) })))
+                       combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 5) })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_qasymm8);
diff --git a/tests/validation/NEON/ReduceMean.cpp b/tests/validation/NEON/ReduceMean.cpp
index 210c3670d4..782b97220e 100644
--- a/tests/validation/NEON/ReduceMean.cpp
+++ b/tests/validation/NEON/ReduceMean.cpp
@@ -46,7 +46,7 @@ constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value f
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 constexpr AbsoluteTolerance<float> tolerance_f16(0.03f);   /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */
 #endif                                                     // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(2); /**< Tolerance value for comparing reference's output against implementation's output for 8-bit asymmetric quantized type */
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for 8-bit asymmetric quantized type */
 
 const auto axis_keep = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1, 0), Coordinates(1, 2), Coordinates(0, 2), Coordinates(1, 3), Coordinates(0, 1, 2, 3) }),
                                framework::dataset::make("KeepDims", { true }));
@@ -159,7 +159,7 @@ TEST_SUITE(QASYMM8)
 FIXTURE_DATA_TEST_CASE(RunSmall,
                        NEReduceMeanQuantizedFixture<uint8_t>,
                        framework::DatasetMode::PRECOMMIT,
-                       combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 0) })))
+                       combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 5) })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8);
@@ -168,7 +168,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall,
 FIXTURE_DATA_TEST_CASE(RunLarge,
                        NEReduceMeanQuantizedFixture<uint8_t>,
                        framework::DatasetMode::NIGHTLY,
-                       combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 0) })))
+                       combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 5) })))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8);
diff --git a/tests/validation/reference/ReductionOperation.cpp b/tests/validation/reference/ReductionOperation.cpp
index 9b35cdf6f5..a0b6a8eaed 100644
--- a/tests/validation/reference/ReductionOperation.cpp
+++ b/tests/validation/reference/ReductionOperation.cpp
@@ -64,7 +64,7 @@ OT reduce_operation(const T *ptr, int reduce_elements, ReductionOperation op, in
 
     if(std::is_integral<type>::value)
     {
-        auto int_res = static_cast<uint32_t>(res);
+        auto int_res = static_cast<int32_t>(res);
         for(int i = 0; i < reduce_elements; ++i)
         {
             auto elem = *(ptr + stride * i);
@@ -101,7 +101,7 @@ OT reduce_operation(const T *ptr, int reduce_elements, ReductionOperation op, in
         {
             int_res /= reduce_elements;
         }
-        res = saturate_cast<type>(int_res);
+        res = static_cast<type>(int_res);
     }
     else
     {
@@ -279,9 +279,17 @@ SimpleTensor<uint8_t> reduction_operation(const SimpleTensor<uint8_t> &src, cons
 {
     if(src.data_type() == DataType::QASYMM8)
     {
-        SimpleTensor<float> src_f = convert_from_asymmetric(src);
-        SimpleTensor<float> dst_f = reference::reduction_operation<float, float>(src_f, dst_shape, axis, op);
-        return convert_to_asymmetric<uint8_t>(dst_f, src.quantization_info());
+        // If the operation is MEAN_SUM, we can directly use the uint8 implementation without taking into account scale and offset
+        if(op == ReductionOperation::MEAN_SUM)
+        {
+            return compute_reduction_operation<uint8_t, uint8_t>(src, dst_shape, axis, op);
+        }
+        else
+        {
+            SimpleTensor<float> src_f = convert_from_asymmetric(src);
+            SimpleTensor<float> dst_f = reference::reduction_operation<float, float>(src_f, dst_shape, axis, op);
+            return convert_to_asymmetric<uint8_t>(dst_f, src.quantization_info());
+        }
     }
     else
     {
@@ -294,9 +302,17 @@ SimpleTensor<int8_t> reduction_operation(const SimpleTensor<int8_t> &src, const
 {
     if(src.data_type() == DataType::QASYMM8_SIGNED)
     {
-        SimpleTensor<float> src_f = convert_from_asymmetric(src);
-        SimpleTensor<float> dst_f = reference::reduction_operation<float, float>(src_f, dst_shape, axis, op);
-        return convert_to_asymmetric<int8_t>(dst_f, src.quantization_info());
+        // If the operation is MEAN_SUM, we can directly use the int8 implementation without taking into account scale and offset
+        if(op == ReductionOperation::MEAN_SUM)
+        {
+            return compute_reduction_operation<int8_t, int8_t>(src, dst_shape, axis, op);
+        }
+        else
+        {
+            SimpleTensor<float> src_f = convert_from_asymmetric(src);
+            SimpleTensor<float> dst_f = reference::reduction_operation<float, float>(src_f, dst_shape, axis, op);
+            return convert_to_asymmetric<int8_t>(dst_f, src.quantization_info());
+        }
     }
     else
     {