Revert "Rework gemm_mm_reshaped_only_rhs_ kernels with new macros"

This reverts commit 10e88a7351 "Rework gemm_mm_reshaped_only_rhs_ kernels with new macros"

Resolves: COMPMID-5095

Signed-off-by: Ramy Elgammal<ramy.elgammal@arm.com>
Change-Id: I46e167882f072e7508b6101d295accb6e089e740
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7045
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

14 files changed, 3078 insertions, 1191 deletions

diff --git a/Android.bp b/Android.bp
index 5c48a70251..136714b260 100644
--- a/Android.bp
+++ b/Android.bp
@@ -30,6 +30,7 @@ opencl_srcs = [
         "src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/fp_post_ops_act_eltwise_op_act.h",
         "src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl",
         "src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl",
+        "src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl",
         "src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/fp_elementwise_op_helpers.h",
         "src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/fp_mixed_precision_helpers.h",
         "src/core/CL/cl_kernels/common/fft.cl",
@@ -39,7 +40,6 @@ opencl_srcs = [
         "src/core/CL/cl_kernels/common/floor.cl",
         "src/core/CL/cl_kernels/common/gather.cl",
         "src/core/CL/cl_kernels/common/gemm.cl",
-        "src/core/CL/cl_kernels/common/gemm_reshaped_rhs_only.cl",
         "src/core/CL/cl_kernels/common/gemm_utils.cl",
         "src/core/CL/cl_kernels/common/gemmlowp.cl",
         "src/core/CL/cl_kernels/common/gemv.cl",
diff --git a/SConscript b/SConscript
index d8dcaadd15..45e88184e4 100644
--- a/SConscript
+++ b/SConscript
@@ -346,9 +346,9 @@ if env['opencl'] and env['embed_kernels']:
                        'src/core/CL/cl_kernels/common/gather.cl',
                        'src/core/CL/cl_kernels/common/gemm.cl',
                        'src/core/CL/cl_kernels/common/gemm_utils.cl',
-                       'src/core/CL/cl_kernels/common/gemm_reshaped_rhs_only.cl',
                        'src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl',
                        'src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl',
+                       'src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl',
                        'src/core/CL/cl_kernels/common/gemv.cl',
                        'src/core/CL/cl_kernels/common/gemmlowp.cl',
                        'src/core/CL/cl_kernels/common/generate_proposals.cl',
diff --git a/SConstruct b/SConstruct
index 425c77643a..ff53229282 100644
--- a/SConstruct
+++ b/SConstruct
@@ -491,4 +491,4 @@ if env['exceptions']:
 unknown = vars.UnknownVariables()
 if unknown:
     print("Unknown variables: %s" % " ".join(unknown.keys()))
-    Exit(1)
\ No newline at end of file
+    Exit(1)
diff --git a/src/core/CL/CLUtils.cpp b/src/core/CL/CLUtils.cpp
index 34ffbb7c6c..8f39c2d700 100644
--- a/src/core/CL/CLUtils.cpp
+++ b/src/core/CL/CLUtils.cpp
@@ -127,7 +127,6 @@ void PostOpCLKernelUtils::set_post_ops_cl_build_options(CLBuildOptions &build_op
     {
         const auto &post_op     = post_ops.get_list().at(post_op_id);
         const auto  slot_prefix = "-DP" + support::cpp11::to_string(slots[post_op_id]);
-        build_opts.add_option("-DPOST_OP" + support::cpp11::to_string(slots[post_op_id]));
         if(post_op->type() == experimental::PostOpType::Activation)
         {
             const auto _post_op  = utils::cast::polymorphic_downcast<const experimental::PostOpAct<ITensorInfo *> *>(post_op.get());
diff --git a/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl
new file mode 100644
index 0000000000..09ddcde043
--- /dev/null
+++ b/src/core/CL/cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl
@@ -0,0 +1,1399 @@
+/*
+ * Copyright (c) 2021-2022 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 "fp_post_ops_act_eltwise_op_act.h"
+#include "gemm_helpers.h"
+#include "repeat.h"
+
+/** (EXPERIMENTAL_POST_OPS) gemm_mm_reshaped_only_rhs kernel */
+#if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE)
+#if defined(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH)
+
+#define CONCAT(a, b) a##b
+
+#define ARM_DOT1(a, b, c) \
+    ({                    \
+        c = fma(a, b, c); \
+    })
+#define ARM_DOT2(a, b, c)       \
+    ({                          \
+        c = fma(a.s0, b.s0, c); \
+        c = fma(a.s1, b.s1, c); \
+    })
+#define ARM_DOT3(a, b, c)           \
+    ({                              \
+        ARM_DOT2(a, b, c);          \
+        c = fma((a.s2), (b.s2), c); \
+    })
+#define ARM_DOT4(a, b, c)           \
+    ({                              \
+        ARM_DOT3(a, b, c);          \
+        c = fma((a.s3), (b.s3), c); \
+    })
+#define ARM_DOT8(a, b, c)            \
+    ({                               \
+        ARM_DOT4((a.lo), (b.lo), c); \
+        ARM_DOT4((a.hi), (b.hi), c); \
+    })
+#define ARM_DOT16(a, b, c)           \
+    ({                               \
+        ARM_DOT8((a.lo), (b.lo), c); \
+        ARM_DOT8((a.hi), (b.hi), c); \
+    })
+
+#if N0 == 2
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+    })
+#elif N0 == 3 // N0 == 3
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+    })
+#elif N0 == 4 // N0 == 4
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##3), (c.s3));     \
+    })
+#elif N0 == 8 // N0 == 8
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##3), (c.s3));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##4), (c.s4));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##5), (c.s5));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##6), (c.s6));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##7), (c.s7));     \
+    })
+#elif N0 == 16 // N0 == 16
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##3), (c.s3));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##4), (c.s4));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##5), (c.s5));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##6), (c.s6));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##7), (c.s7));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##8), (c.s8));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##9), (c.s9));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##A), (c.sA));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##B), (c.sB));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##C), (c.sC));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##D), (c.sD));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##E), (c.sE));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##F), (c.sF));     \
+    })
+#else // N0 not supported
+#error "N0 value not supported"
+#endif // N0 conditions
+
+#if defined(GEMM_MM_RESHAPED_ONLY_RHS_T_POST_ACT_ELTWISE_OP_ACT)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops:
+ * Post op 1: activation (optional)
+ * Post op 2: elementwise op
+ * Post op 3: activation (optional)
+ *
+ * @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
+ * @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform
+ * @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2
+ * @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2
+ * @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
+ *
+ * All parameters are similarly defined in kernel gemm_mm_reshaped_only_rhs_t, with these additions:
+ *
+ * @param[in] eltwise_operand_ptr      Pointer to the eltwise operand matrix. Supported data type: F16/F32
+ * @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes)
+ * @param[in] eltwise_operand_step_x   eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] eltwise_operand_stride_y Stride of the eltwise operand matrix in Y dimension (in bytes)
+ * @param[in] eltwise_operand_step_y   eltwise_operand_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] eltwise_operand_stride_z Stride of the eltwise operand tensor in Z dimension (in bytes)
+ */
+__kernel void gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
+                                                                  IMAGE_DECLARATION(rhs),
+#if defined(BETA)
+                                                                  IMAGE_DECLARATION(bias),
+#endif // defined(BETA)
+                                                                  IMAGE_DECLARATION(dst),
+                                                                  // Post Op arguments
+                                                                  IMAGE_DECLARATION(eltwise_operand),
+                                                                  uint lhs_stride_z,
+                                                                  uint rhs_stride_z,
+#if defined(BETA)
+                                                                  uint bias_stride_z,
+#endif //defined(BETA)
+                                                                  uint dst_stride_z,
+                                                                  uint eltwise_operand_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
+                                                                  ,
+                                                                  const int M,
+                                                                  const int N,
+                                                                  const int K)
+{
+    // Block size
+#define RHS_BLOCK_SIZE ((K0) * (N0))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (K0)
+#define RHS_STEP_X ((K0) * (H0))
+#define RHS_STEP_LOOP (1)
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X (K0)
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
+#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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+
+    // Compute RHS reshaped matrix address
+    uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
+
+#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, zero, 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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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(M0-1)=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 reshaped matrix
+        LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero);
+
+        // Accumulate
+        ARM_DOT_K0XN0(K0, a0, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(K0, a1, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(K0, a2, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(K0, a3, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(K0, a4, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(K0, a5, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(K0, a6, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(K0, a7, b, c7);
+#endif // M0 > 7
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        rhs_offset += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE);
+    }
+
+    // Left-over accumulations
+    for(; i < K; ++i)
+    {
+        // Load values from LHS matrix
+        LOAD_BLOCK(M0, 1, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
+
+        // Load values from RHS reshaped matrix
+        LOAD_BLOCK(N0, 1, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero);
+
+        // Accumulate
+        ARM_DOT_K0XN0(1, a0, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(1, a1, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(1, a2, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(1, a3, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(1, a4, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(1, a5, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(1, a6, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(1, a7, b, c7);
+#endif // M0 > 7
+
+        lhs_offset += sizeof(DATA_TYPE);
+        rhs_offset += sizeof(DATA_TYPE);
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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
+#if defined(ALPHA)
+    SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
+#endif // defined(ALPHA)
+
+    // Add beta*bias
+#if defined(BETA)
+#if defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
+
+    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias[broadcasted]
+    ADD_BLOCK_BROADCAST(M0, c, bias0);
+
+#else // defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
+
+    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias
+    ADD_BLOCK(M0, c, bias);
+
+#endif // defined(BROADCAST_BIAS)
+#endif // defined(BETA)
+
+    // c = act(c)
+    POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
+    // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+    // c = act(c)
+    POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
+
+    // Store output block
+    STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+}
+#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_T_POST_ACT_ELTWISE_OP_ACT)
+
+#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE_POST_ACT_ELTWISE_OP_ACT)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops. The RHS matrix is stored in OpenCL image object.
+ * Post op 1: activation (optional)
+ * Post op 2: elementwise op
+ * Post op 3: activation (optional)
+ *
+ * @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
+ * @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform
+ * @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2
+ * @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2
+ * @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
+ *
+ * All parameters are similarly defined in kernel gemm_mm_reshaped_only_rhs_t_texture, with these additions:
+ *
+ * @param[in] eltwise_operand_ptr      Pointer to the eltwise operand matrix. Supported data type: F16/F32
+ * @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes)
+ * @param[in] eltwise_operand_step_x   eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] eltwise_operand_stride_y Stride of the eltwise operand matrix in Y dimension (in bytes)
+ * @param[in] eltwise_operand_step_y   eltwise_operand_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] eltwise_operand_stride_z Stride of the eltwise operand tensor in Z dimension (in bytes)
+ * @param[in] M                        Number of rows in LHS matrix not reshaped.
+ * @param[in] N                        Number of columns in RHS matrix not reshaped.
+ * @param[in] K                        Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ */
+__kernel void gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
+                                                                          __read_only image2d_t rhs_img,
+#if defined(BETA)
+                                                                          IMAGE_DECLARATION(bias),
+#endif // defined(BETA)
+                                                                          IMAGE_DECLARATION(dst),
+                                                                          // Post Op arguments
+                                                                          IMAGE_DECLARATION(eltwise_operand),
+                                                                          uint lhs_stride_z,
+                                                                          uint rhs_stride_z,
+#if defined(BETA)
+                                                                          uint bias_stride_z,
+#endif //defined(BETA)
+                                                                          uint dst_stride_z,
+                                                                          uint eltwise_operand_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
+                                                                          ,
+                                                                          const int M,
+                                                                          const int N,
+                                                                          const int K)
+{
+    // Pixel unit
+#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0)
+
+    const uint LEFTOVER_K = K % K0;
+
+    // Block size
+#define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (PIXEL_UNIT)
+#define RHS_STEP_X (PIXEL_UNIT * (H0))
+#define RHS_STEP_LOOP (1)
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X PIXEL_UNIT
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
+#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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    const uint z_rhs = (get_global_id(2) % MATRIX_B_DEPTH);
+#else  // defined(MATRIX_B_DEPTH)
+    const uint z_rhs = get_global_id(2);
+#endif // defined(MATRIX_B_DEPTH)
+
+    // Compute RHS matrix coordinates
+    uint       x_rhs = (get_global_id(0) % H0) * (uint)RHS_OFFSET_X;
+    const uint y_rhs = (get_global_id(0) / (uint)H0) + z_rhs * RHS_HEIGHT;
+
+    REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0);
+    REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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);
+
+    int i = 0;
+    for(; i <= (K - K0); i += K0)
+    {
+        // 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 stored in a cl_image
+        REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);
+        LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);
+
+        // Accumulate
+        ARM_DOT_K0XN0(K0, a0, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(K0, a1, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(K0, a2, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(K0, a3, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(K0, a4, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(K0, a5, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(K0, a6, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(K0, a7, b, c7);
+#endif // M0 > 7
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP;
+    }
+
+    if(LEFTOVER_K != 0)
+    {
+        // Note: We cannot read out-of-bound elements from the RHS matrix because
+        // the RHS width is always multiple of K0. This is not be true for the LHS matrix
+
+        union UNION_VEC_TYPE
+        {
+            DATA_TYPE s[K0];
+            VEC_DATA_TYPE(DATA_TYPE, K0)
+            v;
+        };
+
+        union UNION_VEC_TYPE a0 = {.v = 0 };
+#if M0 > 1
+        union UNION_VEC_TYPE a1 = {.v = 0 };
+#endif // M0 > 1
+#if M0 > 2
+        union UNION_VEC_TYPE a2 = {.v = 0 };
+#endif // M0 > 2
+#if M0 > 3
+        union UNION_VEC_TYPE a3 = {.v = 0 };
+#endif // M0 > 3
+#if M0 > 4
+        union UNION_VEC_TYPE a4 = {.v = 0 };
+#endif // M0 > 4
+#if M0 > 5
+        union UNION_VEC_TYPE a5 = {.v = 0 };
+#endif // M0 > 5
+#if M0 > 6
+        union UNION_VEC_TYPE a6 = {.v = 0 };
+#endif // M0 > 6
+#if M0 > 7
+        union UNION_VEC_TYPE a7 = {.v = 0 };
+#endif // M0 > 7
+
+        REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);
+
+        // Load from RHS matrix
+        LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);
+
+        // Load from LHS matrix
+        for(int k = 0; k < LEFTOVER_K; ++k)
+        {
+            a0.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0);
+#if M0 > 1
+            a1.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1);
+#endif // M0 > 1
+#if M0 > 2
+            a2.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2);
+#endif // M0 > 2
+#if M0 > 3
+            a3.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3);
+#endif // M0 > 3
+#if M0 > 4
+            a4.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4);
+#endif // M0 > 4
+#if M0 > 5
+            a5.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5);
+#endif // M0 > 5
+#if M0 > 6
+            a6.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6);
+#endif // M0 > 6
+#if M0 > 7
+            a7.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7);
+#endif // M0 > 7
+
+            lhs_offset += sizeof(DATA_TYPE);
+        }
+
+        // Accumulate
+        ARM_DOT_K0XN0(K0, a0.v, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(K0, a1.v, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(K0, a2.v, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(K0, a3.v, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(K0, a4.v, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(K0, a5.v, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(K0, a6.v, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(K0, a7.v, b, c7);
+#endif // M0 > 7
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
+
+    REPEAT_VAR_INIT_TO_CONST(M0, 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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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
+#if defined(ALPHA)
+    SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
+#endif // defined(ALPHA)
+
+    // Add beta*bias
+#if defined(BETA)
+#if defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
+
+    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias[broadcasted]
+    ADD_BLOCK_BROADCAST(M0, c, bias0);
+
+#else // defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
+
+    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias
+    ADD_BLOCK(M0, c, bias);
+
+#endif // defined(BROADCAST_BIAS)
+#endif // defined(BETA)
+
+    // c = act(c)
+    POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
+    // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+    // c = act(c)
+    POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
+
+    // Store output block
+    STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+#undef PIXEL_UNIT
+}
+#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE_POST_ACT_ELTWISE_OP_ACT)
+
+#define VFMA(a, b, c)     \
+    ({                    \
+        c = fma(a, b, c); \
+    })
+
+#if M0 == 1
+#define 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 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 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 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 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 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 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 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
+
+#if defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_POST_ACT_ELTWISE_OP_ACT)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops:
+ * Post op 1: activation (optional)
+ * Post op 2: elementwise op
+ * Post op 3: activation (optional)
+ *
+ * @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
+ * @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform
+ * @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2
+ * @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2
+ * @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
+ *
+ * All parameters are similarly defined in kernel gemm_mm_reshaped_only_rhs_nt, with these additions:
+ *
+ * @param[in] eltwise_operand_ptr      Pointer to the eltwise operand matrix. Supported data type: F16/F32
+ * @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes)
+ * @param[in] eltwise_operand_step_x   eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] eltwise_operand_stride_y Stride of the eltwise operand matrix in Y dimension (in bytes)
+ * @param[in] eltwise_operand_step_y   eltwise_operand_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] eltwise_operand_stride_z Stride of the eltwise operand tensor in Z dimension (in bytes)
+ * @param[in] M                        Number of rows in LHS matrix not reshaped.
+ * @param[in] N                        Number of columns in RHS matrix not reshaped.
+ * @param[in] K                        Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ */
+__kernel void gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
+                                                                   IMAGE_DECLARATION(rhs),
+#if defined(BETA)
+                                                                   IMAGE_DECLARATION(bias),
+#endif // defined(BETA)
+                                                                   IMAGE_DECLARATION(dst),
+                                                                   // Post Op arguments
+                                                                   IMAGE_DECLARATION(eltwise_operand),
+                                                                   uint lhs_stride_z,
+                                                                   uint rhs_stride_z,
+#if defined(BETA)
+                                                                   uint bias_stride_z,
+#endif //defined(BETA)
+                                                                   uint dst_stride_z,
+                                                                   uint eltwise_operand_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
+                                                                   ,
+                                                                   const int M,
+                                                                   const int N,
+                                                                   const int K)
+{
+    // Block size
+#define RHS_BLOCK_SIZE ((K0) * (N0))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (N0)
+#define RHS_STEP_X ((N0) * (H0))
+#define RHS_STEP_LOOP (1)
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X (N0)
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
+#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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+
+    // Compute RHS reshaped matrix address
+    uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
+
+#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, zin, 0);   //uint zin0=0,zin1=0,zin2=0,... zin7=0;
+    REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); //uint zero0=0,zero1=0,zero2=0,... zero7=0;
+
+#if defined(REINTERPRET_INPUT_AS_3D)
+
+    // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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, zin);
+
+        VEC_DATA_TYPE(DATA_TYPE, N0)
+        b0;
+
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(0, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 1 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(1, a, b0, c);
+#if K0 > 2
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 2 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(2, a, b0, c);
+#endif // K0 > 2
+#if K0 > 3
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 3 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(3, a, b0, c);
+#endif // K0 > 3
+#if K0 > 4
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 4 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(4, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 5 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(5, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 6 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(6, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 7 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(7, a, b0, c);
+#endif // K0 > 4
+#if K0 > 8
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 8 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(8, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 9 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(9, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 10 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(A, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 11 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(B, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 12 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(C, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 13 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(D, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 14 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(E, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 15 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(F, a, b0, c);
+#endif // K0 > 8
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        rhs_offset += K0 * RHS_STEP_X * RHS_STEP_LOOP * sizeof(DATA_TYPE);
+    }
+
+    // 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 + zin0));
+#if M0 > 1
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
+#endif // M0 > 1
+#if M0 > 2
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
+#endif // M0 > 2
+#if M0 > 3
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
+#endif // M0 > 3
+#if M0 > 4
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
+#endif // M0 > 4
+#if M0 > 5
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
+#endif // M0 > 5
+#if M0 > 6
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
+#endif // M0 > 6
+#if M0 > 7
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
+#endif // M0 > 7
+
+        VEC_DATA_TYPE(DATA_TYPE, N0)
+        b0;
+
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(0, a, b0, c);
+
+        lhs_offset += sizeof(DATA_TYPE);
+        rhs_offset += RHS_STEP_X * sizeof(DATA_TYPE);
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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
+#if defined(ALPHA)
+    SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
+#endif // defined(ALPHA)
+
+    // Add beta*bias
+#if defined(BETA)
+#if defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
+
+    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias[broadcasted]
+    ADD_BLOCK_BROADCAST(M0, c, bias0);
+
+#else // defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
+
+    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias
+    ADD_BLOCK(M0, c, bias);
+
+#endif // defined(BROADCAST_BIAS)
+#endif // defined(BETA)
+
+    // c = act(c)
+    POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
+    // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+    // c = act(c)
+    POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
+
+    // Store output block
+    STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+#undef RHS_STEP_LOOP
+}
+#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_POST_ACT_ELTWISE_OP_ACT)
+
+#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE_POST_ACT_ELTWISE_OP_ACT)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops. The RHS matrix is stored in OpenCL image object.
+ * Post op 1: activation (optional)
+ * Post op 2: elementwise op
+ * Post op 3: activation (optional)
+ *
+ * @note (Optional) -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
+ * @note (Required) -DP2_ELTWISE_OP: The (binary) elementwise post op to perform
+ * @note (Required) -DP2_ELTWISE_ARG1_HEIGHT: The height (Y dimension) of the eltwise operand matrix of the eltwise post op at slot 2
+ * @note (Required) -DP2_ELTWISE_ARG1_WIDTH: The width (X dimension) of the eltwise operand matrix of the eltwise post op at slot 2
+ * @note (Optional) -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
+ *
+ * All parameters are similarly defined in kernel gemm_mm_reshaped_only_rhs_nt_texture, with these additions:
+ *
+ * @param[in] eltwise_operand_ptr      Pointer to the eltwise operand matrix. Supported data type: F16/F32
+ * @param[in] eltwise_operand_stride_x Stride of the eltwise operand matrix in X dimension (in bytes)
+ * @param[in] eltwise_operand_step_x   eltwise_operand_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] eltwise_operand_stride_y Stride of the eltwise operand matrix in Y dimension (in bytes)
+ * @param[in] eltwise_operand_step_y   eltwise_operand_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] eltwise_operand_stride_z Stride of the eltwise operand tensor in Z dimension (in bytes)
+ * @param[in] M                        Number of rows in LHS matrix not reshaped.
+ * @param[in] N                        Number of columns in RHS matrix not reshaped.
+ * @param[in] K                        Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ */
+__kernel void gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act(IMAGE_DECLARATION(lhs),
+                                                                           __read_only image2d_t rhs_img,
+#if defined(BETA)
+                                                                           IMAGE_DECLARATION(bias),
+#endif // defined(BETA)
+                                                                           IMAGE_DECLARATION(dst),
+                                                                           // Post Op arguments
+                                                                           IMAGE_DECLARATION(eltwise_operand),
+                                                                           uint lhs_stride_z,
+                                                                           uint rhs_stride_z,
+#if defined(BETA)
+                                                                           uint bias_stride_z,
+#endif //defined(BETA)
+                                                                           uint dst_stride_z,
+                                                                           uint eltwise_operand_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
+                                                                           ,
+                                                                           const int M,
+                                                                           const int N,
+                                                                           const int K)
+{
+    // Pixel unit
+#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0)
+
+    // Block size
+#define RHS_BLOCK_SIZE ((K0) * (PIXEL_UNIT))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (PIXEL_UNIT)
+#define RHS_STEP_X ((PIXEL_UNIT) * (H0))
+#define RHS_STEP_LOOP (1)
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X (PIXEL_UNIT)
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
+#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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    const uint z_rhs = (z % MATRIX_B_DEPTH);
+#else  // defined(MATRIX_B_DEPTH)
+    const uint z_rhs = z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    // Compute RHS matrix coordinates
+    uint       x_rhs = (x % H0) * (uint)RHS_OFFSET_X;
+    const uint y_rhs = (x / (uint)H0) + z_rhs * RHS_HEIGHT;
+
+    REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0);
+    REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);
+
+#if defined(REINTERPRET_INPUT_AS_3D)
+
+    // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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);
+
+    int i = 0;
+    for(; i <= (K - K0); i += K0)
+    {
+        // Load values from LHS matrix
+        LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin);
+
+        VEC_DATA_TYPE(DATA_TYPE, N0)
+        b0;
+
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(0, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 1 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(1, a, b0, c);
+#if K0 > 2
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 2 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(2, a, b0, c);
+#endif // K0 > 2
+#if K0 > 3
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 3 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(3, a, b0, c);
+#endif // K0 > 3
+#if K0 > 4
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 4 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(4, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 5 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(5, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 6 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(6, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 7 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(7, a, b0, c);
+#endif // K0 > 4
+#if K0 > 8
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 8 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(8, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 9 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(9, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 10 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(A, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 11 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(B, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 12 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(C, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 13 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(D, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 14 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(E, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 15 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(F, a, b0, c);
+#endif // K0 > 8
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        x_rhs += K0 * RHS_STEP_X * RHS_STEP_LOOP;
+    }
+
+    // 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 + zin0));
+#if M0 > 1
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
+#endif // M0 > 1
+#if M0 > 2
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
+#endif // M0 > 2
+#if M0 > 3
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
+#endif // M0 > 3
+#if M0 > 4
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
+#endif // M0 > 4
+#if M0 > 5
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
+#endif // M0 > 5
+#if M0 > 6
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
+#endif // M0 > 6
+#if M0 > 7
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
+#endif // M0 > 7
+
+        VEC_DATA_TYPE(DATA_TYPE, N0)
+        b0;
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));
+
+        VFMA_M0xN0(0, a, b0, c);
+
+        lhs_offset += sizeof(DATA_TYPE);
+        x_rhs += RHS_STEP_X;
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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
+#if defined(ALPHA)
+    SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
+#endif // defined(ALPHA)
+
+    // Add beta*bias
+#if defined(BETA)
+#if defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
+
+    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias[broadcasted]
+    ADD_BLOCK_BROADCAST(M0, c, bias0);
+
+#else // defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
+
+    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias
+    ADD_BLOCK(M0, c, bias);
+
+#endif // defined(BROADCAST_BIAS)
+#endif // defined(BETA)
+
+    // c = act(c)
+    POST_OP1_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
+    // c = c + eltwise_operand (mix-precision, broadcast, boundary aware)
+    POST_OP2_ELTWISE_OP(P2_ELTWISE_OP, M0, N0, c, eltwise_operand, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), DATA_TYPE, DATA_TYPE_ACCUMULATOR, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+    // c = act(c)
+    POST_OP3_ACTIVATION_OPTIONAL(M0, DATA_TYPE, DATA_TYPE_ACCUMULATOR, N0, c);
+
+    // Store output block
+    STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+#undef RHS_STEP_LOOP
+}
+#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE_POST_ACT_ELTWISE_OP_ACT)
+#endif // defined(P2_ELTWISE_OP) && defined(P2_ELTWISE_ARG1_HEIGHT) && defined(P2_ELTWISE_ARG1_WIDTH)
+#endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE)
diff --git a/src/core/CL/cl_kernels/common/gemm.cl b/src/core/CL/cl_kernels/common/gemm.cl
index 74e2e5097e..33ab25cad0 100644
--- a/src/core/CL/cl_kernels/common/gemm.cl
+++ b/src/core/CL/cl_kernels/common/gemm.cl
@@ -24,6 +24,1512 @@
 #include "gemm_helpers.h"
 #include "repeat.h"
 
+#if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE)
+
+#define CONCAT(a, b) a##b
+
+#define ARM_DOT1(a, b, c) \
+    ({                    \
+        c = fma(a, b, c); \
+    })
+#define ARM_DOT2(a, b, c)       \
+    ({                          \
+        c = fma(a.s0, b.s0, c); \
+        c = fma(a.s1, b.s1, c); \
+    })
+#define ARM_DOT3(a, b, c)           \
+    ({                              \
+        ARM_DOT2(a, b, c);          \
+        c = fma((a.s2), (b.s2), c); \
+    })
+#define ARM_DOT4(a, b, c)           \
+    ({                              \
+        ARM_DOT3(a, b, c);          \
+        c = fma((a.s3), (b.s3), c); \
+    })
+#define ARM_DOT8(a, b, c)            \
+    ({                               \
+        ARM_DOT4((a.lo), (b.lo), c); \
+        ARM_DOT4((a.hi), (b.hi), c); \
+    })
+#define ARM_DOT16(a, b, c)           \
+    ({                               \
+        ARM_DOT8((a.lo), (b.lo), c); \
+        ARM_DOT8((a.hi), (b.hi), c); \
+    })
+
+#if N0 == 2
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+    })
+#elif N0 == 3 // N0 == 3
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+    })
+#elif N0 == 4 // N0 == 4
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##3), (c.s3));     \
+    })
+#elif N0 == 8 // N0 == 8
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##3), (c.s3));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##4), (c.s4));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##5), (c.s5));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##6), (c.s6));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##7), (c.s7));     \
+    })
+#elif N0 == 16 // N0 == 16
+#define ARM_DOT_K0XN0(k0, a, b, c) \
+    ({                             \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##0), (c.s0));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##1), (c.s1));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##2), (c.s2));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##3), (c.s3));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##4), (c.s4));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##5), (c.s5));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##6), (c.s6));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##7), (c.s7));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##8), (c.s8));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##9), (c.s9));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##A), (c.sA));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##B), (c.sB));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##C), (c.sC));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##D), (c.sD));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##E), (c.sE));     \
+        CONCAT(ARM_DOT, k0)        \
+        ((a), (b##F), (c.sF));     \
+    })
+#else // N0 not supported
+#error "N0 value not supported"
+#endif // N0 conditions
+
+#if defined(GEMM_MM_RESHAPED_ONLY_RHS_T)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
+ *  The LHS matrix is NOT reshaped
+ *  The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed
+ * @note This kernel is duplicated in /experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl
+ *
+ * @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 runtime as kernel parameters.
+ * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
+ * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
+ * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
+ * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
+ * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
+ * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
+ * @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
+ *  - H0 >= 1
+ *
+ * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
+ *       The activation function is performed after the bias addition
+ * @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 matrix. Supported data type: F16/F32
+ * @param[in]  lhs_stride_x                       Stride of the LHS 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 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 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[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
+ * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
+ * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias 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 matrix in Z dimension (in bytes)
+ * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  bias_stride_z                      (Optional) Stride of the bias 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)
+ * @param[in]  M                                  Number of rows in LHS matrix not reshaped.
+ * @param[in]  N                                  Number of columns in RHS matrix not reshaped.
+ * @param[in]  K                                  Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ */
+__kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
+                                          IMAGE_DECLARATION(rhs),
+#if defined(BETA)
+                                          IMAGE_DECLARATION(bias),
+#endif // defined(BETA)
+                                          IMAGE_DECLARATION(dst),
+                                          uint lhs_stride_z,
+                                          uint rhs_stride_z,
+#if defined(BETA)
+                                          uint bias_stride_z,
+#endif //defined(BETA)
+                                          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
+                                          ,
+                                          const int M,
+                                          const int N,
+                                          const int K)
+{
+    // Block size
+#define RHS_BLOCK_SIZE ((K0) * (N0))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (K0)
+#define RHS_STEP_X ((K0) * (H0))
+#define RHS_STEP_LOOP (1)
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X (K0)
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
+#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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+
+    // Compute RHS reshaped matrix address
+    uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
+
+#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, zero, 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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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(M0-1)=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 reshaped matrix
+        LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero);
+
+        // Accumulate
+        ARM_DOT_K0XN0(K0, a0, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(K0, a1, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(K0, a2, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(K0, a3, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(K0, a4, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(K0, a5, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(K0, a6, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(K0, a7, b, c7);
+#endif // M0 > 7
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        rhs_offset += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE);
+    }
+
+    // Left-over accumulations
+    for(; i < K; ++i)
+    {
+        // Load values from LHS matrix
+        LOAD_BLOCK(M0, 1, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
+
+        // Load values from RHS reshaped matrix
+        LOAD_BLOCK(N0, 1, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero);
+
+        // Accumulate
+        ARM_DOT_K0XN0(1, a0, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(1, a1, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(1, a2, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(1, a3, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(1, a4, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(1, a5, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(1, a6, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(1, a7, b, c7);
+#endif // M0 > 7
+
+        lhs_offset += sizeof(DATA_TYPE);
+        rhs_offset += sizeof(DATA_TYPE);
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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
+#if defined(ALPHA)
+    SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
+#endif // defined(ALPHA)
+
+    // Add beta*bias
+#if defined(BETA)
+#if defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
+
+    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias[broadcasted]
+    ADD_BLOCK_BROADCAST(M0, c, bias0);
+
+#else // defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
+
+    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias
+    ADD_BLOCK(M0, c, bias);
+
+#endif // defined(BROADCAST_BIAS)
+#endif // defined(BETA)
+
+#if defined(ACTIVATION_TYPE)
+    ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
+#endif // defined(ACTIVATION_TYPE)
+
+    // Store output block
+    STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+#undef RHS_STEP_LOOP
+}
+#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_T)
+
+#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image
+ *  The LHS matrix is NOT reshaped
+ *  The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed
+ * @note This kernel is duplicated in /experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl
+ *
+ * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
+ * @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 runtime as kernel parameters.
+ * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
+ *       Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
+ *       could be different from the value returned by get_image_height(rhs_img).
+ * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
+ * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
+ * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
+ * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
+ * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
+ * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
+ * @note Only the following configurations of M0, N0 and K0 are currently supported:
+ *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
+ *  - N0 = 4, 8, 16
+ *  - K0 = 4, 8, 16
+ *  - H0 >= 1
+ *
+ * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
+ *       The activation function is performed after the bias addition
+ * @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 matrix. Supported data type: F32
+ * @param[in]  lhs_stride_x                       Stride of the LHS 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 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 matrix
+ * @param[in]  rhs_img                            The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr
+ * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
+ * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
+ * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias 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 matrix in Z dimension (in bytes)
+ * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  bias_stride_z                      (Optional) Stride of the bias 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)
+ * @param[in]  M                                  Number of rows in LHS matrix not reshaped.
+ * @param[in]  N                                  Number of columns in RHS matrix not reshaped.
+ * @param[in]  K                                  Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ */
+__kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs),
+                                                  __read_only image2d_t rhs_img,
+#if defined(BETA)
+                                                  IMAGE_DECLARATION(bias),
+#endif // defined(BETA)
+                                                  IMAGE_DECLARATION(dst),
+                                                  uint lhs_stride_z,
+                                                  uint rhs_stride_z,
+#if defined(BETA)
+                                                  uint bias_stride_z,
+#endif //defined(BETA)
+                                                  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
+                                                  ,
+                                                  const int M,
+                                                  const int N,
+                                                  const int K)
+{
+    // Pixel unit
+#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0)
+
+    const uint LEFTOVER_K = K % K0;
+
+    // Block size
+#define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (PIXEL_UNIT)
+#define RHS_STEP_X (PIXEL_UNIT * (H0))
+#define RHS_STEP_LOOP (1)
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X PIXEL_UNIT
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
+#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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    const uint z_rhs = (get_global_id(2) % MATRIX_B_DEPTH);
+#else  // defined(MATRIX_B_DEPTH)
+    const uint z_rhs = get_global_id(2);
+#endif // defined(MATRIX_B_DEPTH)
+
+    // Compute RHS matrix coordinates
+    uint       x_rhs = (get_global_id(0) % H0) * (uint)RHS_OFFSET_X;
+    const uint y_rhs = (get_global_id(0) / (uint)H0) + z_rhs * RHS_HEIGHT;
+
+    REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0);
+    REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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);
+
+    int i = 0;
+    for(; i <= (K - K0); i += K0)
+    {
+        // 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 stored in a cl_image
+        REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);
+        LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);
+
+        // Accumulate
+        ARM_DOT_K0XN0(K0, a0, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(K0, a1, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(K0, a2, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(K0, a3, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(K0, a4, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(K0, a5, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(K0, a6, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(K0, a7, b, c7);
+#endif // M0 > 7
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP;
+    }
+
+    if(LEFTOVER_K != 0)
+    {
+        // Note: We cannot read out-of-bound elements from the RHS matrix because
+        // the RHS width is always multiple of K0. This is not be true for the LHS matrix
+        // Left-over accumulations for LHS matrix
+
+        union UNION_VEC_TYPE
+        {
+            DATA_TYPE s[K0];
+            VEC_DATA_TYPE(DATA_TYPE, K0)
+            v;
+        };
+
+        union UNION_VEC_TYPE a0 = {.v = 0 };
+#if M0 > 1
+        union UNION_VEC_TYPE a1 = {.v = 0 };
+#endif // M0 > 1
+#if M0 > 2
+        union UNION_VEC_TYPE a2 = {.v = 0 };
+#endif // M0 > 2
+#if M0 > 3
+        union UNION_VEC_TYPE a3 = {.v = 0 };
+#endif // M0 > 3
+#if M0 > 4
+        union UNION_VEC_TYPE a4 = {.v = 0 };
+#endif // M0 > 4
+#if M0 > 5
+        union UNION_VEC_TYPE a5 = {.v = 0 };
+#endif // M0 > 5
+#if M0 > 6
+        union UNION_VEC_TYPE a6 = {.v = 0 };
+#endif // M0 > 6
+#if M0 > 7
+        union UNION_VEC_TYPE a7 = {.v = 0 };
+#endif // M0 > 7
+
+        REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);
+
+        // Load from RHS matrix
+        LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);
+
+        // Load from LHS matrix
+        for(int k = 0; k < LEFTOVER_K; ++k)
+        {
+            a0.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0);
+#if M0 > 1
+            a1.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1);
+#endif // M0 > 1
+#if M0 > 2
+            a2.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2);
+#endif // M0 > 2
+#if M0 > 3
+            a3.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3);
+#endif // M0 > 3
+#if M0 > 4
+            a4.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4);
+#endif // M0 > 4
+#if M0 > 5
+            a5.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5);
+#endif // M0 > 5
+#if M0 > 6
+            a6.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6);
+#endif // M0 > 6
+#if M0 > 7
+            a7.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7);
+#endif // M0 > 7
+
+            lhs_offset += sizeof(DATA_TYPE);
+        }
+
+        // Accumulate
+        ARM_DOT_K0XN0(K0, a0.v, b, c0);
+#if M0 > 1
+        ARM_DOT_K0XN0(K0, a1.v, b, c1);
+#endif // M0 > 1
+#if M0 > 2
+        ARM_DOT_K0XN0(K0, a2.v, b, c2);
+#endif // M0 > 2
+#if M0 > 3
+        ARM_DOT_K0XN0(K0, a3.v, b, c3);
+#endif // M0 > 3
+#if M0 > 4
+        ARM_DOT_K0XN0(K0, a4.v, b, c4);
+#endif // M0 > 4
+#if M0 > 5
+        ARM_DOT_K0XN0(K0, a5.v, b, c5);
+#endif // M0 > 5
+#if M0 > 6
+        ARM_DOT_K0XN0(K0, a6.v, b, c6);
+#endif // M0 > 6
+#if M0 > 7
+        ARM_DOT_K0XN0(K0, a7.v, b, c7);
+#endif // M0 > 7
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
+
+    REPEAT_VAR_INIT_TO_CONST(M0, 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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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
+#if defined(ALPHA)
+    SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
+#endif // defined(ALPHA)
+
+    // Add beta*bias
+#if defined(BETA)
+#if defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
+
+    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias[broadcasted]
+    ADD_BLOCK_BROADCAST(M0, c, bias0);
+
+#else // defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
+
+    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias
+    ADD_BLOCK(M0, c, bias);
+
+#endif // defined(BROADCAST_BIAS)
+#endif // defined(BETA)
+
+#if defined(ACTIVATION_TYPE)
+    ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
+#endif // defined(ACTIVATION_TYPE)
+
+    // Store output block
+    STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+#undef RHS_STEP_LOOP
+#undef PIXEL_UNIT
+}
+#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE)
+
+#define VFMA(a, b, c)     \
+    ({                    \
+        c = fma(a, b, c); \
+    })
+
+#if M0 == 1
+#define 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 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 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 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 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 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 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 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
+
+#if defined(GEMM_MM_RESHAPED_ONLY_RHS_NT)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
+ *  The LHS matrix is NOT reshaped
+ *  The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed
+ * @note This kernel is duplicated in /experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl
+ *
+ * @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 runtime as kernel parameters.
+ * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
+ * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
+ * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
+ * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
+ * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
+ * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
+ * @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
+ *  - H0 >= 1
+ *
+ * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
+ *       The activation function is performed after the bias addition
+ * @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 matrix. Supported data type: F16/F32
+ * @param[in]  lhs_stride_x                       Stride of the LHS 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 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 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[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
+ * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
+ * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias 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 matrix in Z dimension (in bytes)
+ * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  bias_stride_z                      (Optional) Stride of the bias 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)
+ * @param[in]  M                                  Number of rows in LHS matrix not reshaped.
+ * @param[in]  N                                  Number of columns in RHS matrix not reshaped.
+ * @param[in]  K                                  Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ */
+__kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs),
+                                           IMAGE_DECLARATION(rhs),
+#if defined(BETA)
+                                           IMAGE_DECLARATION(bias),
+#endif // defined(BETA)
+                                           IMAGE_DECLARATION(dst),
+                                           uint lhs_stride_z,
+                                           uint rhs_stride_z,
+#if defined(BETA)
+                                           uint bias_stride_z,
+#endif //defined(BETA)
+                                           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
+                                           ,
+                                           const int M,
+                                           const int N,
+                                           const int K)
+{
+    // Block size
+#define RHS_BLOCK_SIZE ((K0) * (N0))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (N0)
+#define RHS_STEP_X ((N0) * (H0))
+#define RHS_STEP_LOOP (1)
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X (N0)
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
+#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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+
+    // Compute RHS reshaped matrix address
+    uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
+
+#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, zin, 0);   //uint zin0=0,zin1=0,zin2=0,... zin7=0;
+    REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); //uint zero0=0,zero1=0,zero2=0,... zero7=0;
+
+#if defined(REINTERPRET_INPUT_AS_3D)
+
+    // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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, zin);
+
+        VEC_DATA_TYPE(DATA_TYPE, N0)
+        b0;
+
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(0, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 1 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(1, a, b0, c);
+#if K0 > 2
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 2 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(2, a, b0, c);
+#endif // K0 > 2
+#if K0 > 3
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 3 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(3, a, b0, c);
+#endif // K0 > 3
+#if K0 > 4
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 4 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(4, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 5 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(5, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 6 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(6, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 7 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(7, a, b0, c);
+#endif // K0 > 4
+#if K0 > 8
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 8 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(8, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 9 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(9, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 10 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(A, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 11 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(B, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 12 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(C, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 13 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(D, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 14 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(E, a, b0, c);
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 15 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(F, a, b0, c);
+#endif // K0 > 8
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        rhs_offset += K0 * RHS_STEP_X * RHS_STEP_LOOP * sizeof(DATA_TYPE);
+    }
+
+    // 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 + zin0));
+#if M0 > 1
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
+#endif // M0 > 1
+#if M0 > 2
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
+#endif // M0 > 2
+#if M0 > 3
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
+#endif // M0 > 3
+#if M0 > 4
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
+#endif // M0 > 4
+#if M0 > 5
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
+#endif // M0 > 5
+#if M0 > 6
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
+#endif // M0 > 6
+#if M0 > 7
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
+#endif // M0 > 7
+
+        VEC_DATA_TYPE(DATA_TYPE, N0)
+        b0;
+
+        b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
+        VFMA_M0xN0(0, a, b0, c);
+
+        lhs_offset += sizeof(DATA_TYPE);
+        rhs_offset += RHS_STEP_X * sizeof(DATA_TYPE);
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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
+#if defined(ALPHA)
+    SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
+#endif // defined(ALPHA)
+
+    // Add beta*bias
+#if defined(BETA)
+#if defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
+
+    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias[broadcasted]
+    ADD_BLOCK_BROADCAST(M0, c, bias0);
+
+#else // defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
+
+    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias
+    ADD_BLOCK(M0, c, bias);
+
+#endif // defined(BROADCAST_BIAS)
+#endif // defined(BETA)
+
+#if defined(ACTIVATION_TYPE)
+    ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
+#endif // defined(ACTIVATION_TYPE)
+
+    // Store output block
+    STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+#undef RHS_STEP_LOOP
+}
+#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_NT)
+
+#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
+ *  The LHS matrix is NOT reshaped
+ *  The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed
+ * @note This kernel is duplicated in /experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl
+ *
+ * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
+ * @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 runtime as kernel parameters.
+ * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
+ *       Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
+ *       could be different from the value returned by get_image_height(rhs_img).
+ * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
+ * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
+ * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
+ * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
+ * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
+ * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
+ * @note Only the following configurations of M0, N0 and K0 are currently supported:
+ *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
+ *  - N0 = 4, 8, 16
+ *  - K0 = 4, 8, 16
+ *  - H0 >= 1
+ *
+ * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
+ *       The activation function is performed after the bias addition
+ * @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 matrix. Supported data type: F32
+ * @param[in]  lhs_stride_x                       Stride of the LHS 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 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 matrix
+ * @param[in]  rhs_img                            The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr
+ * @param[in]  bias_ptr                           (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  bias_stride_x                      (Optional) Stride of the bias matrix in X dimension (in bytes)
+ * @param[in]  bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  bias_stride_y                      (Optional) Stride of the bias matrix in Y dimension (in bytes)
+ * @param[in]  bias_step_y                        (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias 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 matrix in Z dimension (in bytes)
+ * @param[in]  rhs_stride_z                       Stride of the RHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  bias_stride_z                      (Optional) Stride of the bias 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)
+ * @param[in]  M                                  Number of rows in LHS matrix not reshaped.
+ * @param[in]  N                                  Number of columns in RHS matrix not reshaped.
+ * @param[in]  K                                  Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ */
+__kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs),
+                                                   __read_only image2d_t rhs_img,
+#if defined(BETA)
+                                                   IMAGE_DECLARATION(bias),
+#endif // defined(BETA)
+                                                   IMAGE_DECLARATION(dst),
+                                                   uint lhs_stride_z,
+                                                   uint rhs_stride_z,
+#if defined(BETA)
+                                                   uint bias_stride_z,
+#endif //defined(BETA)
+                                                   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
+                                                   ,
+                                                   const int M,
+                                                   const int N,
+                                                   const int K)
+{
+    // Pixel unit
+#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0)
+
+    // Block size
+#define RHS_BLOCK_SIZE ((K0) * (PIXEL_UNIT))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (PIXEL_UNIT)
+#define RHS_STEP_X ((PIXEL_UNIT) * (H0))
+#define RHS_STEP_LOOP 1
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X (PIXEL_UNIT)
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+    const bool cond_y = y == 0;
+    const bool cond_x = ((x + 1) * N0 >= N);
+
+#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 + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    const uint z_rhs = (z % MATRIX_B_DEPTH);
+#else  // defined(MATRIX_B_DEPTH)
+    const uint z_rhs = z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    // Compute RHS matrix coordinates
+    uint       x_rhs = (x % H0) * (uint)RHS_OFFSET_X;
+    const uint y_rhs = (x / (uint)H0) + z_rhs * RHS_HEIGHT;
+
+    REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0);
+    REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);
+
+#if defined(REINTERPRET_INPUT_AS_3D)
+
+    // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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);
+
+    int i = 0;
+    for(; i <= (K - K0); i += K0)
+    {
+        // Load values from LHS matrix
+        LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin);
+
+        VEC_DATA_TYPE(DATA_TYPE, N0)
+        b0;
+
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(0, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 1 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(1, a, b0, c);
+#if K0 > 2
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 2 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(2, a, b0, c);
+#endif // K0 > 2
+#if K0 > 3
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 3 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(3, a, b0, c);
+#endif // K0 > 3
+#if K0 > 4
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 4 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(4, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 5 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(5, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 6 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(6, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 7 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(7, a, b0, c);
+#endif // K0 > 4
+#if K0 > 8
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 8 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(8, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 9 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(9, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 10 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(A, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 11 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(B, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 12 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(C, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 13 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(D, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 14 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(E, a, b0, c);
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 15 * RHS_STEP_X), (y_rhs));
+        VFMA_M0xN0(F, a, b0, c);
+#endif // K0 > 8
+
+        lhs_offset += K0 * sizeof(DATA_TYPE);
+        x_rhs += K0 * RHS_STEP_X * RHS_STEP_LOOP;
+    }
+
+    // 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 + zin0));
+#if M0 > 1
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
+#endif // M0 > 1
+#if M0 > 2
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
+#endif // M0 > 2
+#if M0 > 3
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
+#endif // M0 > 3
+#if M0 > 4
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
+#endif // M0 > 4
+#if M0 > 5
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
+#endif // M0 > 5
+#if M0 > 6
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
+#endif // M0 > 6
+#if M0 > 7
+        VEC_DATA_TYPE(DATA_TYPE, 2)
+        a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
+#endif // M0 > 7
+
+        VEC_DATA_TYPE(DATA_TYPE, N0)
+        b0;
+        b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));
+
+        VFMA_M0xN0(0, a, b0, c);
+
+        lhs_offset += sizeof(DATA_TYPE);
+        x_rhs += RHS_STEP_X;
+    }
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_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, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), 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
+#if defined(ALPHA)
+    SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
+#endif // defined(ALPHA)
+
+    // Add beta*bias
+#if defined(BETA)
+#if defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
+
+    LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias[broadcasted]
+    ADD_BLOCK_BROADCAST(M0, c, bias0);
+
+#else // defined(BROADCAST_BIAS)
+    __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
+
+    LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#ifndef UNIT_BETA
+    SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+    // c = c + bias
+    ADD_BLOCK(M0, c, bias);
+
+#endif // defined(BROADCAST_BIAS)
+#endif // defined(BETA)
+
+#if defined(ACTIVATION_TYPE)
+    ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, c, A_VAL, B_VAL);
+#endif // defined(ACTIVATION_TYPE)
+
+    // Store output block
+    STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+#undef RHS_STEP_LOOP
+}
+#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE)
+#endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE)
+
 #if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR)
 
 #if defined(MIXED_PRECISION)
diff --git a/src/core/CL/cl_kernels/common/gemm_reshaped_rhs_only.cl b/src/core/CL/cl_kernels/common/gemm_reshaped_rhs_only.cl
deleted file mode 100644
index 1d6560a1c2..0000000000
--- a/src/core/CL/cl_kernels/common/gemm_reshaped_rhs_only.cl
+++ /dev/null
@@ -1,953 +0,0 @@
-/*
- * Copyright (c) 2022 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 "activation_float_helpers.h"
-#include "helpers.h"
-#include "tile_helpers.h"
-
-// *INDENT-OFF*
-// clang-format off
-#if defined(GEMM_MM_RESHAPED_ONLY_RHS_T)
-//! @cond Doxygen_Suppress
-/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops:
- * Post op 1: activation (optional)
- * Post op 2: elementwise op
- * Post op 3: activation (optional)
- *
- *  The LHS matrix is NOT reshaped
- *  The RHS is reshaped with @ref ClGemmMatrixMultiplyReshapedOnlyRhsKernel and the block K0xN0 is transposed
- *
- * @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 block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
- * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
- * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
- * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @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
- *  - H0 >= 1
- *
- * @note In case of post ops, the following information must be passed at compile time:
- * @note -DPOST_OP1, -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 1
- * @note -DPOST_OP2: The arithmetic addition post op to perform at slot 2
- * @note -DPOST_OP3, -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
- *
- * @param[in]  lhs_ptr                           Pointer to the LHS tensor. Supported data types: F16/F32
- * @param[in]  lhs_stride_y                      Stride of the LHS tensor in Y dimension (in bytes)
- * @param[in]  lhs_stride_z                      Stride of the LHS tensor in Z dimension (in bytes)
- * @param[in]  lhs_w                             The size of the width dimension of the LHS tensor
- * @param[in]  lhs_h                             The size of the height dimension of the LHS tensor
- * @param[in]  lhs_n                             The size of the depth dimension of the LHS tensor
- * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the LHS tensor
- * @param[in]  rhs_ptr                           Pointer to the RHS reshaped tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  rhs_stride_y                      Stride of the RHS tensor in Y dimension (in bytes)
- * @param[in]  rhs_stride_z                      Stride of the RHS tensor in Z dimension (in bytes)
- * @param[in]  rhs_w                             The size of the width dimension of the RHS tensor
- * @param[in]  rhs_h                             The size of the height dimension of the RHS tensor
- * @param[in]  rhs_n                             The size of the depth dimension of the RHS tensor
- * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS tensor
- * @param[in]  bia_ptr                           (Optional) Pointer to the bias tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  bia_stride_y                      (Optional) Stride of the bias tensor in Y dimension (in bytes)
- * @param[in]  bia_stride_z                      (Optional) Stride of the bias tensor in Z dimension (in bytes)
- * @param[in]  bia_w                             (Optional) The size of the width dimension of the bias tensor
- * @param[in]  bia_h                             (Optional) The size of the height dimension of the bias tensor
- * @param[in]  bia_n                             (Optional) The size of the depth dimension of the bias tensor
- * @param[in]  bia_offset_first_element_in_bytes (Optional) The offset of the first element in the bias tensor
- * @param[in]  ex0_ptr                           (Optional) Pointer to the tensor added with POST_OP2. Supported data type: same as @p lhs_ptr
- * @param[in]  ex0_stride_y                      (Optional) Stride of the tensor added with POST_OP2 in Y dimension (in bytes)
- * @param[in]  ex0_stride_z                      (Optional) Stride of the tensor added with POST_OP2 in Z dimension (in bytes)
- * @param[in]  ex0_w                             (Optional) The size of the width dimension of the tensor added with POST_OP2
- * @param[in]  ex0_h                             (Optional) The size of the height dimension of the tensor added with POST_OP2
- * @param[in]  ex0_n                             (Optional) The size of the depth dimension of the tensor added with POST_OP2
- * @param[in]  ex0_offset_first_element_in_bytes (Optional) The offset of the first element in the tensor added with POST_OP2
- * @param[out] dst_ptr                           (Optional) Pointer to the destination tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  dst_stride_y                      (Optional) Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_stride_z                      (Optional) Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  dst_w                             (Optional) The size of the width dimension of the destination tensor
- * @param[in]  dst_h                             (Optional) The size of the height dimension of the destination tensor
- * @param[in]  dst_n                             (Optional) The size of the depth dimension of the destination tensor
- * @param[in]  dst_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
- * @param[in]  M                                 Number of rows in LHS matrix not reshaped
- * @param[in]  N                                 Number of columns in RHS matrix not reshaped
- * @param[in]  K                                 Number of columns in LHS matrix and rows in RHS matrix not reshaped
- */
-//! @endcond
-__kernel void gemm_mm_reshaped_only_rhs_t(
-    TENSOR3D_T(lhs, BUFFER),
-    TENSOR3D_T(rhs, BUFFER),
-#if defined(BETA)
-    TENSOR3D_T(bia, BUFFER),
-#endif // defined(BETA)
-#if defined(POST_OP2)
-    TENSOR3D_T(ex0, BUFFER),
-#endif // defined(POST_OP_ADD)
-    TENSOR3D_T(dst, BUFFER),
-    const int M,
-    const int N,
-    const int K
-)
-{
-    // Block size
-#define RHS_BLOCK_SIZE ((K0) * (N0))
-
-    // RHS offset and step X
-#if defined(RHS_INTERLEAVE)
-#define RHS_OFFSET_X (K0)
-#define RHS_STEP_X ((K0) * (H0))
-#define RHS_STEP_LOOP (1)
-#else // defined(RHS_INTERLEAVE)
-#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
-#define RHS_STEP_X (K0)
-#define RHS_STEP_LOOP (H0)
-#endif // defined(RHS_INTERLEAVE)
-
-    const uint x = GET_SPATIAL_IDX(0, N0, 0);
-    const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0);
-    const uint z = GET_SPATIAL_IDX(2, 1, 0);
-
-#if defined(DUMMY_WORK_ITEMS)
-    if((x >= N) || (y >= M))
-    {
-        return;
-    }
-#endif // defined(DUMMY_WORK_ITEMS)
-
-    bool x_cond = PARTIAL_STORE_N0 != 0 && ((x + N0) > N);
-    bool y_cond = PARTIAL_STORE_M0 != 0 && y == 0;
-
-    TILE(uint, M0, 1, dst_indirect_y);
-    INITIALIZE_INDIRECT_Y(M0, PARTIAL_STORE_M0, y_cond, dst_indirect_y);
-
-    const uint x_rhs = x / N0;
-
-    lhs_offset_first_element_in_bytes += y * (uint)lhs_stride_y + z * (uint)lhs_stride_y * M;
-    rhs_offset_first_element_in_bytes += (x_rhs % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x_rhs / (uint)H0) * rhs_stride_y;
-
-#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_first_element_in_bytes += (z % MATRIX_B_DEPTH) * rhs_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    rhs_offset_first_element_in_bytes += z * rhs_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    // Initialize the accumulators
-    TILE(DATA_TYPE, M0, N0, c);
-
-    LOOP_UNROLLING(int, i, 0, 1, M0,
-    {
-        c[i].v = 0;
-    })
-
-    int i = 0;
-    for(; i <= (K - K0); i+=K0)
-    {
-        TILE(DATA_TYPE, M0, K0, a);
-        TILE(DATA_TYPE, N0, K0, b);
-
-        // Load tile from the lhs/rhs tensors
-        T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
-        T_LOAD(DATA_TYPE, N0, K0, BUFFER, rhs, 0, 0, 1, RHS_STEP_X * sizeof(DATA_TYPE), b);
-
-        // Compute the matrix multiplication between the two tiles
-        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, T, a, b, c);
-
-        lhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE);
-        rhs_offset_first_element_in_bytes += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE);
-    }
-#if defined(RUN_LEFTOVER_K0)
-    for(; i < K; ++i)
-    {
-        TILE(DATA_TYPE, M0, 1, a);
-        TILE(DATA_TYPE, N0, 1, b);
-
-        // Load tile from the lhs/rhs tensors
-        T_LOAD(DATA_TYPE, M0, 1, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
-        T_LOAD(DATA_TYPE, N0, 1, BUFFER, rhs, 0, 0, 1, RHS_STEP_X * sizeof(DATA_TYPE), b);
-
-        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, T, a, b, c);
-
-        lhs_offset_first_element_in_bytes += sizeof(DATA_TYPE);
-        rhs_offset_first_element_in_bytes += sizeof(DATA_TYPE);
-    }
-#endif // defined(RUN_LEFTOVER_K0)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    T_SCALE_CONSTANT(DATA_TYPE, M0, N0, c, (DATA_TYPE)ALPHA, c);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-#if defined(BROADCAST_BIAS)
-    TILE(DATA_TYPE, 1, N0, bias0);
-
-    T_LOAD_WIDTH_SELECT(DATA_TYPE, 1, N0, PARTIAL_STORE_N0, BUFFER, bia, x, 0, 0, x_cond, bias0);
-
-#ifndef UNIT_BETA
-    T_SCALE_CONSTANT(DATA_TYPE, 1, N0, bias0, (DATA_TYPE)BETA, bias0);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    T_ADD_BROADCAST_X(DATA_TYPE, M0, N0, c, bias0, c);
-#else // defined(BROADCAST_BIAS)
-    TILE(DATA_TYPE, M0, N0, bias0);
-
-    bia_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + (y * bia_stride_y) + (z * bia_stride_y * M);
-
-    T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, bia, 0, bia_stride_y, x_cond, bias0, dst_indirect_y);
-
-#ifndef UNIT_BETA
-    T_SCALE_CONSTANT(DATA_TYPE, M0, N0, bias0, (DATA_TYPE)BETA, bias0);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    T_ADD(DATA_TYPE, M0, N0, c, bias0, c);
-    // c = c + bias
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(POST_OP1)
-    T_ACTIVATION(DATA_TYPE, M0, N0, P1_ACTIVATION_TYPE, P1_ACTIVATION_A_VAL, P1_ACTIVATION_B_VAL, c, c);
-#endif // defined(POST_OP1)
-
-#if defined(POST_OP2)
-    TILE(DATA_TYPE, M0, N0, extra0);
-
-    ex0_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + (y * ex0_stride_y) + (z * ex0_stride_y * M);
-
-    T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, ex0, 0, ex0_stride_y, x_cond, extra0, dst_indirect_y);
-
-    T_ADD(DATA_TYPE, M0, N0, c, extra0, c);
-#endif // defined(POST_OP2)
-
-#if defined(POST_OP3)
-    T_ACTIVATION(DATA_TYPE, M0, N0, P3_ACTIVATION_TYPE, P3_ACTIVATION_A_VAL, P3_ACTIVATION_B_VAL, c, c);
-#endif // defined(POST_OP3)
-
-    dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + (y * dst_stride_y) + (z * dst_stride_y * M);
-
-    // Store the tile in reverse order so that the invalid values are overwritten with the valid ones
-    T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, c, dst_indirect_y);
-
-#undef RHS_BLOCK_SIZE
-#undef RHS_OFFSET_X
-#undef RHS_STEP_X
-}
-#endif // defined(GEMM_RESHAPED_RHS_ONLY_T)
-
-#if defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE)
-//! @cond Doxygen_Suppress
-/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops:
- * Post op 1: activation (optional)
- * Post op 2: elementwise op
- * Post op 3: activation (optional)
- *
- *  The LHS matrix is NOT reshaped
- *  The RHS is reshaped with @ref ClGemmMatrixMultiplyReshapedOnlyRhsKernel and the block K0xN0 is transposed
- *
- * @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 block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
- * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
- * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
- * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @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
- *  - H0 >= 1
- *
- * @note In case of post ops, the following information must be passed at compile time:
- * @note -DPOST_OP1, -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 1
- * @note -DPOST_OP2: The arithmetic addition post op to perform at slot 2
- * @note -DPOST_OP3, -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
- *
- * @param[in]  lhs_ptr                           Pointer to the LHS tensor. Supported data types: F16/F32
- * @param[in]  lhs_stride_y                      Stride of the LHS tensor in Y dimension (in bytes)
- * @param[in]  lhs_stride_z                      Stride of the LHS tensor in Z dimension (in bytes)
- * @param[in]  lhs_w                             The size of the width dimension of the LHS tensor
- * @param[in]  lhs_h                             The size of the height dimension of the LHS tensor
- * @param[in]  lhs_n                             The size of the depth dimension of the LHS tensor
- * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the LHS tensor
- * @param[in]  rhs_ptr                           Pointer to the RHS reshaped tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  rhs_stride_y                      Stride of the RHS tensor in Y dimension (in bytes)
- * @param[in]  rhs_stride_z                      Stride of the RHS tensor in Z dimension (in bytes)
- * @param[in]  rhs_w                             The size of the width dimension of the RHS tensor
- * @param[in]  rhs_h                             The size of the height dimension of the RHS tensor
- * @param[in]  rhs_n                             The size of the depth dimension of the RHS tensor
- * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS tensor
- * @param[in]  bia_ptr                           (Optional) Pointer to the bias tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  bia_stride_y                      (Optional) Stride of the bias tensor in Y dimension (in bytes)
- * @param[in]  bia_stride_z                      (Optional) Stride of the bias tensor in Z dimension (in bytes)
- * @param[in]  bia_w                             (Optional) The size of the width dimension of the bias tensor
- * @param[in]  bia_h                             (Optional) The size of the height dimension of the bias tensor
- * @param[in]  bia_n                             (Optional) The size of the depth dimension of the bias tensor
- * @param[in]  bia_offset_first_element_in_bytes (Optional) The offset of the first element in the bias tensor
- * @param[in]  ex0_ptr                           (Optional) Pointer to the tensor added with POST_OP2. Supported data type: same as @p lhs_ptr
- * @param[in]  ex0_stride_y                      (Optional) Stride of the tensor added with POST_OP2 in Y dimension (in bytes)
- * @param[in]  ex0_stride_z                      (Optional) Stride of the tensor added with POST_OP2 in Z dimension (in bytes)
- * @param[in]  ex0_w                             (Optional) The size of the width dimension of the tensor added with POST_OP2
- * @param[in]  ex0_h                             (Optional) The size of the height dimension of the tensor added with POST_OP2
- * @param[in]  ex0_n                             (Optional) The size of the depth dimension of the tensor added with POST_OP2
- * @param[in]  ex0_offset_first_element_in_bytes (Optional) The offset of the first element in the tensor added with POST_OP2
- * @param[out] dst_ptr                           (Optional) Pointer to the destination tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  dst_stride_y                      (Optional) Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_stride_z                      (Optional) Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  dst_w                             (Optional) The size of the width dimension of the destination tensor
- * @param[in]  dst_h                             (Optional) The size of the height dimension of the destination tensor
- * @param[in]  dst_n                             (Optional) The size of the depth dimension of the destination tensor
- * @param[in]  dst_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
- * @param[in]  M                                 Number of rows in LHS matrix not reshaped
- * @param[in]  N                                 Number of columns in RHS matrix not reshaped
- * @param[in]  K                                 Number of columns in LHS matrix and rows in RHS matrix not reshaped
- */
-//! @endcond
-__kernel void gemm_mm_reshaped_only_rhs_t_texture(
-    TENSOR3D_T(lhs, BUFFER),
-    TENSOR3D_T(rhs, IMAGE),
-#if defined(BETA)
-    TENSOR3D_T(bia, BUFFER),
-#endif // defined(BETA)
-#if defined(POST_OP2)
-    TENSOR3D_T(ex0, BUFFER),
-#endif // defined(POST_OP_ADD)
-    TENSOR3D_T(dst, BUFFER),
-    const int M,
-    const int N,
-    const int K
-)
-{
-    // Block size
-#define RHS_BLOCK_SIZE (K0 * (N0))
-
-    // RHS offset and step X
-#if defined(RHS_INTERLEAVE)
-#define RHS_OFFSET_X (K0)
-#define RHS_STEP_X (K0 * (H0))
-#define RHS_STEP_LOOP (1)
-#else // defined(RHS_INTERLEAVE)
-#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
-#define RHS_STEP_X K0
-#define RHS_STEP_LOOP (H0)
-#endif // defined(RHS_INTERLEAVE)
-
-    const uint x = GET_SPATIAL_IDX(0, N0, 0);
-    const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0);
-    const uint z = GET_SPATIAL_IDX(2, 1, 0);
-
-#if defined(DUMMY_WORK_ITEMS)
-    if((x >= N) || (y >= M))
-    {
-        return;
-    }
-#endif // defined(DUMMY_WORK_ITEMS)
-
-    bool x_cond = PARTIAL_STORE_N0 != 0 && ((x + N0) > N);
-    bool y_cond = PARTIAL_STORE_M0 != 0 && y == 0;
-
-    TILE(uint, M0, 1, dst_indirect_y);
-    INITIALIZE_INDIRECT_Y(M0, PARTIAL_STORE_M0, y_cond, dst_indirect_y);
-
-    lhs_offset_first_element_in_bytes += y * (uint)lhs_stride_y + z * lhs_stride_y * M;
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    const uint z_rhs = z % MATRIX_B_DEPTH;
-#else  // defined(MATRIX_B_DEPTH)
-    const uint z_rhs = z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    uint x_rhs = ((x / N0) % H0) * (uint)RHS_OFFSET_X;
-    const uint y_rhs = ((x / N0) / H0) + z_rhs * rhs_h;
-
-    // Initialize the accumulators
-    TILE(DATA_TYPE, M0, N0, c);
-
-    LOOP_UNROLLING(int, i, 0, 1, M0,
-    {
-        c[i].v = 0;
-    })
-
-    TILE(DATA_TYPE, M0, K0, a);
-    TILE(DATA_TYPE, N0, K0, b);
-
-    int i = 0;
-    for(; i <= (K - K0); i+=K0)
-    {
-        // Load tile from the lhs/rhs tensors
-        T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
-        T_LOAD_DILATED(DATA_TYPE, N0, K0, IMAGE, rhs, x_rhs, y_rhs, RHS_STEP_X, 0, 1, b);
-
-        // Compute the matrix multiplication between the two tiles
-        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, T, a, b, c);
-
-        lhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE);
-        x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP;
-    }
-#if defined(RUN_LEFTOVER_K0)
-    T_LOAD_DILATED(DATA_TYPE, N0, K0, IMAGE, rhs, x_rhs, y_rhs, RHS_STEP_X, 0, 1, b);
-
-        LOOP_UNROLLING(int, k0, 0, 1, PARTIAL_K,
-        {
-            LOOP_UNROLLING(int, m0, 0, 1, M0,
-            {
-                DATA_TYPE a0 = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset_first_element_in_bytes + m0 * lhs_stride_y);
-                LOOP_UNROLLING(int, n0, 0, 1, N0,
-                {
-                    c[m0].s[n0] += a0 * b[n0].s[k0];
-                })
-            })
-            lhs_offset_first_element_in_bytes += sizeof(DATA_TYPE);
-        })
-#endif // defined(RUN_LEFTOVER_K0)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    T_SCALE_CONSTANT(DATA_TYPE, M0, N0, c, (DATA_TYPE)ALPHA, c);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-#if defined(BROADCAST_BIAS)
-    TILE(DATA_TYPE, 1, N0, bias0);
-
-    T_LOAD_WIDTH_SELECT(DATA_TYPE, 1, N0, PARTIAL_STORE_N0, BUFFER, bia, x, 0, 0, x_cond, bias0);
-
-#ifndef UNIT_BETA
-    T_SCALE_CONSTANT(DATA_TYPE, 1, N0, bias0, (DATA_TYPE)BETA, bias0);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    T_ADD_BROADCAST_X(DATA_TYPE, M0, N0, c, bias0, c);
-#else // defined(BROADCAST_BIAS)
-    TILE(DATA_TYPE, M0, N0, bias0);
-
-    bia_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + (y * bia_stride_y) + (z * bia_stride_y * M);
-
-    T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, bia, 0, bia_stride_y, x_cond, bias0, dst_indirect_y);
-
-#ifndef UNIT_BETA
-    T_SCALE_CONSTANT(DATA_TYPE, M0, N0, bias0, (DATA_TYPE)BETA, bias0);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    T_ADD(DATA_TYPE, M0, N0, c, bias0, c);
-    // c = c + bias
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(POST_OP1)
-    T_ACTIVATION(DATA_TYPE, M0, N0, P1_ACTIVATION_TYPE, P1_ACTIVATION_A_VAL, P1_ACTIVATION_B_VAL, c, c);
-#endif // defined(POST_OP1)
-
-#if defined(POST_OP2)
-    TILE(DATA_TYPE, M0, N0, extra0);
-
-    ex0_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + (y * ex0_stride_y) + (z * ex0_stride_y * M);
-
-    T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, ex0, 0, ex0_stride_y, x_cond, extra0, dst_indirect_y);
-
-    T_ADD(DATA_TYPE, M0, N0, c, extra0, c);
-#endif // defined(POST_OP2)
-
-#if defined(POST_OP3)
-    T_ACTIVATION(DATA_TYPE, M0, N0, P3_ACTIVATION_TYPE, P3_ACTIVATION_A_VAL, P3_ACTIVATION_B_VAL, c, c);
-#endif // defined(POST_OP3)
-
-    dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_y * M;
-
-    // Store the tile in reverse order so that the invalid values are overwritten with the valid ones
-    T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, c, dst_indirect_y);
-
-#undef RHS_BLOCK_SIZE
-#undef RHS_OFFSET_X
-#undef RHS_STEP_X
-}
-#endif // defined(GEMM_RESHAPED_RHS_ONLY_T_TEXTURE)
-
-#if defined(GEMM_MM_RESHAPED_ONLY_RHS_NT)
-//! @cond Doxygen_Suppress
-/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops:
- * Post op 1: activation (optional)
- * Post op 2: elementwise op
- * Post op 3: activation (optional)
- *
- *  The LHS matrix is NOT reshaped
- *  The RHS is reshaped with @ref ClGemmMatrixMultiplyReshapedOnlyRhsKernel and the block K0xN0 is not transposed
- *
- * @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 block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
- * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
- * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
- * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @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
- *  - H0 >= 1
- *
- * @note In case of post ops, the following information must be passed at compile time:
- * @note -DPOST_OP1, -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 1
- * @note -DPOST_OP2: The arithmetic addition post op to perform at slot 2
- * @note -DPOST_OP3, -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
- *
- * @param[in]  lhs_ptr                           Pointer to the LHS tensor. Supported data types: F16/F32
- * @param[in]  lhs_stride_y                      Stride of the LHS tensor in Y dimension (in bytes)
- * @param[in]  lhs_stride_z                      Stride of the LHS tensor in Z dimension (in bytes)
- * @param[in]  lhs_w                             The size of the width dimension of the LHS tensor
- * @param[in]  lhs_h                             The size of the height dimension of the LHS tensor
- * @param[in]  lhs_n                             The size of the depth dimension of the LHS tensor
- * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the LHS tensor
- * @param[in]  rhs_ptr                           Pointer to the RHS reshaped tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  rhs_stride_y                      Stride of the RHS tensor in Y dimension (in bytes)
- * @param[in]  rhs_stride_z                      Stride of the RHS tensor in Z dimension (in bytes)
- * @param[in]  rhs_w                             The size of the width dimension of the RHS tensor
- * @param[in]  rhs_h                             The size of the height dimension of the RHS tensor
- * @param[in]  rhs_n                             The size of the depth dimension of the RHS tensor
- * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS tensor
- * @param[in]  bia_ptr                           (Optional) Pointer to the bias tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  bia_stride_y                      (Optional) Stride of the bias tensor in Y dimension (in bytes)
- * @param[in]  bia_stride_z                      (Optional) Stride of the bias tensor in Z dimension (in bytes)
- * @param[in]  bia_w                             (Optional) The size of the width dimension of the bias tensor
- * @param[in]  bia_h                             (Optional) The size of the height dimension of the bias tensor
- * @param[in]  bia_n                             (Optional) The size of the depth dimension of the bias tensor
- * @param[in]  bia_offset_first_element_in_bytes (Optional) The offset of the first element in the bias tensor
- * @param[in]  ex0_ptr                           (Optional) Pointer to the tensor added with POST_OP2. Supported data type: same as @p lhs_ptr
- * @param[in]  ex0_stride_y                      (Optional) Stride of the tensor added with POST_OP2 in Y dimension (in bytes)
- * @param[in]  ex0_stride_z                      (Optional) Stride of the tensor added with POST_OP2 in Z dimension (in bytes)
- * @param[in]  ex0_w                             (Optional) The size of the width dimension of the tensor added with POST_OP2
- * @param[in]  ex0_h                             (Optional) The size of the height dimension of the tensor added with POST_OP2
- * @param[in]  ex0_n                             (Optional) The size of the depth dimension of the tensor added with POST_OP2
- * @param[in]  ex0_offset_first_element_in_bytes (Optional) The offset of the first element in the tensor added with POST_OP2
- * @param[out] dst_ptr                           (Optional) Pointer to the destination tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  dst_stride_y                      (Optional) Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_stride_z                      (Optional) Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  dst_w                             (Optional) The size of the width dimension of the destination tensor
- * @param[in]  dst_h                             (Optional) The size of the height dimension of the destination tensor
- * @param[in]  dst_n                             (Optional) The size of the depth dimension of the destination tensor
- * @param[in]  dst_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
- * @param[in]  M                                 Number of rows in LHS matrix not reshaped
- * @param[in]  N                                 Number of columns in RHS matrix not reshaped
- * @param[in]  K                                 Number of columns in LHS matrix and rows in RHS matrix not reshaped
- */
-//! @endcond
-__kernel void gemm_mm_reshaped_only_rhs_nt(
-    TENSOR3D_T(lhs, BUFFER),
-    TENSOR3D_T(rhs, BUFFER),
-#if defined(BETA)
-    TENSOR3D_T(bia, BUFFER),
-#endif // defined(BETA)
-#if defined(POST_OP2)
-    TENSOR3D_T(ex0, BUFFER),
-#endif // defined(POST_OP_ADD)
-    TENSOR3D_T(dst, BUFFER),
-    const int M,
-    const int N,
-    const int K
-)
-{
-    // Block size
-#define RHS_BLOCK_SIZE ((K0) * (N0))
-
-    // RHS offset and step X
-#if defined(RHS_INTERLEAVE)
-#define RHS_OFFSET_X (N0)
-#define RHS_STEP_X ((N0) * (H0))
-#define RHS_STEP_LOOP (1)
-#else // defined(RHS_INTERLEAVE)
-#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
-#define RHS_STEP_X (N0)
-#define RHS_STEP_LOOP (H0)
-#endif // defined(RHS_INTERLEAVE)
-
-    const uint x = GET_SPATIAL_IDX(0, N0, 0);
-    const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0);
-    const uint z = GET_SPATIAL_IDX(2, 1, 0);
-
-#if defined(DUMMY_WORK_ITEMS)
-    if((x >= N) || (y >= M))
-    {
-        return;
-    }
-#endif // defined(DUMMY_WORK_ITEMS)
-
-    bool x_cond = PARTIAL_STORE_N0 != 0 && ((x + N0) > N);
-    bool y_cond = PARTIAL_STORE_M0 != 0 && y == 0;
-
-    TILE(uint, M0, 1, dst_indirect_y);
-    INITIALIZE_INDIRECT_Y(M0, PARTIAL_STORE_M0, y_cond, dst_indirect_y);
-
-    const uint x_rhs = x / N0;
-
-    lhs_offset_first_element_in_bytes += y * (uint)lhs_stride_y + z * lhs_stride_y * M;
-    rhs_offset_first_element_in_bytes += (x_rhs % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x_rhs / (uint)H0) * rhs_stride_y;
-
-#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_first_element_in_bytes += (z % MATRIX_B_DEPTH) * rhs_stride_z;
-#else  // defined(MATRIX_B_DEPTH)
-    rhs_offset_first_element_in_bytes += z * rhs_stride_z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    // Initialize the accumulators
-    TILE(DATA_TYPE, M0, N0, c);
-
-    LOOP_UNROLLING(int, i, 0, 1, M0,
-    {
-        c[i].v = 0;
-    })
-
-    int i = 0;
-    for(; i <= (K - K0); i+=K0)
-    {
-        TILE(DATA_TYPE, M0, K0, a);
-        TILE(DATA_TYPE, K0, N0, b);
-
-        // Load tile from the lhs/rhs tensors
-        T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
-        T_LOAD(DATA_TYPE, K0, N0, BUFFER, rhs, 0, 0, 1, RHS_STEP_X * sizeof(DATA_TYPE), b);
-
-        // Compute the matrix multiplication between the two tiles
-        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, NT, a, b, c);
-
-        lhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE);
-        rhs_offset_first_element_in_bytes += K0 * RHS_STEP_X * RHS_STEP_LOOP * sizeof(DATA_TYPE);
-    }
-#if defined(RUN_LEFTOVER_K0)
-    for(; i < K; ++i)
-    {
-        TILE(DATA_TYPE, M0, 1, a);
-        TILE(DATA_TYPE, 1, N0, b);
-
-        // Load tile from the lhs/rhs tensors
-        T_LOAD(DATA_TYPE, M0, 1, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
-        T_LOAD(DATA_TYPE, 1, N0, BUFFER, rhs, 0, 0, 1, RHS_STEP_X * sizeof(DATA_TYPE), b);
-
-        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, NT, a, b, c);
-
-        lhs_offset_first_element_in_bytes += sizeof(DATA_TYPE);
-        rhs_offset_first_element_in_bytes += RHS_STEP_X * sizeof(DATA_TYPE);
-    }
-#endif // defined(RUN_LEFTOVER_K0)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    T_SCALE_CONSTANT(DATA_TYPE, M0, N0, c, (DATA_TYPE)ALPHA, c);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-#if defined(BROADCAST_BIAS)
-    TILE(DATA_TYPE, 1, N0, bias0);
-
-    T_LOAD_WIDTH_SELECT(DATA_TYPE, 1, N0, PARTIAL_STORE_N0, BUFFER, bia, x, 0, 0, x_cond, bias0);
-
-#ifndef UNIT_BETA
-    T_SCALE_CONSTANT(DATA_TYPE, 1, N0, bias0, (DATA_TYPE)BETA, bias0);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    T_ADD_BROADCAST_X(DATA_TYPE, M0, N0, c, bias0, c);
-#else // defined(BROADCAST_BIAS)
-    TILE(DATA_TYPE, M0, N0, bias0);
-
-    bia_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + (y * bia_stride_y) + (z * bia_stride_y * M);
-
-    T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, bia, 0, bia_stride_y, x_cond, bias0, dst_indirect_y);
-
-#ifndef UNIT_BETA
-    T_SCALE_CONSTANT(DATA_TYPE, M0, N0, bias0, (DATA_TYPE)BETA, bias0);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    T_ADD(DATA_TYPE, M0, N0, c, bias0, c);
-    // c = c + bias
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(POST_OP1)
-    T_ACTIVATION(DATA_TYPE, M0, N0, P1_ACTIVATION_TYPE, P1_ACTIVATION_A_VAL, P1_ACTIVATION_B_VAL, c, c);
-#endif // defined(POST_OP1)
-
-#if defined(POST_OP2)
-    TILE(DATA_TYPE, M0, N0, extra0);
-
-    ex0_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + (y * ex0_stride_y) + (z * ex0_stride_y * M);
-
-    T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, ex0, 0, ex0_stride_y, x_cond, extra0, dst_indirect_y);
-
-    T_ADD(DATA_TYPE, M0, N0, c, extra0, c);
-#endif // defined(POST_OP2)
-
-#if defined(POST_OP3)
-    T_ACTIVATION(DATA_TYPE, M0, N0, P3_ACTIVATION_TYPE, P3_ACTIVATION_A_VAL, P3_ACTIVATION_B_VAL, c, c);
-#endif // defined(POST_OP3)
-
-    dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_y * M;
-
-    // Store the tile in reverse order so that the invalid values are overwritten with the valid ones
-    T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, c, dst_indirect_y);
-
-#undef RHS_BLOCK_SIZE
-#undef RHS_OFFSET_X
-#undef RHS_STEP_X
-}
-#endif // defined(GEMM_RESHAPED_RHS_ONLY_NT)
-
-#if defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE)
-//! @cond Doxygen_Suppress
-/** This OpenCL kernel computes the matrix multiplication between 2 matrices plus 3 post ops:
- * Post op 1: activation (optional)
- * Post op 2: elementwise op
- * Post op 3: activation (optional)
- *
- *  The LHS matrix is NOT reshaped
- *  The RHS is reshaped with @ref ClGemmMatrixMultiplyReshapedOnlyRhsKernel and the block K0xN0 is not transposed
- *
- * @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 block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
- * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
- * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
- * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
- * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
- * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
- * @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
- *  - H0 >= 1
- *
- * @note In case of post ops, the following information must be passed at compile time:
- * @note -DPOST_OP1, -DP1_ACTIVATION_TYPE, -DP1_ACTIVATION_A_VAL, -DP1_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 1
- * @note -DPOST_OP2: The arithmetic addition post op to perform at slot 2
- * @note -DPOST_OP3, -DP3_ACTIVATION_TYPE, -DP3_ACTIVATION_A_VAL, -DP3_ACTIVATION_B_VAL: The activation type, alpha and beta values of the activation post op at slot 3
- *
- * @param[in]  lhs_ptr                           Pointer to the LHS tensor. Supported data types: F16/F32
- * @param[in]  lhs_stride_y                      Stride of the LHS tensor in Y dimension (in bytes)
- * @param[in]  lhs_stride_z                      Stride of the LHS tensor in Z dimension (in bytes)
- * @param[in]  lhs_w                             The size of the width dimension of the LHS tensor
- * @param[in]  lhs_h                             The size of the height dimension of the LHS tensor
- * @param[in]  lhs_n                             The size of the depth dimension of the LHS tensor
- * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the LHS tensor
- * @param[in]  rhs_ptr                           Pointer to the RHS reshaped tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  rhs_stride_y                      Stride of the RHS tensor in Y dimension (in bytes)
- * @param[in]  rhs_stride_z                      Stride of the RHS tensor in Z dimension (in bytes)
- * @param[in]  rhs_w                             The size of the width dimension of the RHS tensor
- * @param[in]  rhs_h                             The size of the height dimension of the RHS tensor
- * @param[in]  rhs_n                             The size of the depth dimension of the RHS tensor
- * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS tensor
- * @param[in]  bia_ptr                           (Optional) Pointer to the bias tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  bia_stride_y                      (Optional) Stride of the bias tensor in Y dimension (in bytes)
- * @param[in]  bia_stride_z                      (Optional) Stride of the bias tensor in Z dimension (in bytes)
- * @param[in]  bia_w                             (Optional) The size of the width dimension of the bias tensor
- * @param[in]  bia_h                             (Optional) The size of the height dimension of the bias tensor
- * @param[in]  bia_n                             (Optional) The size of the depth dimension of the bias tensor
- * @param[in]  bia_offset_first_element_in_bytes (Optional) The offset of the first element in the bias tensor
- * @param[in]  ex0_ptr                           (Optional) Pointer to the tensor added with POST_OP2. Supported data type: same as @p lhs_ptr
- * @param[in]  ex0_stride_y                      (Optional) Stride of the tensor added with POST_OP2 in Y dimension (in bytes)
- * @param[in]  ex0_stride_z                      (Optional) Stride of the tensor added with POST_OP2 in Z dimension (in bytes)
- * @param[in]  ex0_w                             (Optional) The size of the width dimension of the tensor added with POST_OP2
- * @param[in]  ex0_h                             (Optional) The size of the height dimension of the tensor added with POST_OP2
- * @param[in]  ex0_n                             (Optional) The size of the depth dimension of the tensor added with POST_OP2
- * @param[in]  ex0_offset_first_element_in_bytes (Optional) The offset of the first element in the tensor added with POST_OP2
- * @param[out] dst_ptr                           (Optional) Pointer to the destination tensor. Supported data type: same as @p lhs_ptr
- * @param[in]  dst_stride_y                      (Optional) Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_stride_z                      (Optional) Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  dst_w                             (Optional) The size of the width dimension of the destination tensor
- * @param[in]  dst_h                             (Optional) The size of the height dimension of the destination tensor
- * @param[in]  dst_n                             (Optional) The size of the depth dimension of the destination tensor
- * @param[in]  dst_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
- * @param[in]  M                                 Number of rows in LHS matrix not reshaped
- * @param[in]  N                                 Number of columns in RHS matrix not reshaped
- * @param[in]  K                                 Number of columns in LHS matrix and rows in RHS matrix not reshaped
- */
-//! @endcond
-__kernel void gemm_mm_reshaped_only_rhs_nt_texture(
-    TENSOR3D_T(lhs, BUFFER),
-    TENSOR3D_T(rhs, IMAGE),
-#if defined(BETA)
-    TENSOR3D_T(bia, BUFFER),
-#endif // defined(BETA)
-#if defined(POST_OP2)
-    TENSOR3D_T(ex0, BUFFER),
-#endif // defined(POST_OP_ADD)
-    TENSOR3D_T(dst, BUFFER),
-    const int M,
-    const int N,
-    const int K
-)
-{
-    // Block size
-#define RHS_BLOCK_SIZE ((K0) * (N0))
-
-    // RHS offset and step X
-#if defined(RHS_INTERLEAVE)
-#define RHS_OFFSET_X (N0)
-#define RHS_STEP_X ((N0) * (H0))
-#define RHS_STEP_LOOP (1)
-#else // defined(RHS_INTERLEAVE)
-#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
-#define RHS_STEP_X (N0)
-#define RHS_STEP_LOOP (H0)
-#endif // defined(RHS_INTERLEAVE)
-
-    const uint x = GET_SPATIAL_IDX(0, N0, 0);
-    const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0);
-    const uint z = GET_SPATIAL_IDX(2, 1, 0);
-
-#if defined(DUMMY_WORK_ITEMS)
-    if((x >= N) || (y >= M))
-    {
-        return;
-    }
-#endif // defined(DUMMY_WORK_ITEMS)
-
-    bool x_cond = PARTIAL_STORE_N0 != 0 && ((x + N0) > N);
-    bool y_cond = PARTIAL_STORE_M0 != 0 && y == 0;
-
-    TILE(uint, M0, 1, dst_indirect_y);
-    INITIALIZE_INDIRECT_Y(M0, PARTIAL_STORE_M0, y_cond, dst_indirect_y);
-
-    lhs_offset_first_element_in_bytes += y * (uint)lhs_stride_y + z * lhs_stride_y * M;
-
-#if defined(MATRIX_B_DEPTH)
-    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
-    const uint z_rhs = z % MATRIX_B_DEPTH;
-#else  // defined(MATRIX_B_DEPTH)
-    const uint z_rhs = z;
-#endif // defined(MATRIX_B_DEPTH)
-
-    uint x_rhs = ((x / N0) % H0) * (uint)RHS_OFFSET_X;
-    const uint y_rhs = ((x / N0) / H0) + z_rhs * rhs_h;
-
-    // Initialize the accumulators
-    TILE(DATA_TYPE, M0, N0, c);
-
-    LOOP_UNROLLING(int, i, 0, 1, M0,
-    {
-        c[i].v = 0;
-    })
-
-    int i = 0;
-    for(; i <= (K - K0); i+=K0)
-    {
-        TILE(DATA_TYPE, M0, K0, a);
-        TILE(DATA_TYPE, K0, N0, b);
-
-        // Load tile from the lhs/rhs tensors
-        T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
-        T_LOAD_DILATED(DATA_TYPE, K0, N0, IMAGE, rhs, x_rhs, y_rhs, RHS_STEP_X, 0, 1, b);
-
-        // Compute the matrix multiplication between the two tiles
-        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, NT, a, b, c);
-
-        lhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE);
-        x_rhs += K0 * RHS_STEP_X * RHS_STEP_LOOP;
-    }
-
-#if defined(RUN_LEFTOVER_K0)
-    for(; i < K; ++i)
-    {
-        TILE(DATA_TYPE, M0, 1, a);
-        TILE(DATA_TYPE, 1, N0, b);
-
-        // Load tile from the lhs/rhs tensors
-        T_LOAD(DATA_TYPE, M0, 1, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
-        T_LOAD_DILATED(DATA_TYPE, 1, N0, IMAGE, rhs, x_rhs, y_rhs, RHS_STEP_X, 0, 1, b);
-
-        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, NT, a, b, c);
-
-        lhs_offset_first_element_in_bytes += sizeof(DATA_TYPE);
-        x_rhs += RHS_STEP_X;
-    }
-#endif // defined(RUN_LEFTOVER_K0)
-
-    // Multiply by the weight of matrix-matrix product and store the result
-#if defined(ALPHA)
-    T_SCALE_CONSTANT(DATA_TYPE, M0, N0, c, (DATA_TYPE)ALPHA, c);
-#endif // defined(ALPHA)
-
-    // Add beta*bias
-#if defined(BETA)
-#if defined(BROADCAST_BIAS)
-    TILE(DATA_TYPE, 1, N0, bias0);
-
-    T_LOAD_WIDTH_SELECT(DATA_TYPE, 1, N0, PARTIAL_STORE_N0, BUFFER, bia, x, 0, 0, x_cond, bias0);
-
-#ifndef UNIT_BETA
-    T_SCALE_CONSTANT(DATA_TYPE, 1, N0, bias0, (DATA_TYPE)BETA, bias0);
-#endif // UNIT_BIAS
-
-    // c = c + bias[broadcasted]
-    T_ADD_BROADCAST_X(DATA_TYPE, M0, N0, c, bias0, c);
-#else // defined(BROADCAST_BIAS)
-    TILE(DATA_TYPE, M0, N0, bias0);
-
-    bia_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + (y * bia_stride_y) + (z * bia_stride_y * M);
-
-    T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, bia, 0, bia_stride_y, x_cond, bias0, dst_indirect_y);
-
-#ifndef UNIT_BETA
-    T_SCALE_CONSTANT(DATA_TYPE, M0, N0, bias0, (DATA_TYPE)BETA, bias0);
-#endif // UNIT_BIAS
-
-    // c = c + bias
-    T_ADD(DATA_TYPE, M0, N0, c, bias0, c);
-    // c = c + bias
-#endif // defined(BROADCAST_BIAS)
-#endif // defined(BETA)
-
-#if defined(POST_OP1)
-    T_ACTIVATION(DATA_TYPE, M0, N0, P1_ACTIVATION_TYPE, P1_ACTIVATION_A_VAL, P1_ACTIVATION_B_VAL, c, c);
-#endif // defined(POST_OP1)
-
-#if defined(POST_OP2)
-    TILE(DATA_TYPE, M0, N0, extra0);
-
-    ex0_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + (y * ex0_stride_y) + (z * ex0_stride_y * M);
-
-    T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, ex0, 0, ex0_stride_y, x_cond, extra0, dst_indirect_y);
-
-    T_ADD(DATA_TYPE, M0, N0, c, extra0, c);
-#endif // defined(POST_OP2)
-
-#if defined(POST_OP3)
-    T_ACTIVATION(DATA_TYPE, M0, N0, P3_ACTIVATION_TYPE, P3_ACTIVATION_A_VAL, P3_ACTIVATION_B_VAL, c, c);
-#endif // defined(POST_OP3)
-
-    dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_y * M;
-
-    // Store the tile in reverse order so that the invalid values are overwritten with the valid ones
-    T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, c, dst_indirect_y);
-
-#undef RHS_BLOCK_SIZE
-#undef RHS_OFFSET_X
-#undef RHS_STEP_X
-}
-#endif // defined(GEMM_RESHAPED_RHS_ONLY_NT_TEXTURE)
\ No newline at end of file
diff --git a/src/core/CL/cl_kernels/tile_helpers.h b/src/core/CL/cl_kernels/tile_helpers.h
index 6a77463325..eba2316243 100644
--- a/src/core/CL/cl_kernels/tile_helpers.h
+++ b/src/core/CL/cl_kernels/tile_helpers.h
@@ -475,106 +475,6 @@
         })                                                                                                             \
     })
 
-/** Load a tile from global memory (tensor) and conditionally use a different length for the load
- *
- * @note If WIDTH1_CONDITION is true, the load will use the WIDTH1 length
- * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones
- *
- * @param[in]  DATA_TYPE        Data type
- * @param[in]  HEIGHT           Number of dst rows
- * @param[in]  WIDTH0           Load width to use if WIDTH1_CONDITION = false
- * @param[in]  WIDTH1           Load width to use if WIDTH1_CONDITION = true
- * @param[in]  TENSOR_TYPE      Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
- *                              In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
- * @param[in]  TENSOR           Tensor basename
- * @param[in]  X                Starting X position
- * @param[in]  Y                Starting Y position
- * @param[in]  STRIDE_Y         Stride Y (in bytes) used to load each row.
- * @param[in]  WIDTH1_CONDITION Condition to select the WIDTH1 store
- * @param[out] dst              Output tile
- */
-#define T_LOAD_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, WIDTH1_CONDITION, dst)                                                      \
-    ({                                                                                                                                                                                             \
-        if(WIDTH1_CONDITION)                                                                                                                                                                       \
-        {                                                                                                                                                                                          \
-            LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                                                                                                  \
-            {                                                                                                                                                                                      \
-                VLOAD_PARTIAL(WIDTH0, WIDTH1)                                                         \
-                (dst[HEIGHT - 1 - _i].v, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (Y) * STRIDE_Y));               \
-            })                                                                                                                                                                                     \
-        }                                                                                                                                                                                          \
-        else                                                                                                                                                                                       \
-        {                                                                                                                                                                                          \
-            LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                                                                                                  \
-            {                                                                                                                                                                                      \
-                dst[HEIGHT - 1 - _i].v = V_LOAD(DATA_TYPE, WIDTH0, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y); \
-            })                                                                                                                                                                                     \
-        }                                                                                                                                                                                          \
-    })
-
-/** Load a tile from global memory (tensor) using an indirect Y index tile and conditionally use a different length for the load
- *
- * @note If WIDTH1_CONDITION is true, the load will use the WIDTH1 length
- * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones
- *
- * @param[in]  DATA_TYPE        Data type
- * @param[in]  HEIGHT           Number of dst rows
- * @param[in]  WIDTH0           Load width to use if WIDTH1_CONDITION = false
- * @param[in]  WIDTH1           Load width to use if WIDTH1_CONDITION = true
- * @param[in]  TENSOR_TYPE      Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
- *                              In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
- * @param[in]  TENSOR           Tensor basename
- * @param[in]  X                Starting X position
- * @param[in]  STRIDE_Y         Stride Y (in bytes) used to load each row.
- * @param[in]  WIDTH1_CONDITION Condition to select the WIDTH1 store
- * @param[out] dst              Output tile
- * @param[in]  indirect_y       Indirect Y index tile
- */
-#define T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, dst, indirect_y)                                                      \
-    ({                                                                                                                                                                                             \
-        if(WIDTH1_CONDITION)                                                                                                                                                                       \
-        {                                                                                                                                                                                          \
-            LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                                                                                                  \
-            {                                                                                                                                                                                      \
-                VLOAD_PARTIAL(WIDTH0, WIDTH1)                                                         \
-                (dst[HEIGHT - 1 - _i].v, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y));               \
-            })                                                                                                                                                                                     \
-        }                                                                                                                                                                                          \
-        else                                                                                                                                                                                       \
-        {                                                                                                                                                                                          \
-            LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                                                                                                  \
-            {                                                                                                                                                                                      \
-                dst[HEIGHT - 1 - _i].v = V_LOAD(DATA_TYPE, WIDTH0, TENSOR_TYPE, TENSOR, X, (indirect_y[HEIGHT - 1 - _i].v), STRIDE_Y); \
-            })                                                                                                                                                                                     \
-        }                                                                                                                                                                                          \
-    })
-
-/** Load a tile from global memory (tensor) with dilation for the X and Y direction
- *
- * @note If WIDTH1_CONDITION is true, the load will use the WIDTH1 length
- * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones
- *
- * @param[in]  DATA_TYPE     Data type
- * @param[in]  HEIGHT        Number of dst rows
- * @param[in]  WIDTH         Number of dst columns
- * @param[in]  TENSOR_TYPE   Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
- *                           In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
- * @param[in]  TENSOR        Tensor basename
- * @param[in]  X             Starting X position
- * @param[in]  Y             Starting Y position
- * @param[in]  XI_MULTIPLIER Dilation for the X increment
- * @param[in]  YI_MULTIPLIER Dilation for the Y increment
- * @param[in]  STRIDE_Y      Stride Y (in bytes) used to load each row.
- * @param[out] dst           Output tile
- */
-#define T_LOAD_DILATED(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, Y, XI_MULTIPLIER, YI_MULTIPLIER, STRIDE_Y, dst)                      \
-    ({                                                                                                                 \
-        LOOP_UNROLLING(int, _i, 0, 1, HEIGHT,                                                                          \
-        {                                                                                                              \
-            dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, ((X) + _i * (int)(XI_MULTIPLIER)), ((Y) + _i * (int)(YI_MULTIPLIER)), STRIDE_Y); \
-        })                                                                                                             \
-    })
-
 /** Load a tile from global memory (tensor) using an indirect Y index tile
  *
  * @param[in]  DATA_TYPE   Data type
@@ -1086,25 +986,6 @@
         })                                                                                          \
     })
 
-/** Element-wise addition between two tiles
- *
- * @note Performs: LHS + RHS = DST
- *
- * @param[in]  DATA_TYPE LHS/RHS/DST data type
- * @param[in]  M0        Number of LHS rows
- * @param[in]  N0        Number of LHS columns
- * @param[in]  lhs       LHS tile
- * @param[in]  rhs       Constant LHS tile
- * @param[out] dst       DST tile
- */
-#define T_ADD(DATA_TYPE, M0, N0, lhs, rhs, dst) \
-    ({                                                            \
-        LOOP_UNROLLING(int, _m0, 0, 1, M0,                        \
-        {                                                         \
-            dst[_m0].v = lhs[_m0].v + rhs[_m0].v; \
-        })                                                        \
-    })
-
 /** Element-wise addition with a constant value
  *
  * @note Performs: LHS + constant = DST
@@ -1120,26 +1001,10 @@
     ({                                                            \
         LOOP_UNROLLING(int, _m0, 0, 1, M0,                        \
         {                                                         \
-            dst[_m0].v = lhs[_m0].v + (DATA_TYPE)rhs_constant;               \
-        })                                                        \
-    })
-
-/** Element-wise scale with a constant value
- *
- * @note Performs: LHS * constant = DST
- *
- * @param[in]  DATA_TYPE    LHS/RHS/DST data type
- * @param[in]  M0           Number of LHS rows
- * @param[in]  N0           Number of LHS columns
- * @param[in]  lhs          LHS tile
- * @param[in]  rhs_constant Constant value
- * @param[out] dst          DST tile
- */
-#define T_SCALE_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \
-    ({                                                            \
-        LOOP_UNROLLING(int, _m0, 0, 1, M0,                        \
-        {                                                         \
-            dst[_m0].v = lhs[_m0].v * (DATA_TYPE)rhs_constant; \
+            LOOP_UNROLLING(int, _n0, 0, 1, N0,                    \
+            {                                                     \
+                dst[_m0].s[_n0] = lhs[_m0].s[_n0] + rhs_constant; \
+            })                                                    \
         })                                                        \
     })
 
@@ -1201,26 +1066,6 @@
         })                                                                                \
     }
 
-#define T_MMUL_NT_NT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_NT_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
-#define T_MMUL_NT_NT_float_float_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
-#define T_MMUL_NT_NT_half_half_half(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
-#define T_MMUL_NT_NT_char_char_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_NT_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
-#define T_MMUL_NT_NT_uchar_uchar_uint(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_NT_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
-#define T_MMUL_NT_NT_uchar_uchar_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_NT_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
-#define T_MMUL_NT_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)                       \
-    {                                                                                     \
-        LOOP_UNROLLING(int, _m, 0, 1, M0,                                                 \
-        {                                                                                 \
-            LOOP_UNROLLING(int, _n, 0, 1, N0,                                             \
-            {                                                                             \
-                LOOP_UNROLLING(int, _k, 0, 1, K0,                                         \
-                {                                                                         \
-                    dst[_m].s[_n] = fma((lhs[_m].s[_k]), (rhs[_k].s[_n]), dst[_m].s[_n]); \
-                })                                                                        \
-            })                                                                            \
-        })                                                                                \
-    }
-
 #define T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)                            \
     ({ \
         LOOP_UNROLLING(int, _m, 0, 1, M0, \
@@ -1231,26 +1076,3 @@
             })                                                                                             \
         })                                                                                             \
     })
-
-/** Initialize indirect Y for avoiding out-of-bound reads/writes
- *
- * @param[in]  M0         Tile height to use if CONDITION = false
- * @param[in]  M1         Tile height to use if CONDITION = true
- * @param[in]  COND       Condition to select the M1 tile height
- * @param[out] indirect_y Indirect tile
- */
-#define INITIALIZE_INDIRECT_Y(M0, M1, COND, indirect_y)     \
-    if(COND)                                                \
-    {                                                       \
-        LOOP_UNROLLING(int, _i, 0, 1, M0,                    \
-        {                                                   \
-            indirect_y[_i].v = min(_i, (int)M1 - 1);          \
-        })                                                  \
-    }                                                       \
-    else                                                    \
-    {                                                       \
-        LOOP_UNROLLING(int, _i, 0, 1, M0,                    \
-        {                                                   \
-            indirect_y[_i].v = _i;                            \
-        })                                                  \
-    }
diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp
index 856d37766a..2c43a20908 100644
--- a/src/gpu/cl/ClKernelLibrary.cpp
+++ b/src/gpu/cl/ClKernelLibrary.cpp
@@ -281,10 +281,14 @@ const std::map<std::string, std::string> ClKernelLibrary::_kernel_program_map =
     { "gemm_mm_reshaped_lhs_nt_rhs_t_texture_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl" },
     { "gemm_mm_reshaped_lhs_t_rhs_nt_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl" },
     { "gemm_mm_reshaped_lhs_t_rhs_nt_texture_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.cl" },
-    { "gemm_mm_reshaped_only_rhs_nt", "common/gemm_reshaped_rhs_only.cl" },
-    { "gemm_mm_reshaped_only_rhs_nt_texture", "common/gemm_reshaped_rhs_only.cl" },
-    { "gemm_mm_reshaped_only_rhs_t", "common/gemm_reshaped_rhs_only.cl" },
-    { "gemm_mm_reshaped_only_rhs_t_texture", "common/gemm_reshaped_rhs_only.cl" },
+    { "gemm_mm_reshaped_only_rhs_nt", "common/gemm.cl" },
+    { "gemm_mm_reshaped_only_rhs_nt_texture", "common/gemm.cl" },
+    { "gemm_mm_reshaped_only_rhs_t", "common/gemm.cl" },
+    { "gemm_mm_reshaped_only_rhs_t_texture", "common/gemm.cl" },
+    { "gemm_mm_reshaped_only_rhs_nt_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl" },
+    { "gemm_mm_reshaped_only_rhs_nt_texture_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl" },
+    { "gemm_mm_reshaped_only_rhs_t_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl" },
+    { "gemm_mm_reshaped_only_rhs_t_texture_post_act_eltwise_op_act", "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl" },
     { "gemm_lc_vm_f32", "common/gemm.cl" },
     { "gemm_reshape_lhs_matrix_nt", "common/gemm_utils.cl" },
     { "gemm_reshape_lhs_matrix_t", "common/gemm_utils.cl" },
@@ -586,10 +590,6 @@ const std::map<std::string, std::string> ClKernelLibrary::_program_source_map =
 #include "./cl_kernels/common/gemm_utils.clembed"
     },
     {
-        "common/gemm_reshaped_rhs_only.cl",
-#include "./cl_kernels/common/gemm_reshaped_rhs_only.clembed"
-    },
-    {
         "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.cl",
 #include "./cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_native.clembed"
     },
@@ -598,6 +598,10 @@ const std::map<std::string, std::string> ClKernelLibrary::_program_source_map =
 #include "./cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped.clembed"
     },
     {
+        "common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.cl",
+#include "./cl_kernels/common/experimental/gemm_fused_post_ops/act_eltwise_op_act/gemm_mm_reshaped_only_rhs.clembed"
+    },
+    {
         "common/gemmlowp.cl",
 #include "./cl_kernels/common/gemmlowp.clembed"
     },
diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp
index 546a61e264..a8bcf8d6a1 100644
--- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp
+++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.cpp
@@ -83,7 +83,6 @@ Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, cons
                                     && (!gemm_info.broadcast_bias),
                                     "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.has_pad_y, "Tensors cannot have padding along the Y direction");
     ARM_COMPUTE_RETURN_ON_ERROR(gemm::validate_image2d_support_on_rhs(*src1, rhs_info));
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!post_op_utils.is_post_op_sequence_supported(gemm_info.post_ops), "The sequence of Post Ops is not supported");
 
@@ -143,8 +142,17 @@ Window validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITens
     ARM_COMPUTE_UNUSED(src0, src1, src2);
     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;
 
+    // In case both input and dst have to be reinterpreted as 3D tensors,
+    // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+    // This approach should only be used when the input/dst tensors have pad on the y direction
+    if((reinterpret_input_as_3d == reinterpret_output_as_3d) && gemm_info.has_pad_y)
+    {
+        reinterpret_output_as_3d = false;
+    }
+
     TensorInfo tmp_info(*dst);
 
     if(reinterpret_output_as_3d)
@@ -192,10 +200,19 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext
     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
     _add_bias                 = src2 != nullptr;
     _export_to_cl_image       = rhs_info.export_to_cl_image;
+    _has_pad_y                = gemm_info.has_pad_y;
     _num_post_op_args         = gemm_info.post_ops.total_num_arguments();
 
     auto padding_info = get_padding_info({ src0, src1, src2, dst });
 
+    // In case both input and dst 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) && _has_pad_y)
+    {
+        _reinterpret_input_as_3d  = false;
+        _reinterpret_output_as_3d = false;
+    }
+
     // Check if we need to slide the matrix B
     const unsigned int num_dimensions_src0 = src0->num_dimensions();
     _slide_matrix_b                        = (src1->num_dimensions() >= num_dimensions_src0);
@@ -212,6 +229,10 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext
     // This means that the actual m used by the kernel is given by dst->dimension(1) and not by gemm_info.m
     const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : dst->dimension(1);
 
+    // These variables are used only if gemm_info.has_pad_y == true
+    const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(1) : src0->dimension(1);
+    const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(2) : src0->dimension(2);
+
     // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
     // NOTE: This might have implications on heuristics and performance
     const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
@@ -222,26 +243,31 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext
     _m                                  = internal_m;
     _n                                  = gemm_info.n;
     _k                                  = gemm_info.k;
-
     // Create build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type()));
-    build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_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));
-    build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
-    build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
-    build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
-    build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
     build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
     build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
     build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
     build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
     build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2)));
     build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
-    build_opts.add_option_if(gemm_info.k % rhs_info.k0, "-DRUN_LEFTOVER_K0");
-    build_opts.add_option_if((gemm_info.k % rhs_info.k0) && rhs_info.transpose, "-DPARTIAL_K=" + support::cpp11::to_string(gemm_info.k % rhs_info.k0));
-
+    build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
+    build_opts.add_option_if(rhs_info.export_to_cl_image, "-DOPENCL_IMAGE_SUPPORT");
+    build_opts.add_option("-DRHS_HEIGHT=" + support::cpp11::to_string(src1->dimension(1)));
+    build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_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));
+    build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
+    build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
+    build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
+    if(_has_pad_y)
+    {
+        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(h_gemm_3d));
+        build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d));
+    }
     // If post_ops are used, then we disable the use of gemm_info.activation_info
     if(gemm_info.post_ops.size() > 0)
     {
@@ -249,15 +275,15 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext
     }
     else
     {
-        build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DPOST_OP1");
-        build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DP1_ACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
-        build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DP1_ACTIVATION_A_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
-        build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DP1_ACTIVATION_B_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
+        build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
+        build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
+        build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
     }
 
     std::string kernel_name("gemm_mm_reshaped_only_rhs_");
     kernel_name += rhs_info.transpose ? "t" : "nt";
-    kernel_name += _export_to_cl_image ? "_texture" : "";
+    kernel_name += rhs_info.export_to_cl_image ? "_texture" : "";
+    post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.post_ops);
 
     // A macro guard to compile ONLY the kernel of interest
     build_opts.add_option("-D" + upper_string(kernel_name));
@@ -268,8 +294,11 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext
     // Set config_id for enabling LWS tuning
     _config_id = kernel_name;
     _config_id += "_";
+    _config_id += (_has_pad_y ? "" : "no_pad_y_");
     _config_id += (_add_bias ? "add_bias_" : "");
     _config_id += (gemm_info.broadcast_bias ? "broadcast_bias_" : "");
+    _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
+    _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
     _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
     _config_id += lower_string(string_from_data_type(src0->data_type()));
     _config_id += "_";
@@ -321,12 +350,24 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con
         ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0);
     }
 
+    const size_t lhs_idx_batch_size = _reinterpret_input_as_3d && !_has_pad_y ? 3u : 2u;
+    const size_t rhs_idx_batch_size = 2u;
+    const size_t bia_idx_batch_size = 2u;
+    const size_t out_idx_batch_size = _reinterpret_output_as_3d && !_has_pad_y ? 3u : 2u;
+
     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));
 
+    // Get cross plane pads
+    const unsigned int total_cross_plane_pad_lhs = src0->info()->padding().top + src0->info()->padding().bottom;
+    const unsigned int total_cross_plane_pad_out = dst->info()->padding().top + dst->info()->padding().bottom;
+
+    // The execution should fail if we try to run with has_pad_y = false but we have padding in either the LHS or DST tensor
+    ARM_COMPUTE_ERROR_ON(!_has_pad_y && ((total_cross_plane_pad_lhs != 0) || (total_cross_plane_pad_out != 0)));
+
     cl::Image2D src1_image2d;
 
     if(_export_to_cl_image)
@@ -350,30 +391,63 @@ void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, con
         unsigned int idx = 0;
 
         // LHS buffer
-        add_3d_tensor_nhw_argument(idx, src0);
+        add_2D_tensor_argument(idx, src0, slice);
 
         // RHS buffer or RHS OpenCL image (_export_to_cl_image == true)
         if(_export_to_cl_image)
         {
             _kernel.setArg(idx++, src1_image2d);
         }
-        add_3d_tensor_nhw_argument(idx, src1);
+        else
+        {
+            add_2D_tensor_argument(idx, src1, slice_b);
+        }
 
         // Bias buffer (_add_bias == true)
+        add_2D_tensor_argument_if(_add_bias, idx, src2, slice);
+
+        // dst buffer
+        add_2D_tensor_argument(idx, dst, slice);
+
+        // post op argument buffers
+        for(size_t i = 0; i < _num_post_op_args; ++i)
+        {
+            const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
+            add_2D_tensor_argument(idx, post_op_arg, slice);
+        }
+
+        // LHS stride_z
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[lhs_idx_batch_size]));
+
+        // RHS stride_z (not used if _export_to_cl_image == true)
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[rhs_idx_batch_size]));
+
+        // Bias stride_z (if _add_bias == true)
         if(_add_bias)
         {
-            add_3d_tensor_nhw_argument(idx, src2);
+            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[bia_idx_batch_size]));
         }
 
-        // post op argument buffers
+        // dst stride_z
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[out_idx_batch_size]));
+        // post op argument stride_z
         for(size_t i = 0; i < _num_post_op_args; ++i)
         {
             const auto post_op_arg = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(experimental::get_post_op_arg_type(i)));
-            add_3d_tensor_nhw_argument(idx, post_op_arg);
+            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2]));
         }
 
-        // dst buffer
-        add_3d_tensor_nhw_argument(idx, dst);
+        // Cross-plan padding (if _reinterpret_input_as_3d = true)
+        if(_reinterpret_input_as_3d && _has_pad_y)
+        {
+            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_lhs));
+        }
+
+        // Cross-plan padding (if reinterpret_output_as_3d = true)
+        if(_reinterpret_output_as_3d && _has_pad_y)
+        {
+            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_out));
+        }
 
         // Pass m, n and k at runtime as signed ints, to ensure results of any subractions they could be operand in, would still be signed.
         _kernel.setArg<cl_int>(idx++, _m);
diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h
index 297e681895..00cdb299ce 100644
--- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h
+++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h
@@ -96,6 +96,7 @@ private:
     bool         _use_dummy_work_items{ false };
     bool         _add_bias{ false };
     bool         _export_to_cl_image{ false };
+    bool         _has_pad_y{ false };
     signed int   _m{ 1 };
     signed int   _n{ 1 };
     signed int   _k{ 1 };
diff --git a/src/gpu/cl/operators/ClGemm.cpp b/src/gpu/cl/operators/ClGemm.cpp
index 766aec339e..88f6b79b56 100644
--- a/src/gpu/cl/operators/ClGemm.cpp
+++ b/src/gpu/cl/operators/ClGemm.cpp
@@ -452,6 +452,9 @@ Status ClGemm::validate_reshaped_only_rhs(const ITensorInfo *a, const ITensorInf
     kernel_info.has_pad_y = false;
     ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info));
 
+    kernel_info.has_pad_y = true;
+    ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info));
+
     return Status{};
 }
 
diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
index 860082f32b..cfd98bd8f0 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp
@@ -107,12 +107,6 @@ const auto act_values = framework::dataset::make("Activation",
     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 10.f),
 });
 
-/** Activation values to test */
-const auto act_identity = framework::dataset::make("Activation",
-{
-    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::IDENTITY),
-});
-
 /** M0 values to test - precommit */
 const auto m0_values_precommit = framework::dataset::make("M0", { 4 });
 
@@ -164,8 +158,8 @@ const auto boundary_handling_cases = combine(combine(combine(combine(combine(com
                                     framework::dataset::make("export_to_cl_image_rhs", {true, false})),
                                     // Only need to test F32 as F16 shares identical boundary handling logics
                                     framework::dataset::make("DataType", DataType::F32)),
-                                    framework::dataset::make("alpha", 1.0f )),
-                                    framework::dataset::make("beta", 0.0f )),
+                                    framework::dataset::make("alpha", -0.75f )),
+                                    framework::dataset::make("beta", -0.35f )),
                                     broadcast_bias_values),
                                     framework::dataset::make("Activation", ActivationLayerInfo()));
 
@@ -176,7 +170,7 @@ experimental::PostOpList<PostOpArgBroadcast> post_ops_1()
     experimental::PostOpList<PostOpArgBroadcast> post_ops{};
     post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F});
     post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>(
-        std::make_tuple(false, false, false),
+        std::make_tuple(true, true, false),   // If broadcast in dims 0, 1 and 2
         0,
         ConvertPolicy::SATURATE);
     post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
@@ -186,7 +180,7 @@ experimental::PostOpList<PostOpArgBroadcast> post_ops_2()
 {
     experimental::PostOpList<PostOpArgBroadcast> post_ops{};
     post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>(
-        std::make_tuple(false, false, false),
+        std::make_tuple(false, true, true),   // If broadcast in dims 0, 1 and 2
         1,
         ConvertPolicy::SATURATE);
     post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
@@ -195,18 +189,44 @@ experimental::PostOpList<PostOpArgBroadcast> post_ops_2()
 experimental::PostOpList<PostOpArgBroadcast> post_ops_3()
 {
     experimental::PostOpList<PostOpArgBroadcast> post_ops{};
+    post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
     post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>(
-        std::make_tuple(false, false, false),
+        std::make_tuple(false, false, true),  // If broadcast in dims 0, 1 and 2
         1,
         ConvertPolicy::SATURATE);
     return post_ops;
 }
-
+// To test that the output of the main op is the first parameter in prelu post op
+experimental::PostOpList<PostOpArgBroadcast> post_ops_4()
+{
+    experimental::PostOpList<PostOpArgBroadcast> post_ops{};
+    post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F});
+    post_ops.push_back_op<experimental::PostOpEltwisePRelu<PostOpArgBroadcast>>(
+        std::make_tuple(false, false, true),   // If true, broadcast in corresponding dim: 0, 1 or 2
+        0,
+        ConvertPolicy::SATURATE);
+    post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
+    return post_ops;
+}
+// To test that the output of the main op is the second parameter in prelu post op i.e. it is the alpha_param
+experimental::PostOpList<PostOpArgBroadcast> post_ops_5()
+{
+    experimental::PostOpList<PostOpArgBroadcast> post_ops{};
+    post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F});
+    post_ops.push_back_op<experimental::PostOpEltwisePRelu<PostOpArgBroadcast>>(
+        std::make_tuple(false, false, false),   // If true, broadcast in corresponding dim: 0, 1 or 2
+        1,
+        ConvertPolicy::SATURATE);
+    post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F});
+    return post_ops;
+}
 /** Different Post Op Lists */
 const auto post_op_lists = framework::dataset::make("post_op_lists", {
     post_ops_1(),
     post_ops_2(),
-    post_ops_3()
+    post_ops_3(),
+    post_ops_4(),
+    post_ops_5()
  } );
 
  bool is_post_op_list_valid(unsigned int m, unsigned int n, unsigned int k, unsigned int batch, DataType data_type, const experimental::PostOpList<ITensorInfo*>& post_ops)
@@ -446,7 +466,20 @@ TEST_CASE(BroadcastInBothXandYDims, framework::DatasetMode::ALL)
 
     ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
 }
+TEST_CASE(BroadcastInAllDims, framework::DatasetMode::ALL)
+{
+    const auto data_type = DataType::F32;
+    const unsigned int m = 22;
+    const unsigned int n = 16;
+    const unsigned int k = 15;
+    const unsigned int batch = 3;
+    TensorShape post_op_arg_shape(1, 1, 1);
+    TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type);
+    experimental::PostOpList<ITensorInfo*> post_ops{};
+    post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE);
 
+    ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS);
+}
 TEST_SUITE_END() // Valid
 TEST_SUITE_END() // ValidateFusedPostOps
 TEST_SUITE(Float)
@@ -600,7 +633,7 @@ FIXTURE_DATA_TEST_CASE(RunPrecommit3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixt
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("export_to_cl_image_rhs", {false, true})),
-                                                                   framework::dataset::make("has_pad_y", {false})),
+                                                                   framework::dataset::make("has_pad_y", {false, true})),
                                                                    framework::dataset::make("DataType", DataType::F32)),
                                                                    a_values),
                                                                    beta_values),
@@ -632,7 +665,7 @@ FIXTURE_DATA_TEST_CASE(RunNightly3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixtur
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("export_to_cl_image_rhs", {false, true})),
-                                                                   framework::dataset::make("has_pad_y", {false})),
+                                                                   framework::dataset::make("has_pad_y", {false, true})),
                                                                    framework::dataset::make("DataType", DataType::F32)),
                                                                    a_values),
                                                                    beta_values),
@@ -669,7 +702,7 @@ FIXTURE_DATA_TEST_CASE(RunPrecommit, CLGEMMMatrixMultiplyReshapedOnlyRHSWithPost
                                                                    a_values),
                                                                    beta_values),
                                                                    framework::dataset::make("broadcast_bias", { false } )),
-                                                                   act_identity),
+                                                                   act_values),
                                                                    post_op_lists)
                                                                    )
 {
@@ -766,7 +799,7 @@ FIXTURE_DATA_TEST_CASE(RunPrecommit3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixt
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("export_to_cl_image_rhs", true)),
-                                                                   framework::dataset::make("has_pad_y", {false})),
+                                                                   framework::dataset::make("has_pad_y", {false, true})),
                                                                    framework::dataset::make("DataType", DataType::F16)),
                                                                    a_values),
                                                                    beta_values),
@@ -798,7 +831,7 @@ FIXTURE_DATA_TEST_CASE(RunNightly3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixtur
                                                                    i_values_rhs),
                                                                    t_values_rhs),
                                                                    framework::dataset::make("export_to_cl_image_rhs", true)),
-                                                                   framework::dataset::make("has_pad_y", {false})),
+                                                                   framework::dataset::make("has_pad_y", {false, true})),
                                                                    framework::dataset::make("DataType", DataType::F16)),
                                                                    a_values),
                                                                    beta_values),
@@ -834,7 +867,7 @@ FIXTURE_DATA_TEST_CASE(RunPrecommit, CLGEMMMatrixMultiplyReshapedOnlyRHSWithPost
                                                                    a_values),
                                                                    beta_values),
                                                                    framework::dataset::make("broadcast_bias", { false } )),
-                                                                   act_identity),
+                                                                   act_values),
                                                                    post_op_lists)
                                                                    )
 {
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index 95dcd70104..8b748032fe 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -1551,7 +1551,6 @@ public:
         const TensorShape bias_shape(n,
                                      broadcast_bias ? 1 : m,
                                      broadcast_bias ? 1 : batch_size);
-
         auto post_ops_with_shapes = experimental::transform_post_op_list_arguments<PostOpArgBroadcast, TensorShape>(post_ops,
                                                                                                                     [ = ](auto broadcast)
         {