COMPMID-662: Integrated the new a64_s8_gemm_12x8 + dot product kernel into ACL.

Change-Id: Id8f919e486a132fc58346c9f84fccbeeb83d19b3
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/94233
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

4 files changed, 40 insertions, 42 deletions

diff --git a/tests/validation/CPP/GEMMLowp.cpp b/tests/validation/CPP/GEMMLowp.cpp
index 06926e631e..e1d76503cd 100644
--- a/tests/validation/CPP/GEMMLowp.cpp
+++ b/tests/validation/CPP/GEMMLowp.cpp
@@ -34,7 +34,7 @@ namespace validation
 {
 namespace reference
 {
-SimpleTensor<uint32_t> gemmlowp(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, SimpleTensor<uint32_t> &c)
+SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, SimpleTensor<int32_t> &c)
 {
     ARM_COMPUTE_UNUSED(a);
     ARM_COMPUTE_UNUSED(b);
@@ -99,15 +99,15 @@ SimpleTensor<T> gemmlowp(const SimpleTensor<T> &a, const SimpleTensor<T> &b, Sim
         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<uint8_t>(std::min(255, std::max(0, result)));
+            c[j + i * cols]      = static_cast<int8_t>(std::min(127, std::max(-128, result)));
         }
     }
 
     return c;
 }
 
-template SimpleTensor<uint8_t> gemmlowp(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, SimpleTensor<uint8_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<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);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/CPP/GEMMLowp.h b/tests/validation/CPP/GEMMLowp.h
index 0428e9e34f..2f903f2fe2 100644
--- a/tests/validation/CPP/GEMMLowp.h
+++ b/tests/validation/CPP/GEMMLowp.h
@@ -35,7 +35,7 @@ namespace validation
 {
 namespace reference
 {
-SimpleTensor<uint32_t> gemmlowp(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, SimpleTensor<uint32_t> &c);
+SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, SimpleTensor<int32_t> &c);
 
 template <typename T>
 SimpleTensor<T> gemmlowp(const SimpleTensor<T> &a, const SimpleTensor<T> &b, SimpleTensor<T> &c,
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index de7495bb6a..4924f98ea6 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -45,14 +45,15 @@ 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_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);
 } // namespace
 
 TEST_SUITE(NEON)
 TEST_SUITE(GEMMLowp)
 
-TEST_SUITE(U8)
+TEST_SUITE(S8)
 
 TEST_SUITE(INTERLEAVE_BLOCKED)
 
@@ -79,7 +80,7 @@ TEST_SUITE_END()
 using NEGEMMLowpOffsetFixture = GEMMLowpOffsetValidationFixture<Tensor, Accessor, NEGEMMLowp>;
 
 DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallGEMMLowpDataset(), datasets::LargeGEMMLowpDataset()), framework::dataset::make("DataType",
-                                                                   DataType::U8)),
+                                                                   DataType::S8)),
                shape_a, shape_b, shape_c, a_offset, b_offset, c_offset, c_mult_int, out_shift, data_type)
 {
     // Create tensors
@@ -96,31 +97,28 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da
     gemmlowp.configure(&a, &b, &c, a_offset, b_offset, c_offset, c_mult_int, out_shift);
 }
 
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpOffsetFixture, framework::DatasetMode::ALL, combine(datasets::SmallGEMMLowpDataset(), framework::dataset::make("DataType", DataType::U8)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpOffsetFixture, framework::DatasetMode::ALL, combine(datasets::SmallGEMMLowpDataset(), framework::dataset::make("DataType", DataType::S8)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpOffsetFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMLowpDataset(), framework::dataset::make("DataType", DataType::U8)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpOffsetFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMLowpDataset(), framework::dataset::make("DataType", DataType::S8)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
+TEST_SUITE_END() // U8
 
-//FIXME: This is in the process of being updated, for more info please refer to COMPMID-624.
-#if 0  // defined(__aarch64__)
-TEST_SUITE(U32)
-using NEGEMMLowpFixture = GEMMLowpValidationFixture<Tensor, Accessor, NEGEMMLowp>;
-FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpFixture, framework::DatasetMode::PRECOMMIT, framework::dataset::make("M", 12, 20) * framework::dataset::make("N", 12, 20) * framework::dataset::make("K",
-                       16))
+TEST_SUITE(S32)
+using NEGEMMLowpMatrixMultiplyFixture = GEMMLowpMatrixMultiplyValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>;
+FIXTURE_DATA_TEST_CASE(MatrixMultiply, NEGEMMLowpMatrixMultiplyFixture, framework::DatasetMode::PRECOMMIT, data_matrix_multiply)
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
 TEST_SUITE_END()
-#endif // defined(__aarch64__)
-TEST_SUITE_END()
+
 TEST_SUITE_END()
 TEST_SUITE_END()
 } // namespace validation
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index 1a4084fdc2..fba44008ba 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -57,7 +57,7 @@ protected:
     template <typename U>
     void fill(U &&tensor, int i)
     {
-        ARM_COMPUTE_ERROR_ON(tensor.data_type() != DataType::U8);
+        ARM_COMPUTE_ERROR_ON(tensor.data_type() != DataType::S8);
         std::uniform_int_distribution<> distribution(0, 3);
         library->fill(tensor, distribution, i);
     }
@@ -97,28 +97,28 @@ protected:
         return c;
     }
 
-    SimpleTensor<uint8_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<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)
     {
         // Create reference
-        SimpleTensor<uint8_t> a{ shape_a, data_type, 1 };
-        SimpleTensor<uint8_t> b{ shape_b, data_type, 1 };
-        SimpleTensor<uint8_t> c{ shape_c, data_type, 1 };
+        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 };
 
         // Fill reference
         fill(a, 0);
         fill(b, 1);
         fill(c, 2);
 
-        return reference::gemmlowp<uint8_t>(a, b, c, a_offset, b_offset, c_offset, c_mult_int, out_shift);
+        return reference::gemmlowp<int8_t>(a, b, c, a_offset, b_offset, c_offset, c_mult_int, out_shift);
     }
 
-    TensorType            _target{};
-    SimpleTensor<uint8_t> _reference{};
+    TensorType           _target{};
+    SimpleTensor<int8_t> _reference{};
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType>
-class GEMMLowpValidationFixture : public framework::Fixture
+class GEMMLowpMatrixMultiplyValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
@@ -142,9 +142,9 @@ protected:
     TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c)
     {
         // Create tensors
-        TensorType a = create_tensor<TensorType>(shape_a, DataType::U8, 1);
-        TensorType b = create_tensor<TensorType>(shape_b, DataType::U8, 1);
-        TensorType c = create_tensor<TensorType>(shape_c, DataType::U32, 1);
+        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);
 
         // Create and configure function
         FunctionType gemmlowp;
@@ -164,8 +164,8 @@ protected:
         ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Fill tensors
-        fill(AccessorType(a), 0, 0, 3);
-        fill(AccessorType(b), 1, 0, 3);
+        fill(AccessorType(a), 0, -128, 127);
+        fill(AccessorType(b), 1, -128, 127);
         fill(AccessorType(c), 2, 0, 0);
 
         // Compute GEMM function
@@ -173,23 +173,23 @@ protected:
         return c;
     }
 
-    SimpleTensor<uint32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c)
+    SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c)
     {
         // Create reference
-        SimpleTensor<uint8_t>  a{ shape_a, DataType::U8, 1 };
-        SimpleTensor<uint8_t>  b{ shape_b, DataType::U8, 1 };
-        SimpleTensor<uint32_t> c{ shape_c, DataType::U32, 1 };
+        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 };
 
         // Fill reference
-        fill(a, 0, 0, 3);
-        fill(b, 1, 0, 3);
+        fill(a, 0, -128, 127);
+        fill(b, 1, -128, 127);
         fill(c, 2, 0, 0);
 
         return reference::gemmlowp(a, b, c);
     }
 
-    TensorType             _target{};
-    SimpleTensor<uint32_t> _reference{};
+    TensorType            _target{};
+    SimpleTensor<int32_t> _reference{};
 };
 
 } // namespace validation