COMPMID-2172: Fuse bias addition with CLGEMMMatrixMultiplyNativeKernel

Change-Id: I714b92ec001fc71172719b67fb66d490538b6948
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1399
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 274 insertions, 93 deletions

diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.h
index c611dc4c1f..79689a2894 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.h
@@ -48,8 +48,10 @@ public:
      *
      * @param[in]  input0    Input tensor for the LHS matrix. Data type supported: F32/F16. The number of dimensions for the LHS matrix must be less or equal than 4.
      * @param[in]  input1    Input tensor for the RHS matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3.
+     * @param[in]  input2    Input tensor containing the bias matrix. Data type supported: same as @p input0.
      * @param[out] output    Output tensor info. Data type supported: same as @p input0
      * @param[in]  alpha     Weight of the matrix product
+     * @param[in]  beta      Weight of the matrix bias
      * @param[in]  lhs_info  LHS matrix information used to retrieve the number of rows and accumulations to be processed by each thread. Only the following values are supported:
      *                       lhs_info.m0: 1,2,3,4,5,6,7,8
      *                       lhs_info.k0: 2,3,4,8,16
@@ -58,14 +60,17 @@ public:
      *                       rhs_info.k0: same of lhs_info.k0
      * @param[in]  gemm_info GEMM information used to retrieve the original dimensions of the input matrices
      */
-    void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+    void configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
+                   const GEMMRHSMatrixInfo &rhs_info,
                    const GEMMReshapeInfo &gemm_info);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMMatrixMultiplyNativeKernel
      *
      * @param[in] input0    Input tensor info for the LHS matrix. Data type supported: F32/F16. The number of dimensions for the LHS matrix must be less or equal than 4.
      * @param[in] input1    Input tensor info for the RHS matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3.
+     * @param[in] input2    Input tensor info containing the bias matrix. Data type supported: same as @p input0.
      * @param[in] output    Output tensor info. Data type supported: same as @p input0
      * @param[in] alpha     Weight of the matrix product
+     * @param[in] beta      Weight of the matrix bias
      * @param[in] lhs_info  LHS matrix information used to retrieve the number of rows and accumulations to be processed by each thread. Only the following values are supported:
      *                      lhs_info.m0: 1,2,3,4,5,6,7,8
      *                      lhs_info.k0: 2,3,4,8,16
@@ -76,7 +81,8 @@ public:
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+    static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
+                           const GEMMRHSMatrixInfo &rhs_info,
                            const GEMMReshapeInfo &gemm_info);
 
     // Inherited methods overridden:
@@ -85,11 +91,14 @@ public:
 private:
     const ICLTensor *_input0;
     const ICLTensor *_input1;
+    const ICLTensor *_input2;
     ICLTensor       *_output;
     bool             _slide_matrix_b;
     bool             _reinterpret_input_as_3d;
     bool             _reinterpret_output_as_3d;
     bool             _use_dummy_work_items;
+    bool             _add_bias;
+    bool             _broadcast_bias;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLGEMMMATRIXMULTIPLYNATIVEKERNEL_H__*/
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index 7ada14c774..854d0092d9 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -2122,35 +2122,49 @@ __kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs),
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
  *
- * @param[in]  lhs_ptr                           Pointer to the LHS reshaped matrix. Supported data type: F16/F32
- * @param[in]  lhs_stride_x                      Stride of the LHS reshaped matrix in X dimension (in bytes)
- * @param[in]  lhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  lhs_stride_y                      Stride of the LHS reshaped matrix in Y dimension (in bytes)
- * @param[in]  lhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
- * @param[in]  rhs_ptr                           Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
- * @param[in]  rhs_stride_x                      Stride of the RHS reshaped matrix in X dimension (in bytes)
- * @param[in]  rhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  rhs_stride_y                      Stride of the RHS reshaped matrix in Y dimension (in bytes)
- * @param[in]  rhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
- * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: same as @p lhs_ptr
- * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
- * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
- * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
- * @param[in]  lhs_stride_z                      Stride of the LHS reshaped matrix in Z dimension (in bytes)
- * @param[in]  rhs_stride_z                      Stride of the RHS reshaped matrix in Z dimension (in bytes)
- * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
- * @param[in]  lhs_cross_plane_pad               (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
- * @param[in]  dst_cross_plane_pad               (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
+ * @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                         lhs_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                         lhs_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 matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  rhs_stride_x                       Stride of the RHS matrix in X dimension (in bytes)
+ * @param[in]  rhs_step_x                         rhs_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  rhs_stride_y                       Stride of the RHS matrix in Y dimension (in bytes)
+ * @param[in]  rhs_step_y                         rhs_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 matrix
+ * @param[in]  bias_ptr                           (Optional)Pointer to the bias reshaped matrix. 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 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)
  */
 __kernel void gemm_mm_native(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)
                              ,
@@ -2192,8 +2206,8 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
     rhs_offset += z * rhs_stride_z;
 #endif // defined(MATRIX_B_DEPTH)
 
-    REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
-    REPEAT_VAR_INIT_TO_CONST(16, uint, zrhs, 0);
+    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
@@ -2211,7 +2225,7 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
 #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;
+    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)
@@ -2229,7 +2243,7 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
         LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
 
         // Load values from RHS matrix
-        LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zrhs);
+        LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zero);
 
         RHS_VFMA_M0xN0(0, a, b0, c);
         RHS_VFMA_M0xN0(1, a, b1, c);
@@ -2305,7 +2319,7 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
 
     __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * dst_stride_y);
 
-    REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
+    REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
 
 #if defined(REINTERPRET_OUTPUT_AS_3D)
     // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
@@ -2323,11 +2337,40 @@ __kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
 #endif // defined(REINTERPRET_OUTPUT_AS_3D)
 
     // Multiply by the weight of matrix-matrix product and store the result
-    // Multiply by the weight of matrix-matrix product and store the result
 #if defined(ALPHA)
     SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
 #endif // defined(ALPHA)
 
+    // 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(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
+
+#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 + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
+                                    2) * bias_stride_z;
+
+    LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
+
+#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)
+
     // Store output block
     STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout);
 
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp
index a3de6e0853..0b9359e610 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp
@@ -51,7 +51,8 @@ namespace
 {
 using ElementsProcessed = Steps;
 
-Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
+                          const GEMMRHSMatrixInfo &rhs_info,
                           const GEMMReshapeInfo &gemm_info)
 {
     ARM_COMPUTE_UNUSED(alpha);
@@ -85,6 +86,22 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1,
         ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != static_cast<unsigned int>(m));
     }
 
+    if(input2 != nullptr && !(helpers::float_ops::is_zero(beta)))
+    {
+        const int input2_dim0 = static_cast<int>(input2->dimension(0));
+        const int input2_dim1 = static_cast<int>(input2->dimension(1));
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input2, input1);
+        if(gemm_info.broadcast_bias())
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim1 != 1 || input2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim0 != n || input2_dim1 != m), "Incorrect dimension of bias matrix");
+        }
+    }
+
     if(output->total_size() != 0)
     {
         const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
@@ -95,7 +112,8 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1,
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,
+                                                        const GEMMRHSMatrixInfo &rhs_info,
                                                         const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed)
 {
     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
@@ -150,8 +168,24 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
                                      ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x),
                                      output->dimension(1) + bottom_pad);
 
-    window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
-                     update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor
+    if(input2 != nullptr)
+    {
+        const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
+
+        const int bias_processed_per_iteration_y = gemm_info.broadcast_bias() ? 1 : num_elems_processed_per_iteration_y;
+
+        AccessWindowStatic input2_access(input2, 0, 0,
+                                         ceil_to_multiple(input2->dimension(0), bias_processed_per_iteration_x),
+                                         ceil_to_multiple(input2->dimension(1), bias_processed_per_iteration_y));
+
+        window_changed = update_window_and_padding(win, input0_access, input1_access, input2_access) || // window used by the execute_window_loop
+                         update_window_and_padding(win_out, output_access);                             // window used to update the padding requirements of output tensor
+    }
+    else
+    {
+        window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
+                         update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor
+    }
 
     output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape()));
 
@@ -167,23 +201,28 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
 } // namespace
 
 CLGEMMMatrixMultiplyNativeKernel::CLGEMMMatrixMultiplyNativeKernel()
-    : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false)
+    : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false),
+      _add_bias(false), _broadcast_bias(false)
 {
 }
 
-void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, const GEMMLHSMatrixInfo &lhs_info,
+void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta,
+                                                 const GEMMLHSMatrixInfo &lhs_info,
                                                  const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
 
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), alpha, lhs_info, rhs_info, gemm_info));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), (input2 != nullptr ? input2->info() : nullptr), output->info(), alpha, beta, lhs_info, rhs_info, gemm_info));
 
     _input0                   = input0;
     _input1                   = input1;
+    _input2                   = helpers::float_ops::is_zero(beta) ? nullptr : input2;
     _output                   = output;
     _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d();
     _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
+    _add_bias                 = _input2 != nullptr;
+    _broadcast_bias           = gemm_info.broadcast_bias();
 
     // In case both input and output have to be reinterpreted as 3D tensors,
     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
@@ -200,7 +239,7 @@ void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const
     ElementsProcessed num_elements_processed{};
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
+    auto win_config = validate_and_configure_window(input0->info(), input1->info(), input2 != nullptr ? input2->info() : nullptr, output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure_internal(win_config.second);
 
@@ -208,6 +247,9 @@ void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
     build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
+    build_opts.add_option_if(_input2 != 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(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
     build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
     build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
@@ -229,6 +271,8 @@ void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const
     // Set config_id for enabling LWS tuning
     _config_id = kernel_name;
     _config_id += "_";
+    _config_id += (_add_bias ? "add_bias_" : "");
+    _config_id += (_broadcast_bias ? "broadcast_bias_" : "");
     _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
     _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
     _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
@@ -248,13 +292,15 @@ void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const
     _config_id += support::cpp11::to_string(rhs_info.k0);
 }
 
-Status CLGEMMMatrixMultiplyNativeKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info,
+Status CLGEMMMatrixMultiplyNativeKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta,
+                                                  const GEMMLHSMatrixInfo &lhs_info,
                                                   const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
 {
     ElementsProcessed num_elements_processed{};
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, alpha, lhs_info, rhs_info, gemm_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
                                                               input1->clone().get(),
+                                                              input2 != nullptr ? input2->clone().get() : nullptr,
                                                               output->clone().get(),
                                                               lhs_info,
                                                               rhs_info,
@@ -285,7 +331,15 @@ void CLGEMMMatrixMultiplyNativeKernel::run(const Window &window, cl::CommandQueu
     if(_reinterpret_input_as_3d)
     {
         // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
-        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3;
+        unsigned int idx0;
+        if(_add_bias)
+        {
+            idx0 = 4 * num_arguments_per_2D_tensor() + 4;
+        }
+        else
+        {
+            idx0 = 3 * num_arguments_per_2D_tensor() + 3;
+        }
         const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom;
         _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
     }
@@ -293,7 +347,15 @@ void CLGEMMMatrixMultiplyNativeKernel::run(const Window &window, cl::CommandQueu
     if(_reinterpret_output_as_3d)
     {
         // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
-        const unsigned int idx0                  = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
+        unsigned int idx0;
+        if(_add_bias)
+        {
+            idx0 = 4 * num_arguments_per_2D_tensor() + 4 + (_reinterpret_input_as_3d ? 1 : 0);
+        }
+        else
+        {
+            idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
+        }
         const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
         _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
     }
@@ -311,9 +373,17 @@ void CLGEMMMatrixMultiplyNativeKernel::run(const Window &window, cl::CommandQueu
         unsigned int idx = 0;
         add_2D_tensor_argument(idx, _input0, slice);
         add_2D_tensor_argument(idx, _input1, slice_b);
+        if(_add_bias)
+        {
+            add_2D_tensor_argument(idx, _input2, slice);
+        }
         add_2D_tensor_argument(idx, _output, slice);
         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
+        if(_add_bias)
+        {
+            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input2->info()->strides_in_bytes()[2]));
+        }
         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
         enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
     }
diff --git a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
index c7c390353a..b0d1fd2ad1 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp
@@ -68,6 +68,9 @@ constexpr float         tolerance_num_f16 = 0.02f;
 /** Alpha values to test - Precommit */
 const auto a_values = framework::dataset::make("alpha", {1.0f, -0.75f} );
 
+/** Beta values to test - Precommit */
+const auto beta_values = framework::dataset::make("beta", {-0.75f, 0.0f} );
+
 /** M values to test */
 const auto m_values = framework::dataset::make("M", 37);
 
@@ -107,8 +110,11 @@ const auto n0_values_nightly = framework::dataset::make("N0", { 2, 3, 4, 8 });
 /** K0 values to test - Nightly */
 const auto k0_values_nightly = framework::dataset::make("K0", { 2, 3, 4, 8 });
 
+/** Broadcast bias from vector to matrix */
+const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", {false, true} );
+
 /** Configuration test */
-void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, DataType data_type)
+void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, bool broadcast_bias, DataType data_type)
 {
     const unsigned int M = m_value;
     const unsigned int N = n_value;
@@ -122,27 +128,30 @@ void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned
     rhs_info.n0         = n0_value;
     rhs_info.k0         = k0_value;
 
-    GEMMReshapeInfo gemm_info(M, N, K);
+    GEMMReshapeInfo gemm_info(M, N, K, false, false, 0, false, broadcast_bias);
 
     const TensorShape lhs_shape(K, M, b_value);
     const TensorShape rhs_shape(N, K, b_value);
-
+    const TensorShape bias_shape(N,
+                                 broadcast_bias? 1 : M,
+                                 broadcast_bias? 1 : b_value);
     const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape, 1, data_type),
                                                    TensorInfo(rhs_shape, 1, data_type),
                                                    gemm_info);
-
     // Create tensors
-    CLTensor lhs          = create_tensor<CLTensor>(lhs_shape, data_type);
+    CLTensor lhs  = create_tensor<CLTensor>(lhs_shape, data_type);
     CLTensor rhs  = create_tensor<CLTensor>(rhs_shape, data_type);
-    CLTensor dst          = create_tensor<CLTensor>(dst_shape, data_type);
+    CLTensor bias = create_tensor<CLTensor>(bias_shape, data_type);
+    CLTensor dst  = create_tensor<CLTensor>(dst_shape, data_type);
 
     ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
 
     // Create and configure function
     CLGEMMMatrixMultiplyNative gemm;
-    gemm.configure(&lhs, &rhs, &dst, 1.0f, lhs_info, rhs_info, gemm_info);
+    gemm.configure(&lhs, &rhs, &bias, &dst, 1.0f, 1.0f, lhs_info, rhs_info, gemm_info);
 }
 } // namespace
 
@@ -150,7 +159,7 @@ TEST_SUITE(CL)
 TEST_SUITE(GEMMMatrixMultiplyNative)
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -158,13 +167,14 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi
                                                                    m0_values_precommit),
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
-m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value)
+                                                                   broadcast_bias_values),
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias)
 {
-    validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, DataType::F32);
+    validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias, DataType::F32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -173,14 +183,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<float>, frame
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -189,14 +201,16 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<float>, frame
                                                                    n0_values_nightly),
                                                                    k0_values_nightly),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_w_values,
                                                                    m_h_values),
                                                                    n_values),
@@ -206,14 +220,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<float>, f
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_w_values,
                                                                    m_h_values),
                                                                    n_values),
@@ -223,7 +238,8 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<float>, f
                                                                    n0_values_nightly),
                                                                    k0_values_nightly),
                                                                    framework::dataset::make("DataType", DataType::F32)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
@@ -232,7 +248,7 @@ TEST_SUITE_END() // FP32
 
 TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -241,14 +257,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<half>, framew
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_values,
                                                                    n_values),
                                                                    k_values),
@@ -257,14 +275,16 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<half>, framew
                                                                    n0_values_nightly),
                                                                    k0_values_nightly),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<half>, framework::DatasetMode::ALL,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_w_values,
                                                                    m_h_values),
                                                                    n_values),
@@ -274,14 +294,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<half>, fr
                                                                    n0_values_precommit),
                                                                    k0_values_precommit),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
 }
 
 FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<half>, framework::DatasetMode::NIGHTLY,
-                combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
                                                                    m_w_values,
                                                                    m_h_values),
                                                                    n_values),
@@ -291,7 +312,8 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<half>, fr
                                                                    n0_values_nightly),
                                                                    k0_values_nightly),
                                                                    framework::dataset::make("DataType", DataType::F16)),
-                                                                   a_values))
+                                                                   a_values),
+                                                                   beta_values))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index fcb41bb0ba..b721d841f7 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -684,7 +684,7 @@ class GEMMMatrixMultiplyNativeValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, DataType data_type, float alpha)
+    void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, DataType data_type, float alpha, float beta, bool broadcast_bias)
     {
         GEMMLHSMatrixInfo lhs_info;
         lhs_info.m0 = m0;
@@ -697,9 +697,12 @@ public:
         // Set the tensor shapes for LHS and RHS matrices
         const TensorShape lhs_shape(k, m, batch_size);
         const TensorShape rhs_shape(n, k, batch_size);
+        const TensorShape bias_shape(n,
+                                     broadcast_bias ? 1 : m,
+                                     broadcast_bias ? 1 : batch_size);
 
-        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type, alpha);
-        _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha);
+        _target    = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias);
+        _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias);
     }
 
 protected:
@@ -714,11 +717,13 @@ protected:
         library->fill_borders_with_garbage(tensor, distribution_inf, i);
     }
 
-    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, DataType data_type, float alpha)
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                              DataType data_type, float alpha, float beta, bool broadcast_bias)
     {
         // Create tensors
-        TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
-        TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType lhs  = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs  = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1);
         TensorType dst;
 
         const unsigned int M = lhs_shape[1];
@@ -727,23 +732,27 @@ protected:
 
         // Create and configure function
         GEMMFunctionType gemm;
-        gemm.configure(&lhs, &rhs, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K));
+        gemm.configure(&lhs, &rhs, &bias, &dst, alpha, beta, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, 0, false, broadcast_bias));
 
         ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Allocate tensors
         lhs.allocator()->allocate();
         rhs.allocator()->allocate();
+        bias.allocator()->allocate();
         dst.allocator()->allocate();
 
         ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Fill tensors
         fill(AccessorType(lhs), 0);
         fill(AccessorType(rhs), 1);
+        fill(AccessorType(bias), 2);
 
         // Compute GEMM
         gemm.run();
@@ -751,7 +760,7 @@ protected:
         return dst;
     }
 
-    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha)
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, bool broadcast_bias)
     {
         TensorShape dst_shape = lhs_shape;
         dst_shape[0]          = rhs_shape[0];
@@ -760,13 +769,27 @@ protected:
         // Create reference
         SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
         SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
-        SimpleTensor<T> c{ dst_shape, data_type, 1 };
+        SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+
+        const int n          = rhs_shape[0];
+        const int m          = lhs_shape[1];
+        const int batch_size = lhs_shape[2];
 
         // Fill reference
         fill(lhs, 0);
         fill(rhs, 1);
+        fill(bias, 2);
 
-        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+        if(broadcast_bias)
+        {
+            // In case of broadcast, we need simply copy the first into the following "M" ones
+            for(int i = 1; i < m * batch_size; i++)
+            {
+                memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
+            }
+        }
+
+        return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
     }
 
     TensorType      _target{};
@@ -778,7 +801,7 @@ class GEMMMatrixMultiplyNative3DValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, DataType data_type, float alpha)
+    void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, DataType data_type, float alpha, float beta)
     {
         GEMMLHSMatrixInfo lhs_info;
         lhs_info.m0 = m0;
@@ -794,9 +817,10 @@ public:
         // Set the tensor shapes for LHS and RHS matrices
         const TensorShape lhs_shape(k, m, batch_size);
         const TensorShape rhs_shape(n, k, batch_size);
+        const TensorShape bias_shape(n, 1, 1);
 
-        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type, alpha, m_h);
-        _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, m_h);
+        _target    = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h);
+        _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h);
     }
 
 protected:
@@ -807,13 +831,13 @@ protected:
         library->fill(tensor, distribution, i);
     }
 
-    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, DataType data_type, float alpha,
-                              unsigned int m_h)
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                              DataType data_type, float alpha, float beta, unsigned int m_h)
     {
         // Create tensors
-        TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1);
-        TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1);
-        TensorType rhs_reshaped;
+        TensorType lhs  = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs  = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1);
         TensorType dst;
 
         const unsigned int M = lhs_shape[1];
@@ -824,25 +848,27 @@ protected:
 
         // Create and configure function
         GEMMFunctionType gemm;
-        gemm.configure(&lhs, &rhs, &dst, alpha, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h));
+        gemm.configure(&lhs, &rhs, &bias, &dst, alpha, beta, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h, false, true));
 
         ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Allocate tensors
         lhs.allocator()->allocate();
         rhs.allocator()->allocate();
-        rhs_reshaped.allocator()->allocate();
+        bias.allocator()->allocate();
         dst.allocator()->allocate();
 
         ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
-        ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Fill tensors
         fill(AccessorType(lhs), 0);
         fill(AccessorType(rhs), 1);
+        fill(AccessorType(bias), 2);
 
         // Compute GEMM
         gemm.run();
@@ -850,7 +876,7 @@ protected:
         return dst;
     }
 
-    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, unsigned int m_h)
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, unsigned int m_h)
     {
         TensorShape dst_shape = lhs_shape;
         dst_shape.set(0, rhs_shape[0]);
@@ -861,13 +887,24 @@ protected:
         // Create reference
         SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
         SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
-        SimpleTensor<T> c{ dst_shape, data_type, 1 };
+        SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+
+        const int n          = rhs_shape[0];
+        const int m          = lhs_shape[1];
+        const int batch_size = lhs_shape[2];
 
         // Fill reference
         fill(lhs, 0);
         fill(rhs, 1);
+        fill(bias, 2);
 
-        return reference::gemm<T>(lhs, rhs, c, alpha, 0.0f);
+        // In case of broadcast, we need simply copy the first into the following "M" ones
+        for(int i = 1; i < m * batch_size; i++)
+        {
+            memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
+        }
+
+        return reference::gemm<T>(lhs, rhs, bias, alpha, beta);
     }
 
     TensorType      _target{};