COMPMID-811 Add NHWC data format support for CL depthwise convolution

Change-Id: I574f7945f0be009c638d860028bce8b52b4120fd
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/136484
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

8 files changed, 520 insertions, 77 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 2bcacad7f0..fb688b5ee9 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -197,6 +197,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "deconvolution_upsample", "deconvolution_layer.cl" },
     { "depthwise_convolution_3x3", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_f16", "depthwise_convolution.cl" },
+    { "depthwise_convolution_3x3_nhwc", "depthwise_convolution.cl" },
+    { "depthwise_convolution_3x3_nhwc_stride1", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_quantized_nchw", "depthwise_convolution_quantized.cl" },
     { "depthwise_convolution_3x3_quantized_nhwc_stride1", "depthwise_convolution_quantized.cl" },
     { "depthwise_convolution_3x3_quantized_nhwc_stride2", "depthwise_convolution_quantized.cl" },
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index 5f4247e5d3..f3aa0d6dd8 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -451,6 +451,22 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32(
 
 #endif // defined(DEPTH_MULTIPLIER)
 
+#if defined(NCHW)
+#define in_stride_x src_stride_x
+#define in_stride_y src_stride_y
+#define in_stride_z src_stride_z
+#define out_stride_x dst_stride_x
+#define out_stride_y dst_stride_y
+#define out_stride_z dst_stride_z
+#else //defined(NCHW)
+#define in_stride_x src_stride_y
+#define in_stride_y src_stride_z
+#define in_stride_z src_stride_x
+#define out_stride_x dst_stride_y
+#define out_stride_y dst_stride_z
+#define out_stride_z dst_stride_x
+#endif //defined(NCHW)
+
 #if defined(SRC_WIDTH) && defined(DATA_TYPE)
 /** This kernel reshapes each of the tensor's low three dimensions to single rows.
  *
@@ -484,17 +500,16 @@ __kernel void depthwise_weights_reshape(
 #endif /* HAS_BIAS */
 )
 {
-    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
 #ifdef HAS_BIAS
     Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
 #endif /* HAS_BIAS */
 
-    __global DATA_TYPE *input_ptr = (__global DATA_TYPE *)src.ptr;
-    __global uchar *output_ptr    = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * SRC_WIDTH * dst_stride_x + get_global_id(2) * dst_stride_y;
+    __global uchar *input_ptr  = src_ptr + src_offset_first_element_in_bytes + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z;
+    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * SRC_WIDTH * dst_stride_x + get_global_id(2) * dst_stride_y;
 
-    for(int i = 0; i < SRC_WIDTH; ++i, ++input_ptr)
+    for(int i = 0; i < SRC_WIDTH; ++i, input_ptr += in_stride_x)
     {
-        *((__global DATA_TYPE *)(output_ptr + i * dst_stride_x)) = *input_ptr;
+        *((__global DATA_TYPE *)(output_ptr + i * dst_stride_x)) = *((__global DATA_TYPE *)input_ptr);
     }
 
 #if defined(HAS_BIAS)
@@ -541,7 +556,7 @@ __kernel void depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(d
     const int src_y = -PAD_TOP + src_pixel_linear / max_initial_x * STRIDE_Y;
     const int src_z = get_global_id(2) / DEPTH_MULTIPLIER;
 
-    __global uchar *input_ptr      = src_ptr + src_offset_first_element_in_bytes + src_z * src_stride_z;
+    __global uchar *input_ptr      = src_ptr + src_offset_first_element_in_bytes + src_z * in_stride_z;
     __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst.ptr));
 
     for(int y = src_y; y < src_y + KERNEL_HEIGHT; ++y)
@@ -554,7 +569,7 @@ __kernel void depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(d
             }
             else
             {
-                *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
+                *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * in_stride_x + y * in_stride_y));
             }
         }
     }
@@ -596,7 +611,7 @@ __kernel void depthwise_vector_to_tensor(
     const int z          = id0 / patch_size;
     const int index2D    = id0 - z * patch_size;
 
-    __global uchar *out_ptr          = dst_ptr + dst_offset_first_element_in_bytes + index2D % CONV_WIDTH * dst_stride_x + index2D / CONV_WIDTH * dst_stride_y + z * dst_stride_z;
+    __global uchar *out_ptr          = dst_ptr + dst_offset_first_element_in_bytes + index2D % CONV_WIDTH * out_stride_x + index2D / CONV_WIDTH * out_stride_y + z * out_stride_z;
     *((__global DATA_TYPE *)out_ptr) = *((__global DATA_TYPE *)src.ptr);
 }
 
@@ -980,3 +995,335 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16(
     vstore4(pixels1, 0, (__global half *)(dst.ptr + 1 * dst_stride_y));
 }
 #endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(DEPTH_MULTIPLIER)
+
+#if defined(VEC_SIZE) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT)
+
+#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
+
+#if defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y)
+/** This function computes the depthwise convolution for NHWC data layout when the stride along the width or height is not 1.
+ *
+ * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2)
+ * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1)
+ * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X)
+ * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1)
+ *
+ * @param[in] src_ptr                               Pointer to the source image. Supported data types: FP32
+ * @param[in] src_stride_x                          Stride of the source image 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 image 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_offset_first_element_in_bytes     The offset of the first element in the source image
+ * @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 Y processed per workitem(in bytes)
+ * @param[in] dst_ptr                               Pointer to the destination tensor. Supported data types: same as 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] 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 Y 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: QASYMM8
+ * @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 Y 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] max_offset                            Max offset for the input tensor
+ * @param[in] biases_ptr                            (Optional) Pointer to the biases vector. Supported data types: same as src_ptr
+ * @param[in] biases_stride_x                       (Optional) Stride of the biases vector in X dimension (in bytes)
+ * @param[in] biases_step_x                         (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] biases_offset_first_element_in_bytes  (Optional) The offset of the first element in the biases vector
+ */
+__kernel void depthwise_convolution_3x3_nhwc(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#if defined(HAS_BIAS)
+    VECTOR_DECLARATION(biases),
+#endif /* defined(HAS_BIAS) */
+    int max_offset)
+{
+    int x = get_global_id(0); // channels
+    int y = get_global_id(1); // spatial coordinate x
+    int z = get_global_id(2); // spatial coordinate y
+
+    Vector weights = CONVERT_TO_VECTOR_STRUCT(weights);
+
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(float) * VEC_SIZE;
+
+    int  z_coord  = 0;
+    int4 offset   = 0;
+    int4 y_offset = ((int4)(y * CONV_STRIDE_X) + (int4)(0, 1, 2, 3) - CONV_PAD_LEFT) * (int4)src_stride_y;
+
+    // We compute 2x1x1 [C,W,H] elements
+    VEC_FLOAT acc = 0;
+
+    // Load weights
+    VEC_FLOAT w0 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 0 * weights_stride_y + 0 * weights_stride_z));
+    VEC_FLOAT w1 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 1 * weights_stride_y + 0 * weights_stride_z));
+    VEC_FLOAT w2 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 2 * weights_stride_y + 0 * weights_stride_z));
+    VEC_FLOAT w3 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 0 * weights_stride_y + 1 * weights_stride_z));
+    VEC_FLOAT w4 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 1 * weights_stride_y + 1 * weights_stride_z));
+    VEC_FLOAT w5 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 2 * weights_stride_y + 1 * weights_stride_z));
+    VEC_FLOAT w6 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 0 * weights_stride_y + 2 * weights_stride_z));
+    VEC_FLOAT w7 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 1 * weights_stride_y + 2 * weights_stride_z));
+    VEC_FLOAT w8 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 2 * weights_stride_y + 2 * weights_stride_z));
+
+    // Load input values
+    // z == 0
+    // Clamp z_coord as for z = 0, it can be negative
+    // z_coord is casted to unsigned int in order to use just a min() operation
+    // A "-1" 32 bit signed variable converted to unsigned gives 4294967295
+    z_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP;
+    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
+    offset  = y_offset + (int4)(z_coord * src_stride_z);
+    offset  = min(offset, max_offset);
+
+    VEC_FLOAT values0 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s0));
+    VEC_FLOAT values1 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s1));
+    VEC_FLOAT values2 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s2));
+
+    // z == 1
+    // z_coord can be only negative for z = 0 so we do not need to clamp it
+    // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset
+    z_coord           = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP + 1;
+    offset            = y_offset + (int4)(z_coord * src_stride_z);
+    VEC_FLOAT values3 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s0));
+    VEC_FLOAT values4 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s1));
+    VEC_FLOAT values5 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s2));
+
+    // z == 2
+    // After z = 1 we can simply add src_stride_z to offset without updating z_coord
+    // However offset can be out-of-bound so we need to check if it is greater than max_offset
+    offset += (int4)src_stride_z;
+    offset            = min(offset, max_offset);
+    VEC_FLOAT values6 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s0));
+    VEC_FLOAT values7 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s1));
+    VEC_FLOAT values8 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s2));
+
+    acc = fma(values0, w0, acc);
+    acc = fma(values1, w1, acc);
+    acc = fma(values2, w2, acc);
+
+    acc = fma(values3, w3, acc);
+    acc = fma(values4, w4, acc);
+    acc = fma(values5, w5, acc);
+
+    acc = fma(values6, w6, acc);
+    acc = fma(values7, w7, acc);
+    acc = fma(values8, w8, acc);
+
+#if defined(HAS_BIAS)
+    Vector    biases      = CONVERT_TO_VECTOR_STRUCT(biases);
+    VEC_FLOAT bias_values = VLOAD(VEC_SIZE)(0, (__global float *)biases.ptr);
+    acc += bias_values;
+#endif // defined(HAS_BIAS)
+
+    Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+    VSTORE(VEC_SIZE)
+    (acc, 0, (__global float *)(dst.ptr));
+}
+#endif // defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y)
+
+#if defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED)
+/** This function computes the depthwise convolution for NHWC data layout when the stride along the width and height is 1.
+ *
+ * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2)
+ * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112)
+ * @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2)
+ * @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1)
+ *
+ * @param[in] src_ptr                               Pointer to the source image. Supported data types: FP32
+ * @param[in] src_stride_x                          Stride of the source image 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 image 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_offset_first_element_in_bytes     The offset of the first element in the source image
+ * @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 Y processed per workitem(in bytes)
+ * @param[in] dst_ptr                               Pointer to the destination tensor. Supported data types: same as 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] 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 Y 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: QASYMM8
+ * @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 Y 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] max_offset                            Max offset for the input tensor
+ * @param[in] biases_ptr                            (Optional) Pointer to the biases vector. Supported data types: same as src_ptr
+ * @param[in] biases_stride_x                       (Optional) Stride of the biases vector in X dimension (in bytes)
+ * @param[in] biases_step_x                         (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] biases_offset_first_element_in_bytes  (Optional) The offset of the first element in the biases vector
+ */
+__kernel void depthwise_convolution_3x3_nhwc_stride1(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#if defined(HAS_BIAS)
+    VECTOR_DECLARATION(biases),
+#endif /* defined(HAS_BIAS) */
+    int max_offset)
+{
+    int x = get_global_id(0); // channels
+    int y = get_global_id(1); // spatial coordinate x
+    int z = get_global_id(2); // spatial coordinate y
+
+    Vector weights = CONVERT_TO_VECTOR_STRUCT(weights);
+
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(float) * VEC_SIZE;
+
+    int  z_coord  = 0;
+    int4 offset   = 0;
+    int4 y_offset = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3) - CONV_PAD_LEFT) * (int4)src_stride_y;
+
+    // We compute 2x2x2 [C,W,H] elements
+    VEC_FLOAT acc0 = 0;
+    VEC_FLOAT acc1 = 0;
+    VEC_FLOAT acc2 = 0;
+    VEC_FLOAT acc3 = 0;
+
+    // Load weights
+    VEC_FLOAT w0 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 0 * weights_stride_y + 0 * weights_stride_z));
+    VEC_FLOAT w1 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 1 * weights_stride_y + 0 * weights_stride_z));
+    VEC_FLOAT w2 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 2 * weights_stride_y + 0 * weights_stride_z));
+    VEC_FLOAT w3 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 0 * weights_stride_y + 1 * weights_stride_z));
+    VEC_FLOAT w4 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 1 * weights_stride_y + 1 * weights_stride_z));
+    VEC_FLOAT w5 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 2 * weights_stride_y + 1 * weights_stride_z));
+    VEC_FLOAT w6 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 0 * weights_stride_y + 2 * weights_stride_z));
+    VEC_FLOAT w7 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 1 * weights_stride_y + 2 * weights_stride_z));
+    VEC_FLOAT w8 = VLOAD(VEC_SIZE)(0, (__global float *)(weights.ptr + 2 * weights_stride_y + 2 * weights_stride_z));
+
+    // Load input values
+    // z == 0
+    // Clamp z_coord as for z = 0, it can be negative
+    // z_coord is casted to unsigned int in order to use just a min() operation
+    // A "-1" 32 bit signed variable converted to unsigned gives 4294967295
+    z_coord = z * NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP;
+    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
+    offset  = y_offset + (int4)(z_coord * src_stride_z);
+    offset  = min(offset, max_offset);
+
+    VEC_FLOAT values0 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s0));
+    VEC_FLOAT values1 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s1));
+    VEC_FLOAT values2 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s2));
+    VEC_FLOAT values3 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s3));
+
+    // z == 1
+    // z_coord can be only negative for z = 0 so we do not need to clamp it
+    // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset
+    z_coord           = z * NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 1;
+    offset            = y_offset + (int4)(z_coord * src_stride_z);
+    VEC_FLOAT values4 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s0));
+    VEC_FLOAT values5 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s1));
+    VEC_FLOAT values6 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s2));
+    VEC_FLOAT values7 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s3));
+
+    // z == 2
+    // After z = 1 we can simply add src_stride_z to offset without updating z_coord
+    // However offset can be out-of-bound so we need to check if it is greater than max_offset
+    offset += (int4)src_stride_z;
+    offset             = min(offset, max_offset);
+    VEC_FLOAT values8  = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s0));
+    VEC_FLOAT values9  = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s1));
+    VEC_FLOAT values10 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s2));
+    VEC_FLOAT values11 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s3));
+
+    // z == 3
+    // After z = 1 we can simply add src_stride_z to offset without updating z_coord
+    // However offset can be out-of-bound so we need to check if it is greater than max_offset
+    offset += (int4)(src_stride_z);
+    offset             = min(offset, max_offset);
+    VEC_FLOAT values12 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s0));
+    VEC_FLOAT values13 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s1));
+    VEC_FLOAT values14 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s2));
+    VEC_FLOAT values15 = VLOAD(VEC_SIZE)(0, (__global float *)(src_addr + offset.s3));
+
+    acc0 = fma(values0, w0, acc0);
+    acc0 = fma(values1, w1, acc0);
+    acc0 = fma(values2, w2, acc0);
+    acc1 = fma(values1, w0, acc1);
+    acc1 = fma(values2, w1, acc1);
+    acc1 = fma(values3, w2, acc1);
+
+    acc0 = fma(values4, w3, acc0);
+    acc0 = fma(values5, w4, acc0);
+    acc0 = fma(values6, w5, acc0);
+    acc1 = fma(values5, w3, acc1);
+    acc1 = fma(values6, w4, acc1);
+    acc1 = fma(values7, w5, acc1);
+
+    acc0 = fma(values8, w6, acc0);
+    acc0 = fma(values9, w7, acc0);
+    acc0 = fma(values10, w8, acc0);
+    acc1 = fma(values9, w6, acc1);
+    acc1 = fma(values10, w7, acc1);
+    acc1 = fma(values11, w8, acc1);
+
+    acc2 = fma(values4, w0, acc2);
+    acc2 = fma(values5, w1, acc2);
+    acc2 = fma(values6, w2, acc2);
+    acc3 = fma(values5, w0, acc3);
+    acc3 = fma(values6, w1, acc3);
+    acc3 = fma(values7, w2, acc3);
+
+    acc2 = fma(values8, w3, acc2);
+    acc2 = fma(values9, w4, acc2);
+    acc2 = fma(values10, w5, acc2);
+    acc3 = fma(values9, w3, acc3);
+    acc3 = fma(values10, w4, acc3);
+    acc3 = fma(values11, w5, acc3);
+
+    acc2 = fma(values12, w6, acc2);
+    acc2 = fma(values13, w7, acc2);
+    acc2 = fma(values14, w8, acc2);
+    acc3 = fma(values13, w6, acc3);
+    acc3 = fma(values14, w7, acc3);
+    acc3 = fma(values15, w8, acc3);
+
+#if defined(HAS_BIAS)
+    Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
+
+    VEC_FLOAT bias_values = VLOAD(VEC_SIZE)(0, (__global float *)biases.ptr);
+
+    acc0 += bias_values;
+    acc1 += bias_values;
+    acc2 += bias_values;
+    acc3 += bias_values;
+#endif // defined(HAS_BIAS)
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z;
+
+    VSTORE(VEC_SIZE)
+    (acc0, 0, (__global float *)(dst_addr + 0 * dst_stride_y));
+    VSTORE(VEC_SIZE)
+    (acc1, 0, (__global float *)(dst_addr + 1 * dst_stride_y));
+
+#if((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0)
+    if((z * NUM_PLANES_PROCESSED + 1) < DST_DIM_2)
+#endif // ((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0)
+    {
+        VSTORE(VEC_SIZE)
+        (acc2, 0, (__global float *)(dst_addr + 0 * dst_stride_y + 1 * dst_stride_z));
+        VSTORE(VEC_SIZE)
+        (acc3, 0, (__global float *)(dst_addr + 1 * dst_stride_y + 1 * dst_stride_z));
+    }
+}
+
+#endif // defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED)
+#endif // defined(VEC_SIZE) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT)
\ No newline at end of file
diff --git a/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl b/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl
index b58dc7af72..ae87420774 100644
--- a/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl
+++ b/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl
@@ -296,7 +296,12 @@ __kernel void output_stage_quantized(
 
 #if defined(HAS_BIAS)
     // Load and add bias
+#if defined(NCHW)
     int bias_value = *((__global int *)(vector_offset(&bias, get_global_id(2))));
+#else  // defined(NCHW)
+    int16 bias_value = vload16(0, ((__global int *)(vector_offset(&bias, get_global_id(0) * 16))));
+#endif // defined(NCHW)
+
     vals += (int16)(bias_value);
 #endif //defined(HAS_BIAS)
 
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index d24ef0f496..1de08aa1a2 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -44,18 +44,27 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
                           const ActivationLayerInfo &act_info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.enabled()) && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
-                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
-                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU),
-                                    "For QASYMM8 only relu, lower bounded relu and lower-upper bounded relu are supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::QASYMM8);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.enabled()) && (input->data_type() == DataType::F32 || ((act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+                                                                                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+                                                                                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU))),
+                                    "For QASYMM8 only relu, lower bounded relu and lower-upper bounded relu are supported"); //COMPMID-1317 add fused activation for F32
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1); // COMPMID-1071 Add depth multiplier support for NHWC
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(1) != 3 || weights->dimension(2) != 3);
 
+    const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type());
+
     if(biases != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+        if(is_qasymm)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+        }
         ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
     }
@@ -72,12 +81,23 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *output,
                                                         const PadStrideInfo &conv_info)
 {
-    const unsigned int num_rows_processed_per_iteration = 4;
-    const unsigned int num_elems_accessed_per_iteration = 4;
+    const bool is_qasymm   = is_data_type_quantized_asymmetric(input->data_type());
+    const bool is_stride_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
+
+    const unsigned int num_rows_processed_per_iteration = is_qasymm ? 4 : (is_stride_1 ? 2 : 1);
+    const unsigned int num_elems_accessed_per_iteration = is_qasymm ? 4 : 2;
     const unsigned int num_rows_read_per_iteration      = num_rows_processed_per_iteration + 2;
-    const unsigned int num_rows_written_per_iteration   = num_rows_processed_per_iteration / conv_info.stride().first;
+    const unsigned int num_rows_written_per_iteration   = std::ceil(num_rows_processed_per_iteration / static_cast<float>(conv_info.stride().first));
 
-    const BorderSize border_size(std::max(conv_info.pad_left(), conv_info.pad_top()), 0, std::max(conv_info.pad_right(), conv_info.pad_bottom()), 0);
+    BorderSize border_size;
+    if(is_qasymm)
+    {
+        border_size = BorderSize(std::max(conv_info.pad_left(), conv_info.pad_top()), 0, std::max(conv_info.pad_right(), conv_info.pad_bottom()), 0);
+    }
+    else
+    {
+        border_size = BorderSize(conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
+    }
 
     // Configure kernel window
     Window win = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_written_per_iteration));
@@ -103,7 +123,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 } // namespace
 
 CLDepthwiseConvolutionLayer3x3NHWCKernel::CLDepthwiseConvolutionLayer3x3NHWCKernel()
-    : _num_rows_processed_per_iteration(1)
+    : _num_rows_processed_per_iteration(1), _num_planes_processed_per_iteration(1)
 {
 }
 
@@ -135,66 +155,97 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input,
     ARM_COMPUTE_ERROR_ON(conv_stride_x < 1 || conv_stride_x > 2);
     ARM_COMPUTE_ERROR_ON(std::max(conv_info.pad_top(), conv_info.pad_bottom()) > 1);
 
-    _input                            = input;
-    _output                           = output;
-    _weights                          = weights;
-    _biases                           = biases;
-    _conv_stride_y                    = conv_info.stride().second;
-    _num_rows_processed_per_iteration = 4;
+    const bool is_qasymm   = is_data_type_quantized_asymmetric(input->info()->data_type());
+    const bool is_stride_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
 
-    const unsigned int num_elems_accessed_per_iteration = 4;
+    _input                              = input;
+    _output                             = output;
+    _weights                            = weights;
+    _biases                             = biases;
+    _conv_stride_y                      = conv_info.stride().second;
+    _num_rows_processed_per_iteration   = is_qasymm ? 4 : (is_stride_1 ? 2 : 1);
+    _num_planes_processed_per_iteration = (is_stride_1 && !is_qasymm) ? 2 : 1;
 
-    _border_size = BorderSize(std::max(conv_info.pad_left(), conv_info.pad_top()), 0, std::max(conv_info.pad_right(), conv_info.pad_bottom()), 0);
+    const unsigned int num_elems_accessed_per_iteration = is_qasymm ? 4 : 2;
 
-    float multiplier        = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale;
-    int   output_multiplier = 0;
-    int   output_shift      = 0;
-    quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+    if(is_qasymm)
+    {
+        _border_size = BorderSize(std::max(conv_info.pad_left(), conv_info.pad_top()), 0, std::max(conv_info.pad_right(), conv_info.pad_bottom()), 0);
+    }
+    else
+    {
+        _border_size = BorderSize(conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
+    }
 
     CLBuildOptions build_opts;
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
-    build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-_input->info()->quantization_info().offset));
-    build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-_weights->info()->quantization_info().offset));
-    build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(_output->info()->quantization_info().offset));
-    build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * input->info()->quantization_info().offset * weights->info()->quantization_info().offset));
-    build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
-    build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_accessed_per_iteration));
-    build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
     build_opts.add_option("-DSRC_DIM_2=" + support::cpp11::to_string(_input->info()->dimension(2)));
     build_opts.add_option("-DCONV_PAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
     build_opts.add_option("-DCONV_PAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
 
-    if(act_info.enabled())
+    if(is_qasymm)
     {
-        const int a_val = input->info()->quantization_info().quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
-        const int b_val = input->info()->quantization_info().quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
-        const int o1    = input->info()->quantization_info().offset;
-
-        build_opts.add_option("-DFUSED_ACTIVATION=" + lower_string(string_from_activation_func(act_info.activation())));
-        build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
-        build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
-        build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
-
-        if(output != nullptr)
+        float multiplier        = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale;
+        int   output_multiplier = 0;
+        int   output_shift      = 0;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+        build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
+        build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-_input->info()->quantization_info().offset));
+        build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-_weights->info()->quantization_info().offset));
+        build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(_output->info()->quantization_info().offset));
+        build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * input->info()->quantization_info().offset * weights->info()->quantization_info().offset));
+        build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
+        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
+
+        if(act_info.enabled())
         {
-            const float s1 = input->info()->quantization_info().scale;
-            const float s2 = output->info()->quantization_info().scale;
-            const int   o2 = output->info()->quantization_info().offset;
+            const int a_val = input->info()->quantization_info().quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
+            const int b_val = input->info()->quantization_info().quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
+            const int o1    = input->info()->quantization_info().offset;
+
+            build_opts.add_option("-DFUSED_ACTIVATION=" + lower_string(string_from_activation_func(act_info.activation())));
+            build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
+            build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
+            build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
 
-            if(o1 != o2 || s1 != s2)
+            if(output != nullptr)
             {
-                build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
-                build_opts.add_option("-DS2_VAL=" + float_to_string_with_full_precision(s2));
-                build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
-                build_opts.add_option("-DO2_VAL=" + support::cpp11::to_string(o2));
+                const float s1 = input->info()->quantization_info().scale;
+                const float s2 = output->info()->quantization_info().scale;
+                const int   o2 = output->info()->quantization_info().offset;
+
+                if(o1 != o2 || s1 != s2)
+                {
+                    build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
+                    build_opts.add_option("-DS2_VAL=" + float_to_string_with_full_precision(s2));
+                    build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
+                    build_opts.add_option("-DO2_VAL=" + support::cpp11::to_string(o2));
+                }
             }
         }
     }
+    else if(is_stride_1)
+    {
+        build_opts.add_option("-DNUM_ROWS_PROCESSED=" + support::cpp11::to_string(_num_rows_processed_per_iteration));
+        build_opts.add_option("-DNUM_PLANES_PROCESSED=" + support::cpp11::to_string(_num_planes_processed_per_iteration));
+        build_opts.add_option("-DDST_DIM_2=" + support::cpp11::to_string(_output->info()->dimension(2)));
+    }
+    else
+    {
+        build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_stride_x));
+        build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(_conv_stride_y));
+    }
 
     // Create kernel
-    std::string kernel_name = std::string("depthwise_convolution_3x3_quantized_nhwc_stride") + support::cpp11::to_string(conv_stride_x);
-    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+    std::string kernel_name = std::string("depthwise_convolution_3x3") + (is_qasymm ? std::string("_quantized") : std::string()) + std::string("_nhwc");
+    if(is_qasymm || is_stride_1)
+    {
+        kernel_name += std::string("_stride") + support::cpp11::to_string(conv_stride_x);
+    }
+
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Configure kernel window
     auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info);
@@ -213,6 +264,8 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input,
     _config_id += support::cpp11::to_string(output->info()->dimension(0));
     _config_id += "_";
     _config_id += support::cpp11::to_string(output->info()->dimension(1));
+    _config_id += "_";
+    _config_id += string_from_data_type(input->info()->data_type());
 }
 
 Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
@@ -233,15 +286,18 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::Com
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
 
+    Window win = window;
+    win.set(Window::DimZ, Window::Dimension(0, std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)), 1));
+
     // Create input window and adjust
-    Window win_in = window;
+    Window win_in = win;
     win_in.set_dimension_step(Window::DimY, _num_rows_processed_per_iteration);
     win_in.set_dimension_step(Window::DimZ, _conv_stride_y);
 
     ARM_COMPUTE_ERROR_ON((win_in.y().step() < window.y().step()) || (win_in.z().step() < window.z().step()));
 
     Window slice_in  = win_in.first_slice_window_3D();
-    Window slice_out = window.first_slice_window_3D();
+    Window slice_out = win.first_slice_window_3D();
 
     if(_biases != nullptr)
     {
@@ -252,6 +308,15 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::Com
         add_1D_tensor_argument(idx, _biases, win_biases);
     }
 
+    if(!(is_data_type_quantized_asymmetric(_input->info()->data_type())))
+    {
+        unsigned int idx        = 3 * num_arguments_per_3D_tensor() + ((_biases != nullptr) ? num_arguments_per_1D_tensor() : 0);
+        const int    max_offset = _input->info()->strides_in_bytes().z() * _input->info()->dimension(2) - (_input->info()->padding().bottom + _input->info()->padding().top) *
+                                  _input->info()->strides_in_bytes().y();
+
+        _kernel.setArg(idx, max_offset);
+    }
+
     do
     {
         unsigned int idx = 0;
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index c89b16eedc..bef13f9b1c 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -47,13 +47,15 @@ namespace
 {
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier)
 {
+    const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
+
     ARM_COMPUTE_UNUSED(conv_info);
     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, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && has_bias);
-    ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(2) * depth_multiplier) != output->dimension(2));
+    ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(idx_c) * depth_multiplier) != output->dimension(2));
     ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0)));
 
     return Status{};
@@ -68,6 +70,10 @@ void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *outpu
     _input  = input;
     _output = output;
 
+    const DataLayout data_layout = input->info()->data_layout();
+    const size_t     idx_w       = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const size_t     idx_h       = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+
     // Create kernel
     CLBuildOptions build_opts;
 
@@ -78,11 +84,12 @@ void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *outpu
     build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
     build_opts.add_option("-DPAD_RIGHT=" + support::cpp11::to_string(conv_info.pad_right()));
     build_opts.add_option("-DPAD_BOTTOM=" + support::cpp11::to_string(conv_info.pad_bottom()));
-    build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
-    build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
+    build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_w)));
+    build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(idx_h)));
     build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
     build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
     build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
+    build_opts.add_option("-D" + string_from_data_layout(input->info()->data_layout()));
     build_opts.add_option_if(has_bias, "-DHAS_BIAS");
     build_opts.add_option_if_else(is_data_type_quantized_asymmetric(input->info()->data_type()),
                                   "-DPAD_VALUE=" + support::cpp11::to_string(input->info()->quantization_info().offset),
diff --git a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
index 0d158f1dab..c97ecaf8e0 100644
--- a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
@@ -37,12 +37,16 @@ using namespace arm_compute;
 
 namespace
 {
-TensorShape compute_output_shape(const TensorShape &input, size_t conv_w, size_t conv_h)
+TensorShape compute_output_shape(const TensorShape &input, size_t conv_w, size_t conv_h, const DataLayout &data_layout)
 {
+    const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+    const size_t idx_c = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
     TensorShape output_shape(input);
-    output_shape.set(0, conv_w);
-    output_shape.set(1, conv_h);
-    output_shape.set(2, input.x() / (conv_w * conv_h));
+    output_shape.set(idx_w, conv_w);
+    output_shape.set(idx_h, conv_h);
+    output_shape.set(idx_c, input.x() / (conv_w * conv_h));
 
     return output_shape;
 }
@@ -54,7 +58,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, s
 
     if(output->total_size() != 0)
     {
-        TensorShape output_shape = compute_output_shape(input->tensor_shape(), conv_w, conv_h);
+        TensorShape output_shape = compute_output_shape(input->tensor_shape(), conv_w, conv_h, output->data_layout());
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
@@ -74,7 +78,7 @@ void CLDepthwiseVectorToTensorKernel::configure(const ICLTensor *input, ICLTenso
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
     // Output auto inizialitation if not yet initialized
-    TensorShape output_shape = compute_output_shape(input->info()->tensor_shape(), conv_w, conv_h);
+    TensorShape output_shape = compute_output_shape(input->info()->tensor_shape(), conv_w, conv_h, output->info()->data_layout());
     auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
 
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), conv_w, conv_h));
@@ -87,6 +91,7 @@ void CLDepthwiseVectorToTensorKernel::configure(const ICLTensor *input, ICLTenso
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option("-DCONV_WIDTH=" + support::cpp11::to_string(conv_w));
     build_opts.add_option("-DCONV_HEIGHT=" + support::cpp11::to_string(conv_h));
+    build_opts.add_option("-D" + string_from_data_layout(output->info()->data_layout()));
 
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_vector_to_tensor", build_opts.options()));
 
diff --git a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
index 59c45adc72..fd3b75484a 100644
--- a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
@@ -39,19 +39,23 @@ namespace
 {
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases)
 {
+    const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
+    const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
+
     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, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && (biases != nullptr));
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) != output->dimension(1));
-    ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (input->dimension(0) * input->dimension(1) + ((biases != nullptr) ? 1 : 0)));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_c) != output->dimension(1));
+    ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (input->dimension(idx_w) * input->dimension(idx_h) + ((biases != nullptr) ? 1 : 0)));
 
     if(biases != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != input->dimension(2));
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != input->dimension(idx_c));
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
     }
 
@@ -73,11 +77,14 @@ void CLDepthwiseWeightsReshapeKernel::configure(const ICLTensor *input, ICLTenso
     _biases = biases;
     _output = output;
 
+    const size_t idx_w = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
+
     // Create kernel
     std::set<std::string> build_opts;
 
     build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-    build_opts.emplace("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
+    build_opts.emplace("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_w)));
+    build_opts.emplace("-D" + string_from_data_layout(input->info()->data_layout()));
     if(_biases != nullptr)
     {
         build_opts.emplace("-DHAS_BIAS");
@@ -107,10 +114,14 @@ void CLDepthwiseWeightsReshapeKernel::run(const Window &window, cl::CommandQueue
     Window slice     = window.first_slice_window_3D();
     Window slice_out = window.first_slice_window_2D();
 
+    const size_t idx_w = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::HEIGHT);
+    const size_t idx_c = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::CHANNEL);
+
     // Setup slice
-    slice.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _input->info()->dimension(0)));
-    slice.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), 1));
-    slice.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), 1));
+    slice.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(idx_w), _input->info()->dimension(idx_w)));
+    slice.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(idx_h), 1));
+    slice.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(idx_c), 1));
 
     // Setup output slice
     // The first two dimensions of the output are increased by the inner loops
diff --git a/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp
index 1a6dc14850..c6147ee318 100644
--- a/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp
+++ b/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp
@@ -168,6 +168,7 @@ void CLDirectConvolutionLayerOutputStageKernel::configure(ICLTensor *input, cons
     // Create kernel
     CLBuildOptions build_opts;
     build_opts.add_option_if(bias != nullptr, "-DHAS_BIAS");
+    build_opts.add_option("-D" + string_from_data_layout(input->info()->data_layout()));
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("output_stage_quantized", build_opts.options()));
 
     // Set static kernel arguments