COMPMID-2093: Implement CLGEMMNative

Change-Id: I347130f6b5ae8d08b7c5c101b523b158565874a1
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1114
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 640 insertions, 1 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index e102f44b88..dfcbfa7cc5 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -314,6 +314,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemm_mm_floating_point_f16_bifrost_acc32", "gemm.cl" },
     { "gemm_mm_floating_point_f32_bifrost", "gemm.cl" },
     { "gemm_mm_floating_point_f32_bifrost_1000", "gemm.cl" },
+    { "gemm_mm_native", "gemm.cl" },
     { "gemm_mm_reshaped_lhs_nt_rhs_t", "gemm.cl" },
     { "gemm_mm_reshaped_only_rhs_nt", "gemm.cl" },
     { "gemm_mm_reshaped_only_rhs_t", "gemm.cl" },
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index c3107a20f2..da45d0fc18 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -1535,7 +1535,7 @@ __kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs),
 #endif // M0 > 6
 #if M0 > 7
         VEC_DATA_TYPE(DATA_TYPE, 2)
-        a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin));
+        a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
 #endif // M0 > 7
 
         LD_RHS_VFMA_M0xN0(0, a, c);
@@ -1886,8 +1886,325 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs),
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 }
+
 #endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(K) && defined(DATA_TYPE)
 
+#if defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE)
+
+#define VFMA(a, b, c)     \
+    ({                    \
+        c = fma(a, b, c); \
+    })
+
+#if M0 == 1
+#define RHS_VFMA_M0xN0(i, a, b, c)                                    \
+    ({                                                                \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
+    })
+#elif M0 == 2 // M0 == 2
+#define RHS_VFMA_M0xN0(i, a, b, c)                                    \
+    ({                                                                \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
+    })
+#elif M0 == 3 // M0 == 3
+#define RHS_VFMA_M0xN0(i, a, b, c)                                    \
+    ({                                                                \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
+    })
+#elif M0 == 4 // M0 == 4
+#define RHS_VFMA_M0xN0(i, a, b, c)                                    \
+    ({                                                                \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
+    })
+#elif M0 == 5 // M0 == 5
+#define RHS_VFMA_M0xN0(i, a, b, c)                                    \
+    ({                                                                \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
+    })
+#elif M0 == 6 // M0 == 6
+#define RHS_VFMA_M0xN0(i, a, b, c)                                    \
+    ({                                                                \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
+    })
+#elif M0 == 7 // M0 == 7
+#define RHS_VFMA_M0xN0(i, a, b, c)                                    \
+    ({                                                                \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
+    })
+#elif M0 == 8 // M0 == 8
+#define RHS_VFMA_M0xN0(i, a, b, c)                                    \
+    ({                                                                \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
+        VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \
+    })
+#else // M0 not supported
+#error "M0 not supported"
+#endif // M0 not supported
+
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
+ *  The LHS matrix is NOT reshaped
+ *  The RHS matrix is NOT reshaped
+ *
+ * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
+ * @note The GEMM's dimensions (M,N and K) must be passed at compile time using -DM, -DN and and -DK (i.e. -DM=52, -DN=30 and -DK=90)
+ * @note The number of columns of LHS matrix must be passed at compile time using -DK (i.e. -DK=64)
+ * @note The number of M0 rows to process must be passed at compile time using -DM0 (i.e. -DM0=2)
+ * @note The number of K0 partial accumulations must be passed at compile time using -DK0 (i.e., -DK0=2)
+ * @note The number of N0 columns to process must be passed at compile time using -DN0 (i.e. -DN0=2)
+ * @note Only the following configurations of M0, N0 and K0 are currently supported:
+ *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
+ *  - N0 = 2, 3, 4, 8, 16
+ *  - K0 = 2, 3, 4, 8, 16
+ *
+ * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
+ *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
+ *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
+ *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
+ *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
+ *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
+ *
+ * @param[in]  lhs_ptr                           Pointer to the LHS reshaped matrix. Supported data type: F16/F32
+ * @param[in]  lhs_stride_x                      Stride of the LHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  lhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  lhs_stride_y                      Stride of the LHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  lhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
+ * @param[in]  rhs_ptr                           Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  rhs_stride_x                      Stride of the RHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  rhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  rhs_stride_y                      Stride of the RHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  rhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: same as @p lhs_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
+ * @param[in]  lhs_stride_z                      Stride of the LHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  rhs_stride_z                      Stride of the RHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  lhs_cross_plane_pad               (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
+ * @param[in]  dst_cross_plane_pad               (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
+ */
+__kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
+                             IMAGE_DECLARATION(rhs),
+                             IMAGE_DECLARATION(dst),
+                             uint lhs_stride_z,
+                             uint rhs_stride_z,
+                             uint dst_stride_z
+#if defined(REINTERPRET_INPUT_AS_3D)
+                             ,
+                             uint lhs_cross_plane_pad
+#endif // REINTERPRET_INPUT_AS_3D
+#if defined(REINTERPRET_OUTPUT_AS_3D)
+                             ,
+                             uint dst_cross_plane_pad
+#endif // REINTERPRET_OUTPUT_AS_3D
+                            )
+{
+    // Block size
+#define RHS_BLOCK_SIZE ((K0) * (N0))
+
+    // RHS offset and step X
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+#if defined(DUMMY_WORK_ITEMS)
+    if((x * N0 >= N) || (y * M0 >= M))
+    {
+        return;
+    }
+#endif // defined(DUMMY_WORK_ITEMS)
+
+    // Compute LHS matrix address
+    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+
+    // Compute RHS matrix address
+    uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE);
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
+    rhs_offset += z * rhs_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
+    REPEAT_VAR_INIT_TO_CONST(16, uint, zrhs, 0);
+
+#if defined(REINTERPRET_INPUT_AS_3D)
+    // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+
+    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
+    // multiply lhs_stride_z by DEPTH_GEMM3D
+    lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;
+
+#else // defined(REINTERPRET_INPUT_AS_3D)
+
+    // Add offset for batched GEMM
+    lhs_offset += z * lhs_stride_z;
+
+#endif // defined(REINTERPRET_INPUT_AS_3D)
+
+    // Initialize the accumulators
+    REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0)    c0=0,c1=0,c2=0,... c(N0-1)=0;
+
+    int i = 0;
+    for(; i <= (K - K0); i += K0)
+    {
+        // Supported cases (M0, K0):
+        // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
+        // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
+        // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
+        // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
+        // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
+        // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
+        // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
+        // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
+        // Load values from LHS matrix
+        LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
+
+        // Load values from RHS matrix
+        LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zrhs);
+
+        RHS_VFMA_M0xN0(0, a, b0, c);
+        RHS_VFMA_M0xN0(1, a, b1, c);
+#if K0 > 2
+        RHS_VFMA_M0xN0(2, a, b2, c);
+#endif // K0 > 2
+#if K0 > 3
+        RHS_VFMA_M0xN0(3, a, b3, c);
+#endif // K0 > 3
+#if K0 > 4
+        RHS_VFMA_M0xN0(4, a, b4, c);
+        RHS_VFMA_M0xN0(5, a, b5, c);
+        RHS_VFMA_M0xN0(6, a, b6, c);
+        RHS_VFMA_M0xN0(7, a, b7, c);
+#endif // K0 > 4
+#if K0 > 8
+        RHS_VFMA_M0xN0(8, a, b8, c);
+        RHS_VFMA_M0xN0(9, a, b9, c);
+        RHS_VFMA_M0xN0(A, a, b10, c);
+        RHS_VFMA_M0xN0(B, a, b11, c);
+        RHS_VFMA_M0xN0(C, a, b12, c);
+        RHS_VFMA_M0xN0(D, a, b13, c);
+        RHS_VFMA_M0xN0(E, a, b14, c);
+        RHS_VFMA_M0xN0(F, a, b15, c);
+#endif // K0 > 8
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        rhs_offset += K0 * rhs_stride_y;
+    }
+
+    // Left-over accumulations
+    for(; i < K; ++i)
+    {
+        // Load values from LHS matrix
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0));
+#if M0 > 1
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1));
+#endif // M0 > 1
+#if M0 > 2
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2));
+#endif // M0 > 2
+#if M0 > 3
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3));
+#endif // M0 > 3
+#if M0 > 4
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4));
+#endif // M0 > 4
+#if M0 > 5
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5));
+#endif // M0 > 5
+#if M0 > 6
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6));
+#endif // M0 > 6
+#if M0 > 7
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7));
+#endif // M0 > 7
+
+        VEC_DATA_TYPE(DATA_TYPE, N0)
+        b = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * rhs_stride_y));
+        RHS_VFMA_M0xN0(0, a, b, c);
+
+        lhs_offset += sizeof(DATA_TYPE);
+        rhs_offset += rhs_stride_y;
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * dst_stride_y);
+
+    REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
+
+#if defined(REINTERPRET_OUTPUT_AS_3D)
+    // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+
+    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
+    // multiply dst_stride_z by DEPTH_GEMM3D
+    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
+
+#else // defined(REINTERPRET_OUTPUT_AS_3D)
+
+    // Add offset for batched GEMM
+    dst_addr += z * dst_stride_z;
+
+#endif // defined(REINTERPRET_OUTPUT_AS_3D)
+
+    // Multiply by the weight of matrix-matrix product and store the result
+    // Multiply by the weight of matrix-matrix product and store the result
+#if defined(ALPHA)
+    SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
+#endif // defined(ALPHA)
+
+    // Store output block
+    STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+}
+#endif // defined(M0) && defined(N0) && defined(K0) && defined(K) && defined(DATA_TYPE)
+
 #if defined(TRANSPOSE_W) && defined(MULT_TRANSPOSE1XW_WIDTH)
 
 #if ELEMENT_SIZE == 1
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp
new file mode 100644
index 0000000000..85d882c865
--- /dev/null
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp
@@ -0,0 +1,321 @@
+/*
+ * Copyright (c) 2019 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/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "support/ToolchainSupport.h"
+
+#include <cstddef>
+#include <cstdint>
+#include <tuple>
+
+using namespace arm_compute::misc::shape_calculator;
+
+namespace arm_compute
+{
+namespace
+{
+using ElementsProcessed = Steps;
+
+Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                          const GEMMReshapeInfo &gemm_info)
+{
+    ARM_COMPUTE_UNUSED(alpha);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");
+    ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16);
+    ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");
+
+    const int m = gemm_info.m();
+    const int n = gemm_info.n();
+    const int k = gemm_info.k();
+
+    ARM_COMPUTE_UNUSED(m);
+    ARM_COMPUTE_UNUSED(n);
+    ARM_COMPUTE_UNUSED(k);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != static_cast<unsigned int>(k));
+    ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != static_cast<unsigned int>(n));
+    ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(1) != static_cast<unsigned int>(k));
+    if(gemm_info.reinterpret_input_as_3d())
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) * input0->dimension(2) != static_cast<unsigned int>(m));
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != static_cast<unsigned int>(m));
+    }
+
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                                                        const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed)
+{
+    unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
+    unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
+    bool          reinterpret_input_as_3d             = gemm_info.reinterpret_input_as_3d();
+    bool          reinterpret_output_as_3d            = (gemm_info.depth_output_gemm3d() != 0);
+
+    Window win{};
+    Window win_out{};
+    bool   window_changed = false;
+
+    // In case both input and output have to be reinterpreted as 3D tensors,
+    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+    if(reinterpret_input_as_3d == reinterpret_output_as_3d)
+    {
+        reinterpret_output_as_3d = false;
+    }
+
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)));
+
+    TensorInfo tmp_info(*output);
+
+    if(reinterpret_output_as_3d)
+    {
+        // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
+        // the window needs to be constructed on the 2D collapsed version of the tensor
+        TensorShape tmp_shape(output->tensor_shape());
+        tmp_shape.collapse(2U, 1U);
+        tmp_info.set_tensor_shape(tmp_shape);
+    }
+
+    // Configure kernel window
+    num_elems_processed_per_iteration_x = rhs_info.n0;
+    num_elems_processed_per_iteration_y = lhs_info.m0;
+
+    // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor
+    // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic
+    const int m          = gemm_info.m();
+    const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y;
+
+    win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+    win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+    AccessWindowStatic input0_access(input0, 0, 0,
+                                     input0->dimension(0),
+                                     input0->dimension(1) + bottom_pad);
+    AccessWindowStatic input1_access(input1, 0, 0,
+                                     ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x),
+                                     input1->dimension(1));
+    AccessWindowStatic output_access(output, 0, 0,
+                                     ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x),
+                                     output->dimension(1) + bottom_pad);
+
+    window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
+                     update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor
+
+    output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape()));
+
+    // Collapse along the Z direction
+    // This collapse needs to be here in order to tune the Z dimension of LWS
+    Window             collapsed             = win;
+    const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
+    collapsed                                = win.collapse(win, dimension_to_collapse);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, collapsed);
+}
+} // namespace
+
+CLGEMMMatrixMultiplyNativeKernel::CLGEMMMatrixMultiplyNativeKernel()
+    : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false)
+{
+}
+
+void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, const GEMMLHSMatrixInfo &lhs_info,
+                                                 const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), alpha, lhs_info, rhs_info, gemm_info));
+
+    _input0                   = input0;
+    _input1                   = input1;
+    _output                   = output;
+    _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d();
+    _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
+    _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
+
+    // In case both input and output have to be reinterpreted as 3D tensors,
+    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+    if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
+    {
+        _reinterpret_input_as_3d  = false;
+        _reinterpret_output_as_3d = false;
+    }
+
+    // Check if we need to slide the matrix B
+    const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
+    _slide_matrix_b                          = (_input1->info()->num_dimensions() >= num_dimensions_input0);
+
+    ElementsProcessed num_elements_processed{};
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
+    // Create build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
+    build_opts.add_option_if(std::abs(1.0f - alpha) > 0.00001f, "-DALPHA=" + float_to_string_with_full_precision(alpha));
+    build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
+    build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
+    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
+    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));
+    build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
+    build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
+    build_opts.add_option("-DM=" + support::cpp11::to_string(input0->info()->dimension(1)));
+    build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n()));
+    build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k()));
+    build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
+    build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
+    build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
+
+    std::string kernel_name("gemm_mm_native");
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+
+    // Set config_id for enabling LWS tuning
+    _config_id = kernel_name;
+    _config_id += "_";
+    _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
+    _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
+    _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(gemm_info.k());
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(2));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(lhs_info.m0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(rhs_info.n0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(rhs_info.k0);
+}
+
+Status CLGEMMMatrixMultiplyNativeKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info,
+                                                  const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
+{
+    ElementsProcessed num_elements_processed{};
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, alpha, lhs_info, rhs_info, gemm_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
+                                                              input1->clone().get(),
+                                                              output->clone().get(),
+                                                              lhs_info,
+                                                              rhs_info,
+                                                              gemm_info,
+                                                              num_elements_processed)
+                                .first);
+
+    return Status{};
+}
+
+void CLGEMMMatrixMultiplyNativeKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    if(_input1->info()->num_dimensions() < 3)
+    {
+        // The stride_z for matrix B must be zero if we do not slice
+        ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
+    }
+
+    Window slice          = window.first_slice_window_3D();
+    Window slice_matrix_b = slice;
+
+    slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
+    slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+    if(_reinterpret_input_as_3d)
+    {
+        // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
+        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3;
+        const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom;
+        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
+    }
+
+    if(_reinterpret_output_as_3d)
+    {
+        // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
+        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
+        const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
+        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
+    }
+
+    do
+    {
+        Window slice_b = slice;
+        // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
+        // This scenario can happen when the matrix multiplication is used to perform a convolution operation
+        if(!_slide_matrix_b)
+        {
+            slice_b = slice_matrix_b;
+        }
+
+        unsigned int idx = 0;
+        add_2D_tensor_argument(idx, _input0, slice);
+        add_2D_tensor_argument(idx, _input1, slice_b);
+        add_2D_tensor_argument(idx, _output, slice);
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
+        enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
+    }
+    while(window.slide_window_slice_3D(slice));
+}
+} // namespace arm_compute