COMPMID-1710: Improve test coverage for CLGEMMMatrixMultiplyReshapedKernel

Added test for:
1) Fp16
2) GEMM3D

Change-Id: I17c03fe04fe49fba71685d33a6fd8572c91e1a56
Reviewed-on: https://review.mlplatform.org/416
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

2 files changed, 339 insertions, 67 deletions

diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
index e2fa194765..1a41e459bd 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
@@ -44,16 +44,47 @@ namespace test
 {
 namespace validation
 {
+using namespace arm_compute::misc::shape_calculator;
+
+// Create function for CLGEMMReshapeLHSMatrixKernel
+using CLGEMMReshapeLHSMatrix = CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder<CLGEMMReshapeLHSMatrixKernel, 16>;
+
+// Create function for CLGEMMReshapeRHSMatrixKernel
+using CLGEMMReshapeRHSMatrix = CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder<CLGEMMReshapeRHSMatrixKernel, 16>;
+
+// Create function for CLGEMMMatrixMultiplyReshapedKernel
+using CLGEMMMatrixMultiplyReshaped = CLSynthetizeFunction<CLGEMMMatrixMultiplyReshapedKernel>;
+
+// Fixture for CLGEMMMatrixMultiplyReshaped
+template <typename T>
+using CLGEMMMatrixMultiplyReshapedFixture = GEMMMatrixMultiplyReshapedValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>;
+
+// Fixture for CLGEMMMatrixMultiplyReshaped3D
+template <typename T>
+using CLGEMMMatrixMultiplyReshaped3DFixture = GEMMMatrixMultiplyReshaped3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>;
+
 namespace
 {
 // *INDENT-OFF*
 // clang-format off
-RelativeTolerance<float> tolerance_f32(0.001f);
+RelativeTolerance<float> rel_tolerance_f32(0.001f);
 constexpr float          abs_tolerance_f32(0.0001f);
 
+RelativeTolerance<half> rel_tolerance_f16(half(0.2));
+constexpr float         tolerance_num_f16 = 0.02f;
+
+/** Alpha values to test - Precommit */
+const auto a_values_precommit = framework::dataset::make("alpha", {1.0f, -0.75f} );
+
 /** M values to test */
 const auto m_values = framework::dataset::make("M", 37);
 
+/** M_W values to test */
+const auto m_w_values = framework::dataset::make("M_W", 5);
+
+/** M_H values to test */
+const auto m_h_values = framework::dataset::make("M_H", 7);
+
 /** N values to test */
 const auto n_values = framework::dataset::make("N", 51);
 
@@ -64,13 +95,13 @@ const auto k_values = framework::dataset::make("K", 43);
 const auto b_values = framework::dataset::make("batch_size", 1, 3);
 
 /** M0 values to test - Precommit */
-const auto m0_values_precommit = framework::dataset::make("M0", {4, 5, 6});
+const auto m0_values_precommit = framework::dataset::make("M0", {4, 6});
 
 /** N0 values to test - Precommit */
-const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4, 8 });
+const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 });
 
 /** K0 values to test - Precommit */
-const auto k0_values_precommit = framework::dataset::make("K0", { 4, 8 });
+const auto k0_values_precommit = framework::dataset::make("K0", { 4 });
 
 /** V0 values to test - Precommit */
 const auto v0_values_precommit = framework::dataset::make("V0", 1, 3);
@@ -78,6 +109,9 @@ const auto v0_values_precommit = framework::dataset::make("V0", 1, 3);
 /** H0 values to test - Precommit */
 const auto h0_values_precommit = framework::dataset::make("H0", 1, 3);
 
+/** Alpha values to test - Nightly */
+const auto a_values_nightly = framework::dataset::make("alpha", {1.0f, -0.75f, 0.85f} );
+
 /** M0 values to test - Nightly */
 const auto m0_values_nightly = framework::dataset::make("M0", 2, 8);
 
@@ -99,37 +133,8 @@ const auto i_values_lhs = framework::dataset::make("interleave_lhs", { true, fal
 /** Interleave values to test with RHS matrix */
 const auto i_values_rhs = framework::dataset::make("interleave_rhs", { true, false });
 
-} // namespace
-
-using namespace arm_compute::misc::shape_calculator;
-
-// Create function for CLGEMMReshapeLHSMatrixKernel
-using CLGEMMReshapeLHSMatrix = CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder<CLGEMMReshapeLHSMatrixKernel, 16>;
-
-// Create function for CLGEMMReshapeRHSMatrixKernel
-using CLGEMMReshapeRHSMatrix = CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder<CLGEMMReshapeRHSMatrixKernel, 16>;
-
-// Create function for CLGEMMMatrixMultiplyReshapedKernel
-using CLGEMMMatrixMultiplyReshaped = CLSynthetizeFunction<CLGEMMMatrixMultiplyReshapedKernel>;
-
-// Fixture for CLGEMMMatrixMultiplyReshaped
-using CLGEMMMatrixMultiplyReshapedFixture = GEMMMatrixMultiplyReshapedValidationFixture<CLTensor, CLAccessor, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>;
-
-TEST_SUITE(CL)
-TEST_SUITE(GEMMMatrixMultiplyReshaped)
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
-                                                                   m_values,
-                                                                   n_values),
-                                                                   k_values),
-                                                                   framework::dataset::make("batch_size", 1)),
-                                                                   m0_values_precommit),
-                                                                   n0_values_precommit),
-                                                                   k0_values_precommit),
-                                                                   v0_values_precommit),
-                                                                   h0_values_precommit),
-                                                                   i_values_lhs),
-                                                                   i_values_rhs),
-m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs)
+/** Configuration test */
+void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, unsigned int v0_value, unsigned int h0_value, bool i_value_lhs, bool i_value_rhs, DataType data_type)
 {
     const unsigned int M = m_value;
     const unsigned int N = n_value;
@@ -152,22 +157,22 @@ m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_v
     GEMMReshapeInfo gemm_info(M, N, K);
 
     const TensorShape lhs_shape(K, M, b_value);
-    const TensorShape lhs_shape_reshaped = compute_lhs_reshaped_shape(TensorInfo(lhs_shape, 1, DataType::F32),
+    const TensorShape lhs_shape_reshaped = compute_lhs_reshaped_shape(TensorInfo(lhs_shape, 1, data_type),
                                                                       lhs_info,
                                                                       false);
 
     const TensorShape rhs_shape(N, K, b_value);
-    const TensorShape rhs_shape_reshaped = compute_rhs_reshaped_shape(TensorInfo(rhs_shape, 1, DataType::F32),
+    const TensorShape rhs_shape_reshaped = compute_rhs_reshaped_shape(TensorInfo(rhs_shape, 1, data_type),
                                                                       rhs_info);
 
-    const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape_reshaped, 1, DataType::F32),
-                                                   TensorInfo(rhs_shape_reshaped, 1, DataType::F32),
+    const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape_reshaped, 1, data_type),
+                                                   TensorInfo(rhs_shape_reshaped, 1, data_type),
                                                    gemm_info);
 
     // Create tensors
-    CLTensor lhs_reshaped = create_tensor<CLTensor>(lhs_shape_reshaped, DataType::F32);
-    CLTensor rhs_reshaped = create_tensor<CLTensor>(rhs_shape_reshaped, DataType::F32);
-    CLTensor dst          = create_tensor<CLTensor>(dst_shape, DataType::F32);
+    CLTensor lhs_reshaped = create_tensor<CLTensor>(lhs_shape_reshaped, data_type);
+    CLTensor rhs_reshaped = create_tensor<CLTensor>(rhs_shape_reshaped, data_type);
+    CLTensor dst          = create_tensor<CLTensor>(dst_shape, data_type);
 
     ARM_COMPUTE_EXPECT(lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -177,11 +182,31 @@ m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_v
     CLGEMMMatrixMultiplyReshaped gemm;
     gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, 1.0f, lhs_info, rhs_info, gemm_info);
 }
+} // namespace
 
+TEST_SUITE(CL)
+TEST_SUITE(GEMMMatrixMultiplyReshaped)
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   framework::dataset::make("batch_size", 1)),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs),
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs)
+{
+    validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs, DataType::F32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<float>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -192,14 +217,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture, framework:
                                                                    v0_values_precommit),
                                                                    h0_values_precommit),
                                                                    i_values_lhs),
-                                                                   i_values_rhs))
+                                                                   i_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   a_values_precommit))
 {
     // Validate output
-    validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<float>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -210,13 +237,141 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture, framework:
                                                                    v0_values_nightly),
                                                                    h0_values_nightly),
                                                                    i_values_lhs),
-                                                                   i_values_rhs))
+                                                                   i_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   a_values_nightly))
 {
     // Validate output
-    validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   a_values_precommit))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_nightly),
+                                                                   n0_values_nightly),
+                                                                   k0_values_nightly),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   a_values_nightly))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
 }
-TEST_SUITE_END() // Float
 TEST_SUITE_END() // FP32
+
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<half>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   a_values_precommit))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<half>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_nightly),
+                                                                   n0_values_nightly),
+                                                                   k0_values_nightly),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   a_values_nightly))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   a_values_precommit))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values,
+                                                                   m_h_values),
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_nightly),
+                                                                   n0_values_nightly),
+                                                                   k0_values_nightly),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   a_values_nightly))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+TEST_SUITE_END() // FP16
+TEST_SUITE_END() // Float
 TEST_SUITE_END() // GEMMMatrixMulipltyReshaped
 TEST_SUITE_END() // CL
 } // namespace validation
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index ce2b177ce9..24c9d96611 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -151,13 +151,13 @@ protected:
     SimpleTensor<T> _reference{};
 };
 
-template <typename TensorType, typename AccessorType, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
 class GEMMMatrixMultiplyReshapedValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
     void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool interleave_lhs,
-               bool interleave_rhs)
+               bool interleave_rhs, DataType data_type, float alpha)
     {
         GEMMLHSMatrixInfo lhs_info;
         lhs_info.m0         = m0;
@@ -177,8 +177,8 @@ public:
         const TensorShape lhs_shape(k, m, batch_size);
         const TensorShape rhs_shape(n, k, batch_size);
 
-        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info);
-        _reference = compute_reference(lhs_shape, rhs_shape);
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type, alpha);
+        _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha);
     }
 
 protected:
@@ -189,11 +189,11 @@ protected:
         library->fill(tensor, distribution, i);
     }
 
-    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info)
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, DataType data_type, float alpha)
     {
         // Create tensors
-        TensorType lhs = create_tensor<TensorType>(lhs_shape, DataType::F32, 1);
-        TensorType rhs = create_tensor<TensorType>(rhs_shape, DataType::F32, 1);
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1);
         TensorType lhs_reshaped;
         TensorType rhs_reshaped;
         TensorType dst;
@@ -210,7 +210,7 @@ protected:
         GEMMFunctionType       gemm;
         reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info);
         reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
-        gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, 1.0f, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K));
+        gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K));
 
         ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -240,27 +240,144 @@ protected:
         return dst;
     }
 
-    SimpleTensor<float> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape)
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha)
     {
         TensorShape dst_shape = lhs_shape;
         dst_shape[0]          = rhs_shape[0];
         dst_shape[1]          = lhs_shape[1];
 
         // Create reference
-        SimpleTensor<float> lhs{ lhs_shape, DataType::F32, 1 };
-        SimpleTensor<float> rhs{ rhs_shape, DataType::F32, 1 };
-        SimpleTensor<float> c{ dst_shape, DataType::F32, 1 };
+        SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> c{ dst_shape, data_type, 1 };
 
         // Fill reference
         fill(lhs, 0);
         fill(rhs, 1);
-        fill(c, 2);
 
-        return reference::gemm<float>(lhs, rhs, c, 1.0f, 0.0f);
+        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
     }
 
-    TensorType          _target{};
-    SimpleTensor<float> _reference{};
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMMatrixMultiplyReshaped3DValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0,
+               bool interleave_lhs,
+               bool interleave_rhs, DataType data_type, float alpha)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0         = m0;
+        lhs_info.k0         = k0;
+        lhs_info.v0         = v0;
+        lhs_info.interleave = interleave_lhs;
+        lhs_info.transpose  = false;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0         = n0;
+        rhs_info.k0         = k0;
+        rhs_info.h0         = h0;
+        rhs_info.interleave = interleave_rhs;
+        rhs_info.transpose  = true;
+
+        // In case of GEMM3D, m is the product between m_w and m_h
+        const unsigned int m = m_w * m_h;
+
+        // Set the tensor shapes for LHS and RHS matrices
+        const TensorShape lhs_shape(k, m, batch_size);
+        const TensorShape rhs_shape(n, k, batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type, alpha, m_h);
+        _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, m_h);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, DataType data_type, float alpha,
+                              unsigned int m_h)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType lhs_reshaped;
+        TensorType rhs_reshaped;
+        TensorType dst;
+
+        const unsigned int M = lhs_shape[1];
+        const unsigned int N = rhs_shape[0];
+        const unsigned int K = lhs_shape[0];
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        ReshapeLHSFunctionType reshape_lhs;
+        ReshapeRHSFunctionType reshape_rhs;
+        GEMMFunctionType       gemm;
+        reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info);
+        reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
+        gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h));
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        lhs_reshaped.allocator()->allocate();
+        rhs_reshaped.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(lhs), 0);
+        fill(AccessorType(rhs), 1);
+
+        // Compute GEMM
+        reshape_lhs.run();
+        reshape_rhs.run();
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, unsigned int m_h)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape.set(0, rhs_shape[0]);
+        dst_shape.set(1, lhs_shape[1] / m_h);
+        dst_shape.set(2, m_h);
+        dst_shape.set(3, lhs_shape[2]);
+
+        // Create reference
+        SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> c{ dst_shape, data_type, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+
+        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
 };
 } // namespace validation
 } // namespace test