Implement MatMul Function

Resolves: COMPMID-5949
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: Idd8cfe6ea94a14f0b23178f6781251b5f0955563
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9390
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

11 files changed, 643 insertions, 5 deletions

diff --git a/Android.bp b/Android.bp
index 5617812539..f315def2e6 100644
--- a/Android.bp
+++ b/Android.bp
@@ -741,6 +741,7 @@ cc_library_static {
         "src/gpu/cl/operators/ClGemmLowpOutputStage.cpp",
         "src/gpu/cl/operators/ClIndirectConv2d.cpp",
         "src/gpu/cl/operators/ClLogicalNot.cpp",
+        "src/gpu/cl/operators/ClMatMul.cpp",
         "src/gpu/cl/operators/ClMul.cpp",
         "src/gpu/cl/operators/ClPRelu.cpp",
         "src/gpu/cl/operators/ClPermute.cpp",
@@ -823,6 +824,7 @@ cc_library_static {
         "src/runtime/CL/functions/CLLogicalAnd.cpp",
         "src/runtime/CL/functions/CLLogicalNot.cpp",
         "src/runtime/CL/functions/CLLogicalOr.cpp",
+        "src/runtime/CL/functions/CLMatMul.cpp",
         "src/runtime/CL/functions/CLMaxUnpoolingLayer.cpp",
         "src/runtime/CL/functions/CLMeanStdDevNormalizationLayer.cpp",
         "src/runtime/CL/functions/CLNormalizationLayer.cpp",
diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index e37134d454..26e459680c 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2022 Arm Limited.
+ * Copyright (c) 2016-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -77,6 +77,7 @@
 #include "arm_compute/runtime/CL/functions/CLLogicalAnd.h"
 #include "arm_compute/runtime/CL/functions/CLLogicalNot.h"
 #include "arm_compute/runtime/CL/functions/CLLogicalOr.h"
+#include "arm_compute/runtime/CL/functions/CLMatMul.h"
 #include "arm_compute/runtime/CL/functions/CLMaxUnpoolingLayer.h"
 #include "arm_compute/runtime/CL/functions/CLMeanStdDevNormalizationLayer.h"
 #include "arm_compute/runtime/CL/functions/CLNormalizationLayer.h"
diff --git a/arm_compute/runtime/CL/functions/CLMatMul.h b/arm_compute/runtime/CL/functions/CLMatMul.h
new file mode 100644
index 0000000000..56dd9c5655
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLMatMul.h
@@ -0,0 +1,99 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_RUNTIME_CL_FUNCTIONS_CLMATMUL
+#define ARM_COMPUTE_RUNTIME_CL_FUNCTIONS_CLMATMUL
+
+#include "arm_compute/runtime/IFunction.h"
+#include <memory>
+namespace arm_compute
+{
+// Forward declarations for used types instead of including their header, that could minimize compile time
+class CLCompileContext;
+class ICLTensor;
+class ITensorInfo;
+class MatMulInfo;
+class Status;
+
+/** Basic function to execute MatMul (Matrix Multiplication) on OpenCL */
+class CLMatMul : public IFunction
+{
+public:
+    /** Default constructor.*/
+    CLMatMul();
+    /** Default destructor */
+    ~CLMatMul();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLMatMul(const CLMatMul &) = delete;
+    /** Default move constructor */
+    CLMatMul(CLMatMul &&);
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLMatMul &operator=(const CLMatMul &) = delete;
+    /** Default move assignment operator */
+    CLMatMul &operator=(CLMatMul &&);
+    /** Initialise the kernel's inputs and output
+     *
+     * Valid data layouts:
+     * - All
+     *
+     * Valid data type configurations:
+     * |lhs          |rhs          |output         |
+     * |:------------|:------------|:--------------|
+     * |F32          |F32          |F32            |
+     * |F16          |F16          |F16            |
+     *
+     * @note BatchMatMul: Batched Matrix Multiply - [A * B], Multiplies all slices (slice is an element of a batch) of Tensors A and B
+     *                    and stores the result in the dst tensor of the same batch size.
+     *                    Batch here is number of slices from A and B multiplied at a time, do not confuse with the batch dimension 'N' of NHWC/NCHW
+     *                    For NHWC for example: the batch is the higher dimensions H * N, and in general it is H * all higher dimensions.
+     * @note All tensors must have the same data type.
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  lhs             LHS input tensor (Matrix or Vector A). Data types supported: F16/F32
+     * @param[in]  rhs             RHS input tensor (Matrix B). Data type supported: same as @p lhs.
+     * @param[out] output          Output tensor. Data type supported: same as @p lhs.
+     * @param[in]  matmul_info     Attributes for MatMul
+     */
+    void configure(const CLCompileContext &compile_context, ICLTensor *rhs, ICLTensor *lhs, ICLTensor *output, const MatMulInfo &matmul_info);
+    /** Initialise the kernel's inputs and output
+     *
+     * Similar to @ref CLMatMul::configure()
+     */
+    void configure(ICLTensor *lhs, ICLTensor *rhs, ICLTensor *output, const MatMulInfo &matmul_info);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLMatMul.
+     *
+     * Similar to @ref CLMatMul::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *output, const MatMulInfo &matmul_info);
+    // Inherited methods overridden:
+    void run() override;
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
+};
+} // namespace arm_compute
+
+#endif /* ARM_COMPUTE_RUNTIME_CL_FUNCTIONS_CLMATMUL */
diff --git a/filelist.json b/filelist.json
index 1e59adfc8e..c8e1ce0b9b 100644
--- a/filelist.json
+++ b/filelist.json
@@ -512,7 +512,9 @@
     "MatMul": {
       "files": {
         "common": [
-          "src/gpu/cl/kernels/ClNativeMatMulKernel.cpp"
+          "src/gpu/cl/kernels/ClNativeMatMulKernel.cpp",
+          "src/gpu/cl/operators/ClMatMul.cpp",
+          "src/runtime/CL/functions/CLMatMul.cpp"
         ]
       }
     },
diff --git a/src/gpu/cl/kernels/ClNativeMatMulKernel.h b/src/gpu/cl/kernels/ClNativeMatMulKernel.h
index 021292a4ae..3d0f18ec84 100644
--- a/src/gpu/cl/kernels/ClNativeMatMulKernel.h
+++ b/src/gpu/cl/kernels/ClNativeMatMulKernel.h
@@ -24,8 +24,6 @@
 #ifndef ACL_SRC_GPU_CL_KERNELS_CLNATIVEMATMULKERNEL
 #define ACL_SRC_GPU_CL_KERNELS_CLNATIVEMATMULKERNEL
 
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/KernelDescriptors.h"
 #include "src/core/common/Macros.h"
 #include "src/gpu/cl/ClCompileContext.h"
@@ -65,7 +63,7 @@ public:
     void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
 
 private:
-    bool _export_rhs_to_cl_image { false };
+    bool _export_rhs_to_cl_image{ false };
 };
 } // namespace kernels
 } // namespace opencl
diff --git a/src/gpu/cl/operators/ClMatMul.cpp b/src/gpu/cl/operators/ClMatMul.cpp
new file mode 100644
index 0000000000..dadaa1f779
--- /dev/null
+++ b/src/gpu/cl/operators/ClMatMul.cpp
@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "src/gpu/cl/operators/ClMatMul.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+#include "src/common/utils/Log.h"
+#include "src/gpu/cl/kernels/ClNativeMatMulKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+using namespace arm_compute::opencl::kernels;
+ClMatMul::ClMatMul()
+    : _native_matmul_kernel(std::make_unique<ClNativeMatMulKernel>())
+{
+}
+ClMatMul::~ClMatMul()
+{
+}
+Status ClMatMul::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *output, const MatMulInfo &matmul_info)
+{
+    MatMulKernelInfo kernel_info;
+    kernel_info.adj_lhs = matmul_info.adj_lhs();
+    kernel_info.adj_rhs = matmul_info.adj_rhs();
+    return ClNativeMatMulKernel::validate(lhs, rhs, output, kernel_info);
+}
+void ClMatMul::configure(const CLCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *output, const MatMulInfo &matmul_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, output);
+    ARM_COMPUTE_LOG_PARAMS(lhs, rhs, output, matmul_info);
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, output, matmul_info));
+    const GPUTarget gpu_target = CLScheduler::get().target();
+
+    // Placeholder: Getting the heuristics calculated values for M0, N0, K0, and whether to export RHS to texture pipe
+
+    // Filling the MatMul Kernel info
+    MatMulKernelInfo kernel_info;
+    kernel_info.adj_lhs                = matmul_info.adj_lhs();
+    kernel_info.adj_rhs                = matmul_info.adj_rhs();
+    kernel_info.m0                     = 1;     // to be properly calculated from heuristics
+    kernel_info.n0                     = 4;     // to be properly calculated from heuristics
+    kernel_info.k0                     = 4;     // to be properly calculated from heuristics
+    kernel_info.export_rhs_to_cl_image = false; // to be properly determined from heuristics
+
+    // Set the target for the kernels
+    _native_matmul_kernel->set_target(gpu_target);
+
+    // Configure the native matrix multiply kernel
+    _native_matmul_kernel->configure(compile_context, lhs, rhs, output, kernel_info);
+}
+void ClMatMul::run(ITensorPack &tensors)
+{
+    CLScheduler::get().enqueue_op(*_native_matmul_kernel, tensors, true);
+}
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/gpu/cl/operators/ClMatMul.h b/src/gpu/cl/operators/ClMatMul.h
new file mode 100644
index 0000000000..894b8d5816
--- /dev/null
+++ b/src/gpu/cl/operators/ClMatMul.h
@@ -0,0 +1,84 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_SRC_GPU_CL_OPERATORS_ClMatMul
+#define ARM_COMPUTE_SRC_GPU_CL_OPERATORS_ClMatMul
+
+#include "src/gpu/cl/IClOperator.h"
+#include "src/gpu/cl/kernels/ClNativeMatMulKernel.h"
+#include <memory>
+
+namespace arm_compute
+{
+namespace opencl
+{
+/** Basic operator to execute BatchMatMul on OpenCL. This operator calls the following OpenCL kernels:
+ *
+ *  -# @ref kernels::ClNativeMatMulKernel
+ */
+class ClMatMul : public IClOperator
+{
+public:
+    /** Constructor */
+    ClMatMul();
+    ~ClMatMul();
+    /** Initialise the kernel's inputs and output
+     *
+     * Valid data layouts:
+     * - All
+     *
+     * Valid data type configurations:
+     * |lhs          |rhs          |output       |
+     * |:------------|:------------|:------------|
+     * |F32          |F32          |F32          |
+     * |F16          |F16          |F16          |
+     *
+     * @note BatchMatMul: Batched Matrix Multiply - [A * B], Multiplies all slices (slice is an element of a batch) of Tensors A and B
+     *                    and stores the result in the dst tensor of the same batch size.
+     *                    Batch here is number of slices from A and B multiplied at a time, do not confuse with the batch dimension 'N' of NHWC/NCHW
+     *                    For NHWC for example: the batch is the higher dimensions H * N, and in general it is H * all higher dimensions.
+     * @note All tensors must have the same data type.
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  lhs             LHS input tensor info (Matrix A). Data types supported: F16/F32
+     * @param[in]  rhs             RHS input tensor info (Matrix B). Data types supported: same as @p lhs.
+     * @param[out] output          Output tensor info. Data types supported: same as @p lhs
+     * @param[in]  matmul_info     Attributes for MatMul
+     */
+    void configure(const CLCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *output, const MatMulInfo &matmul_info);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to @ref ClMatMul::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *output, const MatMulInfo &matmul_info);
+    // Inherited methods overridden:
+    void run(ITensorPack &tensors) override;
+
+private:
+    std::unique_ptr<kernels::ClNativeMatMulKernel> _native_matmul_kernel;
+};
+} // namespace opencl
+} // namespace arm_compute
+#endif // ARM_COMPUTE_SRC_GPU_CL_OPERATORS_ClMatMul
diff --git a/src/runtime/CL/functions/CLMatMul.cpp b/src/runtime/CL/functions/CLMatMul.cpp
new file mode 100644
index 0000000000..f42e4ff309
--- /dev/null
+++ b/src/runtime/CL/functions/CLMatMul.cpp
@@ -0,0 +1,69 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/CL/functions/CLMatMul.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTypes.h"
+#include "src/gpu/cl/operators/ClMatMul.h"
+
+namespace arm_compute
+{
+using OperatorType = opencl::ClMatMul;
+
+struct CLMatMul::Impl
+{
+    std::unique_ptr<OperatorType> op{ nullptr };
+    ITensorPack                   run_pack{};
+};
+CLMatMul::CLMatMul()
+    : _impl(std::make_unique<Impl>())
+{
+}
+
+CLMatMul::~CLMatMul() = default;
+
+void CLMatMul::configure(ICLTensor *lhs, ICLTensor *rhs, ICLTensor *output, const MatMulInfo &matmul_info)
+{
+    configure(CLKernelLibrary::get().get_compile_context(), lhs, rhs, output, matmul_info);
+}
+
+void CLMatMul::configure(const CLCompileContext &compile_context, ICLTensor *lhs, ICLTensor *rhs, ICLTensor *output, const MatMulInfo &matmul_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, output);
+
+    _impl->op = std::make_unique<OperatorType>();
+    _impl->op->configure(compile_context, lhs->info(), rhs->info(), output->info(), matmul_info);
+    _impl->run_pack = { { ACL_SRC_0, lhs }, { ACL_SRC_1, rhs }, { ACL_DST, output } };
+}
+
+Status CLMatMul::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *output, const MatMulInfo &matmul_info)
+{
+    return OperatorType::validate(lhs, rhs, output, matmul_info);
+}
+
+void CLMatMul::run()
+{
+    _impl->op->run(_impl->run_pack);
+}
+
+} // namespace arm_compute
diff --git a/tests/validation/CL/MatMul.cpp b/tests/validation/CL/MatMul.cpp
new file mode 100644
index 0000000000..bd259f785e
--- /dev/null
+++ b/tests/validation/CL/MatMul.cpp
@@ -0,0 +1,94 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/functions/CLMatMul.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/datasets/LargeMatMulDataset.h"
+#include "tests/datasets/SmallMatMulDataset.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/MatMulFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for fp32 data type */
+constexpr float          abs_tolerance_f32(
+    0.0001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp32 data type in case using relative tolerance fails because of small values */
+constexpr float abs_tolerance_f16(
+    0.001f);                                                   /**< Absolute tolerance value for comparing reference's output against implementation's output for fp16  data type in case using relative tolerance fails because of small values */
+RelativeTolerance<half_float::half> tolerance_f16(half(0.01)); /**< Tolerance value for comparing reference's output against implementation's output for fp16 data type */
+} // namespace
+template <typename T>
+using MatMulFixture = MatMulValidationFixture<CLTensor, CLAccessor, CLMatMul, T>;
+
+TEST_SUITE(CL)
+TEST_SUITE(MatMul)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, MatMulFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallMatMulDataset(),
+                                                                                                                  framework::dataset::make("pretransose_A", { false, true })),
+                                                                                                          framework::dataset::make("pretransose_B", { false, true })),
+                                                                                                  framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, MatMulFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeMatMulDataset(),
+                                                                                                                framework::dataset::make("pretransose_A", { false, true })),
+                                                                                                        framework::dataset::make("pretransose_B", { false, true })),
+                                                                                                framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, MatMulFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallMatMulDataset(),
+                                                                                                                 framework::dataset::make("pretransose_A", { false, true })),
+                                                                                                         framework::dataset::make("pretransose_B", { false, true })),
+                                                                                                 framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, MatMulFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeMatMulDataset(),
+                                                                                                               framework::dataset::make("pretransose_A", { false, true })),
+                                                                                                       framework::dataset::make("pretransose_B", { false, true })),
+                                                                                               framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END() // FP16
+TEST_SUITE_END() // MatMul
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/MatMulFixture.h b/tests/validation/fixtures/MatMulFixture.h
new file mode 100644
index 0000000000..1112dcb2fb
--- /dev/null
+++ b/tests/validation/fixtures/MatMulFixture.h
@@ -0,0 +1,180 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef TESTS_VALIDATION_FIXTURES_MATMULFIXTURE
+#define TESTS_VALIDATION_FIXTURES_MATMULFIXTURE
+
+#include "arm_compute/core/Types.h"
+#include "tests/framework/Fixture.h"
+#include "tests/validation/reference/GEMM.h"
+#include "tests/validation/reference/Permute.h"
+#include "tests/validation/reference/Permute.h"
+#include "tests/validation/reference/ReshapeLayer.h"
+#include <random>
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class MatMulValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape output_shape, bool pretranspose_a, bool pretranspose_b, DataType data_type)
+    {
+        // For brevity, the input shapes are assumed to be not-transposed for both Lhs and Rhs matrices.
+        if(pretranspose_a)
+        {
+            permute(shape_a, PermutationVector(1U, 0U));
+        }
+        if(pretranspose_b)
+        {
+            permute(shape_b, PermutationVector(1U, 0U));
+        }
+        _target    = compute_target(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, data_type);
+        _reference = compute_reference(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, data_type);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i, float lo = -1.f, float hi = 1.f)
+    {
+        switch(tensor.data_type())
+        {
+            case DataType::F16:
+            {
+                arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ float(lo), float(hi) };
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            case DataType::F32:
+            {
+                std::uniform_real_distribution<float> distribution(lo, hi);
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            default:
+                library->fill_tensor_uniform(tensor, i);
+        }
+    }
+    TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &output_shape, bool pretranspose_a, bool pretranspose_b, DataType data_type)
+    {
+        // 1. Create Classes and configure function
+        // Create tensors
+        TensorType   a   = create_tensor<TensorType>(shape_a, data_type, 1);
+        TensorType   b   = create_tensor<TensorType>(shape_b, data_type, 1);
+        TensorType   dst = create_tensor<TensorType>(output_shape, data_type, 1);
+        FunctionType matmul;
+        // Configure MatMulInfo class
+        MatMulInfo info;
+        info.adj_lhs(pretranspose_a);
+        info.adj_rhs(pretranspose_b);
+        matmul.configure(&a, &b, &dst, info);
+        // Assertions
+        ARM_COMPUTE_ASSERT(a.info()->is_resizable());
+        ARM_COMPUTE_ASSERT(b.info()->is_resizable());
+        ARM_COMPUTE_ASSERT(dst.info()->is_resizable());
+        // Allocate tensors
+        a.allocator()->allocate();
+        b.allocator()->allocate();
+        dst.allocator()->allocate();
+        ARM_COMPUTE_ASSERT(!a.info()->is_resizable());
+        ARM_COMPUTE_ASSERT(!b.info()->is_resizable());
+        ARM_COMPUTE_ASSERT(!dst.info()->is_resizable());
+
+        // 2. Fill tensors and run once
+        // Fill tensors
+        fill(AccessorType(a), 0);
+        fill(AccessorType(b), 1);
+        matmul.run(); // First run
+
+        return dst;
+    }
+    SimpleTensor<T> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &output_shape, bool pretranspose_a, bool pretranspose_b, DataType data_type)
+    {
+        // We collapse dimensions > 3 onto dimension 3, i.e. 5D+ tensors will look like 4D
+        // This is necessary unless we choose to extend gemm reference for 5D+ tensors
+        TensorShape output_shape_collapsed = output_shape.collapsed_from(Window::DimW);
+        TensorShape a_shape_collapsed      = shape_a.collapsed_from(Window::DimW);
+        TensorShape b_shape_collapsed      = shape_b.collapsed_from(Window::DimW);
+
+        // Create reference
+        SimpleTensor<T> a{ a_shape_collapsed, data_type, 1 };
+        SimpleTensor<T> b{ b_shape_collapsed, data_type, 1 };
+        SimpleTensor<T> c{ output_shape_collapsed, data_type, 1 };
+
+        // Fill reference
+        fill(a, 0);
+        fill(b, 1);
+
+        /* Note: Assuming the usual batch matmul dimensions A = (B x M x K), B = (B x K x N), if pretranspose_a is set to true, then A is assumed to be (B x K x M),
+           therefore, A must be pre-transposed before passing it to the fixture. And, we transpose A again in the fixture to make it (B x M x K)
+           in order to be able to call reference implementation that works with (B x M x K) input.
+           Similarly, if pretranspose_b is set to true, then B is assumed to be (B x N x K), B must be pre-transposed before passing it to the fixture. */
+
+        // Define transposed shapes
+        TensorShape a_transposed_shape(a.shape());
+        a_transposed_shape.set(0, a.shape().y());
+        a_transposed_shape.set(1, a.shape().x());
+        TensorShape b_transposed_shape(b.shape());
+        b_transposed_shape.set(0, b.shape().y());
+        b_transposed_shape.set(1, b.shape().x());
+
+        // Define transposed tensors
+        SimpleTensor<T> a_transposed{ a_transposed_shape, data_type };
+        SimpleTensor<T> b_transposed{ b_transposed_shape, data_type };
+
+        // pretranspose a if necessary
+        if(pretranspose_a)
+        {
+            a_transposed = reference::permute<T>(a, PermutationVector(1U, 0U));
+        }
+
+        // pretranspose b if necessary
+        if(pretranspose_b)
+        {
+            b_transposed = reference::permute<T>(b, PermutationVector(1U, 0U));
+        }
+
+        // Setting beta to 0 will effectively disable C for the
+        // computation of the reference: alpha * A * B + 0 * C
+        // Use transposed tensors if boolean enabled else use original tensors
+        SimpleTensor<T> result = reference::gemm<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, 1.0f, 0.f);
+
+        // We reshape the gemm output back if the tensor is high dimensional
+        if(output_shape_collapsed != output_shape)
+        {
+            result = reference::reshape_layer(result, output_shape);
+        }
+
+        return result;
+    }
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* TESTS_VALIDATION_FIXTURES_MATMULFIXTURE */
diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h
index c3af0a2419..9b9c7b5b34 100644
--- a/utils/TypePrinter.h
+++ b/utils/TypePrinter.h
@@ -3678,6 +3678,35 @@ inline std::string to_string(const experimental::dynamic_fusion::SoftmaxAttribut
     str << softmax_attr;
     return str.str();
 }
+/** Formatted output of the arm_compute::MatMulInfo type.
+ *
+ * @param[out] os          Output stream.
+ * @param[in]  matmul_info arm_compute::MatMulInfo  type to output.
+ *
+ * @return Modified output stream.
+ */
+inline ::std::ostream &operator<<(::std::ostream &os, const arm_compute::MatMulInfo &matmul_info)
+{
+    os << "MatMulKernelInfo="
+       << "["
+       << "adj_lhs=" << matmul_info.adj_lhs() << ", "
+       << "adj_rhs=" << matmul_info.adj_rhs() << ", "
+       << "fused_activation=" << matmul_info.fused_activation().activation() << "]";
+
+    return os;
+}
+/** Formatted output of the arm_compute::MatMulInfo type.
+ *
+ * @param[in] matmul_info arm_compute::MatMulInfo type to output.
+ *
+ * @return Formatted string.
+ */
+inline std::string to_string(const arm_compute::MatMulInfo &matmul_info)
+{
+    std::stringstream str;
+    str << matmul_info;
+    return str.str();
+}
 
 /** Formatted output of the arm_compute::MatMulKernelInfo type.
  *