COMPMID-2406: Create a new GEMMLowpQuantizeDownInt32ToInt16ScaleKernel for NEON

Change-Id: I3f3e247728fd6dafca066e41835f0ef9442d9b7a
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1379
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 269 insertions, 2 deletions

diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index f0460b4a23..2f604c95ea 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -294,6 +294,9 @@ const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases = framewo
 using NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture =
     GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint>;
 
+using NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture =
+    GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>;
+
 // *INDENT-OFF*
 // clang-format off
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
@@ -404,6 +407,117 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedP
 TEST_SUITE_END() // BoundedReLu
 
 TEST_SUITE_END() // QuantizeDownInt32ToUint8ScaleByFixedPoint
+
+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,
+                                                                    2)
+                                                                    * framework::dataset::make("min", 0) * framework::dataset::make("max", 0) * framework::dataset::make("addBias", { false, true });
+
+const auto quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1,
+                                                                         2)
+                                                                         * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true });
+
+using NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture =
+    GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>;
+
+// *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::S32), // Input not a multiple of 16
+                                             TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Invalid min and max
+                                             TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type
+                                          }),
+    framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32),
+                                            TensorInfo(TensorShape(21U), 1, DataType::S32),
+                                            TensorInfo(TensorShape(20U), 1, DataType::S32),
+                                          })),
+    framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QSYMM16),
+                                            TensorInfo(TensorShape(21U, 13U), 1, DataType::QSYMM16),
+                                            TensorInfo(TensorShape(20U, 13U), 1, DataType::S32),
+                                           })),
+    framework::dataset::make("Min",{        -205,
+                                            -60000,
+                                            -180,
+                                           })),
+    framework::dataset::make("Max",{        205,
+                                            60000,
+                                            180,
+                                           })),
+    framework::dataset::make("Expected", { true, false, false })),
+    a_info, b_info, output_info, min, max, expected)
+{
+    // Lock tensors
+    Status status =  NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::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_int16_scale_by_fixedpoint_cases),
+               shape, result_fixedpoint_multiplier, result_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::QSYMM16);
+
+    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
+    NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint output_stage;
+    output_stage.configure(&in, add_bias ? &bias : nullptr, &out, result_fixedpoint_multiplier, result_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, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int16_scale_by_fixedpoint_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+TEST_SUITE(BoundedReLu)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(),
+                       quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+TEST_SUITE_END() // BoundedReLu
+
+TEST_SUITE_END() // QuantizeDownInt32ToInt16ScaleByFixedPoint
+
 TEST_SUITE_END() // OutputStage
 
 TEST_SUITE_END() // GEMMLowp
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index ad5acfc418..8385221c78 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -385,6 +385,98 @@ protected:
     SimpleTensor<uint8_t> _reference{};
 };
 
+template <typename TensorType, typename AccessorType, typename FunctionType>
+class GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t min, int32_t max, bool add_bias)
+    {
+        _target    = compute_target(shape, result_fixedpoint_multiplier, result_shift, min, max, add_bias);
+        _reference = compute_reference(shape, result_fixedpoint_multiplier, result_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 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::QSYMM16, 1);
+
+        // Create and configure function
+        FunctionType output_stage;
+        output_stage.configure(&a, add_bias ? &b : nullptr, &c, result_fixedpoint_multiplier, result_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<int16_t> compute_reference(const TensorShape &shape, int32_t result_fixed_point_multiplier, int32_t result_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);
+
+        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);
+        }
+        else
+        {
+            return reference::gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint<int32_t>(a, result_fixed_point_multiplier, result_shift, min, max);
+        }
+    }
+
+    TensorType            _target{};
+    SimpleTensor<int16_t> _reference{};
+};
+
 template <typename TensorType, typename AccessorType, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
 class GEMMLowpMatrixMultiplyReshapedValidationFixture : public framework::Fixture
 {
diff --git a/tests/validation/reference/GEMMLowp.cpp b/tests/validation/reference/GEMMLowp.cpp
index 9a7e409e8a..97d05327e7 100644
--- a/tests/validation/reference/GEMMLowp.cpp
+++ b/tests/validation/reference/GEMMLowp.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -95,6 +95,34 @@ void quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> *in,
         (*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
+        result = asymm_rounding_divide_by_pow2(asymm_int_mult(result, result_fixedpoint_multiplier), result_shift);
+
+        // Bounded ReLu
+        if(min != max)
+        {
+            result = std::max(min, std::min(max, result));
+        }
+
+        (*dst)[i] = static_cast<int16_t>(std::max(-32768, std::min(32767, result)));
+    }
+}
 } // namespace
 
 template <typename T_out, typename T_in>
@@ -201,10 +229,36 @@ SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(
     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)
+{
+    SimpleTensor<int16_t> dst(in.shape(), DataType::QSYMM16);
+
+    quantize_down_int32_to_int16_scale_by_fixedpoint<T>(&in, &bias, &dst, result_fixedpoint_multiplier, result_shift, min, max);
+
+    return dst;
+}
+
 template SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<int32_t> &a, int32_t result_fixedpoint_multiplier, 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, int32_t result_fixedpoint_multiplier,
                                                                                          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_int32_to_uint8_scale(const SimpleTensor<int32_t> &a, int32_t result_offset, int32_t result_mult_int, 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, int32_t result_mult_int,
diff --git a/tests/validation/reference/GEMMLowp.h b/tests/validation/reference/GEMMLowp.h
index 4396155b96..5581f67652 100644
--- a/tests/validation/reference/GEMMLowp.h
+++ b/tests/validation/reference/GEMMLowp.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -58,6 +58,13 @@ SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(
 template <typename T>
 SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(const SimpleTensor<T> &in, const SimpleTensor<T> &bias, int32_t result_fixedpoint_multiplier, 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);
 } // namespace reference
 } // namespace validation
 } // namespace test