COMPMID-807: NHWC support in CLDirectConvolution.

Change-Id: I8738aca2cc0104e4c4d7c9605762ab59fce10a33
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/137333
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

7 files changed, 735 insertions, 125 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index ba6a629e0d..475352456c 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -213,10 +213,13 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "derivative", "derivative.cl" },
     { "dilate", "dilate.cl" },
     { "direct_convolution1x1", "direct_convolution1x1.cl" },
+    { "direct_convolution1x1_nhwc", "direct_convolution1x1.cl" },
     { "direct_convolution1x1_f32_bifrost", "direct_convolution1x1.cl" },
     { "direct_convolution3x3", "direct_convolution3x3.cl" },
+    { "direct_convolution3x3_nhwc", "direct_convolution3x3.cl" },
     { "direct_convolution3x3_f32_bifrost", "direct_convolution3x3.cl" },
     { "direct_convolution5x5", "direct_convolution5x5.cl" },
+    { "direct_convolution5x5_nhwc", "direct_convolution5x5.cl" },
     { "direct_convolution5x5_f32_bifrost", "direct_convolution5x5.cl" },
     { "direct_convolution_1x1_3x3_5x5_quantized", "direct_convolution_1x1_3x3_5x5_quantized.cl" },
     { "erode", "erode.cl" },
diff --git a/src/core/CL/cl_kernels/direct_convolution1x1.cl b/src/core/CL/cl_kernels/direct_convolution1x1.cl
index 7a308c99e2..cceeb0f9c4 100644
--- a/src/core/CL/cl_kernels/direct_convolution1x1.cl
+++ b/src/core/CL/cl_kernels/direct_convolution1x1.cl
@@ -31,6 +31,122 @@
 
 #if defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
 
+#if defined(DATA_LAYOUT_NHWC)
+
+#define PTR_TO_VALUE(PTR, DATA_TYPE) *((__global DATA_TYPE *)(PTR))
+
+/** This kernel performs a direct convolution to convolve the low three dimensions of a tensor with data layout NHWC
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
+ * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1
+ * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
+ * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16/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_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 Z processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                          Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                            dst_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
+ * @param[in]  weights_ptr                           Pointer to the weights tensor. Supported data types: same as @p src_ptr
+ * @param[in]  weights_stride_x                      Stride of the weights tensor in X dimension (in bytes)
+ * @param[in]  weights_step_x                        weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  weights_stride_y                      Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in]  weights_step_y                        weights_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in]  weights_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ */
+__kernel void direct_convolution1x1_nhwc(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#ifdef HAS_BIAS
+    VECTOR_DECLARATION(biases),
+#endif /* defined(HAS_BIAS) */
+    unsigned int weights_stride_w)
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef HAS_BIAS
+    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif /* defined(HAS_BIAS) */
+
+    VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)
+    values        = 0;
+    const int id0 = get_global_id(0);
+    const int id1 = get_global_id(1);
+    const int id2 = get_global_id(2);
+    weights.ptr += id0 * weights_stride_w;
+    __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0) - src_stride_x * id0 + id2 * STRIDE_Y * (int)src_stride_z;
+
+    for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
+    {
+        DATA_TYPE weight = *(__global DATA_TYPE *)weights.ptr;
+#if STRIDE_X == 1
+        VEC_DATA_TYPE(DATA_TYPE, 8)
+        col0 = (VEC_DATA_TYPE(DATA_TYPE, 8))(
+                   PTR_TO_VALUE(src_addr + 0 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 1 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 2 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 3 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 4 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 5 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 6 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 7 * src_stride_y, DATA_TYPE));
+#elif STRIDE_X == 2 /* STRIDE_X == 1 */
+        VEC_DATA_TYPE(DATA_TYPE, 8)
+        col0 = (VEC_DATA_TYPE(DATA_TYPE, 8))(
+                   PTR_TO_VALUE(src_addr + 0 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 2 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 4 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 6 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 8 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 10 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 12 * src_stride_y, DATA_TYPE),
+                   PTR_TO_VALUE(src_addr + 14 * src_stride_y, DATA_TYPE));
+#else               /* STRIDE_X not equals 1 or 2 */
+#error "STRIDE_X larger than 2 is not supported"
+#endif /* STRIDE_X == 2 */
+        values = ADD_OP(values, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))weight, col0));
+
+        src_addr += src_stride_x;
+        weights.ptr += weights_stride_x;
+    }
+
+#ifdef HAS_BIAS
+    values = ADD_OP(values, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, id0))));
+#endif /* defined(HAS_BIAS) */
+
+    *((__global DATA_TYPE *)dst.ptr)                      = values.s0;
+    *((__global DATA_TYPE *)(dst.ptr + 1 * dst_stride_y)) = values.s1;
+    *((__global DATA_TYPE *)(dst.ptr + 2 * dst_stride_y)) = values.s2;
+    *((__global DATA_TYPE *)(dst.ptr + 3 * dst_stride_y)) = values.s3;
+    *((__global DATA_TYPE *)(dst.ptr + 4 * dst_stride_y)) = values.s4;
+    *((__global DATA_TYPE *)(dst.ptr + 5 * dst_stride_y)) = values.s5;
+    *((__global DATA_TYPE *)(dst.ptr + 6 * dst_stride_y)) = values.s6;
+    *((__global DATA_TYPE *)(dst.ptr + 7 * dst_stride_y)) = values.s7;
+}
+#endif // defined(DATA_LAYOUT_NHWC)
+
 #if STRIDE_X == 3
 #define INPUT_PIXEL_STR(data_size) extract_input_stride3_##data_size
 #define INPUT_PIXEL(data_size) INPUT_PIXEL_STR(data_size)
@@ -46,7 +162,7 @@
  *
  * @param[in] input_pixel Pointer to the first pixel.
  *
- * @return extracted input pixels.
+ * @return extracted input values.
  */
 inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_pixel)
 {
@@ -57,7 +173,7 @@ inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYP
  *
  * @param[in] input_pixel Pointer to the first pixel.
  *
- * @return extracted input pixels.
+ * @return extracted input values.
  */
 inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_pixel)
 {
@@ -70,7 +186,7 @@ inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYP
  *
  * @param[in] input_pixel Pointer to the first pixel.
  *
- * @return extracted input pixels.
+ * @return extracted input values.
  */
 inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_32(__global const DATA_TYPE *input_pixel)
 {
@@ -89,7 +205,7 @@ inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_32(__global const DATA_
  *
  * @param[in] input_pixel Pointer to the first pixel.
  *
- * @return extracted input pixels.
+ * @return extracted input values.
  */
 inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_16(__global const DATA_TYPE *input_pixel)
 {
@@ -106,7 +222,7 @@ inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_16(__global const DATA_
  *
  * @param[in] input_pixel Pointer to the first pixel.
  *
- * @return extracted input pixels.
+ * @return extracted input values.
  */
 inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_8(__global const DATA_TYPE *input_pixel)
 {
@@ -173,27 +289,26 @@ __kernel void direct_convolution1x1(
 #endif /* defined(HAS_BIAS) */
 
     VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)
-    pixels = 0;
+    values = 0;
 
     const uint z_index = get_global_id(2);
 
     weights.ptr += z_index * weights_stride_w;
-
     for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
     {
         DATA_TYPE weight = *(__global DATA_TYPE *)weights.ptr;
         VEC_DATA_TYPE(DATA_TYPE, 8)
         input_pixel = INPUT_PIXEL(DATA_SIZE)((__global DATA_TYPE *)src.ptr);
-        pixels      = ADD_OP(pixels, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))weight, input_pixel));
+        values      = ADD_OP(values, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))weight, input_pixel));
         src.ptr += src_stride_z;
         weights.ptr += weights_stride_z;
     }
 
 #ifdef HAS_BIAS
-    pixels = ADD_OP(pixels, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, z_index))));
+    values = ADD_OP(values, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, z_index))));
 #endif /* defined(HAS_BIAS) */
 
-    vstore8(CONVERT_SAT(pixels, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)dst.ptr);
+    vstore8(CONVERT_SAT(values, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)dst.ptr);
 }
 #endif // defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
 
@@ -314,4 +429,4 @@ __kernel void direct_convolution1x1_f32_bifrost(
     vstore4(acc2, 0, (__global float *)(dst.ptr + 2 * dst_stride_y));
     vstore4(acc3, 0, (__global float *)(dst.ptr + 3 * dst_stride_y));
 }
-#endif // defined(WEIGHTS_DEPTH)
\ No newline at end of file
+#endif // defined(WEIGHTS_DEPTH)
diff --git a/src/core/CL/cl_kernels/direct_convolution3x3.cl b/src/core/CL/cl_kernels/direct_convolution3x3.cl
index 824306f2ba..08d25f6741 100644
--- a/src/core/CL/cl_kernels/direct_convolution3x3.cl
+++ b/src/core/CL/cl_kernels/direct_convolution3x3.cl
@@ -66,6 +66,185 @@
         acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1), (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2)); \
     })
 
+#if defined(DATA_LAYOUT_NHWC)
+
+#define PTR_TO_VALUE(PTR, DATA_TYPE) *((__global DATA_TYPE *)(PTR))
+
+#if STRIDE_X == 1
+#define CONVOLUTION1x3_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x3_STRIDE_NHWC_STRIDE1(acc, row_ptr, weights_ptr)
+#elif STRIDE_X == 2 /* STRIDE_X == 1 */
+#define CONVOLUTION1x3_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x3_STRIDE_NHWC_STRIDE2(acc, row_ptr, weights_ptr)
+#else /* STRIDE_X not equals 1 or 2 */
+#error "STRIDE_X larger than 2 is not supported"
+#endif /* STRIDE_X == 2 */
+
+#define CONVOLUTION1x3_STRIDE_NHWC_STRIDE1(acc, row_ptr, weights_ptr)                                                                      \
+    {                                                                                                                                      \
+        VEC_DATA_TYPE(DATA_TYPE, 8)                                                                                                        \
+        src0 = (VEC_DATA_TYPE(DATA_TYPE, 8))(                                                                                              \
+                PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE),                                                                           \
+                PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE),                                                                           \
+                PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE),                                                                           \
+                PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE),                                                                           \
+                PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE),                                                                           \
+                PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE),                                                                           \
+                PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE),                                                                           \
+                PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE));                                                                          \
+        VEC_DATA_TYPE(DATA_TYPE, 2)                                                                                                        \
+        src1 = (VEC_DATA_TYPE(DATA_TYPE, 2))(                                                                                              \
+                PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE),                                                                           \
+                PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE));                                                                          \
+        VEC_DATA_TYPE(DATA_TYPE, 3)                                                                                                        \
+        weights = (VEC_DATA_TYPE(DATA_TYPE, 3))(                                                                                           \
+                  PTR_TO_VALUE((weights_ptr) + 0 * weights_stride_y, DATA_TYPE),                                                                 \
+                  PTR_TO_VALUE((weights_ptr) + 1 * weights_stride_y, DATA_TYPE),                                                                 \
+                  PTR_TO_VALUE((weights_ptr) + 2 * weights_stride_y, DATA_TYPE));                                                                \
+        acc = ADD_OP(acc, MUL_OP(src0, (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s0));                                                          \
+        acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0), (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s1)); \
+        acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01), (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s2)); \
+    }
+
+#define CONVOLUTION1x3_STRIDE_NHWC_STRIDE2(acc, row_ptr, weights_ptr)                                                                   \
+    {                                                                                                                                   \
+        VEC_DATA_TYPE(DATA_TYPE, 16)                                                                                                    \
+        src0 = (VEC_DATA_TYPE(DATA_TYPE, 16))(                                                                                          \
+                PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE),                                                                        \
+                PTR_TO_VALUE(row_ptr + 10 * src_stride_y, DATA_TYPE),                                                                       \
+                PTR_TO_VALUE(row_ptr + 11 * src_stride_y, DATA_TYPE),                                                                       \
+                PTR_TO_VALUE(row_ptr + 12 * src_stride_y, DATA_TYPE),                                                                       \
+                PTR_TO_VALUE(row_ptr + 13 * src_stride_y, DATA_TYPE),                                                                       \
+                PTR_TO_VALUE(row_ptr + 14 * src_stride_y, DATA_TYPE),                                                                       \
+                PTR_TO_VALUE(row_ptr + 15 * src_stride_y, DATA_TYPE));                                                                      \
+        DATA_TYPE src1 = PTR_TO_VALUE(row_ptr + 16 * src_stride_y, DATA_TYPE);                                                          \
+        VEC_DATA_TYPE(DATA_TYPE, 3)                                                                                                     \
+        weights = (VEC_DATA_TYPE(DATA_TYPE, 3))(                                                                                        \
+                  PTR_TO_VALUE((weights_ptr) + 0 * weights_stride_y, DATA_TYPE),                                                              \
+                  PTR_TO_VALUE((weights_ptr) + 1 * weights_stride_y, DATA_TYPE),                                                              \
+                  PTR_TO_VALUE((weights_ptr) + 2 * weights_stride_y, DATA_TYPE));                                                             \
+        \
+        acc = ADD_OP(acc, MUL_OP(src0.s02468ACE, (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s0));                                             \
+        acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF), (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s1));      \
+        acc = ADD_OP(acc, MUL_OP((VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1), (VEC_DATA_TYPE(DATA_TYPE, 8))weights.s2)); \
+    }
+
+/** This kernel performs a direct convolution to convolve the low three dimensions.
+ *
+ * @note This OpenCL kernel works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
+ * @note If biases are used then -DHAS_BIAS has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: QS8/QS16/F16/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_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 Z processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                          Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                            dst_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
+ * @param[in]  weights_ptr                           Pointer to the weights tensor. Supported data types: same as @p src_ptr
+ * @param[in]  weights_stride_x                      Stride of the weights tensor in X dimension (in bytes)
+ * @param[in]  weights_step_x                        weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  weights_stride_y                      Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in]  weights_step_y                        weights_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in]  weights_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ */
+__kernel void direct_convolution3x3_nhwc(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#ifdef HAS_BIAS
+    VECTOR_DECLARATION(biases),
+#endif /* defined(HAS_BIAS) */
+    unsigned int weights_stride_w)
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)
+    values0       = 0;
+    const int id0 = get_global_id(0);
+    const int id1 = get_global_id(1);
+    const int id2 = get_global_id(2);
+
+    __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0);
+    __global uchar *src_addr     = (__global uchar *)offset(&src, 0, 0) - src_stride_x * id0 + ((id2 * STRIDE_Y) - PAD_TOP) * (int)src_stride_z;
+
+    weights_addr += id0 * weights_stride_w;
+
+    const int coordy = ((id2 * STRIDE_Y) - PAD_TOP);
+    for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
+    {
+#if PAD_TOP > 0
+        if(coordy < 0) // special case Z = -1 doesn't exists
+        {
+            //skip first row and load the two next ones
+            CONVOLUTION1x3_NHWC(values0, src_addr + 1 * (int)src_stride_z, (weights_addr + 1 * (int)weights_stride_z));
+            CONVOLUTION1x3_NHWC(values0, src_addr + 2 * (int)src_stride_z, (weights_addr + 2 * (int)weights_stride_z));
+        }
+        else if(coordy == (SRC_HEIGHT - PAD_TOP - 1))
+        {
+            // special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the
+            // Z axis has no padding at all.
+            CONVOLUTION1x3_NHWC(values0, src_addr, (weights_addr + 0 * (int)weights_stride_z));
+            CONVOLUTION1x3_NHWC(values0, src_addr + 1 * (int)src_stride_z, (weights_addr + 1 * (int)weights_stride_z));
+        }
+        else
+        {
+            CONVOLUTION1x3_NHWC(values0, src_addr, (weights_addr + 0 * (int)weights_stride_z));
+            CONVOLUTION1x3_NHWC(values0, src_addr + 1 * (int)src_stride_z, (weights_addr + 1 * (int)weights_stride_z));
+            CONVOLUTION1x3_NHWC(values0, src_addr + 2 * (int)src_stride_z, (weights_addr + 2 * (int)weights_stride_z));
+        }
+#else  // PAD_TOP > 0
+        CONVOLUTION1x3_NHWC(values0, src_addr, (weights_addr + 0 * (int)weights_stride_z));
+        CONVOLUTION1x3_NHWC(values0, src_addr + 1 * (int)src_stride_z, (weights_addr + 1 * (int)weights_stride_z));
+        CONVOLUTION1x3_NHWC(values0, src_addr + 2 * (int)src_stride_z, (weights_addr + 2 * (int)weights_stride_z));
+#endif // PAD_TOP > 0
+        src_addr += src_stride_x;
+        weights_addr += weights_stride_x;
+    }
+
+#ifdef HAS_BIAS
+    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+    values0       = ADD_OP(values0, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, id0))));
+#endif /* defined(HAS_BIAS) */
+
+    *((__global DATA_TYPE *)(dst.ptr + 0 * dst_stride_y)) = values0.s0;
+    *((__global DATA_TYPE *)(dst.ptr + 1 * dst_stride_y)) = values0.s1;
+    *((__global DATA_TYPE *)(dst.ptr + 2 * dst_stride_y)) = values0.s2;
+    *((__global DATA_TYPE *)(dst.ptr + 3 * dst_stride_y)) = values0.s3;
+    *((__global DATA_TYPE *)(dst.ptr + 4 * dst_stride_y)) = values0.s4;
+    *((__global DATA_TYPE *)(dst.ptr + 5 * dst_stride_y)) = values0.s5;
+    *((__global DATA_TYPE *)(dst.ptr + 6 * dst_stride_y)) = values0.s6;
+    *((__global DATA_TYPE *)(dst.ptr + 7 * dst_stride_y)) = values0.s7;
+}
+#endif // defined(DATA_LAYOUT_NHWC)
+
 /** This kernel performs a direct convolution to convolve the low three dimensions.
  *
  * @note This OpenCL kernel works with stride_x = 1 and 2
@@ -117,7 +296,7 @@ __kernel void direct_convolution3x3(
     Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
 
     VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)
-    pixels0 = 0;
+    values0 = 0;
 
     __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0);
     __global uchar *src_addr     = (__global uchar *)offset(&src, 0, 0);
@@ -127,9 +306,9 @@ __kernel void direct_convolution3x3(
 
     for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
     {
-        CONVOLUTION1x3(pixels0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y));
-        CONVOLUTION1x3(pixels0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y));
-        CONVOLUTION1x3(pixels0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y));
+        CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y));
+        CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y));
+        CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y));
 
         src_addr += src_stride_z;
         weights_addr += weights_stride_z;
@@ -138,10 +317,10 @@ __kernel void direct_convolution3x3(
 #ifdef HAS_BIAS
     Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
 
-    pixels0 = ADD_OP(pixels0, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, kernel_index))));
+    values0 = ADD_OP(values0, (VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, kernel_index))));
 #endif /* defined(HAS_BIAS) */
 
-    vstore8(CONVERT_SAT(pixels0, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)dst.ptr);
+    vstore8(CONVERT_SAT(values0, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)dst.ptr);
 }
 #endif //defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
 
@@ -214,9 +393,9 @@ __kernel void direct_convolution3x3_f32_bifrost(
     Image    src = CONVERT_TO_IMAGE_STRUCT(src);
     Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
 
-    float4 pixels0 = 0;
-    float4 pixels1 = 0;
-    float4 pixels2 = 0;
+    float4 values0 = 0;
+    float4 values1 = 0;
+    float4 values2 = 0;
 
     __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + kernel_index * weights_stride_w);
     __global uchar *src_addr     = (__global uchar *)offset(&src, 0, 0);
@@ -236,39 +415,39 @@ __kernel void direct_convolution3x3_f32_bifrost(
         src0 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
         src1 = vload2(0, (__global float *)(src_addr + 0 * src_stride_y) + 4);
 
-        CONVOLUTION1x3_BIFROST(pixels0, src0, src1, weights_row0);
+        CONVOLUTION1x3_BIFROST(values0, src0, src1, weights_row0);
 
         // Load values from row1 of input tensor
         src0 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
         src1 = vload2(0, (__global float *)(src_addr + 1 * src_stride_y) + 4);
 
         // Accumulate
-        CONVOLUTION1x3_BIFROST(pixels0, src0, src1, weights_row1);
-        CONVOLUTION1x3_BIFROST(pixels1, src0, src1, weights_row0);
+        CONVOLUTION1x3_BIFROST(values0, src0, src1, weights_row1);
+        CONVOLUTION1x3_BIFROST(values1, src0, src1, weights_row0);
 
         // Load values from row2 of input tensor
         src0 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
         src1 = vload2(0, (__global float *)(src_addr + 2 * src_stride_y) + 4);
 
         // Accumulate
-        CONVOLUTION1x3_BIFROST(pixels0, src0, src1, weights_row2);
-        CONVOLUTION1x3_BIFROST(pixels1, src0, src1, weights_row1);
-        CONVOLUTION1x3_BIFROST(pixels2, src0, src1, weights_row0);
+        CONVOLUTION1x3_BIFROST(values0, src0, src1, weights_row2);
+        CONVOLUTION1x3_BIFROST(values1, src0, src1, weights_row1);
+        CONVOLUTION1x3_BIFROST(values2, src0, src1, weights_row0);
 
         // Load values from row3 of input tensor
         src0 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
         src1 = vload2(0, (__global float *)(src_addr + 3 * src_stride_y) + 4);
 
         // Accumulate
-        CONVOLUTION1x3_BIFROST(pixels1, src0, src1, weights_row2);
-        CONVOLUTION1x3_BIFROST(pixels2, src0, src1, weights_row1);
+        CONVOLUTION1x3_BIFROST(values1, src0, src1, weights_row2);
+        CONVOLUTION1x3_BIFROST(values2, src0, src1, weights_row1);
 
         // Row4
         src0 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y));
         src1 = vload2(0, (__global float *)(src_addr + 4 * src_stride_y) + 4);
 
         // Accumulate
-        CONVOLUTION1x3_BIFROST(pixels2, src0, src1, weights_row2);
+        CONVOLUTION1x3_BIFROST(values2, src0, src1, weights_row2);
 
         src_addr += src_stride_z;
         weights_addr += weights_stride_z;
@@ -279,13 +458,13 @@ __kernel void direct_convolution3x3_f32_bifrost(
 
     float bias = (float) * ((__global float *)(vector_offset(&biases, kernel_index)));
 
-    pixels0 += (float4)bias;
-    pixels1 += (float4)bias;
-    pixels2 += (float4)bias;
+    values0 += (float4)bias;
+    values1 += (float4)bias;
+    values2 += (float4)bias;
 #endif /* defined(HAS_BIAS) */
 
-    vstore4(pixels0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y));
-    vstore4(pixels1, 0, (__global float *)(dst.ptr + 1 * dst_stride_y));
-    vstore4(pixels2, 0, (__global float *)(dst.ptr + 2 * dst_stride_y));
+    vstore4(values0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y));
+    vstore4(values1, 0, (__global float *)(dst.ptr + 1 * dst_stride_y));
+    vstore4(values2, 0, (__global float *)(dst.ptr + 2 * dst_stride_y));
 }
 #endif // defined(WEIGHTS_DEPTH)
diff --git a/src/core/CL/cl_kernels/direct_convolution5x5.cl b/src/core/CL/cl_kernels/direct_convolution5x5.cl
index e678f6f51b..70be058854 100644
--- a/src/core/CL/cl_kernels/direct_convolution5x5.cl
+++ b/src/core/CL/cl_kernels/direct_convolution5x5.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -69,6 +69,190 @@
         acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s468a, src0.sCE, src1.s02) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1;     \
     })
 
+#if defined(DATA_LAYOUT_NHWC)
+
+#define PTR_TO_VALUE(PTR, DATA_TYPE) *((__global DATA_TYPE *)(PTR))
+
+#if STRIDE_X == 1
+#define CONVOLUTION1x5_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x5_STRIDE1_NHWC(acc, row_ptr, weights_ptr)
+#elif STRIDE_X == 2 /* STRIDE_X == 1 */
+#define CONVOLUTION1x5_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x5_STRIDE2_NHWC(acc, row_ptr, weights_ptr)
+#else /* STRIDE_X not equals 1 or 2 */
+#error "STRIDE_X larger than 2 is not supported"
+#endif /* STRIDE_X == 2 */
+
+#define CONVOLUTION1x5_STRIDE1_NHWC(acc, row_ptr, weights_ptr)                                                                         \
+    ({                                                                                                                                 \
+        VEC_DATA_TYPE(DATA_TYPE, 8)                                                                                                    \
+        src0 = (VEC_DATA_TYPE(DATA_TYPE, 8))(                                                                                          \
+                PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE),                  \
+                PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE),                  \
+                PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE),                  \
+                PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE));                 \
+        VEC_DATA_TYPE(DATA_TYPE, 4)                                                                                                    \
+        src1 = (VEC_DATA_TYPE(DATA_TYPE, 4))(                                                                                          \
+                PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE),                  \
+                PTR_TO_VALUE(row_ptr + 10 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 11 * src_stride_y, DATA_TYPE));               \
+        VEC_DATA_TYPE(DATA_TYPE, 4)                                                                                                    \
+        weights_values0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(                                                                               \
+                          PTR_TO_VALUE(weights_ptr + 0 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 1 * weights_stride_y, DATA_TYPE),  \
+                          PTR_TO_VALUE(weights_ptr + 2 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 3 * weights_stride_y, DATA_TYPE)); \
+        DATA_TYPE weights_value1 = PTR_TO_VALUE(weights_ptr + 4 * weights_stride_y, DATA_TYPE);                                        \
+        acc += src0 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0;                                                                 \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1;        \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2;        \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s345, src0.s67, src1.s012) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3;        \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s45, src0.s67, src1.s0123) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1;            \
+    })
+
+#define CONVOLUTION1x5_STRIDE2_NHWC(acc, row_ptr, weights_ptr)                                                                         \
+    ({                                                                                                                                 \
+        VEC_DATA_TYPE(DATA_TYPE, 16)                                                                                                   \
+        src0 = (VEC_DATA_TYPE(DATA_TYPE, 16))(                                                                                         \
+                PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE),                  \
+                PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE),                  \
+                PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE),                  \
+                PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE),                  \
+                PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE),                  \
+                PTR_TO_VALUE(row_ptr + 10 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 11 * src_stride_y, DATA_TYPE),                \
+                PTR_TO_VALUE(row_ptr + 12 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 13 * src_stride_y, DATA_TYPE),                \
+                PTR_TO_VALUE(row_ptr + 14 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 15 * src_stride_y, DATA_TYPE));               \
+        VEC_DATA_TYPE(DATA_TYPE, 4)                                                                                                    \
+        src1 = (VEC_DATA_TYPE(DATA_TYPE, 4))(                                                                                          \
+                PTR_TO_VALUE(row_ptr + 16 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 17 * src_stride_y, DATA_TYPE),                \
+                PTR_TO_VALUE(row_ptr + 18 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 19 * src_stride_y, DATA_TYPE));               \
+        VEC_DATA_TYPE(DATA_TYPE, 4)                                                                                                    \
+        weights_values0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(                                                                               \
+                          PTR_TO_VALUE(weights_ptr + 0 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 1 * weights_stride_y, DATA_TYPE),  \
+                          PTR_TO_VALUE(weights_ptr + 2 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 3 * weights_stride_y, DATA_TYPE)); \
+        DATA_TYPE weights_value1 = PTR_TO_VALUE(weights_ptr + 4 * weights_stride_y, DATA_TYPE);                                        \
+        acc += src0.s02468ACE * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0;                                                       \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1;                \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2;        \
+        \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s3579, src0.sBDF, src1.s1) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3;        \
+        acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s468a, src0.sCE, src1.s02) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1;            \
+    })
+
+/** This kernel performs a direct convolution to convolve the low three dimensions in a tensor with the NHWC data layout
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
+ * @note If biases are used then -DHAS_BIAS has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16/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_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 Z processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                          Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                            dst_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
+ * @param[in]  weights_ptr                           Pointer to the weights tensor. Supported data types: same as @p src_ptr
+ * @param[in]  weights_stride_x                      Stride of the weights tensor in X dimension (in bytes)
+ * @param[in]  weights_step_x                        weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  weights_stride_y                      Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in]  weights_step_y                        weights_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in]  weights_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ */
+__kernel void direct_convolution5x5_nhwc(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#ifdef HAS_BIAS
+    VECTOR_DECLARATION(biases),
+#endif /* defined(HAS_BIAS) */
+    unsigned int weights_stride_w)
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    values0 = 0;
+
+    const int id0 = get_global_id(0);
+    const int id1 = get_global_id(1);
+    const int id2 = get_global_id(2);
+
+    __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0);
+    __global uchar *src_addr     = (__global uchar *)offset(&src, 0, 0) - src_stride_x * id0 + ((id2 * STRIDE_Y) - PAD_TOP) * (int)src_stride_z;
+
+    weights_addr += id0 * weights_stride_w;
+    const int coordy = id2 - PAD_TOP;
+
+    for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
+    {
+#if(PAD_TOP)
+        if(coordy < 0) // special case Z = -1 doesn't exists
+        {
+            //skip first row and load the two next ones
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z));
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z));
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z));
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z));
+        }
+        else if(coordy == (DST_HEIGHT - PAD_TOP - 1))
+        {
+            // special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the
+            // Z axis has no padding at all.
+            CONVOLUTION1x5_NHWC(values0, src_addr, weights_addr);
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z));
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z));
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z));
+        }
+        else
+        {
+            CONVOLUTION1x5_NHWC(values0, src_addr, weights_addr);
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z));
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z));
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z));
+            CONVOLUTION1x5_NHWC(values0, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z));
+        }
+#else  //PAD_TOP > 0
+        CONVOLUTION1x5_NHWC(values0, src_addr, weights_addr);
+        CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (int)src_stride_z), (weights_addr + 1 * (int)weights_stride_z));
+        CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (int)src_stride_z), (weights_addr + 2 * (int)weights_stride_z));
+        CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (int)src_stride_z), (weights_addr + 3 * (int)weights_stride_z));
+        CONVOLUTION1x5_NHWC(values0, (src_addr + 4 * (int)src_stride_z), (weights_addr + 4 * (int)weights_stride_z));
+#endif // PAD_TOP > 0
+
+        src_addr += src_stride_x;
+        weights_addr += weights_stride_x;
+    }
+
+#ifdef HAS_BIAS
+    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+    values0 += (VEC_DATA_TYPE(DATA_TYPE, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, id0)));
+#endif /* defined(HAS_BIAS) */
+
+    *((__global DATA_TYPE *)(dst.ptr + 0 * dst_stride_y)) = values0.s0;
+    *((__global DATA_TYPE *)(dst.ptr + 1 * dst_stride_y)) = values0.s1;
+    *((__global DATA_TYPE *)(dst.ptr + 2 * dst_stride_y)) = values0.s2;
+    *((__global DATA_TYPE *)(dst.ptr + 3 * dst_stride_y)) = values0.s3;
+    *((__global DATA_TYPE *)(dst.ptr + 4 * dst_stride_y)) = values0.s4;
+    *((__global DATA_TYPE *)(dst.ptr + 5 * dst_stride_y)) = values0.s5;
+    *((__global DATA_TYPE *)(dst.ptr + 6 * dst_stride_y)) = values0.s6;
+    *((__global DATA_TYPE *)(dst.ptr + 7 * dst_stride_y)) = values0.s7;
+}
+
+#endif // defined(DATA_LAYOUT_NHWC)
+
 /** This kernel performs a direct convolution to convolve the low three dimensions.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
@@ -119,7 +303,7 @@ __kernel void direct_convolution5x5(
     Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
 
     VEC_DATA_TYPE(DATA_TYPE, 8)
-    pixels0 = 0;
+    values0 = 0;
 
     __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0);
     __global uchar *src_addr     = (__global uchar *)offset(&src, 0, 0);
@@ -129,11 +313,11 @@ __kernel void direct_convolution5x5(
 
     for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
     {
-        CONVOLUTION1x5(pixels0, (__global DATA_TYPE *)src_addr, (__global DATA_TYPE *)weights_addr);
-        CONVOLUTION1x5(pixels0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y));
-        CONVOLUTION1x5(pixels0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y));
-        CONVOLUTION1x5(pixels0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_y));
-        CONVOLUTION1x5(pixels0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_y));
+        CONVOLUTION1x5(values0, (__global DATA_TYPE *)src_addr, (__global DATA_TYPE *)weights_addr);
+        CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y));
+        CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y));
+        CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_y));
+        CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_y));
 
         src_addr += src_stride_z;
         weights_addr += weights_stride_z;
@@ -142,10 +326,10 @@ __kernel void direct_convolution5x5(
 #ifdef HAS_BIAS
     Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
 
-    pixels0 += (VEC_DATA_TYPE(DATA_TYPE, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, kernel_index)));
+    values0 += (VEC_DATA_TYPE(DATA_TYPE, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, kernel_index)));
 #endif /* defined(HAS_BIAS) */
 
-    vstore8(pixels0, 0, (__global DATA_TYPE *)dst.ptr);
+    vstore8(values0, 0, (__global DATA_TYPE *)dst.ptr);
 }
 #endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
 
@@ -226,8 +410,8 @@ __kernel void direct_convolution5x5_f32_bifrost(
     Image    src = CONVERT_TO_IMAGE_STRUCT(src);
     Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
 
-    float4 pixels0 = 0.0f;
-    float4 pixels1 = 0.0f;
+    float4 values0 = 0.0f;
+    float4 values1 = 0.0f;
 
     __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + kernel_index * weights_stride_w);
     __global uchar *src_addr     = (__global uchar *)offset(&src, 0, 0);
@@ -247,14 +431,14 @@ __kernel void direct_convolution5x5_f32_bifrost(
         src0 = vload8(0, (__global float *)(src_addr + 0 * src_stride_y));
 
         // Accumulate
-        CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row00, weights_row01);
+        CONVOLUTION1x5_BIFROST(values0, src0, weights_row00, weights_row01);
 
         // Load values from row1 of input tensor
         src0 = vload8(0, (__global float *)(src_addr + 1 * src_stride_y));
 
         // Accumulate
-        CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row10, weights_row11);
-        CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row00, weights_row01);
+        CONVOLUTION1x5_BIFROST(values0, src0, weights_row10, weights_row11);
+        CONVOLUTION1x5_BIFROST(values1, src0, weights_row00, weights_row01);
 
         // Load values from row2 of input tensor
         src0 = vload8(0, (__global float *)(src_addr + 2 * src_stride_y));
@@ -264,8 +448,8 @@ __kernel void direct_convolution5x5_f32_bifrost(
         weights_row01 = *((__global float *)(weights_addr + 2 * weights_stride_y) + 4);
 
         // Accumulate
-        CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row00, weights_row01);
-        CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row10, weights_row11);
+        CONVOLUTION1x5_BIFROST(values0, src0, weights_row00, weights_row01);
+        CONVOLUTION1x5_BIFROST(values1, src0, weights_row10, weights_row11);
 
         // Load values from row3 of input tensor
         src0 = vload8(0, (__global float *)(src_addr + 3 * src_stride_y));
@@ -275,8 +459,8 @@ __kernel void direct_convolution5x5_f32_bifrost(
         weights_row11 = *((__global float *)(weights_addr + 3 * weights_stride_y) + 4);
 
         // Accumulate
-        CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row10, weights_row11);
-        CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row00, weights_row01);
+        CONVOLUTION1x5_BIFROST(values0, src0, weights_row10, weights_row11);
+        CONVOLUTION1x5_BIFROST(values1, src0, weights_row00, weights_row01);
 
         // Load values from row4 of input tensor
         src0 = vload8(0, (__global float *)(src_addr + 4 * src_stride_y));
@@ -285,14 +469,14 @@ __kernel void direct_convolution5x5_f32_bifrost(
         weights_row00 = vload4(0, (__global float *)(weights_addr + 4 * weights_stride_y));
         weights_row01 = *((__global float *)(weights_addr + 4 * weights_stride_y) + 4);
 
-        CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row00, weights_row01);
-        CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row10, weights_row11);
+        CONVOLUTION1x5_BIFROST(values0, src0, weights_row00, weights_row01);
+        CONVOLUTION1x5_BIFROST(values1, src0, weights_row10, weights_row11);
 
         // Load values from row5 of input tensor
         src0 = vload8(0, (__global float *)(src_addr + 5 * src_stride_y));
 
         // Accumulate
-        CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row00, weights_row01);
+        CONVOLUTION1x5_BIFROST(values1, src0, weights_row00, weights_row01);
 
         src_addr += src_stride_z;
         weights_addr += weights_stride_z;
@@ -303,11 +487,11 @@ __kernel void direct_convolution5x5_f32_bifrost(
 
     float4 bias = (float4) * ((__global float *)(vector_offset(&biases, kernel_index)));
 
-    pixels0 += bias;
-    pixels1 += bias;
+    values0 += bias;
+    values1 += bias;
 #endif /* defined(HAS_BIAS) */
 
-    vstore4(pixels0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y));
-    vstore4(pixels1, 0, (__global float *)(dst.ptr + 1 * dst_stride_y));
+    vstore4(values0, 0, (__global float *)(dst.ptr + 0 * dst_stride_y));
+    vstore4(values1, 0, (__global float *)(dst.ptr + 1 * dst_stride_y));
 }
 #endif // defined(WEIGHTS_DEPTH)
diff --git a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
index 7f7437d6ef..754f0d8f23 100644
--- a/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDirectConvolutionLayerKernel.cpp
@@ -47,19 +47,20 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != weights->dimension(1),
-                                    "Weights should have same width as length");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != 1 && weights->dimension(0) != 3 && weights->dimension(0) != 5,
+
+    const DataLayout data_layout = input->data_layout();
+    const int        width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int        height_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+    const int        channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != weights->dimension(height_idx), "Weights should have same width and height");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != 1 && weights->dimension(width_idx) != 3 && weights->dimension(width_idx) != 5,
                                     "Kernel sizes other than 1x1, 3x3 or 5x5 are not supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(2) != input->dimension(2),
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(channel_idx) != input->dimension(channel_idx),
                                     "Weights feature map dimension should match the respective input's one");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != weights->dimension(1),
-                                    "Only rectangular weights are supported!");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4,
-                                    "Weights can be at most 4 dimensional");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(0) == 1) && std::get<0>(conv_info.stride()) > 3,
-                                    "Strides larger than 3 not supported for 1x1 convolution.");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(0) == 3 || weights->dimension(0) == 5) && std::get<0>(conv_info.stride()) > 2,
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4, "Weights can be at most 4 dimensional");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 1) && std::get<0>(conv_info.stride()) > 3, "Strides larger than 3 not supported for 1x1 convolution.");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((weights->dimension(width_idx) == 3 || weights->dimension(width_idx) == 5) && std::get<0>(conv_info.stride()) > 2,
                                     "Strides larger than 2 not supported for 3x3 convolution.");
 
     if(biases != nullptr)
@@ -89,36 +90,27 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info, const GPUTarget target)
+inline bool can_run_optimized_kernel_for_bifrost(GPUTarget gpu_target, unsigned int conv_stride_x, unsigned int conv_stride_y, unsigned int kernel_size,
+                                                 DataType data_type, DataLayout data_layout)
 {
-    const unsigned int kernel_size = weights->dimension(0);
-    const DataType     data_type   = input->data_type();
-
-    // Get convolved dimensions
-    TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(*input, *weights, conv_info);
-
-    // Output auto inizialitation if not yet initialized
-    // FIXME: input->clone()->set_tensor_shape(output_shape) doesn't work with subtensors for grouped direct convolutions (AlexNet).
-    auto_init_if_empty(*output, output_shape,
-                       1,
-                       input->data_type(),
-                       input->quantization_info());
+    return gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::G76) && (kernel_size <= 5)
+           && (conv_stride_x == 1) && (conv_stride_y == 1) && (data_type == DataType::F32) && (data_layout == DataLayout::NCHW);
+}
 
-    unsigned int conv_stride_x = std::get<0>(conv_info.stride());
-    unsigned int conv_stride_y = std::get<1>(conv_info.stride());
-    unsigned int conv_pad_left = conv_info.pad_left();
-    unsigned int conv_pad_top  = conv_info.pad_top();
+inline void setup_num_elems(unsigned int &num_elems_read_per_iteration_x, unsigned int &num_elems_read_per_iteration_y,
+                            unsigned int &num_elems_written_per_iteration_x, unsigned int &num_elems_written_per_iteration_y,
+                            unsigned int kernel_size, const PadStrideInfo &conv_info, const GPUTarget target, ITensorInfo *input)
+{
+    const DataType   data_type     = input->data_type();
+    const DataLayout data_layout   = input->data_layout();
+    unsigned int     conv_stride_x = std::get<0>(conv_info.stride());
+    unsigned int     conv_stride_y = std::get<1>(conv_info.stride());
 
-    unsigned int num_elems_read_per_iteration_x    = 0;
-    unsigned int num_elems_read_per_iteration_y    = 0;
-    unsigned int num_elems_written_per_iteration_x = 0;
-    unsigned int num_elems_written_per_iteration_y = 0;
+    const bool run_optimized_bifrost = can_run_optimized_kernel_for_bifrost(target, conv_stride_x, conv_stride_y, kernel_size, data_type, data_layout);
 
-    if(gpu_target_is_in(target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::G76) && (kernel_size <= 5) && (conv_stride_x == 1)
-       && (conv_stride_y == 1) && (data_type == DataType::F32))
+    if(run_optimized_bifrost)
     {
         // Configure kernel window
-
         switch(kernel_size)
         {
             case 1:
@@ -218,22 +210,124 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
         }
     }
 
-    // Create window and update padding
-    bool   window_changed = false;
-    Window win            = calculate_max_window(*output, Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y));
+    if(data_layout == DataLayout::NHWC)
+    {
+        num_elems_written_per_iteration_x = 1;
+        num_elems_read_per_iteration_x    = 1;
+        switch(kernel_size)
+        {
+            case 1:
+                switch(conv_stride_x)
+                {
+                    case 1:
+                        num_elems_read_per_iteration_y    = 8;
+                        num_elems_written_per_iteration_y = 8;
+                        break;
+                    case 2:
+                        num_elems_read_per_iteration_y    = 16;
+                        num_elems_written_per_iteration_y = 8;
+                        break;
+                    default:
+                        ARM_COMPUTE_ERROR("Invalid convolution stride X");
+                }
+                break;
+            case 3:
+                switch(conv_stride_x)
+                {
+                    case 1:
+                        num_elems_read_per_iteration_y    = 10;
+                        num_elems_written_per_iteration_y = 8;
+                        break;
+                    case 2:
+                        num_elems_read_per_iteration_y    = 17;
+                        num_elems_written_per_iteration_y = 8;
+                        break;
+                    default:
+                        ARM_COMPUTE_ERROR("Invalid convolution stride X");
+                }
+                break;
+            case 5:
+                switch(conv_stride_x)
+                {
+                    case 1:
+                        num_elems_read_per_iteration_y    = 12;
+                        num_elems_written_per_iteration_y = 8;
+                        break;
+                    case 2:
+                        num_elems_read_per_iteration_y    = 20;
+                        num_elems_written_per_iteration_y = 8;
+                        break;
+                    default:
+                        ARM_COMPUTE_ERROR("Invalid convolution stride X");
+                }
+                break;
+            default:
+                ARM_COMPUTE_ERROR("Not implemented.");
+                break;
+        }
+    }
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info, const GPUTarget target)
+{
+    const DataLayout   data_layout = input->data_layout();
+    const int          width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const unsigned int kernel_size = weights->dimension(width_idx);
+
+    // Get convolved dimensions
+    TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(*input, *weights, conv_info);
 
-    AccessWindowRectangle input_access(input, -conv_pad_left, -conv_pad_top,
-                                       num_elems_read_per_iteration_x, num_elems_read_per_iteration_y,
-                                       conv_stride_x, conv_stride_y);
-    AccessWindowStatic    weights_access(weights, 0, 0, kernel_size, kernel_size);
-    AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y);
+    // Output auto inizialitation if not yet initialized
+    // FIXME: input->clone()->set_tensor_shape(output_shape) doesn't work with subtensors for grouped direct convolutions (AlexNet).
+    auto_init_if_empty(*output, output_shape,
+                       1,
+                       input->data_type(),
+                       input->quantization_info());
 
-    window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
+    unsigned int num_elems_read_per_iteration_x    = 0;
+    unsigned int num_elems_read_per_iteration_y    = 0;
+    unsigned int num_elems_written_per_iteration_x = 0;
+    unsigned int num_elems_written_per_iteration_y = 0;
+
+    unsigned int conv_pad_left = conv_info.pad_left();
+    unsigned int conv_pad_top  = conv_info.pad_top();
+    unsigned int conv_stride_x = std::get<0>(conv_info.stride());
+    unsigned int conv_stride_y = std::get<1>(conv_info.stride());
 
-    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    setup_num_elems(num_elems_read_per_iteration_x, num_elems_read_per_iteration_y,
+                    num_elems_written_per_iteration_x, num_elems_written_per_iteration_y,
+                    kernel_size, conv_info, target, input);
 
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
+    // Create window and update padding
+    bool   window_changed = false;
+    Window win            = calculate_max_window(*output, Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y));
+
+    if(data_layout == DataLayout::NHWC)
+    {
+        AccessWindowStatic input_access(input, 0, -conv_pad_left,
+                                        num_elems_read_per_iteration_x,
+                                        ceil_to_multiple(input->dimension(1) + conv_info.pad_right(), num_elems_read_per_iteration_y));
+        AccessWindowStatic    weights_access(weights, 0, 0, weights->dimension(0), weights->dimension(1));
+        AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y);
+        window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
+        output_access.set_valid_region(win, 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);
+    }
+    else if(data_layout == DataLayout::NCHW)
+    {
+        AccessWindowRectangle input_access(input, -conv_pad_left, -conv_pad_top, num_elems_read_per_iteration_x, num_elems_read_per_iteration_y, conv_stride_x, conv_stride_y);
+        AccessWindowStatic    weights_access(weights, 0, 0, kernel_size, kernel_size);
+        AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y);
+        window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
+        output_access.set_valid_region(win, 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);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not supported");
+    }
 }
 } // namespace
 
@@ -251,7 +345,12 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 
-    const unsigned int kernel_size = weights->info()->dimension(0);
+    const DataLayout data_layout = input->info()->data_layout();
+    const int        width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int        height_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+    const int        channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+    const unsigned int kernel_size = weights->info()->dimension(width_idx);
     const DataType     data_type   = input->info()->data_type();
 
     // Get convolved dimensions
@@ -274,7 +373,19 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
 
     _conv_stride_x = std::get<0>(conv_info.stride());
     _conv_stride_y = std::get<1>(conv_info.stride());
-    _border_size   = BorderSize(conv_info.pad_top(), conv_info.pad_right(), conv_info.pad_bottom(), conv_info.pad_left());
+
+    if(data_layout == DataLayout::NHWC)
+    {
+        _border_size = BorderSize(conv_info.pad_left(), 0, conv_info.pad_right(), 0);
+    }
+    else if(data_layout == DataLayout::NCHW)
+    {
+        _border_size = BorderSize(conv_info.pad_top(), conv_info.pad_right(), conv_info.pad_bottom(), conv_info.pad_left());
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Not supported");
+    }
 
     _input   = input;
     _weights = weights;
@@ -285,14 +396,19 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
 
     std::stringstream kernel_name;
     kernel_name << "direct_convolution" << kernel_size << "x" << kernel_size;
+    if(data_layout == DataLayout::NHWC)
+    {
+        kernel_name << "_" << lower_string(string_from_data_layout(data_layout));
+    }
 
     CLBuildOptions build_options;
     build_options.add_option_if(_biases != nullptr, std::string("-DHAS_BIAS"));
 
-    if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::G76) && (kernel_size <= 5) && (_conv_stride_x == 1)
-       && (_conv_stride_y == 1) && (data_type == DataType::F32))
+    const bool run_optimized_for_bifrost = can_run_optimized_kernel_for_bifrost(gpu_target, _conv_stride_x, _conv_stride_y, kernel_size, data_type, data_layout);
+
+    if(run_optimized_for_bifrost)
     {
-        build_options.add_option(std::string("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(_weights->info()->dimension(2))));
+        build_options.add_option(std::string("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(_weights->info()->dimension(channel_idx))));
 
         kernel_name << "_f32_bifrost";
         _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name.str(), build_options.options()));
@@ -304,10 +420,20 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
         build_options.add_option_if(is_quantized_asymm, std::string("-DKERNEL_SIZE=" + support::cpp11::to_string(kernel_size)));
         build_options.add_option(std::string("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)));
         build_options.add_option(std::string("-DDATA_SIZE=" + get_data_size_from_data_type(data_type)));
-        build_options.add_option(std::string("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(_weights->info()->dimension(2))));
+        build_options.add_option(std::string("-DWEIGHTS_DEPTH=" + support::cpp11::to_string(_weights->info()->dimension(channel_idx))));
         build_options.add_option(std::string("-DSTRIDE_X=" + support::cpp11::to_string(_conv_stride_x)));
+        if(data_layout == DataLayout::NHWC)
+        {
+            build_options.add_option(std::string("-DDATA_LAYOUT_NHWC=1"));
+            build_options.add_option(std::string("-DDST_HEIGHT=" + support::cpp11::to_string(_output->info()->dimension(height_idx))));
+            build_options.add_option(std::string("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(width_idx))));
+            build_options.add_option(std::string("-DSRC_HEIGHT=" + support::cpp11::to_string(_input->info()->dimension(height_idx))));
+            build_options.add_option(std::string("-DSRC_WIDTH=" + support::cpp11::to_string(_input->info()->dimension(width_idx))));
+            build_options.add_option(std::string("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left())));
+            build_options.add_option(std::string("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top())));
+            build_options.add_option(std::string("-DSTRIDE_Y=" + support::cpp11::to_string(_conv_stride_y)));
+        }
         build_options.add_option(std::string("-DDATA_TYPE_PROMOTED=" + get_cl_type_from_data_type(data_type)));
-
         // Create kernel
         _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(is_quantized_asymm ? "direct_convolution_1x1_3x3_5x5_quantized" : kernel_name.str(),
                                                                                build_options.options()));
@@ -353,9 +479,11 @@ void CLDirectConvolutionLayerKernel::configure(const ICLTensor *input, const ICL
     _config_id += "_";
     _config_id += support::cpp11::to_string(_conv_stride_y);
     _config_id += "_";
-    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    _config_id += support::cpp11::to_string(output->info()->dimension(width_idx));
     _config_id += "_";
-    _config_id += support::cpp11::to_string(output->info()->dimension(1));
+    _config_id += support::cpp11::to_string(output->info()->dimension(height_idx));
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_layout(data_layout));
 }
 
 Status CLDirectConvolutionLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
@@ -378,12 +506,16 @@ void CLDirectConvolutionLayerKernel::run(const Window &window, cl::CommandQueue
 
     win_in.adjust(Window::DimX, -_border_size.left, true);
     win_in.adjust(Window::DimY, -_border_size.top, true);
-    win_in.set_dimension_step(Window::DimX, window.x().step() * _conv_stride_x);
-    win_in.set_dimension_step(Window::DimY, window.y().step() * _conv_stride_y);
 
-    Window slice_in = win_in.first_slice_window_3D();
+    const DataLayout data_layout = _input->info()->data_layout();
+    const int        width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int        height_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+
+    win_in.set_dimension_step(width_idx, window[width_idx].step() * _conv_stride_x);
+    win_in.set_dimension_step(height_idx, window[height_idx].step() * _conv_stride_y);
 
-    unsigned int idx1 = 2 * num_arguments_per_3D_tensor();
+    Window       slice_in = win_in.first_slice_window_3D();
+    unsigned int idx1     = 2 * num_arguments_per_3D_tensor();
     add_3D_tensor_argument(idx1, _weights, slice);
 
     if(_biases != nullptr)
@@ -400,7 +532,6 @@ void CLDirectConvolutionLayerKernel::run(const Window &window, cl::CommandQueue
         unsigned int idx = 0;
         add_3D_tensor_argument(idx, _input, slice_in);
         add_3D_tensor_argument(idx, _output, slice);
-
         enqueue(queue, *this, slice, _lws_hint);
     }
     while(window.slide_window_slice_3D(slice) && win_in.slide_window_slice_3D(slice_in));
diff --git a/tests/validation/CL/DirectConvolutionLayer.cpp b/tests/validation/CL/DirectConvolutionLayer.cpp
index 87f9449359..e29ec7ef4c 100644
--- a/tests/validation/CL/DirectConvolutionLayer.cpp
+++ b/tests/validation/CL/DirectConvolutionLayer.cpp
@@ -172,7 +172,7 @@ TEST_SUITE_END()
 TEST_SUITE(FP32)
 FIXTURE_DATA_TEST_CASE(Run, CLDirectConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(data, framework::dataset::make("DataType", DataType::F32)),
                                                                                                                  ActivationFunctionsDataset),
-                                                                                                         framework::dataset::make("DataLayout", DataLayout::NCHW)))
+                                                                                                         framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_fp32);
diff --git a/tests/validation/fixtures/DirectConvolutionLayerFixture.h b/tests/validation/fixtures/DirectConvolutionLayerFixture.h
index 9a58167605..f9a220b0bb 100644
--- a/tests/validation/fixtures/DirectConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/DirectConvolutionLayerFixture.h
@@ -106,7 +106,7 @@ protected:
             case DataType::F16:
             case DataType::F32:
             {
-                std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+                std::uniform_real_distribution<> distribution(-1.f, 1.f);
                 library->fill(tensor, distribution, i);
                 break;
             }
@@ -182,10 +182,8 @@ protected:
         fill(bias, 2);
 
         SimpleTensor<T> dst = reference::convolution_layer<T>(src, weights, bias, output_shape, info);
-
         return (act_info.enabled()) ? reference::activation_layer<T>(dst, act_info) : dst;
     }
-
     TensorType       _target{};
     SimpleTensor<T>  _reference{};
     QuantizationInfo _quantization_info{};