Add Texture Pipe Support for Matmul Lhs T/NT Rhs NT kernels

Resolves: COMPMID-5945, COMPMID-5954

Change-Id: I7b27021d21f8e08c4896f6b1f595a75125064f9e
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9356
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 223 insertions, 41 deletions

diff --git a/tests/datasets/LargeMatMulDataset.h b/tests/datasets/LargeMatMulDataset.h
index cbc97d5e4a..b5181bc30b 100644
--- a/tests/datasets/LargeMatMulDataset.h
+++ b/tests/datasets/LargeMatMulDataset.h
@@ -54,6 +54,18 @@ public:
     }
 };
 
+class LargeMatMulDatasetRhsExportToCLImageRhsNT final : public MatMulDataset
+{
+public:
+    // For shape choices, please refer to the explanations given in SmallMatMulDatasetRhsExportToCLImageRhsNT
+    LargeMatMulDatasetRhsExportToCLImageRhsNT()
+    {
+        add_config(TensorShape(21U, 13U, 3U, 2U), TensorShape(32U, 21U, 3U, 2U), TensorShape(32U, 13U, 3U, 2U));
+        add_config(TensorShape(38U, 12U, 1U, 5U, 2U), TensorShape(20U, 38U, 1U, 5U, 2U), TensorShape(20U, 12U, 1U, 5U, 2U));
+        add_config(TensorShape(45U, 38U, 3U, 2U, 3U), TensorShape(20U, 45U, 3U, 2U, 3U), TensorShape(20U, 38U, 3U, 2U, 3U));
+    }
+};
+
 } // namespace datasets
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/datasets/SmallMatMulDataset.h b/tests/datasets/SmallMatMulDataset.h
index ae92b9abf5..93e5f7dc2c 100644
--- a/tests/datasets/SmallMatMulDataset.h
+++ b/tests/datasets/SmallMatMulDataset.h
@@ -57,6 +57,22 @@ public:
     }
 };
 
+class SmallMatMulDatasetRhsExportToCLImageRhsNT final : public MatMulDataset
+{
+public:
+    // Some considerations:
+    //  (1) N (Dimension 0 of Rhs matrix) dimension should be a multiple of 4
+    //  (2) Having N=20 enables us to test all possible N0 values, i.e. 4, 8, 16
+    //  (3) It's important to have more than one loop iterations in the K dimension
+    //      K has been chosen in accordance with K0
+    //  (4) The 5-th dimension has been chosen as non-unit because export_to_cl_iamge checks
+    //      were using dim1 * dim2 * dim3 to calculate the CLImage height; however, in our case
+    //      the tensor can be > 4D. To stress that case, the fifth dimension is chosen to be non-unit as well
+    SmallMatMulDatasetRhsExportToCLImageRhsNT()
+    {
+        add_config(TensorShape(7U, 3U, 2U, 1U, 2U), TensorShape(20U, 7U, 2U, 1U, 2U), TensorShape(20U, 3U, 2U, 1U, 2U));
+    }
+};
 } // namespace datasets
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/CL/MatMulKernel.cpp b/tests/validation/CL/MatMulKernel.cpp
index 5d2e59ab4c..59af8dba45 100644
--- a/tests/validation/CL/MatMulKernel.cpp
+++ b/tests/validation/CL/MatMulKernel.cpp
@@ -95,6 +95,12 @@ TEST_CASE(SupportedBlockSizes, framework::DatasetMode::ALL)
         { MatMulKernelInfo(false, false, 9, 1, 2), true },
         { MatMulKernelInfo(false, false, 3, 16, 3), true },
         { MatMulKernelInfo(false, false, 7, 3, 4), true },
+        { MatMulKernelInfo(false, false, 7, 3, 4, true), false }, // N0 not in {4, 8, 16}
+        { MatMulKernelInfo(false, false, 7, 1, 4, true), false }, // N0 not in {4, 8, 16}
+        { MatMulKernelInfo(false, false, 7, 12, 4, true), false }, // N0 not in {4, 8, 16}
+        { MatMulKernelInfo(false, false, 7, 4, 4, true), true },
+        { MatMulKernelInfo(false, false, 7, 8, 4, true), true },
+        { MatMulKernelInfo(false, false, 7, 16, 4, true), true },
 
         // Lhs not-transposed, Rhs transposed
         { MatMulKernelInfo(false, true, 0, 1, 1), false },  // M0 should be > 0
@@ -115,6 +121,12 @@ TEST_CASE(SupportedBlockSizes, framework::DatasetMode::ALL)
         { MatMulKernelInfo(true, false, 4, 1, 22), true },
         { MatMulKernelInfo(true, false, 3, 3, 3), true },
         { MatMulKernelInfo(true, false, 2, 4, 8), true },
+        { MatMulKernelInfo(true, false, 2, 3, 8, true), false }, // N0 not in {4, 8, 16}
+        { MatMulKernelInfo(true, false, 2, 7, 8, true), false }, // N0 not in {4, 8, 16}
+        { MatMulKernelInfo(true, false, 2, 5, 8, true), false }, // N0 not in {4, 8, 16}
+        { MatMulKernelInfo(true, false, 2, 4, 8, true), true },
+        { MatMulKernelInfo(true, false, 2, 8, 8, true), true },
+        { MatMulKernelInfo(true, false, 2, 16, 8, true), true },
 
         // // Lhs transposed, Rhs-transposed
         { MatMulKernelInfo(true, true, 2, 1, 5), false },  // K0 should in {1, 2, 3, 4, 8, 16}
@@ -134,12 +146,65 @@ TEST_CASE(SupportedBlockSizes, framework::DatasetMode::ALL)
     const TensorInfo lhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::F32);
     const TensorInfo rhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::F32);
 
+    const bool export_to_cl_image_supported = image2d_from_buffer_supported(CLKernelLibrary::get().get_device());
     for(auto &pair : supported_block_sizes)
     {
         TensorInfo output_info;
         Status     status = ClNativeMatMulKernel::validate(&lhs_info, &rhs_info, &output_info, pair.first);
 
-        ARM_COMPUTE_EXPECT(bool(status) == pair.second, framework::LogLevel::ERRORS);
+        if(!pair.first.export_rhs_to_cl_image || export_to_cl_image_supported)
+        {
+           ARM_COMPUTE_EXPECT(bool(status) == pair.second, framework::LogLevel::ERRORS);
+        }
+    }
+}
+
+TEST_CASE(ExportToCLImage, framework::DatasetMode::ALL)
+{
+    // We skip this test if the hardware does not support exporting to CL Image
+    if(image2d_from_buffer_supported(CLKernelLibrary::get().get_device()))
+    {
+        constexpr size_t pixel_size = 4;
+        const size_t max_image_w = pixel_size * CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>();
+        const size_t max_image_h = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>();
+
+        using ShapeConfigurationTuple = std::tuple<TensorShape, TensorShape, bool, bool, bool>;
+        const std::vector<ShapeConfigurationTuple> shape_configurations =
+        {
+            // lhs_shape, rhs_shape, adj_lhs, adj_rhs, expected
+            // Lhs t/Nt, Rhs Nt
+            // Transposition of Lhs doesn't add any value to the tests, therefore always assumed false below
+            { TensorShape(5U, 1U), TensorShape(3U, 5U), false, false, false }, // N should be multiple of 4
+            { TensorShape(5U, 1U), TensorShape(14U, 5U), false, false, false }, // N should be multiple of 4
+            { TensorShape(5U, 1U), TensorShape(12U, 5U), false, false, true },
+            { TensorShape(5U, 1U), TensorShape(8U, 5U), false, false, true },
+            { TensorShape(5U, 1U), TensorShape(4U, 5U), false, false, true },
+            { TensorShape(max_image_h + 1, 1U), TensorShape(4U, max_image_h + 1), false, false, false }, // Cannot fit into CL Image memory's height
+            { TensorShape(5U, 1U), TensorShape(max_image_w + 1, 5U), false, false, false }, // Cannot fit into CL Image memory's width
+            { TensorShape(max_image_h, 1U), TensorShape(4U, max_image_h), false, false, true }, // Barely fits into CL Image memory's height
+            { TensorShape(5U, 1U), TensorShape(max_image_w, 5U), false, false, true }, // Barely fits into CL Image memory's width
+        };
+
+        for(auto &tuple : shape_configurations)
+        {
+            TensorShape lhs_shape = std::get<0>(tuple);
+            TensorShape rhs_shape = std::get<1>(tuple);
+
+            const TensorInfo lhs_info = TensorInfo(lhs_shape, 1, DataType::F32);
+            const TensorInfo rhs_info = TensorInfo(rhs_shape, 1, DataType::F32);
+
+            const bool adj_lhs = std::get<2>(tuple);
+            const bool adj_rhs = std::get<3>(tuple);
+
+            // We choose M0, N0, K0 equal to 4 so that they're always valid for CLImage in any combination
+            const MatMulKernelInfo matmul_kernel_info {adj_lhs, adj_rhs, 4, 4, 4, true /* export_rhs_to_cl_image */};
+
+            TensorInfo output_info;
+            Status     status = ClNativeMatMulKernel::validate(&lhs_info, &rhs_info, &output_info, matmul_kernel_info);
+
+            const bool expected = std::get<4>(tuple);
+            ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
+        }
     }
 }
 
@@ -244,68 +309,75 @@ TEST_SUITE_END() // Validate
 
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunTiny, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(),
-                                                                                                                   framework::dataset::make("pretransose_A", { false, true })),
-                                                                                                                   framework::dataset::make("pretransose_B", { false, true })),
-                                                                                                                   m0_values_precommit),
-                                                                                                                   n0_values_precommit),
-                                                                                                           k0_values_precommit),
-                                                                                                   framework::dataset::make("DataType", DataType::F32)))
+TEST_SUITE(Buffer)
+FIXTURE_DATA_TEST_CASE(RunTiny, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(),
+                                                                                                                       framework::dataset::make("pretransose_A", { false, true })),
+                                                                                                                       framework::dataset::make("pretransose_B", { false, true })),
+                                                                                                                       m0_values_precommit),
+                                                                                                                       n0_values_precommit),
+                                                                                                                       k0_values_precommit),
+                                                                                                                       framework::dataset::make("export_rhs_to_cl_image", { false })),
+                                                                                                               framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
 }
-FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(),
-                                                                                                                    framework::dataset::make("pretransose_A", { false, true })),
-                                                                                                                    framework::dataset::make("pretransose_B", { false, true })),
-                                                                                                                    m0_values_precommit),
-                                                                                                                    n0_values_precommit),
-                                                                                                            k0_values_precommit),
-                                                                                                    framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(),
+                                                                                                                       framework::dataset::make("pretransose_A", { false, true })),
+                                                                                                                       framework::dataset::make("pretransose_B", { false, true })),
+                                                                                                                       m0_values_precommit),
+                                                                                                                       n0_values_precommit),
+                                                                                                                       k0_values_precommit),
+                                                                                                                       framework::dataset::make("export_rhs_to_cl_image", { false })),
+                                                                                                               framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
 }
-FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                                                                                                                    framework::dataset::make("pretransose_A", { false })),
                                                                                                                    framework::dataset::make("pretransose_B", { false })),
                                                                                                                    m0_values_nightly_lhs_nt),
                                                                                                                    n0_values_nightly_rhs_nt),
                                                                                                                    k0_values_nightly_lhs_nt_rhs_nt),
+                                                                                                                   framework::dataset::make("export_rhs_to_cl_image", { false })),
                                                                                                                    framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
 }
-FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                                                                                                                      framework::dataset::make("pretransose_A", { false })),
                                                                                                                      framework::dataset::make("pretransose_B", { true })),
                                                                                                                      m0_values_nightly_lhs_nt),
                                                                                                                      n0_values_nightly_rhs_t),
                                                                                                                      k0_values_nightly_rhs_t),
+                                                                                                                     framework::dataset::make("export_rhs_to_cl_image", { false })),
                                                                                                                      framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
 }
-FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                                                                                                                      framework::dataset::make("pretransose_A", { true })),
                                                                                                                      framework::dataset::make("pretransose_B", { false })),
                                                                                                                      m0_values_nightly_lhs_t),
                                                                                                                      n0_values_nightly_rhs_nt),
                                                                                                                      k0_values_nightly_lhs_t_rhs_nt),
+                                                                                                                     framework::dataset::make("export_rhs_to_cl_image", { false })),
                                                                                                                      framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
 }
 FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY,
-                       combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+                       combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                                                                        framework::dataset::make("pretransose_A", { true })),
                                                                framework::dataset::make("pretransose_B", { true })),
                                                        m0_values_nightly_lhs_t),
                                                n0_values_nightly_rhs_t),
                                        k0_values_nightly_rhs_t),
+                                       framework::dataset::make("export_rhs_to_cl_image", { false })),
                                framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
@@ -313,75 +385,150 @@ FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulKernelFixture
 }
 // Running High Dimensional test is enough for FP32, because we're stressing the number of dimensions, not data type or M0/N0/K0
 // It's a good idea to test for each Lhs/Rhs T/NT combinations because they're different CL kernels
-FIXTURE_DATA_TEST_CASE(RunHighDimensional, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::HighDimensionalMatMulDataset(),
-                                                                                                                      framework::dataset::make("pretransose_A", { false, true })),
-                                                                                                                      framework::dataset::make("pretransose_B", { false, true })),
-                                                                                                                      framework::dataset::make("M0", { 2 })),
-                                                                                                                      framework::dataset::make("N0", { 2 })),
-                                                                                                                      framework::dataset::make("K0", { 2 })),
-                                                                                                              framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunHighDimensional, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::HighDimensionalMatMulDataset(),
+                                                                                                                 framework::dataset::make("pretransose_A", { false, true })),
+                                                                                                                 framework::dataset::make("pretransose_B", { false, true })),
+                                                                                                                 framework::dataset::make("M0", { 2 })),
+                                                                                                                 framework::dataset::make("N0", { 2 })),
+                                                                                                                 framework::dataset::make("K0", { 2 })),
+                                                                                                                 framework::dataset::make("export_rhs_to_cl_image", { false })),
+                                                                                                                 framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
 }
+TEST_SUITE_END() // Buffer
+
+TEST_SUITE(ExportRhsToCLImage)
+FIXTURE_DATA_TEST_CASE(RunSmallRhsNotTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDatasetRhsExportToCLImageRhsNT(),
+                                                                                                                       framework::dataset::make("pretransose_A", { true, false })),
+                                                                                                                       framework::dataset::make("pretransose_B", { false })),
+                                                                                                                       framework::dataset::make("M0", { 2 })),
+                                                                                                                       framework::dataset::make("N0", { 4, 8, 16 })),
+                                                                                                                       framework::dataset::make("K0", { 2, 4 })),
+                                                                                                                       framework::dataset::make("export_rhs_to_cl_image", { true })),
+                                                                                                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    if(_device_supports_export_to_cl_image)
+    {
+        validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+    }
+}
+FIXTURE_DATA_TEST_CASE(RunLargeRhsNotTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDatasetRhsExportToCLImageRhsNT(),
+                                                                                                                       framework::dataset::make("pretransose_A", { true, false })),
+                                                                                                                       framework::dataset::make("pretransose_B", { false })),
+                                                                                                                       framework::dataset::make("M0", { 2 })),              // Choices of M0 does not matter much because it's related to Lhs tensor
+                                                                                                                       framework::dataset::make("N0", { 4, 8, 16 })),
+                                                                                                                       framework::dataset::make("K0", { 1, 2, 3, 4 })),
+                                                                                                                       framework::dataset::make("export_rhs_to_cl_image", { true })),
+                                                                                                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    if(_device_supports_export_to_cl_image)
+    {
+        validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+    }
+}
+TEST_SUITE_END() // ExportRhsToCLImage
 TEST_SUITE_END() // FP32
 
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulKernelFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(),
-                                                                                                                   framework::dataset::make("pretransose_A", { false, true })),
-                                                                                                                   framework::dataset::make("pretransose_B", { false, true })),
-                                                                                                                   m0_values_precommit),
-                                                                                                                   n0_values_precommit),
-                                                                                                           k0_values_precommit),
-                                                                                                   framework::dataset::make("DataType", DataType::F16)))
+TEST_SUITE(Buffer)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulKernelFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(),
+                                                                                                                      framework::dataset::make("pretransose_A", { false, true })),
+                                                                                                                      framework::dataset::make("pretransose_B", { false, true })),
+                                                                                                                      m0_values_precommit),
+                                                                                                                      n0_values_precommit),
+                                                                                                                      k0_values_precommit),
+                                                                                                                      framework::dataset::make("export_rhs_to_cl_image", { false })),
+                                                                                                              framework::dataset::make("DataType", DataType::F16)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16);
 }
-FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                                                                                                                   framework::dataset::make("pretransose_A", { false })),
                                                                                                                   framework::dataset::make("pretransose_B", { false })),
                                                                                                                   m0_values_nightly_lhs_nt),
                                                                                                                   n0_values_nightly_rhs_nt),
                                                                                                                   k0_values_nightly_lhs_nt_rhs_nt),
+                                                                                                                  framework::dataset::make("export_rhs_to_cl_image", { false })),
                                                                                                                   framework::dataset::make("DataType", DataType::F16)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16);
 }
-FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                                                                                                                     framework::dataset::make("pretransose_A", { false })),
                                                                                                                     framework::dataset::make("pretransose_B", { true })),
                                                                                                                     m0_values_nightly_lhs_nt),
                                                                                                                     n0_values_nightly_rhs_t),
                                                                                                                     k0_values_nightly_rhs_t),
+                                                                                                                    framework::dataset::make("export_rhs_to_cl_image", { false })),
                                                                                                                     framework::dataset::make("DataType", DataType::F16)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16);
 }
-FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                                                                                                                     framework::dataset::make("pretransose_A", { true })),
                                                                                                                     framework::dataset::make("pretransose_B", { false })),
                                                                                                                     m0_values_nightly_lhs_t),
                                                                                                                     n0_values_nightly_rhs_nt),
                                                                                                                     k0_values_nightly_lhs_t_rhs_nt),
+                                                                                                                    framework::dataset::make("export_rhs_to_cl_image", { false })),
                                                                                                                     framework::dataset::make("DataType", DataType::F16)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16);
 }
-FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                        framework::dataset::make("pretransose_A", { true })),
                        framework::dataset::make("pretransose_B", { true })),
                        m0_values_nightly_lhs_t),
                        n0_values_nightly_rhs_t),
                        k0_values_nightly_rhs_t),
+                       framework::dataset::make("export_rhs_to_cl_image", { false })),
                        framework::dataset::make("DataType", DataType::F16)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16);
 }
+TEST_SUITE_END() // Buffer
+
+TEST_SUITE(ExportRhsToCLImage)
+FIXTURE_DATA_TEST_CASE(RunSmallRhsCLImageRhsNotTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDatasetRhsExportToCLImageRhsNT(),
+                                                                                                                       framework::dataset::make("pretransose_A", { true, false })),
+                                                                                                                       framework::dataset::make("pretransose_B", { false })),
+                                                                                                                       framework::dataset::make("M0", { 2 })),
+                                                                                                                       framework::dataset::make("N0", { 4, 8, 16 })),
+                                                                                                                       framework::dataset::make("K0", { 2, 4 })),
+                                                                                                                       framework::dataset::make("export_rhs_to_cl_image", { true })),
+                                                                                                               framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    if(_device_supports_export_to_cl_image)
+    {
+        validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16);
+    }
+}
+FIXTURE_DATA_TEST_CASE(RunLargeRhsCLImageRhsNotTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDatasetRhsExportToCLImageRhsNT(),
+                                                                                                                       framework::dataset::make("pretransose_A", { true, false })),
+                                                                                                                       framework::dataset::make("pretransose_B", { false })),
+                                                                                                                       framework::dataset::make("M0", { 2 })),              // Choices of M0 does not matter much because it's related to Lhs tensor
+                                                                                                                       framework::dataset::make("N0", { 4, 8, 16 })),
+                                                                                                                       framework::dataset::make("K0", { 1, 2, 3, 4 })),
+                                                                                                                       framework::dataset::make("export_rhs_to_cl_image", { true })),
+                                                                                                               framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    if(_device_supports_export_to_cl_image)
+    {
+        validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16);
+    }
+}
+TEST_SUITE_END() // ExportRhsToCLImage
 TEST_SUITE_END() // FP16
 TEST_SUITE_END() // Float
 TEST_SUITE_END() // MatMulKernel
diff --git a/tests/validation/fixtures/MatMulKernelFixture.h b/tests/validation/fixtures/MatMulKernelFixture.h
index 459564618f..c131fea7fa 100644
--- a/tests/validation/fixtures/MatMulKernelFixture.h
+++ b/tests/validation/fixtures/MatMulKernelFixture.h
@@ -48,7 +48,7 @@ class MatMulKernelValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape output_shape, bool pretranspose_a, bool pretranspose_b, const int M0, const int N0, const int K0, DataType data_type)
+    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape output_shape, bool pretranspose_a, bool pretranspose_b, const int M0, const int N0, const int K0, bool export_rhs_to_cl_image, DataType data_type)
     {
         // For brevity, the input shapes are assumed to be not-transposed for both Lhs and Rhs matrices.
         if(pretranspose_a)
@@ -61,8 +61,13 @@ public:
             permute(shape_b, PermutationVector(1U, 0U));
         }
 
-        _target    = compute_target(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, M0, N0, K0, data_type);
-        _reference = compute_reference(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, data_type);
+        _device_supports_export_to_cl_image = image2d_from_buffer_supported(CLKernelLibrary::get().get_device());
+
+        if(!export_rhs_to_cl_image || _device_supports_export_to_cl_image)
+        {
+            _target    = compute_target(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, M0, N0, K0, export_rhs_to_cl_image, data_type);
+            _reference = compute_reference(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, data_type);
+        }
     }
 
 protected:
@@ -89,7 +94,7 @@ protected:
     }
 
     CLTensor compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &output_shape, bool pretranspose_a, bool pretranspose_b, const int M0, const int N0, const int K0,
-                            DataType data_type)
+                            bool export_rhs_to_cl_image, DataType data_type)
     {
         // Create tensors
         CLTensor a   = create_tensor<CLTensor>(shape_a, data_type, 1);
@@ -103,6 +108,7 @@ protected:
         matmul_info.m0      = M0;
         matmul_info.n0      = N0;
         matmul_info.k0      = K0;
+        matmul_info.export_rhs_to_cl_image = export_rhs_to_cl_image;
 
         matMul.configure(a.info(), b.info(), dst.info(), matmul_info);
         ARM_COMPUTE_ASSERT(a.info()->is_resizable());
@@ -195,6 +201,7 @@ protected:
 
     CLTensor        _target{};
     SimpleTensor<T> _reference{};
+    bool _device_supports_export_to_cl_image { true };
 };
 
 } // namespace validation