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>

4 files changed, 234 insertions, 3 deletions

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