COMPMID-2805: Add QASYMM8_SIGNED support in NEGEMMLowpOutputStage

Add support from requantizing down from S32 to Int8 with fixed point
requantization. This involves the following:
- Compute fixed point multiplication between each entry of input by
  result_fixedpoint_multiplier
- Add bias to final result if bias tensor is not a nullptr
- Round to nearest division by a power-of-two using result_shift
- Add offset to each result
- Clamp the value between the specified min and max bounds
- Cast to int8 data type

Change-Id: I641b3fac0833c568d8565ccb859bbc561a24c17d
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2340
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 282 insertions, 99 deletions

diff --git a/tests/validate_examples/cl_gemm.cpp b/tests/validate_examples/cl_gemm.cpp
index 128c5f6e7f..39fe111448 100644
--- a/tests/validate_examples/cl_gemm.cpp
+++ b/tests/validate_examples/cl_gemm.cpp
@@ -321,11 +321,11 @@ public:
                     SimpleTensor<int32_t> biases{ TensorShape(N), DataType::S32, 1 };
                     // Fill bias
                     fill(biases, 3);
-                    ref_dst = reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(ref_tmp_dst, biases, dst_multiplier_vec, dst_shift_vec, offset_dst);
+                    ref_dst = reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(ref_tmp_dst, biases, dst_multiplier_vec, dst_shift_vec, offset_dst);
                 }
                 else
                 {
-                    ref_dst = reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(ref_tmp_dst, dst_multiplier_vec, dst_shift_vec, offset_dst);
+                    ref_dst = reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(ref_tmp_dst, dst_multiplier_vec, dst_shift_vec, offset_dst);
                 }
                 validate(CLAccessor(dst), ref_dst);
                 break;
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index b79523da1a..78fbc5845f 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -410,6 +410,119 @@ TEST_SUITE_END() // BoundedReLu
 
 TEST_SUITE_END() // QuantizeDownInt32ToUint8ScaleByFixedPoint
 
+TEST_SUITE(QuantizeDownInt32ToInt8ScaleByFixedPoint)
+
+const auto quantize_down_int32_to_int8_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
+                                                                   2)
+                                                                   * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0) * framework::dataset::make("max", 0) * framework::dataset::make("addBias", { false, true });
+
+const auto quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
+                                                                        2)
+                                                                        * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true });
+
+using NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture =
+    GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint>;
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
+        framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::F32), // Invalid input data type
+                                                 TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Invalid min and max
+                                                 TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type
+                                                 TensorInfo(TensorShape(21U, 13U), 1, DataType::S32),
+        }),
+        framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32),
+                                                TensorInfo(TensorShape(21U), 1, DataType::S32),
+                                                TensorInfo(TensorShape(20U), 1, DataType::S32),
+                                                TensorInfo(TensorShape(21U), 1, DataType::S32),
+        })),
+        framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED),
+                                                TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED),
+                                                TensorInfo(TensorShape(20U, 13U), 1, DataType::S32),
+                                                TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED),
+        })),
+        framework::dataset::make("Min",{ -110,
+                                         -130,
+                                         -113,
+                                         -113,
+        })),
+        framework::dataset::make("Max",{ 87,
+                                         140,
+                                         97,
+                                         97,
+        })),
+        framework::dataset::make("Expected", { false, false, false, true })),
+               a_info, b_info, output_info, min, max, expected)
+{
+    // Lock tensors
+    Status status =  NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::validate(&a_info.clone()->set_is_resizable(false),
+                                                                                  &b_info.clone()->set_is_resizable(false),
+                                                                                  &output_info.clone()->set_is_resizable(false),
+                                                                                  min,
+                                                                                  max);
+    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                                                                   quantize_down_int32_to_int8_scale_by_fixedpoint_cases),
+               shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias)
+{
+    TensorShape shape_bias(shape[0]);
+
+    // Create tensors
+    Tensor in   = create_tensor<Tensor>(shape, DataType::S32);
+    Tensor bias = create_tensor<Tensor>(shape_bias, DataType::S32);
+    Tensor out  = create_tensor<Tensor>(shape, DataType::QASYMM8_SIGNED);
+
+    ARM_COMPUTE_EXPECT(in.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(out.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint output_stage;
+    output_stage.configure(&in, add_bias ? &bias : nullptr, &out, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
+
+    // Validate valid region input and output
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(in.info()->valid_region(), valid_region);
+    validate(out.info()->valid_region(), valid_region);
+
+    // Validate valid region bias
+    if(add_bias)
+    {
+        const ValidRegion valid_region_bias = shape_to_valid_region(shape_bias);
+        validate(bias.info()->valid_region(), valid_region_bias);
+    }
+
+    // Validate padding
+    const PaddingSize padding(0);
+    validate(in.info()->padding(), padding);
+    validate(out.info()->padding(), padding);
+
+    if(add_bias)
+    {
+        validate(bias.info()->padding(), padding);
+    }
+}
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int8_scale_by_fixedpoint_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+TEST_SUITE(BoundedReLu)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // BoundedReLu
+TEST_SUITE_END() // QuantizeDownInt32ToInt8ScaleByFixedPoint
+
 TEST_SUITE(QuantizeDownInt32ToInt16ScaleByFixedPoint)
 
 const auto quantize_down_int32_to_int16_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index 5d092ecac2..c17105edad 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -254,8 +254,8 @@ protected:
                                                                                        output_stage.gemmlowp_offset, output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound);
                 break;
             case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
-                return reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(output, bias,
-                                                                                                     output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_offset, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound);
+                return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(output, bias,
+                                                                                               output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_offset, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound);
                 break;
             default:
                 ARM_COMPUTE_ERROR("Not Supported!");
@@ -361,6 +361,101 @@ protected:
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType>
+class GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias)
+    {
+        _target    = compute_target(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias);
+        _reference = compute_reference(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_int_distribution<> distribution(-6000, 6000);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias)
+    {
+        TensorShape shape_bias(shape[0]);
+
+        // Create tensors
+        TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1);
+        TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1);
+        TensorType c = create_tensor<TensorType>(shape, DataType::QASYMM8_SIGNED, 1);
+
+        // Create and configure function
+        FunctionType output_stage;
+        output_stage.configure(&a, add_bias ? &b : nullptr, &c, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
+
+        ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        a.allocator()->allocate();
+        c.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensor
+        fill(AccessorType(a), 0);
+
+        if(add_bias)
+        {
+            ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+            // Allocate bias tensor
+            b.allocator()->allocate();
+
+            ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+            // Fill tensor
+            fill(AccessorType(b), 1);
+        }
+
+        // Compute GEMM function
+        output_stage.run();
+        return c;
+    }
+
+    SimpleTensor<int8_t> compute_reference(const TensorShape &shape, int32_t result_fixed_point_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max,
+                                           bool add_bias)
+    {
+        // Create reference
+        TensorShape shape_bias(shape[0]);
+
+        SimpleTensor<int32_t> a{ shape, DataType::S32, 1 };
+        SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 };
+
+        // Fill reference
+        fill(a, 0);
+
+        const std::vector<int32_t> result_fixed_point_multiplier_vec = { result_fixed_point_multiplier };
+        const std::vector<int32_t> result_shift_vec                  = { result_shift };
+
+        if(add_bias)
+        {
+            // Fill bias
+            fill(b, 1);
+
+            return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int8_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
+        }
+        else
+        {
+            return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int8_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
+        }
+    }
+
+    TensorType           _target{};
+    SimpleTensor<int8_t> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType>
 class GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture : public framework::Fixture
 {
 public:
@@ -443,11 +538,11 @@ protected:
             // Fill bias
             fill(b, 1);
 
-            return reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
+            return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
         }
         else
         {
-            return reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
+            return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
         }
     }
 
@@ -530,16 +625,19 @@ protected:
         // Fill reference
         fill(a, 0);
 
+        const std::vector<int32_t> result_fixed_point_multiplier_vec = { result_fixed_point_multiplier };
+        const std::vector<int32_t> result_shift_vec                  = { result_shift };
+
         if(add_bias)
         {
             // Fill bias
             fill(b, 1);
 
-            return reference::gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint<int32_t>(a, b, result_fixed_point_multiplier, result_shift, min, max);
+            return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int16_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, 0, min, max);
         }
         else
         {
-            return reference::gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint<int32_t>(a, result_fixed_point_multiplier, result_shift, min, max);
+            return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int16_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, 0, min, max);
         }
     }
 
diff --git a/tests/validation/reference/GEMMLowp.cpp b/tests/validation/reference/GEMMLowp.cpp
index 08be4a5182..4529b91a48 100644
--- a/tests/validation/reference/GEMMLowp.cpp
+++ b/tests/validation/reference/GEMMLowp.cpp
@@ -39,6 +39,28 @@ namespace reference
 namespace
 {
 template <typename T>
+struct DataTypeExtractor
+{
+    static DataType data_type()
+    {
+        DataType data_type = DataType::UNKNOWN;
+        if(std::is_same<T, int8_t>::value)
+        {
+            data_type = DataType::QASYMM8_SIGNED;
+        }
+        else if(std::is_same<T, uint8_t>::value)
+        {
+            data_type = DataType::QASYMM8;
+        }
+        else if(std::is_same<T, int16_t>::value)
+        {
+            data_type = DataType::QSYMM16;
+        }
+        return data_type;
+    }
+};
+
+template <typename T>
 void quantize_down_int32_to_uint8_scale(const SimpleTensor<T> *in, const SimpleTensor<T> *bias, SimpleTensor<uint8_t> *dst, int32_t result_offset, std::vector<int32_t> result_mult_int,
                                         std::vector<int32_t> result_shift, int32_t min, int32_t max)
 {
@@ -68,16 +90,16 @@ void quantize_down_int32_to_uint8_scale(const SimpleTensor<T> *in, const SimpleT
     }
 }
 
-template <typename T>
-void quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> *in, const SimpleTensor<T> *bias, SimpleTensor<uint8_t> *dst, std::vector<int32_t> result_fixedpoint_multiplier,
-                                                      std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max)
+template <typename TIn, typename TOut>
+void quantize_down_scale_by_fixedpoint(const SimpleTensor<TIn> *in, const SimpleTensor<TIn> *bias, SimpleTensor<TOut> *dst, std::vector<int32_t> result_fixedpoint_multiplier,
+                                       std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max)
 {
     const int  cols_in        = in->shape().x();
     const bool is_per_channel = result_fixedpoint_multiplier.size() > 1;
 
     for(int i = 0; i < in->num_elements(); ++i)
     {
-        int32_t result = (*in)[i];
+        TIn result = (*in)[i];
 
         if(bias != nullptr)
         {
@@ -88,43 +110,15 @@ void quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> *in,
         const int32_t multiplier = (is_per_channel) ? result_fixedpoint_multiplier[i % cols_in] : result_fixedpoint_multiplier[0];
         const int32_t shift      = (is_per_channel) ? result_shift[i % cols_in] : result_shift[0];
 
-        result = asymm_rounding_divide_by_pow2(asymm_int_mult(result, multiplier), shift);
-        result += result_offset_after_shift;
-
-        // Bounded ReLu
-        if(min != max)
+        if(shift < 0)
         {
-            result = std::max(min, std::min(max, result));
-        }
-
-        (*dst)[i] = static_cast<uint8_t>(std::max(0, std::min(255, result)));
-    }
-}
-
-template <typename T>
-void quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> *in, const SimpleTensor<T> *bias, SimpleTensor<int16_t> *dst, int32_t result_fixedpoint_multiplier, int32_t result_shift,
-                                                      int32_t min, int32_t max)
-{
-    const int cols_in = in->shape().x();
-
-    for(int i = 0; i < in->num_elements(); ++i)
-    {
-        int32_t result = (*in)[i];
-
-        if(bias != nullptr)
-        {
-            result += (*bias)[i % cols_in];
-        }
-
-        // Fixed point multiplication
-        if(result_shift < 0)
-        {
-            result = asymm_int_mult(result * (1 << (-result_shift)), result_fixedpoint_multiplier);
+            result = asymm_int_mult(result * (1 << (-shift)), multiplier);
         }
         else
         {
-            result = asymm_rounding_divide_by_pow2(asymm_int_mult(result, result_fixedpoint_multiplier), result_shift);
+            result = asymm_rounding_divide_by_pow2(asymm_int_mult(result, multiplier), shift);
         }
+        result += result_offset_after_shift;
 
         // Bounded ReLu
         if(min != max)
@@ -132,7 +126,8 @@ void quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> *in,
             result = std::max(min, std::min(max, result));
         }
 
-        (*dst)[i] = static_cast<int16_t>(std::max(-32768, std::min(32767, result)));
+        (*dst)[i] = static_cast<TOut>(std::max<TIn>(std::numeric_limits<TOut>::lowest(),
+                                                    std::min<TIn>(std::numeric_limits<TOut>::max(), result)));
     }
 }
 } // namespace
@@ -219,59 +214,43 @@ SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTe
     return dst;
 }
 
-template <typename T>
-SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> &in, std::vector<int32_t> result_fixedpoint_multiplier, std::vector<int32_t> result_shift,
-                                                                                int32_t result_offset_after_shift, int32_t min, int32_t max)
+template <typename TIn, typename TOut>
+SimpleTensor<TOut> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<TIn> &in, std::vector<int32_t> result_fixedpoint_multiplier, std::vector<int32_t> result_shift,
+                                                              int32_t result_offset_after_shift, int32_t min, int32_t max)
 {
-    SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8);
+    SimpleTensor<TOut> dst(in.shape(), DataTypeExtractor<TOut>::data_type());
 
-    quantize_down_int32_to_uint8_scale_by_fixedpoint<T>(&in, nullptr, &dst, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
+    quantize_down_scale_by_fixedpoint<TIn, TOut>(&in, nullptr, &dst, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
 
     return dst;
 }
 
-template <typename T>
-SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, std::vector<int32_t> result_fixedpoint_multiplier,
-                                                                                std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max)
-{
-    SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8);
-
-    quantize_down_int32_to_uint8_scale_by_fixedpoint<T>(&in, &bias, &dst, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
-
-    return dst;
-}
-
-template <typename T>
-SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t min,
-                                                                                int32_t max)
-{
-    SimpleTensor<int16_t> dst(in.shape(), DataType::QSYMM16);
-
-    quantize_down_int32_to_int16_scale_by_fixedpoint<T>(&in, nullptr, &dst, result_fixedpoint_multiplier, result_shift, min, max);
-
-    return dst;
-}
-
-template <typename T>
-SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_fixedpoint_multiplier, int32_t result_shift,
-                                                                                int32_t min, int32_t max)
+template <typename TIn, typename TOut>
+SimpleTensor<TOut> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<TIn> &in, const SimpleTensor<TIn> &bias, std::vector<int32_t> result_fixedpoint_multiplier,
+                                                              std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max)
 {
-    SimpleTensor<int16_t> dst(in.shape(), DataType::QSYMM16);
+    SimpleTensor<TOut> dst(in.shape(), DataTypeExtractor<TOut>::data_type());
 
-    quantize_down_int32_to_int16_scale_by_fixedpoint<T>(&in, &bias, &dst, result_fixedpoint_multiplier, result_shift, min, max);
+    quantize_down_scale_by_fixedpoint<TIn, TOut>(&in, &bias, &dst, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
 
     return dst;
 }
 
-template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, std::vector<int32_t> result_fixedpoint_multiplier,
-                                                                                         std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
-template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b,
-                                                                                         std::vector<int32_t> result_fixedpoint_multiplier,
-                                                                                         std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
-template SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, int32_t result_fixedpoint_multiplier, int32_t result_shift,
-                                                                                         int32_t min, int32_t max);
-template SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_fixedpoint_multiplier,
-                                                                                         int32_t result_shift, int32_t min, int32_t max);
+template SimpleTensor<uint8_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, std::vector<int32_t> result_fixedpoint_multiplier,
+                                                                          std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
+template SimpleTensor<uint8_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b,
+                                                                          std::vector<int32_t> result_fixedpoint_multiplier,
+                                                                          std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
+template SimpleTensor<int8_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, std::vector<int32_t> result_fixedpoint_multiplier,
+                                                                         std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
+template SimpleTensor<int8_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b,
+                                                                         std::vector<int32_t> result_fixedpoint_multiplier,
+                                                                         std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
+template SimpleTensor<int16_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, std::vector<int32_t> result_fixedpoint_multiplier,
+                                                                          std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
+template SimpleTensor<int16_t> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b,
+                                                                          std::vector<int32_t> result_fixedpoint_multiplier,
+                                                                          std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, int32_t result_offset, std::vector<int32_t> result_mult_int,
                                                                            std::vector<int32_t> result_shift, int32_t min, int32_t max);
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, const SimpleTensor<int32_t> &b, int32_t result_offset, std::vector<int32_t> result_mult_int,
diff --git a/tests/validation/reference/GEMMLowp.h b/tests/validation/reference/GEMMLowp.h
index 815527e1b7..7ff01ef611 100644
--- a/tests/validation/reference/GEMMLowp.h
+++ b/tests/validation/reference/GEMMLowp.h
@@ -52,20 +52,13 @@ template <typename T>
 SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_offset, std::vector<int32_t> result_mult_int,
                                                                   std::vector<int32_t> result_shift, int32_t min = 0, int32_t max = 0);
 
-template <typename T>
-SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> &in, std::vector<int32_t> result_fixedpoint_multiplier, std::vector<int32_t> result_shift,
-                                                                                int32_t result_offset_after_shift, int32_t min = 0, int32_t max = 0);
-
-template <typename T>
-SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, std::vector<int32_t> result_fixedpoint_multiplier,
-                                                                                std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min = 0, int32_t max = 0);
+template <typename TIn, typename TOut>
+SimpleTensor<TOut> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<TIn> &in, std::vector<int32_t> result_fixedpoint_multiplier, std::vector<int32_t> result_shift,
+                                                              int32_t result_offset_after_shift, int32_t min = 0, int32_t max = 0);
 
-template <typename T>
-SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, int32_t result_fixedpoint_multiplier, int32_t result_shift,
-                                                                                int32_t min, int32_t max);
-template <typename T>
-SimpleTensor<int16_t> gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_fixedpoint_multiplier,
-                                                                                int32_t result_shift, int32_t min, int32_t max);
+template <typename TIn, typename TOut>
+SimpleTensor<TOut> gemmlowp_quantize_down_scale_by_fixedpoint(const SimpleTensor<TIn> &in, const SimpleTensor<TIn> &bias, std::vector<int32_t> result_fixedpoint_multiplier,
+                                                              std::vector<int32_t> result_shift, int32_t result_offset_after_shift, int32_t min = 0, int32_t max = 0);
 } // namespace reference
 } // namespace validation
 } // namespace test