COMPMID-935 - Implementing Convolution with Winograd on OpenCL (part 4)

Implemented Winograd Output Transform (2x2,3x3) on OpenCL
Implemented CLWinogradConvolutionLayer on OpenCL

Change-Id: I6a113fc5f052ca07f878d2b800d2ab003f84af65
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125148
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

8 files changed, 498 insertions, 102 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 4b7fa8a3b3..9df2dcbacd 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -354,6 +354,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "winograd_filter_transform_2x2_3x3_nchw", "winograd.cl" },
     { "winograd_input_transform_2x2_3x3_stepz1_nchw", "winograd.cl" },
     { "winograd_input_transform_2x2_3x3_stepz2_nchw", "winograd.cl" },
+    { "winograd_output_transform_2x2_3x3_nchw", "winograd.cl" },
     { "YUYV422_to_IYUV_bt709", "color_convert.cl" },
     { "YUYV422_to_NV12_bt709", "color_convert.cl" },
     { "YUYV422_to_RGB888_bt709", "color_convert.cl" },
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index cba5eea437..a5b0acbe9c 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -162,6 +162,8 @@ __kernel void gemm_interleave4x4(TENSOR3D_DECLARATION(src),
  * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA
  * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2)
  * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2)
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -199,8 +201,18 @@ __kernel void gemm_mm_interleaved_transposed_f32_midgard(IMAGE_DECLARATION(src0)
 
     // src_addr_a = address of matrix A
     // src_addr_b = address of matrix B
-    __global float *src_addr_a = (__global float *)(src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes);
-    __global float *src_addr_b = (__global float *)(src1_ptr + z * src1_stride_z + x * src1_stride_y + src1_offset_first_element_in_bytes);
+    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
+    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
+    src1_addr_in_bytes += z * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    __global float *src_addr_a = (__global float *)(src0_ptr + src0_addr_in_bytes);
+    __global float *src_addr_b = (__global float *)(src1_ptr + src1_addr_in_bytes);
 
     // Compute end row address for matrix B
     __global float *src_end_addr_b = src_addr_b + COLS_B;
@@ -277,6 +289,9 @@ __kernel void gemm_mm_interleaved_transposed_f32_midgard(IMAGE_DECLARATION(src0)
  * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA
  * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2)
  * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2)
+ * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2)
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -314,8 +329,18 @@ __kernel void gemm_mm_interleaved_transposed_f32_bifrost(IMAGE_DECLARATION(src0)
 
     // src_addr_a = address of matrix A
     // src_addr_b = address of matrix B
-    __global float *src_addr_a = (__global float *)(src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes);
-    __global float *src_addr_b = (__global float *)(src1_ptr + z * src1_stride_z + x * src1_stride_y + src1_offset_first_element_in_bytes);
+    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
+    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
+    src1_addr_in_bytes += z * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    __global float *src_addr_a = (__global float *)(src0_ptr + src0_addr_in_bytes);
+    __global float *src_addr_b = (__global float *)(src1_ptr + src1_addr_in_bytes);
 
     // Compute end row address for matrix B
     __global float *src_end_addr_b = src_addr_b + COLS_B;
@@ -510,6 +535,8 @@ __kernel void gemm_mm_interleaved_transposed_f32_bifrost(IMAGE_DECLARATION(src0)
  * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA
  * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2)
  * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2)
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -547,8 +574,18 @@ __kernel void gemm_mm_interleaved_transposed_f16(IMAGE_DECLARATION(src0),
 
     // src_addr_a = address of matrix A
     // src_addr_b = address of matrix B
-    __global half *src_addr_a = (__global half *)(src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes);
-    __global half *src_addr_b = (__global half *)(src1_ptr + z * src1_stride_z + x * src1_stride_y + src1_offset_first_element_in_bytes);
+    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
+    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
+    src1_addr_in_bytes += z * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    __global half *src_addr_a = (__global half *)(src0_ptr + src0_addr_in_bytes);
+    __global half *src_addr_b = (__global half *)(src1_ptr + src1_addr_in_bytes);
 
     // Compute end row address for matrix B
     __global half *src_end_addr_b = src_addr_b + COLS_B;
@@ -627,8 +664,9 @@ __kernel void gemm_mm_interleaved_transposed_f16(IMAGE_DECLARATION(src0),
  * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA
  * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2)
  * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2)
- *
- * @note: ALPHA must be passed in 8 bit fixed point format
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
+ * @note:ALPHA must be passed in 8 bit fixed point format
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: QS8
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -666,8 +704,18 @@ __kernel void gemm_mm_interleaved_transposed_qs8(IMAGE_DECLARATION(src0),
 
     // src_addr_a = address of matrix A
     // src_addr_b = address of matrix B
-    __global char *src_addr_a = src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
-    __global char *src_addr_b = src1_ptr + z * src1_stride_z + x * src1_stride_y + src1_offset_first_element_in_bytes;
+    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
+    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
+    src1_addr_in_bytes += z * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    __global char *src_addr_a = (__global char *)(src0_ptr + src0_addr_in_bytes);
+    __global char *src_addr_b = (__global char *)(src1_ptr + src1_addr_in_bytes);
 
     // Compute end row address for matrix B
     __global char *src_end_addr_b = src_addr_b + COLS_B;
@@ -738,8 +786,9 @@ __kernel void gemm_mm_interleaved_transposed_qs8(IMAGE_DECLARATION(src0),
  * @note The number of columns of matrix B and the optional alpha's value need to be passed at compile time using -DCOLS_B and -DALPHA
  * @note The multiplication factor for the transposition width (mult_transpose1xW_width) must be passed at compile time using -DMULT_TRANSPOSE1XW_WIDTH (i.e. -DMULT_TRANSPOSE1XW_WIDTH=2)
  * @note The multiplication factor for the height of the 4x4 interleaved block must be passed at compile time using -DMULT_INTERLEAVE4X4_HEIGHT (i.e. -DMULT_INTERLEAVE4X4_HEIGHT=2)
- *
- * @note: ALPHA must be passed in 16 bit fixed point format
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
+ * @note:ALPHA must be passed in 16 bit fixed point format
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: QS16
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -777,8 +826,18 @@ __kernel void gemm_mm_interleaved_transposed_qs16(IMAGE_DECLARATION(src0),
 
     // src_addr_a = address of matrix A
     // src_addr_b = address of matrix B
-    __global short *src_addr_a = (__global short *)(src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes);
-    __global short *src_addr_b = (__global short *)(src1_ptr + z * src1_stride_z + x * src1_stride_y + src1_offset_first_element_in_bytes);
+    int src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;
+    int src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
+    src1_addr_in_bytes += z * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    __global short *src_addr_a = (__global short *)(src0_ptr + src0_addr_in_bytes);
+    __global short *src_addr_b = (__global short *)(src1_ptr + src1_addr_in_bytes);
 
     // Compute end row address for matrix B
     __global short *src_end_addr_b = src_addr_b + COLS_B;
@@ -845,6 +904,8 @@ __kernel void gemm_mm_interleaved_transposed_qs16(IMAGE_DECLARATION(src0),
  * @note The floating point data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
  * @note The number of elements processed along the x and y directions must be passed at compile time using -DNUM_ELEMS_PROCESSED_PER_THREAD_X and -DNUM_ELEMS_PROCESSED_PER_THREAD_Y
  * @note The number of matrix A columns and the optional alpha's value need to be passed at compile time using -DCOLS_A and -DALPHA
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16/F32
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -885,7 +946,13 @@ __kernel void gemm_mm_floating_point(IMAGE_DECLARATION(src0),
 
     // Add offset for batched GEMM
     src_addr.s0 += get_global_id(2) * src0_stride_z;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
     src_addr.s1 += get_global_id(2) * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
 
     int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(DATA_TYPE));
 
@@ -1013,6 +1080,8 @@ __kernel void gemm_mm_floating_point(IMAGE_DECLARATION(src0),
  * This kernel optimally uses -DNUM_ELEMS_PROCESSED_PER_THREAD_X=4.
  * @note The number of matrix A columns must be passed at compile time using -DCOLS_A.
  * @note The optional value of scalar alpha is passed at compile time using -DALPHA=alpha
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16/F32
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -1054,8 +1123,12 @@ __kernel void gemm_mm_floating_point_f32_bifrost(IMAGE_DECLARATION(src0),
     // Add offset for batched GEMM
     src_addr.s0 += get_global_id(2) * src0_stride_z;
 
-    // For convolution layer we do not want to slide the matrix B along Z
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
     src_addr.s1 += get_global_id(2) * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
 
     // Address boundary for matrix A
     int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(float));
@@ -1251,6 +1324,8 @@ __kernel void gemm_mm_floating_point_f32_bifrost(IMAGE_DECLARATION(src0),
  * This kernel optimally uses -DNUM_ELEMS_PROCESSED_PER_THREAD_X=2.
  * @note The number of matrix A columns must be passed at compile time using -DCOLS_A.
  * @note The optional value of scalar alpha is passed at compile time using -DALPHA=alpha if alpha!=1.0f.
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F16/F32
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -1293,8 +1368,12 @@ __kernel void gemm_mm_floating_point_f32_bifrost_1000(IMAGE_DECLARATION(src0),
     // Add offset for batched GEMM
     src_addr.s0 += get_global_id(2) * src0_stride_z;
 
-    // For convolution layer we do not want to slide the matrix B along Z
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
     src_addr.s1 += get_global_id(2) * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
 
     // Address boundary for the matrix A
     int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(float));
@@ -1460,6 +1539,8 @@ __kernel void gemm_mm_floating_point_f32_bifrost_1000(IMAGE_DECLARATION(src0),
  * @note The number matrix A columns, the number of elements processed per thread along the Y direction and the alpha's value need to be passed at compile time using -DCOLS_A, -DNUM_ELEMS_PROCESSED_PER_THREAD_Y and -DALPHA
  * @note The fixed point position need to be passed at compile time using -DFIXED_POINT_POSITION
  * @note The optional alpha value must be passed in 8 bit fixed point format using -DALPHA
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: QS8/QS16
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -1500,7 +1581,13 @@ __kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0),
 
     // Add offset for batched GEMM
     src_addr.s0 += get_global_id(2) * src0_stride_z;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
     src_addr.s1 += get_global_id(2) * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
 
     int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(char));
 
@@ -1636,6 +1723,8 @@ __kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0),
  * @note The number of matrix A columns, the number of elements processed per thread along the Y direction and the alpha's value need to be passed at compile time using -DCOLS_A, -DNUM_ELEMS_PROCESSED_PER_THREAD_Y and -DALPHA
  * @note The fixed point position need to be passed at compile time using -DFIXED_POINT_POSITION
  * @note The optional alpha value must be passed in 16 bit fixed point format using -DALPHA
+ * @note In case the matrix B has 3 dimensions and the matrix A more than 3, in order to avoid out-of-bounds reads, the number of channels of matrix B must be passed at compile time using MATRIX_B_DEPTH (i.e. -DMATRIX_B_DEPTH=16)
+ *       This case can happen when GEMM is used to perform the element-wise multiplication through a batched matrix multiplication (2D Winograd) and we have multiple inputs (i.e. a = [K, M, 16, Batches], b = [N, K, 16])
  *
  * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: QS8/QS16
  * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
@@ -1676,7 +1765,13 @@ __kernel void gemm_mm_qs16(IMAGE_DECLARATION(src0),
 
     // Add offset for batched GEMM
     src_addr.s0 += get_global_id(2) * src0_stride_z;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
     src_addr.s1 += get_global_id(2) * src1_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
 
     int end_row_vec_a = src_addr.s0 + (COLS_A * sizeof(short));
 
diff --git a/src/core/CL/cl_kernels/winograd.cl b/src/core/CL/cl_kernels/winograd.cl
index 238e21a18a..25c129d0aa 100644
--- a/src/core/CL/cl_kernels/winograd.cl
+++ b/src/core/CL/cl_kernels/winograd.cl
@@ -23,8 +23,102 @@
  */
 #include "helpers.h"
 
-#if defined(NUM_TILES_X)
+#if defined(NUM_CHANNELS)
+
+/** This OpenCL kernel performs Winograd filter transform 3x3 when the data format is NCHW and the output tile is 2x2
+ *
+ * @note The number of channels must be passed at compile time using -DNUM_CHANNELS: e.g. -DNUM_CHANNELS=64
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  src_step_w                        src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination tensor 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 tensor 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]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void winograd_filter_transform_2x2_3x3_nchw(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, NUM_CHANNELS);
+
+    const __global uchar *src_addr = tensor4D_offset(&src, 0, 0, 0, 0);
+
+    // Load the values from the input tensor
+    float3 w0 = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
+    float3 w1 = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
+    float3 w2 = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
+
+    // Transform the 3x3 tile in a 4x4 tile
+    float4 out0 = 0.0f;
+    float4 out1 = 0.0f;
+    float4 out2 = 0.0f;
+    float4 out3 = 0.0f;
+
+    // Row 0
+    out0.s0 = (w0.s0);
+    out0.s1 = (w0.s0 + w0.s1 + w0.s2) * 0.5f;
+    out0.s2 = (w0.s0 + w0.s2 - w0.s1) * 0.5f;
+    out0.s3 = (w0.s2);
+
+    // Row 1
+    out1.s0 = (w0.s0 + w1.s0 + w2.s0) * 0.5f;
+    out1.s1 = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) * 0.25f;
+    out1.s2 = (w0.s0 + w1.s0 + w2.s0 + w0.s2 + w1.s2 + w2.s2 - w0.s1 - w1.s1 - w2.s1) * 0.25f;
+    out1.s3 = (w0.s2 + w1.s2 + w2.s2) * 0.5f;
+
+    // Row 2
+    out2.s0 = (w0.s0 + w2.s0 - w1.s0) * 0.5f;
+    out2.s1 = (w0.s0 + w2.s0 + w0.s1 + w2.s1 + w0.s2 + w2.s2 - w1.s0 - w1.s1 - w1.s2) * 0.25f;
+    out2.s2 = (w0.s0 + w2.s0 + w1.s1 + w0.s2 + w2.s2 - w1.s0 - w0.s1 - w2.s1 - w1.s2) * 0.25f;
+    out2.s3 = (w0.s2 + w2.s2 - w1.s2) * 0.5f;
+
+    // Row 3
+    out3.s0 = (w2.s0);
+    out3.s1 = (w2.s0 + w2.s1 + w2.s2) * 0.5f;
+    out3.s2 = (w2.s0 + w2.s2 - w2.s1) * 0.5f;
+    out3.s3 = (w2.s2);
 
+    int z  = get_global_id(2);
+    int x0 = z / NUM_CHANNELS; // idx filter
+    int y0 = z % NUM_CHANNELS; // idx channel
+
+    // Get output address
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * dst_stride_x + y0 * dst_stride_y;
+
+    // Store the 16 values across the 16 channels
+    *(__global float *)(dst_addr + 0 * dst_stride_z)  = out0.s0;
+    *(__global float *)(dst_addr + 1 * dst_stride_z)  = out0.s1;
+    *(__global float *)(dst_addr + 2 * dst_stride_z)  = out0.s2;
+    *(__global float *)(dst_addr + 3 * dst_stride_z)  = out0.s3;
+    *(__global float *)(dst_addr + 4 * dst_stride_z)  = out1.s0;
+    *(__global float *)(dst_addr + 5 * dst_stride_z)  = out1.s1;
+    *(__global float *)(dst_addr + 6 * dst_stride_z)  = out1.s2;
+    *(__global float *)(dst_addr + 7 * dst_stride_z)  = out1.s3;
+    *(__global float *)(dst_addr + 8 * dst_stride_z)  = out2.s0;
+    *(__global float *)(dst_addr + 9 * dst_stride_z)  = out2.s1;
+    *(__global float *)(dst_addr + 10 * dst_stride_z) = out2.s2;
+    *(__global float *)(dst_addr + 11 * dst_stride_z) = out2.s3;
+    *(__global float *)(dst_addr + 12 * dst_stride_z) = out3.s0;
+    *(__global float *)(dst_addr + 13 * dst_stride_z) = out3.s1;
+    *(__global float *)(dst_addr + 14 * dst_stride_z) = out3.s2;
+    *(__global float *)(dst_addr + 15 * dst_stride_z) = out3.s3;
+}
+#endif // defined(NUM_CHANNELS)
+
+#if defined(NUM_TILES_X) && defined(PAD_LEFT) && defined(PAD_TOP)
 /** This OpenCL kernel computes the input transform when the kernel size is 3x3 and the output tile is 2x2
  *
  * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5).
@@ -205,13 +299,12 @@ __kernel void winograd_input_transform_2x2_3x3_stepz2_nchw(
     vstore2(out32, 0, (__global float *)(dst_addr + 14 * dst_stride_z));
     vstore2(out33, 0, (__global float *)(dst_addr + 15 * dst_stride_z));
 }
-#endif //defined(NUM_TILES_X)
+#endif // defined(NUM_TILES_X) && defined(PAD_LEFT) && defined(PAD_TOP)
 
-#if defined(NUM_CHANNELS)
-
-/** This OpenCL kernel performs Winograd filter transform 3x3 when the data format is NCHW and the output tile is 2x2
+#if defined(NUM_TILES_X)
+/** This OpenCL kernel performs Winograd output transform when the output tile is 2x2, the filter size 3x3 and the data format is NCHW
  *
- * @note The number of channels must be passed at compile time using -DNUM_CHANNELS: e.g. -DNUM_CHANNELS=64
+ * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16
  *
  * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
  * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
@@ -220,8 +313,6 @@ __kernel void winograd_input_transform_2x2_3x3_stepz2_nchw(
  * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
  * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes)
- * @param[in]  src_step_w                        src_stride_w * number of elements along W processed per workitem(in bytes)
  * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
  * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
  * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
@@ -232,72 +323,84 @@ __kernel void winograd_input_transform_2x2_3x3_stepz2_nchw(
  * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
  */
-__kernel void winograd_filter_transform_2x2_3x3_nchw(
-    TENSOR4D_DECLARATION(src),
-    TENSOR3D_DECLARATION(dst))
+__kernel void winograd_output_transform_2x2_3x3_nchw(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst)
+#if defined(HAS_BIAS)
+    ,
+    VECTOR_DECLARATION(bias)
+#endif // defined(HAS_BIAS)
+)
 {
-    Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, NUM_CHANNELS);
-
-    const __global uchar *src_addr = tensor4D_offset(&src, 0, 0, 0, 0);
-
-    // Load the values from the input tensor
-    float3 w0 = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float3 w1 = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float3 w2 = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
-
-    // Transform the 3x3 tile in a 4x4 tile
-    float4 out0 = 0.0f;
-    float4 out1 = 0.0f;
-    float4 out2 = 0.0f;
-    float4 out3 = 0.0f;
-
-    // Row 0
-    out0.s0 = (w0.s0);
-    out0.s1 = (w0.s0 + w0.s1 + w0.s2) * 0.5f;
-    out0.s2 = (w0.s0 + w0.s2 - w0.s1) * 0.5f;
-    out0.s3 = (w0.s2);
-
-    // Row 1
-    out1.s0 = (w0.s0 + w1.s0 + w2.s0) * 0.5f;
-    out1.s1 = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) * 0.25f;
-    out1.s2 = (w0.s0 + w1.s0 + w2.s0 + w0.s2 + w1.s2 + w2.s2 - w0.s1 - w1.s1 - w2.s1) * 0.25f;
-    out1.s3 = (w0.s2 + w1.s2 + w2.s2) * 0.5f;
-
-    // Row 2
-    out2.s0 = (w0.s0 + w2.s0 - w1.s0) * 0.5f;
-    out2.s1 = (w0.s0 + w2.s0 + w0.s1 + w2.s1 + w0.s2 + w2.s2 - w1.s0 - w1.s1 - w1.s2) * 0.25f;
-    out2.s2 = (w0.s0 + w2.s0 + w1.s1 + w0.s2 + w2.s2 - w1.s0 - w0.s1 - w2.s1 - w1.s2) * 0.25f;
-    out2.s3 = (w0.s2 + w2.s2 - w1.s2) * 0.5f;
-
-    // Row 3
-    out3.s0 = (w2.s0);
-    out3.s1 = (w2.s0 + w2.s1 + w2.s2) * 0.5f;
-    out3.s2 = (w2.s0 + w2.s2 - w2.s1) * 0.5f;
-    out3.s3 = (w2.s2);
-
-    int z  = get_global_id(2);
-    int x0 = z / NUM_CHANNELS; // idx filter
-    int y0 = z % NUM_CHANNELS; // idx channel
+    // Each thread stores a 2x2 tile
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+
+    const __global uchar *src_addr = tensor3D_offset(&src, 0, 0, 0);
+
+    // Load the values across the 16 channels to compose the 4x4 tile
+    float d00 = *((__global float *)(src_addr + 0 * src_stride_z));
+    float d01 = *((__global float *)(src_addr + 1 * src_stride_z));
+    float d02 = *((__global float *)(src_addr + 2 * src_stride_z));
+    float d03 = *((__global float *)(src_addr + 3 * src_stride_z));
+
+    float d10 = *((__global float *)(src_addr + 4 * src_stride_z));
+    float d11 = *((__global float *)(src_addr + 5 * src_stride_z));
+    float d12 = *((__global float *)(src_addr + 6 * src_stride_z));
+    float d13 = *((__global float *)(src_addr + 7 * src_stride_z));
+
+    float d20 = *((__global float *)(src_addr + 8 * src_stride_z));
+    float d21 = *((__global float *)(src_addr + 9 * src_stride_z));
+    float d22 = *((__global float *)(src_addr + 10 * src_stride_z));
+    float d23 = *((__global float *)(src_addr + 11 * src_stride_z));
+
+    float d30 = *((__global float *)(src_addr + 12 * src_stride_z));
+    float d31 = *((__global float *)(src_addr + 13 * src_stride_z));
+    float d32 = *((__global float *)(src_addr + 14 * src_stride_z));
+    float d33 = *((__global float *)(src_addr + 15 * src_stride_z));
+
+    // Compute the 2x2 output tile
+    float k0 = d01 + d11 + d21;
+    float k1 = d02 + d12 + d22;
+    float k2 = d11 - d21 - d31;
+    float k3 = d12 - d22 - d32;
+
+    // out00 = d00 + d10 + d20 + d01 + d11 + d21 + d02 + d12 + d22
+    // out01 = d01 + d11 + d21 - (d02 + d12 + d22) - (d03 + d13 + d23)
+    // out10 = d10 - d20 - d30 + (d11 - d21 - d31) + (d12 - d22 - d32)
+    // out11 = d11 - d21 - d31 - (d12 - d22 - d32) - (d13 - d23 - d33)
+
+    float out00 = d10;
+    float out01 = -d13;
+    float out10 = d10;
+    float out11 = -d13;
+
+    out00 += d00 + d20 + k0 + k1;
+    out01 += k0 - k1 - (d03 + d23);
+    out10 += -d20 - d30 + k2 + k3;
+    out11 += k2 - k3 + d23 + d33;
+
+    int y_in  = get_global_id(1);
+    int x_out = (y_in % NUM_TILES_X) * 2;
+    int y_out = (y_in / NUM_TILES_X) * 2;
+    int z_out = get_global_id(0);
+
+#if defined(HAS_BIAS)
+    // Add bias
+    Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
+
+    float b = (float) * ((__global float *)(vector_offset(&bias, z_out)));
+
+    out00 += (float)b;
+    out01 += (float)b;
+    out10 += (float)b;
+    out11 += (float)b;
+#endif // defined(HAS_BIAS)
 
     // Get output address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * dst_stride_x + y0 * dst_stride_y;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_out * dst_stride_x + y_out * dst_stride_y + z_out * dst_stride_z;
 
-    // Store the 16 values across the 16 channels
-    *(__global float *)(dst_addr + 0 * dst_stride_z)  = out0.s0;
-    *(__global float *)(dst_addr + 1 * dst_stride_z)  = out0.s1;
-    *(__global float *)(dst_addr + 2 * dst_stride_z)  = out0.s2;
-    *(__global float *)(dst_addr + 3 * dst_stride_z)  = out0.s3;
-    *(__global float *)(dst_addr + 4 * dst_stride_z)  = out1.s0;
-    *(__global float *)(dst_addr + 5 * dst_stride_z)  = out1.s1;
-    *(__global float *)(dst_addr + 6 * dst_stride_z)  = out1.s2;
-    *(__global float *)(dst_addr + 7 * dst_stride_z)  = out1.s3;
-    *(__global float *)(dst_addr + 8 * dst_stride_z)  = out2.s0;
-    *(__global float *)(dst_addr + 9 * dst_stride_z)  = out2.s1;
-    *(__global float *)(dst_addr + 10 * dst_stride_z) = out2.s2;
-    *(__global float *)(dst_addr + 11 * dst_stride_z) = out2.s3;
-    *(__global float *)(dst_addr + 12 * dst_stride_z) = out3.s0;
-    *(__global float *)(dst_addr + 13 * dst_stride_z) = out3.s1;
-    *(__global float *)(dst_addr + 14 * dst_stride_z) = out3.s2;
-    *(__global float *)(dst_addr + 15 * dst_stride_z) = out3.s3;
+    // Store the 2x2 output tile
+    vstore2((float2)(out00, out01), 0, (__global float *)(dst_addr + 0 * dst_stride_y));
+    vstore2((float2)(out10, out11), 0, (__global float *)(dst_addr + 1 * dst_stride_y));
 }
-#endif // defined(NUM_CHANNELS)
+#endif // defined(NUM_TILES_X)
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index 9c69800928..7b785bb8da 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -55,6 +55,7 @@ inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *i
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the matrix B must be <= 3");
 
     if(!is_interleaved_transposed)
     {
@@ -174,7 +175,7 @@ inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *inpu
 } // namespace
 
 CLGEMMMatrixMultiplyKernel::CLGEMMMatrixMultiplyKernel()
-    : _input0(nullptr), _input1(nullptr), _output(nullptr)
+    : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true)
 {
 }
 
@@ -192,9 +193,10 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
     // Perform validate step
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info));
 
-    _input0 = input0;
-    _input1 = input1;
-    _output = output;
+    _input0         = input0;
+    _input1         = input1;
+    _output         = output;
+    _slide_matrix_b = _input1->info()->num_dimensions() >= _input0->info()->num_dimensions();
 
     const DataType data_type = input0->info()->data_type();
     const int      fp_pos    = input0->info()->fixed_point_position();
@@ -257,6 +259,9 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
                                       "-DALPHA=" + float_to_string_with_full_precision(alpha));
     }
 
+    // Do not slide matrix B if _slide_matrix_b = false
+    build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
+
     std::string kernel_name;
     if(is_interleaved_transposed)
     {
@@ -365,7 +370,7 @@ void CLGEMMMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &que
         Window slice_b = slice;
         // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
         // This scenario can happen when the matrix multiplication is used to perform a convolution operation
-        if(_input1->info()->num_dimensions() < 3)
+        if(!_slide_matrix_b)
         {
             slice_b = slice_matrix_b;
         }
@@ -374,9 +379,9 @@ void CLGEMMMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &que
         add_2D_tensor_argument(idx, _input0, slice);
         add_2D_tensor_argument(idx, _input1, slice_b);
         add_2D_tensor_argument(idx, _output, slice);
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[3]));
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[3]));
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[3]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
         enqueue(queue, *this, slice, _lws_hint);
     }
     while(window.slide_window_slice_3D(slice));
diff --git a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
index 5489fde818..f69a39e4ad 100644
--- a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
@@ -76,15 +76,18 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     }
 
     AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    window_changed = window_changed || update_window_and_padding(win, input_access);
 
     // Configure window in case of configured output
     if(output->total_size() != 0)
     {
         AccessWindowTranspose output_access(output, 0, 0, num_elems_processed_per_iteration, 1, scale_x, 1.f / scale_x);
-        window_changed = window_changed || update_window_and_padding(win, output_access);
+        window_changed = window_changed || update_window_and_padding(win, input_access, output_access);
         output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), input->tensor_shape()));
     }
+    else
+    {
+        window_changed = window_changed || update_window_and_padding(win, input_access);
+    }
 
     // Collapse along the Z direction
     Window collapsed = win.collapse(win, Window::DimZ);
diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
index 3dbbe157b2..655b82bf66 100644
--- a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
@@ -76,7 +76,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
     AccessWindowStatic    output_access(output, 0, 0, output->dimension(0), output->dimension(1));
     window_changed = update_window_and_padding(win, input_access, output_access);
-    output_access.set_valid_region(win, input->valid_region());
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
 
     Window win_collapsed = win.collapse(win, Window::DimZ);
 
diff --git a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
index 72adb5f358..3b9350f9ba 100644
--- a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
@@ -44,11 +44,11 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_dims.width != 3 || kernel_dims.height != 3, "Winograd input transform only supports 3x3 kernels");
     ARM_COMPUTE_UNUSED(kernel_dims);
 
-    const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, conv_info, Size2D(3U, 3U));
-
     // Validate configured output
     if(output->total_size() != 0)
     {
+        const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, conv_info, kernel_dims);
+
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }
@@ -151,7 +151,8 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
 Status CLWinogradInputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(validate_arguments(input, output, conv_info, kernel_dims));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, conv_info, kernel_dims));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), conv_info, kernel_dims).first);
 
     return Status{};
 }
diff --git a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
new file mode 100644
index 0000000000..c9823275eb
--- /dev/null
+++ b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
@@ -0,0 +1,188 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "support/ToolchainSupport.h"
+
+#include <cmath>
+
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims, const Size2D &num_tiles)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area());
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_dims.width != 3 || kernel_dims.height != 3, "Only 3x3 kernels are supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(static_cast<unsigned int>(std::sqrt(input->dimension(2))) != 4, "Only 2x2 output tile is supported");
+    ARM_COMPUTE_UNUSED(kernel_dims);
+
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+    }
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input, output_convolved_dims, DataLayout::NCHW));
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    constexpr unsigned int num_elems_processed_per_iteration = 1;
+
+    Window win            = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    bool   window_changed = false;
+
+    AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration);
+    AccessWindowStatic    output_access(output, 0, 0, ceil_to_multiple(output->dimension(0), 2), ceil_to_multiple(output->dimension(1), 2));
+
+    if(bias != nullptr)
+    {
+        AccessWindowStatic bias_access(bias, 0, 0, bias->dimension(0), bias->dimension(1));
+        window_changed = update_window_and_padding(win, input_access, bias_access, output_access);
+    }
+    else
+    {
+        window_changed = update_window_and_padding(win, input_access, output_access);
+    }
+    output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+CLWinogradOutputTransformKernel::CLWinogradOutputTransformKernel()
+    : _input(nullptr), _bias(nullptr), _output(nullptr)
+{
+}
+
+void CLWinogradOutputTransformKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims,
+                                                const Size2D &num_tiles)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_UNUSED(kernel_dims);
+
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input->info(), output_convolved_dims, DataLayout::NCHW)));
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info(), kernel_dims, output_convolved_dims, num_tiles));
+
+    _input  = input;
+    _bias   = bias;
+    _output = output;
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option_if(_bias != nullptr, std::string("-DHAS_BIAS"));
+    build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles.width));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("winograd_output_transform_2x2_3x3_nchw", build_opts.options()));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure(win_config.second);
+
+    // Set config_id for enabling LWS tuning
+    _config_id = "winograd_output_transform_2x2_3x3";
+    _config_id += lower_string(string_from_data_type(input->info()->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(1));
+}
+
+Status CLWinogradOutputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims,
+                                                 const Size2D &num_tiles)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, (bias != nullptr ? bias->clone().get() : nullptr), output, kernel_dims, output_convolved_dims, num_tiles));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), output->clone().get()).first);
+
+    return Status{};
+}
+
+void CLWinogradOutputTransformKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    // Get initial windows
+    Window slice = window.first_slice_window_3D();
+    slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
+
+    // Setup output slice
+    Window slice_out(slice);
+    slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    if(_bias != nullptr)
+    {
+        unsigned int idx1 = 2 * num_arguments_per_3D_tensor();
+        Window       slice_biases;
+        slice_biases.use_tensor_dimensions(_bias->info()->tensor_shape());
+        add_1D_tensor_argument(idx1, _bias, slice_biases);
+    }
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice_out);
+        enqueue(queue, *this, slice, _lws_hint);
+    }
+    while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_out));
+}
\ No newline at end of file