COMPMID-1687: Optimize CLGEMMMatrixMultiplyKernel for Mali-G76 - Part1

The current implementation is limited just to FP32

Change-Id: I185ab57e483e879d7c301e9cc3033efc8b41e244
Reviewed-on: https://review.mlplatform.org/389
Reviewed-by: Anthony Barbier <Anthony.barbier@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

4 files changed, 139 insertions, 15 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index 37b92f2d6c..d89426dd32 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -78,6 +78,7 @@
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixAdditionKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
new file mode 100644
index 0000000000..d0f67e6f2c
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (c) 2018 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_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H__
+#define __ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** OpenCL kernel to multiply matrices when both the input matrices LHS (input0) and RHS (input1) have been reshaped
+ *
+ * @note The input matrices @p input0 and @p input1 must be reshaped through @ref CLGEMMReshapeLHSMatrixKernel and  @ref CLGEMMReshapeRHSMatrixKernel
+ */
+class CLGEMMMatrixMultiplyReshapedKernel : public ICLKernel
+{
+public:
+    /** Default Constructor */
+    CLGEMMMatrixMultiplyReshapedKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMMatrixMultiplyReshapedKernel(const CLGEMMMatrixMultiplyReshapedKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLGEMMMatrixMultiplyReshapedKernel &operator=(const CLGEMMMatrixMultiplyReshapedKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLGEMMMatrixMultiplyReshapedKernel(CLGEMMMatrixMultiplyReshapedKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLGEMMMatrixMultiplyReshapedKernel &operator=(CLGEMMMatrixMultiplyReshapedKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  input0    Input tensor containing the LHS reshaped matrix. Data type supported: F32/F16
+     * @param[in]  input1    Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0
+     * @param[out] output    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
+     * @param[in]  alpha     Weight of the matrix product
+     * @param[in]  lhs_info  LHS matrix information used for reshaping the input0 tensor
+     * @param[in]  rhs_info  RHS matrix information used for reshaping the input1 tensor
+     * @param[in]  gemm_info GEMM information used to retrieve the original dimensions of the input matrices
+     */
+    void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                   const GEMMReshapeInfo &gemm_info);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMMatrixMultiplyReshapedKernel
+     *
+     * @param[in] input0    Input tensor containing the LHS reshaped matrix. Data type supported: F32/F16
+     * @param[in] input1    Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0
+     * @param[in] output    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
+     * @param[in] alpha     Weight of the matrix product
+     * @param[in] lhs_info  LHS matrix information used for reshaping the input0 tensor
+     * @param[in] rhs_info  RHS matrix information used for reshaping the input1 tensor
+     * @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                           const GEMMReshapeInfo &gemm_info);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor *_input0;
+    const ICLTensor *_input1;
+    ICLTensor       *_output;
+    bool             _slide_matrix_b;
+    bool             _reinterpret_output_as_3d;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H__*/
\ No newline at end of file
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 33893ad877..f41d00f54d 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -619,6 +619,31 @@ inline TensorShape compute_mm_shape(const ITensorInfo &input0, const ITensorInfo
     return output_shape;
 }
 
+inline TensorShape compute_mm_shape(const ITensorInfo &input0, const ITensorInfo &input1, const GEMMReshapeInfo &gemm_info)
+{
+    ARM_COMPUTE_ERROR_ON_MSG(input0.num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4");
+
+    const bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d() != 0;
+    const int  depth_output_gemm3d      = reinterpret_output_as_3d ? gemm_info.depth_output_gemm3d() : 1;
+
+    // If the output of GEMM has to be reinterpreted as 3D, the number of input0 rows (M) is obtained collapsing the second and third
+    // dimension of the output tensor
+    const int dim0 = gemm_info.n();
+    const int dim1 = gemm_info.m() / depth_output_gemm3d;
+    const int dim2 = input0.tensor_shape()[2];
+    const int dim3 = input0.tensor_shape()[3];
+
+    TensorShape output_shape{ input0.tensor_shape() };
+
+    output_shape.set(0, dim0);
+    output_shape.set(1, dim1);
+    output_shape.set(2, reinterpret_output_as_3d ? depth_output_gemm3d : dim2);
+    output_shape.set(3, reinterpret_output_as_3d ? dim2 : dim3);
+    output_shape.set(4, reinterpret_output_as_3d ? dim3 : 1);
+
+    return output_shape;
+}
+
 inline TensorShape compute_output_stage_shape(const ITensorInfo &input, unsigned int gemm_3d_depth = 1, bool batch_size_on_z = false)
 {
     ARM_COMPUTE_ERROR_ON(input.data_layout() != DataLayout::NHWC && gemm_3d_depth > 1);
diff --git a/arm_compute/runtime/CL/functions/CLGEMM.h b/arm_compute/runtime/CL/functions/CLGEMM.h
index c4513f29d9..7d47194e56 100644
--- a/arm_compute/runtime/CL/functions/CLGEMM.h
+++ b/arm_compute/runtime/CL/functions/CLGEMM.h
@@ -27,6 +27,9 @@
 #include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixAdditionKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h"
 #include "arm_compute/runtime/CL/CLMemoryGroup.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
@@ -39,9 +42,12 @@ class ICLTensor;
 
 /** Basic function to execute GEMM on OpenCL. This function calls the following OpenCL kernels:
  *
- *  -# @ref CLGEMMInterleave4x4Kernel (only if the reshaped GEMM is selected by the heuristic model)
- *  -# @ref CLGEMMTranspose1xWKernel (only if the reshaped GEMM is selected by the heuristic model)
- *  -# @ref CLGEMMMatrixMultiplyKernel
+ *  -# @ref CLGEMMInterleave4x4Kernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target is NOT Mali-G76)
+ *  -# @ref CLGEMMReshapeLHSMatrixKernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target IS Mali-G76)
+ *  -# @ref CLGEMMTranspose1xWKernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target is NOT Mali-G76)
+ *  -# @ref CLGEMMReshapeRHSMatrixKernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target IS Mali-G76)
+ *  -# @ref CLGEMMMatrixMultiplyKernel (if GPU target is NOT G76 or if the reshaped GEMM is NOT selected)
+ *  -# @ref CLGEMMMatrixMultiplyReshapedKernel (only if the reshaped GEMM is selected by the heuristic model and the GPU target IS Mali-G76)
  *  -# @ref CLGEMMMatrixAdditionKernel (if c != nullptr and beta != 0.0)
  *
  */
@@ -100,18 +106,22 @@ public:
     void prepare() override;
 
 private:
-    CLMemoryGroup              _memory_group;
-    CLGEMMInterleave4x4Kernel  _interleave_kernel;
-    CLGEMMTranspose1xWKernel   _transpose_kernel;
-    CLGEMMMatrixMultiplyKernel _mm_kernel;
-    CLGEMMMatrixAdditionKernel _ma_kernel;
-    CLTensor                   _tmp_a;
-    CLTensor                   _tmp_b;
-    const ICLTensor           *_original_b;
-    bool                       _is_interleaved_transposed;
-    bool                       _run_addition;
-    bool                       _reshape_b_only_on_first_run;
-    bool                       _is_prepared;
+    CLMemoryGroup                      _memory_group;
+    CLGEMMInterleave4x4Kernel          _interleave_kernel; // TODO - COMPMID-1835: Remove this kernel and use CLGEMMReshapeLHSMatrixKernel
+    CLGEMMTranspose1xWKernel           _transpose_kernel;  // TODO - COMPMID-1836: Remove this kernel and use CLGEMMReshapeRHSMatrixKernel
+    CLGEMMMatrixMultiplyKernel         _mm_kernel;
+    CLGEMMMatrixAdditionKernel         _ma_kernel;
+    CLGEMMReshapeLHSMatrixKernel       _reshape_lhs_kernel;
+    CLGEMMReshapeRHSMatrixKernel       _reshape_rhs_kernel;
+    CLGEMMMatrixMultiplyReshapedKernel _mm_reshaped_kernel;
+    CLTensor                           _tmp_a;
+    CLTensor                           _tmp_b;
+    const ICLTensor                   *_original_b;
+    bool                               _is_interleaved_transposed;
+    bool                               _run_addition;
+    bool                               _reshape_b_only_on_first_run;
+    bool                               _is_prepared;
+    bool                               _is_G76_path; // TODO: To be removed once completed COMPMID-1835 and COMPMID-1836
 };
 }