COMPMID-675 - Reworked NEGEMMLowp interface/function

The new interface makes NEGEMMLowp able to work with ASYMM8 data types.

Implemented 2 new functions:
- NEGEMMLowpMatrixMultiplyCore
- NEGEMMLowpOutputStage

These functions should make the integration in android NN doable

For more information about GEMMLowp:
https://github.com/google/gemmlowp/blob/master/doc/low-precision.md

Change-Id: Ie2c775f45234f68ca53dba644b3a912b997fd890
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/95504
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

8 files changed, 155 insertions, 166 deletions

diff --git a/tests/benchmark/NEON/GEMMLowp.cpp b/tests/benchmark/NEON/GEMMLowp.cpp
index 8cf143393d..a0e5e694bd 100644
--- a/tests/benchmark/NEON/GEMMLowp.cpp
+++ b/tests/benchmark/NEON/GEMMLowp.cpp
@@ -23,7 +23,6 @@
  */
 #include "arm_compute/core/TensorShape.h"
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMLowp.h"
 #include "arm_compute/runtime/Tensor.h"
 #include "arm_compute/runtime/TensorAllocator.h"
 #include "tests/NEON/Accessor.h"
diff --git a/tests/datasets/GEMMLowpDataset.h b/tests/datasets/GEMMLowpDataset.h
index 4bf2a98d61..062c05b1d9 100644
--- a/tests/datasets/GEMMLowpDataset.h
+++ b/tests/datasets/GEMMLowpDataset.h
@@ -37,7 +37,7 @@ namespace datasets
 class GEMMLowpDataset
 {
 public:
-    using type = std::tuple<TensorShape, TensorShape, TensorShape, int32_t, int32_t, int32_t, int32_t, int32_t>;
+    using type = std::tuple<TensorShape, TensorShape, TensorShape, int32_t, int32_t>;
 
     struct iterator
     {
@@ -45,18 +45,12 @@ public:
                  std::vector<TensorShape>::const_iterator b_it,
                  std::vector<TensorShape>::const_iterator c_it,
                  std::vector<int32_t>::const_iterator     a_offset_it,
-                 std::vector<int32_t>::const_iterator     b_offset_it,
-                 std::vector<int32_t>::const_iterator     c_offset_it,
-                 std::vector<int32_t>::const_iterator     c_mult_int_it,
-                 std::vector<int32_t>::const_iterator     out_shift_it)
+                 std::vector<int32_t>::const_iterator     b_offset_it)
             : _a_it{ std::move(a_it) },
               _b_it{ std::move(b_it) },
               _c_it{ std::move(c_it) },
               _a_offset_it{ std::move(a_offset_it) },
-              _b_offset_it{ std::move(b_offset_it) },
-              _c_offset_it{ std::move(c_offset_it) },
-              _c_mult_int_it{ std::move(c_mult_int_it) },
-              _out_shift_it{ std::move(out_shift_it) }
+              _b_offset_it{ std::move(b_offset_it) }
         {
         }
 
@@ -68,15 +62,12 @@ public:
             description << "C=" << *_c_it << ":";
             description << "a_offset=" << *_a_offset_it << ":";
             description << "b_offset=" << *_b_offset_it << ":";
-            description << "c_offset=" << *_c_offset_it << ":";
-            description << "c_mult_int=" << *_c_mult_int_it << ":";
-            description << "out_shift=" << *_out_shift_it << ":";
             return description.str();
         }
 
         GEMMLowpDataset::type operator*() const
         {
-            return std::make_tuple(*_a_it, *_b_it, *_c_it, *_a_offset_it, *_b_offset_it, *_c_offset_it, *_c_mult_int_it, *_out_shift_it);
+            return std::make_tuple(*_a_it, *_b_it, *_c_it, *_a_offset_it, *_b_offset_it);
         }
 
         iterator &operator++()
@@ -86,9 +77,6 @@ public:
             ++_c_it;
             ++_a_offset_it;
             ++_b_offset_it;
-            ++_c_offset_it;
-            ++_c_mult_int_it;
-            ++_out_shift_it;
 
             return *this;
         }
@@ -99,32 +87,25 @@ public:
         std::vector<TensorShape>::const_iterator _c_it;
         std::vector<int32_t>::const_iterator     _a_offset_it;
         std::vector<int32_t>::const_iterator     _b_offset_it;
-        std::vector<int32_t>::const_iterator     _c_offset_it;
-        std::vector<int32_t>::const_iterator     _c_mult_int_it;
-        std::vector<int32_t>::const_iterator     _out_shift_it;
     };
 
     iterator begin() const
     {
-        return iterator(_a_shapes.begin(), _b_shapes.begin(), _c_shapes.begin(), _a_offset.begin(), _b_offset.begin(), _c_offset.begin(), _c_mult_int.begin(), _out_shift.begin());
+        return iterator(_a_shapes.begin(), _b_shapes.begin(), _c_shapes.begin(), _a_offset.begin(), _b_offset.begin());
     }
 
     int size() const
     {
-        return std::min(_a_shapes.size(), std::min(_b_shapes.size(), std::min(_c_shapes.size(), std::min(_a_offset.size(), std::min(_b_offset.size(), std::min(_c_offset.size(), std::min(_c_mult_int.size(),
-                                                                                                         _out_shift.size())))))));
+        return std::min(_a_shapes.size(), std::min(_b_shapes.size(), std::min(_c_shapes.size(), std::min(_a_offset.size(), _b_offset.size()))));
     }
 
-    void add_config(TensorShape a, TensorShape b, TensorShape c, int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t c_mult_int, int32_t out_shift)
+    void add_config(TensorShape a, TensorShape b, TensorShape c, int32_t a_offset, int32_t b_offset)
     {
         _a_shapes.emplace_back(std::move(a));
         _b_shapes.emplace_back(std::move(b));
         _c_shapes.emplace_back(std::move(c));
         _a_offset.emplace_back(std::move(a_offset));
         _b_offset.emplace_back(std::move(b_offset));
-        _c_offset.emplace_back(std::move(c_offset));
-        _c_mult_int.emplace_back(std::move(c_mult_int));
-        _out_shift.emplace_back(std::move(out_shift));
     }
 
 protected:
@@ -137,9 +118,6 @@ private:
     std::vector<TensorShape> _c_shapes{};
     std::vector<int32_t>     _a_offset{};
     std::vector<int32_t>     _b_offset{};
-    std::vector<int32_t>     _c_offset{};
-    std::vector<int32_t>     _c_mult_int{};
-    std::vector<int32_t>     _out_shift{};
 };
 } // namespace datasets
 } // namespace test
diff --git a/tests/datasets/LargeGEMMLowpDataset.h b/tests/datasets/LargeGEMMLowpDataset.h
index 10f79e423d..cc1feb49a2 100644
--- a/tests/datasets/LargeGEMMLowpDataset.h
+++ b/tests/datasets/LargeGEMMLowpDataset.h
@@ -42,12 +42,12 @@ class LargeGEMMLowpDataset final : public GEMMLowpDataset
 public:
     LargeGEMMLowpDataset()
     {
-        add_config(TensorShape(923U, 429U), TensorShape(871U, 923U), TensorShape(871U, 429U), 0, 0, 0, 1, 0);
-        add_config(TensorShape(873U, 513U), TensorShape(784U, 873U), TensorShape(784U, 513U), 0, 4, 3, 2, 0);
-        add_config(TensorShape(697U, 872U), TensorShape(563U, 697U), TensorShape(563U, 872U), -2, 0, 1, 1, 0);
-        add_config(TensorShape(1021U, 973U), TensorShape(783U, 1021U), TensorShape(783U, 973U), 5, 13, -6, 2, 2);
-        add_config(TensorShape(681U, 1023U), TensorShape(213U, 681U), TensorShape(213U, 1023U), -3, -2, 8, 4, 3);
-        add_config(TensorShape(941U, 1011U), TensorShape(623U, 941U), TensorShape(623U, 1011U), -9, 1, -3, 3, 1);
+        add_config(TensorShape(923U, 429U), TensorShape(871U, 923U), TensorShape(871U, 429U), 0, 0);
+        add_config(TensorShape(873U, 513U), TensorShape(784U, 873U), TensorShape(784U, 513U), 0, 4);
+        add_config(TensorShape(697U, 872U), TensorShape(563U, 697U), TensorShape(563U, 872U), -2, 0);
+        add_config(TensorShape(1021U, 973U), TensorShape(783U, 1021U), TensorShape(783U, 973U), 5, 13);
+        add_config(TensorShape(681U, 1023U), TensorShape(213U, 681U), TensorShape(213U, 1023U), -3, -2);
+        add_config(TensorShape(941U, 1011U), TensorShape(623U, 941U), TensorShape(623U, 1011U), -9, 1);
     }
 };
 } // namespace datasets
diff --git a/tests/datasets/SmallGEMMLowpDataset.h b/tests/datasets/SmallGEMMLowpDataset.h
index b7fe3907ad..881546e70f 100644
--- a/tests/datasets/SmallGEMMLowpDataset.h
+++ b/tests/datasets/SmallGEMMLowpDataset.h
@@ -42,12 +42,12 @@ class SmallGEMMLowpDataset final : public GEMMLowpDataset
 public:
     SmallGEMMLowpDataset()
     {
-        add_config(TensorShape(21U, 13U), TensorShape(33U, 21U), TensorShape(33U, 13U), 0, 0, 0, 1, 0);
-        add_config(TensorShape(52U, 13U), TensorShape(33U, 52U), TensorShape(33U, 13U), 0, 4, 3, 2, 0);
-        add_config(TensorShape(52U, 26U), TensorShape(33U, 52U), TensorShape(33U, 26U), -2, 0, 1, 1, 0);
-        add_config(TensorShape(31U, 27U), TensorShape(23U, 31U), TensorShape(23U, 27U), 5, 13, -6, 2, 2);
-        add_config(TensorShape(38U, 12U), TensorShape(21U, 38U), TensorShape(21U, 12U), -3, -2, 8, 4, 3);
-        add_config(TensorShape(32U, 72U), TensorShape(17U, 32U), TensorShape(17U, 72U), -9, 1, -3, 3, 1);
+        add_config(TensorShape(21U, 13U), TensorShape(33U, 21U), TensorShape(33U, 13U), 0, 0);
+        add_config(TensorShape(52U, 13U), TensorShape(33U, 52U), TensorShape(33U, 13U), 0, 4);
+        add_config(TensorShape(52U, 26U), TensorShape(33U, 52U), TensorShape(33U, 26U), -2, 0);
+        add_config(TensorShape(31U, 27U), TensorShape(23U, 31U), TensorShape(23U, 27U), 5, 13);
+        add_config(TensorShape(38U, 12U), TensorShape(21U, 38U), TensorShape(21U, 12U), -3, -2);
+        add_config(TensorShape(32U, 72U), TensorShape(17U, 32U), TensorShape(17U, 72U), -9, 1);
     }
 };
 } // namespace datasets
diff --git a/tests/validation/CPP/GEMMLowp.cpp b/tests/validation/CPP/GEMMLowp.cpp
index e1d76503cd..bac3a20c8e 100644
--- a/tests/validation/CPP/GEMMLowp.cpp
+++ b/tests/validation/CPP/GEMMLowp.cpp
@@ -21,10 +21,9 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "GEMM.h"
+#include "GEMMLowp.h"
 
 #include "arm_compute/core/Types.h"
-#include "tests/validation/FixedPoint.h"
 
 namespace arm_compute
 {
@@ -34,17 +33,21 @@ namespace validation
 {
 namespace reference
 {
-SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, SimpleTensor<int32_t> &c)
+template <typename T>
+SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<T> &a, const SimpleTensor<T> &b, int32_t a_offset, int32_t b_offset)
 {
-    ARM_COMPUTE_UNUSED(a);
-    ARM_COMPUTE_UNUSED(b);
-    ARM_COMPUTE_UNUSED(c);
-    const int            K       = a.shape().x();
-    const int            b_width = b.shape().x();
-    const int            rows    = c.shape().y(); //M
-    const int            cols    = c.shape().x(); //N
+    TensorShape shape(b.shape()[0], a.shape()[1]);
+
+    SimpleTensor<int32_t> c(shape, DataType::S32);
+
+    const int K       = a.shape().x();
+    const int b_width = b.shape().x();
+    const int rows    = c.shape().y(); //M
+    const int cols    = c.shape().x(); //N
+
     std::vector<int32_t> acc;
     acc.resize(cols);
+
     for(int i = 0; i < rows; ++i)
     {
         for(int j = 0; j < cols; ++j)
@@ -53,10 +56,10 @@ SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor
         }
         for(int k = 0; k < K; ++k)
         {
-            auto tmp_a = static_cast<int32_t>(a[k + i * K]);
+            const int32_t tmp_a = a_offset + static_cast<int32_t>(a[k + i * K]);
             for(int j = 0; j < b_width; ++j)
             {
-                auto          tmp_b       = static_cast<int32_t>(b[j + k * b_width]);
+                const int32_t tmp_b       = b_offset + static_cast<int32_t>(b[j + k * b_width]);
                 const int32_t mult_as_int = tmp_a * tmp_b;
                 acc[j] += mult_as_int;
             }
@@ -71,43 +74,21 @@ SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor
 }
 
 template <typename T>
-SimpleTensor<T> gemmlowp(const SimpleTensor<T> &a, const SimpleTensor<T> &b, SimpleTensor<T> &c,
-                         int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t c_mult_int, int32_t out_shift)
+SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift)
 {
-    const int            K       = a.shape().x();
-    const int            b_width = b.shape().x();
-    const int            rows    = c.shape().y(); //M
-    const int            cols    = c.shape().x(); //N
-    std::vector<int32_t> acc;
-    acc.resize(cols);
-    for(int i = 0; i < rows; ++i)
+    SimpleTensor<uint8_t> dst(in.shape(), DataType::QASYMM8);
+
+    for(int i = 0; i < in.num_elements(); ++i)
     {
-        for(int j = 0; j < cols; ++j)
-        {
-            acc[j] = 0;
-        }
-        for(int k = 0; k < K; ++k)
-        {
-            const int32_t tmp_a = a_offset + static_cast<int32_t>(a[k + i * K]);
-            for(int j = 0; j < b_width; ++j)
-            {
-                const int32_t tmp_b       = b_offset + static_cast<int32_t>(b[j + k * b_width]);
-                const int32_t mult_as_int = tmp_a * tmp_b;
-                acc[j] += mult_as_int;
-            }
-        }
-        for(int j = 0; j < cols; ++j)
-        {
-            const int32_t result = ((c_offset + acc[j]) * c_mult_int) >> out_shift;
-            c[j + i * cols]      = static_cast<int8_t>(std::min(127, std::max(-128, result)));
-        }
+        const int32_t result = ((in[i] + result_offset) * result_mult_int) >> result_shift;
+        dst[i]               = static_cast<uint8_t>(std::max(0, std::min(255, result)));
     }
 
-    return c;
+    return dst;
 }
 
-template SimpleTensor<int8_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, SimpleTensor<int8_t> &c,
-                                       int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t c_mult_int, int32_t out_shift);
+template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, int32_t a_offset, int32_t b_offset);
+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);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/CPP/GEMMLowp.h b/tests/validation/CPP/GEMMLowp.h
index 2f903f2fe2..c09d8f6176 100644
--- a/tests/validation/CPP/GEMMLowp.h
+++ b/tests/validation/CPP/GEMMLowp.h
@@ -35,11 +35,11 @@ namespace validation
 {
 namespace reference
 {
-SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, SimpleTensor<int32_t> &c);
+template <typename T>
+SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<T> &a, const SimpleTensor<T> &b, int32_t a_offset, int32_t b_offset);
 
 template <typename T>
-SimpleTensor<T> gemmlowp(const SimpleTensor<T> &a, const SimpleTensor<T> &b, SimpleTensor<T> &c,
-                         int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t c_mult_int, int32_t out_shift);
+SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index 4924f98ea6..4407eff060 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -23,12 +23,15 @@
  */
 #include "arm_compute/core/NEON/kernels/NEGEMMInterleaveBlockedKernel.h"
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMLowp.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
 #include "arm_compute/runtime/Tensor.h"
 #include "arm_compute/runtime/TensorAllocator.h"
 #include "tests/NEON/Accessor.h"
 #include "tests/NEON/Helper.h"
+#include "tests/PaddingCalculator.h"
 #include "tests/datasets/LargeGEMMLowpDataset.h"
+#include "tests/datasets/ShapeDatasets.h"
 #include "tests/datasets/SmallGEMMLowpDataset.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Macros.h"
@@ -45,16 +48,13 @@ namespace validation
 {
 namespace
 {
-const auto data_int_blk         = framework::dataset::make("M", 8, 12) * framework::dataset::make("N", 8, 12) * framework::dataset::make("by", 8, 13) * framework::dataset::make("block", 4, 9);
-const auto data_int_blk_tr      = framework::dataset::make("M", 8, 17) * framework::dataset::make("N", 8, 14) * framework::dataset::make("by", 12) * framework::dataset::make("block", 4);
-const auto data_matrix_multiply = framework::dataset::make("M", 12, 20) * framework::dataset::make("N", 12, 20) * framework::dataset::make("K", 16);
+const auto data_int_blk    = framework::dataset::make("M", 8, 12) * framework::dataset::make("N", 8, 12) * framework::dataset::make("by", 8, 13) * framework::dataset::make("block", 4, 9);
+const auto data_int_blk_tr = framework::dataset::make("M", 8, 17) * framework::dataset::make("N", 8, 14) * framework::dataset::make("by", 12) * framework::dataset::make("block", 4);
 } // namespace
 
 TEST_SUITE(NEON)
 TEST_SUITE(GEMMLowp)
 
-TEST_SUITE(S8)
-
 TEST_SUITE(INTERLEAVE_BLOCKED)
 
 using NEInterleaveBlocked            = NESynthetizeFunction<NEGEMMInterleaveBlockedKernel>;
@@ -77,50 +77,95 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMInterleaveBlockedTransposedFixture, frame
 
 TEST_SUITE_END()
 
-using NEGEMMLowpOffsetFixture = GEMMLowpOffsetValidationFixture<Tensor, Accessor, NEGEMMLowp>;
+TEST_SUITE(MatrixMultiplyCore)
+using NEGEMMLowpMatrixMultiplyCoreFixture = GEMMLowpMatrixMultiplyCoreValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>;
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallGEMMLowpDataset(), datasets::LargeGEMMLowpDataset()), framework::dataset::make("DataType",
-                                                                   DataType::S8)),
-               shape_a, shape_b, shape_c, a_offset, b_offset, c_offset, c_mult_int, out_shift, data_type)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, framework::dataset::concat(datasets::SmallGEMMLowpDataset(), datasets::LargeGEMMLowpDataset()),
+               shape_a, shape_b, shape_c, a_offset, b_offset)
 {
     // Create tensors
-    Tensor a = create_tensor<Tensor>(shape_a, data_type);
-    Tensor b = create_tensor<Tensor>(shape_b, data_type);
-    Tensor c = create_tensor<Tensor>(shape_c, data_type);
+    Tensor a = create_tensor<Tensor>(shape_a, DataType::QASYMM8);
+    Tensor b = create_tensor<Tensor>(shape_b, DataType::QASYMM8);
+    Tensor c = create_tensor<Tensor>(shape_c, DataType::S32);
+
+    a.info()->set_quantization_info(QuantizationInfo(1.0f / 255, a_offset));
+    b.info()->set_quantization_info(QuantizationInfo(1.0f / 255, b_offset));
 
     ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
 
     // Create and configure function
-    NEGEMMLowp gemmlowp;
-    gemmlowp.configure(&a, &b, &c, a_offset, b_offset, c_offset, c_mult_int, out_shift);
+    NEGEMMLowpMatrixMultiplyCore gemmlowp_mm;
+    gemmlowp_mm.configure(&a, &b, &c);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpDataset())
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpOffsetFixture, framework::DatasetMode::ALL, combine(datasets::SmallGEMMLowpDataset(), framework::dataset::make("DataType", DataType::S8)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpDataset())
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpOffsetFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMLowpDataset(), framework::dataset::make("DataType", DataType::S8)))
+TEST_SUITE_END() // MatrixMultiplyCore
+
+TEST_SUITE(OutputStage)
+
+TEST_SUITE(QuantizeDownInt32ToUint8Scale)
+
+using NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToUint8Scale>;
+
+const auto quantize_down_int32_to_uint8_scale_cases = framework::dataset::make("result_offset", -4, 4) * framework::dataset::make("result_mult_int", 1, 3) * framework::dataset::make("result_shift", 2,
+                                                      4);
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), quantize_down_int32_to_uint8_scale_cases),
+               shape, result_offset, result_mult_int, result_shift)
+{
+    // Create tensors
+    Tensor in  = create_tensor<Tensor>(shape, DataType::S32);
+    Tensor out = create_tensor<Tensor>(shape, DataType::QASYMM8);
+
+    ARM_COMPUTE_EXPECT(in.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(out.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    NEGEMMLowpQuantizeDownInt32ToUint8Scale output_stage;
+    output_stage.configure(&in, &out, result_offset, result_mult_int, result_shift);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(in.info()->valid_region(), valid_region);
+    validate(out.info()->valid_region(), valid_region);
+
+    // Validate padding
+    const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
+    validate(in.info()->padding(), padding);
+    validate(out.info()->padding(), padding);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-TEST_SUITE_END() // U8
 
-TEST_SUITE(S32)
-using NEGEMMLowpMatrixMultiplyFixture = GEMMLowpMatrixMultiplyValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>;
-FIXTURE_DATA_TEST_CASE(MatrixMultiply, NEGEMMLowpMatrixMultiplyFixture, framework::DatasetMode::PRECOMMIT, data_matrix_multiply)
+FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), quantize_down_int32_to_uint8_scale_cases))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-TEST_SUITE_END()
 
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() // QuantizeDownInt32ToUint8Scale
+
+TEST_SUITE_END() // OutputStage
+
+TEST_SUITE_END() // GEMMLowp
+TEST_SUITE_END() // NEON
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index fba44008ba..f9b0dbd959 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -43,36 +43,39 @@ namespace test
 namespace validation
 {
 template <typename TensorType, typename AccessorType, typename FunctionType>
-class GEMMLowpOffsetValidationFixture : public framework::Fixture
+class GEMMLowpMatrixMultiplyCoreValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t c_mult_int, int32_t out_shift, DataType data_type)
+    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, int32_t a_offset, int32_t b_offset)
     {
-        _target    = compute_target(shape_a, shape_b, shape_c, a_offset, b_offset, c_offset, c_mult_int, out_shift, data_type);
-        _reference = compute_reference(shape_a, shape_b, shape_c, a_offset, b_offset, c_offset, c_mult_int, out_shift, data_type);
+        _target    = compute_target(shape_a, shape_b, shape_c, a_offset, b_offset);
+        _reference = compute_reference(shape_a, shape_b, shape_c, a_offset, b_offset);
     }
 
 protected:
     template <typename U>
     void fill(U &&tensor, int i)
     {
-        ARM_COMPUTE_ERROR_ON(tensor.data_type() != DataType::S8);
-        std::uniform_int_distribution<> distribution(0, 3);
+        // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path
+        std::uniform_int_distribution<> distribution(1, 254);
         library->fill(tensor, distribution, i);
     }
 
     TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c,
-                              int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t c_mult_int, int32_t out_shift, DataType data_type)
+                              int32_t a_offset, int32_t b_offset)
     {
         // Create tensors
-        TensorType a = create_tensor<TensorType>(shape_a, data_type, 1);
-        TensorType b = create_tensor<TensorType>(shape_b, data_type, 1);
-        TensorType c = create_tensor<TensorType>(shape_c, data_type, 1);
+        TensorType a = create_tensor<TensorType>(shape_a, DataType::QASYMM8, 1);
+        TensorType b = create_tensor<TensorType>(shape_b, DataType::QASYMM8, 1);
+        TensorType c = create_tensor<TensorType>(shape_c, DataType::S32, 1);
+
+        a.info()->set_quantization_info(QuantizationInfo(1.0f / 255, a_offset));
+        b.info()->set_quantization_info(QuantizationInfo(1.0f / 255, b_offset));
 
         // Create and configure function
         FunctionType gemmlowp;
-        gemmlowp.configure(&a, &b, &c, a_offset, b_offset, c_offset, c_mult_int, out_shift);
+        gemmlowp.configure(&a, &b, &c);
 
         ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -90,108 +93,91 @@ protected:
         // Fill tensors
         fill(AccessorType(a), 0);
         fill(AccessorType(b), 1);
-        fill(AccessorType(c), 2);
 
         // Compute GEMM function
         gemmlowp.run();
         return c;
     }
 
-    SimpleTensor<int8_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c,
-                                           int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t c_mult_int, int32_t out_shift, DataType data_type)
+    SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c,
+                                            int32_t a_offset, int32_t b_offset)
     {
         // Create reference
-        SimpleTensor<int8_t> a{ shape_a, data_type, 1 };
-        SimpleTensor<int8_t> b{ shape_b, data_type, 1 };
-        SimpleTensor<int8_t> c{ shape_c, data_type, 1 };
+        SimpleTensor<uint8_t> a{ shape_a, DataType::QASYMM8, 1 };
+        SimpleTensor<uint8_t> b{ shape_b, DataType::QASYMM8, 1 };
 
         // Fill reference
         fill(a, 0);
         fill(b, 1);
-        fill(c, 2);
 
-        return reference::gemmlowp<int8_t>(a, b, c, a_offset, b_offset, c_offset, c_mult_int, out_shift);
+        return reference::gemmlowp_matrix_multiply_core<uint8_t>(a, b, a_offset, b_offset);
     }
 
-    TensorType           _target{};
-    SimpleTensor<int8_t> _reference{};
+    TensorType            _target{};
+    SimpleTensor<int32_t> _reference{};
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType>
-class GEMMLowpMatrixMultiplyValidationFixture : public framework::Fixture
+class GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(size_t m, size_t n, size_t k)
+    void setup(TensorShape shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift)
     {
-        const TensorShape shape_a(k, m);
-        const TensorShape shape_b(n, k);
-        const TensorShape shape_c(n, m);
-        _target    = compute_target(shape_a, shape_b, shape_c);
-        _reference = compute_reference(shape_a, shape_b, shape_c);
+        _target    = compute_target(shape, result_offset, result_mult_int, result_shift);
+        _reference = compute_reference(shape, result_offset, result_mult_int, result_shift);
     }
 
 protected:
     template <typename U>
-    void fill(U &&tensor, int i, int lo, int hi)
+    void fill(U &&tensor, int i)
     {
-        std::uniform_int_distribution<> distribution(lo, hi);
+        std::uniform_int_distribution<> distribution(-6000, 6000);
         library->fill(tensor, distribution, i);
     }
 
-    TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c)
+    TensorType compute_target(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift)
     {
         // Create tensors
-        TensorType a = create_tensor<TensorType>(shape_a, DataType::S8, 1);
-        TensorType b = create_tensor<TensorType>(shape_b, DataType::S8, 1);
-        TensorType c = create_tensor<TensorType>(shape_c, DataType::S32, 1);
+        TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1);
+        TensorType b = create_tensor<TensorType>(shape, DataType::QASYMM8, 1);
 
         // Create and configure function
-        FunctionType gemmlowp;
-        gemmlowp.configure(&a, &b, &c);
+        FunctionType output_stage;
+        output_stage.configure(&a, &b, result_offset, result_mult_int, result_shift);
 
         ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Allocate tensors
         a.allocator()->allocate();
         b.allocator()->allocate();
-        c.allocator()->allocate();
 
         ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Fill tensors
-        fill(AccessorType(a), 0, -128, 127);
-        fill(AccessorType(b), 1, -128, 127);
-        fill(AccessorType(c), 2, 0, 0);
+        fill(AccessorType(a), 0);
 
         // Compute GEMM function
-        gemmlowp.run();
-        return c;
+        output_stage.run();
+        return b;
     }
 
-    SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c)
+    SimpleTensor<uint8_t> compute_reference(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift)
     {
         // Create reference
-        SimpleTensor<int8_t>  a{ shape_a, DataType::S8, 1 };
-        SimpleTensor<int8_t>  b{ shape_b, DataType::S8, 1 };
-        SimpleTensor<int32_t> c{ shape_c, DataType::S32, 1 };
+        SimpleTensor<int32_t> a{ shape, DataType::S32, 1 };
 
         // Fill reference
-        fill(a, 0, -128, 127);
-        fill(b, 1, -128, 127);
-        fill(c, 2, 0, 0);
+        fill(a, 0);
 
-        return reference::gemmlowp(a, b, c);
+        return reference::gemmlowp_quantize_down_int32_to_uint8_scale<int32_t>(a, result_offset, result_mult_int, result_shift);
     }
 
     TensorType            _target{};
-    SimpleTensor<int32_t> _reference{};
+    SimpleTensor<uint8_t> _reference{};
 };
-
 } // namespace validation
 } // namespace test
 } // namespace arm_compute