diff --git a/src/core/CL/cl_kernels/winograd_filter_transform.cl b/src/core/CL/cl_kernels/winograd_filter_transform.cl
index 73da005..3b9b1e9 100644
--- a/src/core/CL/cl_kernels/winograd_filter_transform.cl
+++ b/src/core/CL/cl_kernels/winograd_filter_transform.cl
@@ -30,8 +30,9 @@
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note If this kernel is used to perform Winograd filter transform 3x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd filter transform 1x3, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -60,45 +61,54 @@
 
     // Load the values from the input tensor
 #if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
-    float3 w0 = vload3(0, (__global float *)(src_addr));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w0 = vload3(0, (__global DATA_TYPE *)(src_addr));
 #elif defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float3 w0 = (float3)(*((__global float *)(src_addr + 0 * src_stride_y)),
-                         *((__global float *)(src_addr + 1 * src_stride_y)),
-                         *((__global float *)(src_addr + 2 * src_stride_y)));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w0 = (VEC_DATA_TYPE(DATA_TYPE, 3))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)),
+                                       *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)),
+                                       *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)));
 #else  // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float3 w0  = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float3 w1  = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float3 w2  = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w0 = vload3(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w1 = vload3(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w2 = vload3(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
 #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
 
     // Row 0
-    float4 out0 = 0.0f;
-    out0.s0     = (w0.s0);
-    out0.s1     = (w0.s0 + w0.s1 + w0.s2) * 0.5f;
-    out0.s2     = (w0.s0 + w0.s2 - w0.s1) * 0.5f;
-    out0.s3     = (w0.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out0    = 0.0f;
+    out0.s0 = (w0.s0);
+    out0.s1 = (w0.s0 + w0.s1 + w0.s2) * 0.5f;
+    out0.s2 = (w0.s0 + w0.s2 - w0.s1) * 0.5f;
+    out0.s3 = (w0.s2);
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
     // Row 1
-    float4 out1 = 0.0f;
-    out1.s0     = (w0.s0 + w1.s0 + w2.s0) * 0.5f;
-    out1.s1     = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) * 0.25f;
-    out1.s2     = (w0.s0 + w1.s0 + w2.s0 + w0.s2 + w1.s2 + w2.s2 - w0.s1 - w1.s1 - w2.s1) * 0.25f;
-    out1.s3     = (w0.s2 + w1.s2 + w2.s2) * 0.5f;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out1    = 0.0f;
+    out1.s0 = (w0.s0 + w1.s0 + w2.s0) * 0.5f;
+    out1.s1 = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) * 0.25f;
+    out1.s2 = (w0.s0 + w1.s0 + w2.s0 + w0.s2 + w1.s2 + w2.s2 - w0.s1 - w1.s1 - w2.s1) * 0.25f;
+    out1.s3 = (w0.s2 + w1.s2 + w2.s2) * 0.5f;
 
     // Row 2
-    float4 out2 = 0.0f;
-    out2.s0     = (w0.s0 + w2.s0 - w1.s0) * 0.5f;
-    out2.s1     = (w0.s0 + w2.s0 + w0.s1 + w2.s1 + w0.s2 + w2.s2 - w1.s0 - w1.s1 - w1.s2) * 0.25f;
-    out2.s2     = (w0.s0 + w2.s0 + w1.s1 + w0.s2 + w2.s2 - w1.s0 - w0.s1 - w2.s1 - w1.s2) * 0.25f;
-    out2.s3     = (w0.s2 + w2.s2 - w1.s2) * 0.5f;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out2    = 0.0f;
+    out2.s0 = (w0.s0 + w2.s0 - w1.s0) * 0.5f;
+    out2.s1 = (w0.s0 + w2.s0 + w0.s1 + w2.s1 + w0.s2 + w2.s2 - w1.s0 - w1.s1 - w1.s2) * 0.25f;
+    out2.s2 = (w0.s0 + w2.s0 + w1.s1 + w0.s2 + w2.s2 - w1.s0 - w0.s1 - w2.s1 - w1.s2) * 0.25f;
+    out2.s3 = (w0.s2 + w2.s2 - w1.s2) * 0.5f;
 
     // Row 3
-    float4 out3 = 0.0f;
-    out3.s0     = (w2.s0);
-    out3.s1     = (w2.s0 + w2.s1 + w2.s2) * 0.5f;
-    out3.s2     = (w2.s0 + w2.s2 - w2.s1) * 0.5f;
-    out3.s3     = (w2.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out3    = 0.0f;
+    out3.s0 = (w2.s0);
+    out3.s1 = (w2.s0 + w2.s1 + w2.s2) * 0.5f;
+    out3.s2 = (w2.s0 + w2.s2 - w2.s1) * 0.5f;
+    out3.s3 = (w2.s2);
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 
     int z  = get_global_id(2);
@@ -111,24 +121,24 @@
     // Store the values across the channels
     // 16 channels for 3x3 kernels
     // 4 channels for 3x1 or 1x3 kernels
-    *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0;
-    *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1;
-    *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2;
-    *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3;
+    *(__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z) = out0.s0;
+    *(__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z) = out0.s1;
+    *(__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z) = out0.s2;
+    *(__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z) = out0.s3;
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    *(__global float *)(dst_addr + 4 * dst_stride_z)  = out1.s0;
-    *(__global float *)(dst_addr + 5 * dst_stride_z)  = out1.s1;
-    *(__global float *)(dst_addr + 6 * dst_stride_z)  = out1.s2;
-    *(__global float *)(dst_addr + 7 * dst_stride_z)  = out1.s3;
-    *(__global float *)(dst_addr + 8 * dst_stride_z)  = out2.s0;
-    *(__global float *)(dst_addr + 9 * dst_stride_z)  = out2.s1;
-    *(__global float *)(dst_addr + 10 * dst_stride_z) = out2.s2;
-    *(__global float *)(dst_addr + 11 * dst_stride_z) = out2.s3;
-    *(__global float *)(dst_addr + 12 * dst_stride_z) = out3.s0;
-    *(__global float *)(dst_addr + 13 * dst_stride_z) = out3.s1;
-    *(__global float *)(dst_addr + 14 * dst_stride_z) = out3.s2;
-    *(__global float *)(dst_addr + 15 * dst_stride_z) = out3.s3;
+    *(__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z)  = out1.s0;
+    *(__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z)  = out1.s1;
+    *(__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z)  = out1.s2;
+    *(__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z)  = out1.s3;
+    *(__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z)  = out2.s0;
+    *(__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z)  = out2.s1;
+    *(__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z) = out2.s2;
+    *(__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z) = out2.s3;
+    *(__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z) = out3.s0;
+    *(__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z) = out3.s1;
+    *(__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z) = out3.s2;
+    *(__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z) = out3.s3;
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 }
 
@@ -137,8 +147,9 @@
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note If this kernel is used to perform Winograd filter transform 3x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd filter transform 1x3, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -167,71 +178,82 @@
 
     // Load the values from the input tensor
 #if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
-    float3 w0 = vload3(0, (__global float *)(src_addr));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w0 = vload3(0, (__global DATA_TYPE *)(src_addr));
 #elif defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float3 w0 = (float3)(*((__global float *)(src_addr + 0 * src_stride_y)),
-                         *((__global float *)(src_addr + 1 * src_stride_y)),
-                         *((__global float *)(src_addr + 2 * src_stride_y)));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w0 = (VEC_DATA_TYPE(DATA_TYPE, 3))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)),
+                                       *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)),
+                                       *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)));
 #else  // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float3 w0  = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float3 w1  = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float3 w2  = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w0 = vload3(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w1 = vload3(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 3)
+    w2 = vload3(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
 #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
 
     // Row 0
-    float8 out0 = 0.0f;
-    out0.s0     = (w0.s0) / 16.f;
-    out0.s1     = (-w0.s0 - w0.s1 - w0.s2) / 24.f;
-    out0.s2     = (-w0.s0 + w0.s1 - w0.s2) / 24.f;
-    out0.s3     = (w0.s0 + 2.f * w0.s1 + 4.f * w0.s2) / 96.f;
-    out0.s4     = (w0.s0 - 2.f * w0.s1 + 4.f * w0.s2) / 96.f;
-    out0.s5     = (w0.s2) / 4.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out0    = 0.0f;
+    out0.s0 = (w0.s0) / 16.f;
+    out0.s1 = (-w0.s0 - w0.s1 - w0.s2) / 24.f;
+    out0.s2 = (-w0.s0 + w0.s1 - w0.s2) / 24.f;
+    out0.s3 = (w0.s0 + 2.f * w0.s1 + 4.f * w0.s2) / 96.f;
+    out0.s4 = (w0.s0 - 2.f * w0.s1 + 4.f * w0.s2) / 96.f;
+    out0.s5 = (w0.s2) / 4.f;
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
     // Row 1
-    float8 out1 = 0.0f;
-    out1.s0     = (-w0.s0 - w1.s0 - w2.s0) / 24.f;
-    out1.s1     = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) / 36.f;
-    out1.s2     = (w0.s0 + w1.s0 + w2.s0 - w0.s1 - w1.s1 - w2.s1 + w0.s2 + w1.s2 + w2.s2) / 36.f;
-    out1.s3     = (-w0.s0 - w1.s0 - w2.s0 + 2.f * (-w0.s1 - w1.s1 - w2.s1) + 4.f * (-w0.s2 - w1.s2 - w2.s2)) / 144.f;
-    out1.s4     = (-w0.s0 - w1.s0 - w2.s0 + 2.f * (w0.s1 + w1.s1 + w2.s1) + 4.f * (-w0.s2 - w1.s2 - w2.s2)) / 144.f;
-    out1.s5     = (-w0.s2 - w1.s2 - w2.s2) / 6.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out1    = 0.0f;
+    out1.s0 = (-w0.s0 - w1.s0 - w2.s0) / 24.f;
+    out1.s1 = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) / 36.f;
+    out1.s2 = (w0.s0 + w1.s0 + w2.s0 - w0.s1 - w1.s1 - w2.s1 + w0.s2 + w1.s2 + w2.s2) / 36.f;
+    out1.s3 = (-w0.s0 - w1.s0 - w2.s0 + 2.f * (-w0.s1 - w1.s1 - w2.s1) + 4.f * (-w0.s2 - w1.s2 - w2.s2)) / 144.f;
+    out1.s4 = (-w0.s0 - w1.s0 - w2.s0 + 2.f * (w0.s1 + w1.s1 + w2.s1) + 4.f * (-w0.s2 - w1.s2 - w2.s2)) / 144.f;
+    out1.s5 = (-w0.s2 - w1.s2 - w2.s2) / 6.f;
 
     // Row 2
-    float8 out2 = 0.0f;
-    out2.s0     = (-w0.s0 + w1.s0 - w2.s0) / 24.f;
-    out2.s1     = (w0.s0 - w1.s0 + w2.s0 + w0.s1 - w1.s1 + w2.s1 + w0.s2 - w1.s2 + w2.s2) / 36.f;
-    out2.s2     = (w0.s0 - w1.s0 + w2.s0 - w0.s1 + w1.s1 - w2.s1 + w0.s2 - w1.s2 + w2.s2) / 36.f;
-    out2.s3     = (-w0.s0 + w1.s0 - w2.s0 + 2.f * (-w0.s1 + w1.s1 - w2.s1) + 4.f * (-w0.s2 + w1.s2 - w2.s2)) / 144.f;
-    out2.s4     = (-w0.s0 + w1.s0 - w2.s0 + 2.f * (w0.s1 - w1.s1 + w2.s1) + 4.f * (-w0.s2 + w1.s2 - w2.s2)) / 144.f;
-    out2.s5     = (-w0.s2 + w1.s2 - w2.s2) / 6.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out2    = 0.0f;
+    out2.s0 = (-w0.s0 + w1.s0 - w2.s0) / 24.f;
+    out2.s1 = (w0.s0 - w1.s0 + w2.s0 + w0.s1 - w1.s1 + w2.s1 + w0.s2 - w1.s2 + w2.s2) / 36.f;
+    out2.s2 = (w0.s0 - w1.s0 + w2.s0 - w0.s1 + w1.s1 - w2.s1 + w0.s2 - w1.s2 + w2.s2) / 36.f;
+    out2.s3 = (-w0.s0 + w1.s0 - w2.s0 + 2.f * (-w0.s1 + w1.s1 - w2.s1) + 4.f * (-w0.s2 + w1.s2 - w2.s2)) / 144.f;
+    out2.s4 = (-w0.s0 + w1.s0 - w2.s0 + 2.f * (w0.s1 - w1.s1 + w2.s1) + 4.f * (-w0.s2 + w1.s2 - w2.s2)) / 144.f;
+    out2.s5 = (-w0.s2 + w1.s2 - w2.s2) / 6.f;
 
     // Row 3
-    float8 out3 = 0.0f;
-    out3.s0     = (w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) / 96.f;
-    out3.s1     = (-w0.s0 - 2.f * w1.s0 - 4.f * w2.s0 - w0.s1 - 2.f * w1.s1 - 4.f * w2.s1 - w0.s2 - 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
-    out3.s2     = (-w0.s0 - 2.f * w1.s0 - 4.f * w2.s0 + w0.s1 + 2.f * w1.s1 + 4.f * w2.s1 - w0.s2 - 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
-    out3.s3     = ((w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (w0.s1 + 2.f * w1.s1 + 4.f * w2.s1) + 4.f * (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
-    out3.s4     = ((w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (-w0.s1 - 2.f * w1.s1 - 4.f * w2.s1) + 4.f * (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
-    out3.s5     = (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2) / 24.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out3    = 0.0f;
+    out3.s0 = (w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) / 96.f;
+    out3.s1 = (-w0.s0 - 2.f * w1.s0 - 4.f * w2.s0 - w0.s1 - 2.f * w1.s1 - 4.f * w2.s1 - w0.s2 - 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
+    out3.s2 = (-w0.s0 - 2.f * w1.s0 - 4.f * w2.s0 + w0.s1 + 2.f * w1.s1 + 4.f * w2.s1 - w0.s2 - 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
+    out3.s3 = ((w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (w0.s1 + 2.f * w1.s1 + 4.f * w2.s1) + 4.f * (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
+    out3.s4 = ((w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (-w0.s1 - 2.f * w1.s1 - 4.f * w2.s1) + 4.f * (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
+    out3.s5 = (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2) / 24.f;
 
     // Row 4
-    float8 out4 = 0.0f;
-    out4.s0     = (w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) / 96.f;
-    out4.s1     = (-w0.s0 + 2.f * w1.s0 - 4.f * w2.s0 - w0.s1 + 2.f * w1.s1 - 4.f * w2.s1 - w0.s2 + 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
-    out4.s2     = (-w0.s0 + 2.f * w1.s0 - 4.f * w2.s0 + w0.s1 - 2.f * w1.s1 + 4.f * w2.s1 - w0.s2 + 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
-    out4.s3     = ((w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (w0.s1 - 2.f * w1.s1 + 4.f * w2.s1) + 4.f * (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
-    out4.s4     = ((w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (-w0.s1 + 2.f * w1.s1 - 4.f * w2.s1) + 4.f * (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
-    out4.s5     = (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2) / 24.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out4    = 0.0f;
+    out4.s0 = (w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) / 96.f;
+    out4.s1 = (-w0.s0 + 2.f * w1.s0 - 4.f * w2.s0 - w0.s1 + 2.f * w1.s1 - 4.f * w2.s1 - w0.s2 + 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
+    out4.s2 = (-w0.s0 + 2.f * w1.s0 - 4.f * w2.s0 + w0.s1 - 2.f * w1.s1 + 4.f * w2.s1 - w0.s2 + 2.f * w1.s2 - 4.f * w2.s2) / 144.f;
+    out4.s3 = ((w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (w0.s1 - 2.f * w1.s1 + 4.f * w2.s1) + 4.f * (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
+    out4.s4 = ((w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (-w0.s1 + 2.f * w1.s1 - 4.f * w2.s1) + 4.f * (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2)) / 576.f;
+    out4.s5 = (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2) / 24.f;
 
     // Row 5
-    float8 out5 = 0.0f;
-    out5.s0     = (w2.s0) / 4.f;
-    out5.s1     = (-w2.s0 - w2.s1 - w2.s2) / 6.f;
-    out5.s2     = (-w2.s0 + w2.s1 - w2.s2) / 6.f;
-    out5.s3     = (w2.s0 + 2.f * w2.s1 + 4.f * w2.s2) / 24.f;
-    out5.s4     = (w2.s0 - 2.f * w2.s1 + 4.f * w2.s2) / 24.f;
-    out5.s5     = (w2.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out5    = 0.0f;
+    out5.s0 = (w2.s0) / 4.f;
+    out5.s1 = (-w2.s0 - w2.s1 - w2.s2) / 6.f;
+    out5.s2 = (-w2.s0 + w2.s1 - w2.s2) / 6.f;
+    out5.s3 = (w2.s0 + 2.f * w2.s1 + 4.f * w2.s2) / 24.f;
+    out5.s4 = (w2.s0 - 2.f * w2.s1 + 4.f * w2.s2) / 24.f;
+    out5.s5 = (w2.s2);
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 
     int z  = get_global_id(2);
@@ -244,44 +266,44 @@
     // Store the values across the channels
     // 36 channels for 3x3 kernels
     // 6 channels for 3x1 or 1x3 kernels
-    *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0;
-    *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1;
-    *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2;
-    *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3;
-    *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4;
-    *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5;
+    *(__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z) = out0.s0;
+    *(__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z) = out0.s1;
+    *(__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z) = out0.s2;
+    *(__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z) = out0.s3;
+    *(__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z) = out0.s4;
+    *(__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z) = out0.s5;
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    *(__global float *)(dst_addr + 6 * dst_stride_z)  = out1.s0;
-    *(__global float *)(dst_addr + 7 * dst_stride_z)  = out1.s1;
-    *(__global float *)(dst_addr + 8 * dst_stride_z)  = out1.s2;
-    *(__global float *)(dst_addr + 9 * dst_stride_z)  = out1.s3;
-    *(__global float *)(dst_addr + 10 * dst_stride_z) = out1.s4;
-    *(__global float *)(dst_addr + 11 * dst_stride_z) = out1.s5;
-    *(__global float *)(dst_addr + 12 * dst_stride_z) = out2.s0;
-    *(__global float *)(dst_addr + 13 * dst_stride_z) = out2.s1;
-    *(__global float *)(dst_addr + 14 * dst_stride_z) = out2.s2;
-    *(__global float *)(dst_addr + 15 * dst_stride_z) = out2.s3;
-    *(__global float *)(dst_addr + 16 * dst_stride_z) = out2.s4;
-    *(__global float *)(dst_addr + 17 * dst_stride_z) = out2.s5;
-    *(__global float *)(dst_addr + 18 * dst_stride_z) = out3.s0;
-    *(__global float *)(dst_addr + 19 * dst_stride_z) = out3.s1;
-    *(__global float *)(dst_addr + 20 * dst_stride_z) = out3.s2;
-    *(__global float *)(dst_addr + 21 * dst_stride_z) = out3.s3;
-    *(__global float *)(dst_addr + 22 * dst_stride_z) = out3.s4;
-    *(__global float *)(dst_addr + 23 * dst_stride_z) = out3.s5;
-    *(__global float *)(dst_addr + 24 * dst_stride_z) = out4.s0;
-    *(__global float *)(dst_addr + 25 * dst_stride_z) = out4.s1;
-    *(__global float *)(dst_addr + 26 * dst_stride_z) = out4.s2;
-    *(__global float *)(dst_addr + 27 * dst_stride_z) = out4.s3;
-    *(__global float *)(dst_addr + 28 * dst_stride_z) = out4.s4;
-    *(__global float *)(dst_addr + 29 * dst_stride_z) = out4.s5;
-    *(__global float *)(dst_addr + 30 * dst_stride_z) = out5.s0;
-    *(__global float *)(dst_addr + 31 * dst_stride_z) = out5.s1;
-    *(__global float *)(dst_addr + 32 * dst_stride_z) = out5.s2;
-    *(__global float *)(dst_addr + 33 * dst_stride_z) = out5.s3;
-    *(__global float *)(dst_addr + 34 * dst_stride_z) = out5.s4;
-    *(__global float *)(dst_addr + 35 * dst_stride_z) = out5.s5;
+    *(__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z)  = out1.s0;
+    *(__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z)  = out1.s1;
+    *(__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z)  = out1.s2;
+    *(__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z)  = out1.s3;
+    *(__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z) = out1.s4;
+    *(__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z) = out1.s5;
+    *(__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z) = out2.s0;
+    *(__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z) = out2.s1;
+    *(__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z) = out2.s2;
+    *(__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z) = out2.s3;
+    *(__global DATA_TYPE *)(dst_addr + 16 * dst_stride_z) = out2.s4;
+    *(__global DATA_TYPE *)(dst_addr + 17 * dst_stride_z) = out2.s5;
+    *(__global DATA_TYPE *)(dst_addr + 18 * dst_stride_z) = out3.s0;
+    *(__global DATA_TYPE *)(dst_addr + 19 * dst_stride_z) = out3.s1;
+    *(__global DATA_TYPE *)(dst_addr + 20 * dst_stride_z) = out3.s2;
+    *(__global DATA_TYPE *)(dst_addr + 21 * dst_stride_z) = out3.s3;
+    *(__global DATA_TYPE *)(dst_addr + 22 * dst_stride_z) = out3.s4;
+    *(__global DATA_TYPE *)(dst_addr + 23 * dst_stride_z) = out3.s5;
+    *(__global DATA_TYPE *)(dst_addr + 24 * dst_stride_z) = out4.s0;
+    *(__global DATA_TYPE *)(dst_addr + 25 * dst_stride_z) = out4.s1;
+    *(__global DATA_TYPE *)(dst_addr + 26 * dst_stride_z) = out4.s2;
+    *(__global DATA_TYPE *)(dst_addr + 27 * dst_stride_z) = out4.s3;
+    *(__global DATA_TYPE *)(dst_addr + 28 * dst_stride_z) = out4.s4;
+    *(__global DATA_TYPE *)(dst_addr + 29 * dst_stride_z) = out4.s5;
+    *(__global DATA_TYPE *)(dst_addr + 30 * dst_stride_z) = out5.s0;
+    *(__global DATA_TYPE *)(dst_addr + 31 * dst_stride_z) = out5.s1;
+    *(__global DATA_TYPE *)(dst_addr + 32 * dst_stride_z) = out5.s2;
+    *(__global DATA_TYPE *)(dst_addr + 33 * dst_stride_z) = out5.s3;
+    *(__global DATA_TYPE *)(dst_addr + 34 * dst_stride_z) = out5.s4;
+    *(__global DATA_TYPE *)(dst_addr + 35 * dst_stride_z) = out5.s5;
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 }
 
@@ -290,8 +312,9 @@
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note If this kernel is used to perform Winograd filter transform 3x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd filter transform 1x3, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -320,25 +343,25 @@
 
     // Load the values from the input tensor
 #if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float w00 = *((__global float *)(src_addr + 0 * src_stride_z));
-    float w01 = *((__global float *)(src_addr + 1 * src_stride_z));
-    float w02 = *((__global float *)(src_addr + 2 * src_stride_z));
+    DATA_TYPE w00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z));
+    DATA_TYPE w01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z));
+    DATA_TYPE w02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z));
 #else // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float  w00 = *((__global float *)(src_addr + 0 * src_stride_z + 0 * src_stride_y));
-    float  w01 = *((__global float *)(src_addr + 0 * src_stride_z + 1 * src_stride_y));
-    float  w02 = *((__global float *)(src_addr + 0 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w01 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w02 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z + 2 * src_stride_y));
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
-    float  w10 = *((__global float *)(src_addr + 1 * src_stride_z + 0 * src_stride_y));
-    float  w11 = *((__global float *)(src_addr + 1 * src_stride_z + 1 * src_stride_y));
-    float  w12 = *((__global float *)(src_addr + 1 * src_stride_z + 2 * src_stride_y));
-    float  w20 = *((__global float *)(src_addr + 2 * src_stride_z + 0 * src_stride_y));
-    float  w21 = *((__global float *)(src_addr + 2 * src_stride_z + 1 * src_stride_y));
-    float  w22 = *((__global float *)(src_addr + 2 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w10 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w11 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w12 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w20 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w21 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w22 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 2 * src_stride_y));
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
 #endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 
     // Row 0
-    float out00, out01, out02, out03, out04, out05;
+    DATA_TYPE out00, out01, out02, out03, out04, out05;
     out00 = (w00) / 16.f;
     out01 = (-w00 - w01 - w02) / 24.f;
     out02 = (-w00 + w01 - w02) / 24.f;
@@ -348,7 +371,7 @@
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
     // Row 1
-    float out10, out11, out12, out13, out14, out15;
+    DATA_TYPE out10, out11, out12, out13, out14, out15;
     out10 = (-w00 - w10 - w20) / 24.f;
     out11 = (w00 + w10 + w20 + w01 + w11 + w21 + w02 + w12 + w22) / 36.f;
     out12 = (w00 + w10 + w20 - w01 - w11 - w21 + w02 + w12 + w22) / 36.f;
@@ -357,7 +380,7 @@
     out15 = (-w02 - w12 - w22) / 6.f;
 
     // Row 2
-    float out20, out21, out22, out23, out24, out25;
+    DATA_TYPE out20, out21, out22, out23, out24, out25;
     out20 = (-w00 + w10 - w20) / 24.f;
     out21 = (w00 - w10 + w20 + w01 - w11 + w21 + w02 - w12 + w22) / 36.f;
     out22 = (w00 - w10 + w20 - w01 + w11 - w21 + w02 - w12 + w22) / 36.f;
@@ -366,7 +389,7 @@
     out25 = (-w02 + w12 - w22) / 6.f;
 
     // Row 3
-    float out30, out31, out32, out33, out34, out35;
+    DATA_TYPE out30, out31, out32, out33, out34, out35;
     out30 = (w00 + 2.f * w10 + 4.f * w20) / 96.f;
     out31 = (-w00 - 2.f * w10 - 4.f * w20 - w01 - 2.f * w11 - 4.f * w21 - w02 - 2.f * w12 - 4.f * w22) / 144.f;
     out32 = (-w00 - 2.f * w10 - 4.f * w20 + w01 + 2.f * w11 + 4.f * w21 - w02 - 2.f * w12 - 4.f * w22) / 144.f;
@@ -375,7 +398,7 @@
     out35 = (w02 + 2.f * w12 + 4.f * w22) / 24.f;
 
     // Row 4
-    float out40, out41, out42, out43, out44, out45;
+    DATA_TYPE out40, out41, out42, out43, out44, out45;
     out40 = (w00 - 2.f * w10 + 4.f * w20) / 96.f;
     out41 = (-w00 + 2.f * w10 - 4.f * w20 - w01 + 2.f * w11 - 4.f * w21 - w02 + 2.f * w12 - 4.f * w22) / 144.f;
     out42 = (-w00 + 2.f * w10 - 4.f * w20 + w01 - 2.f * w11 + 4.f * w21 - w02 + 2.f * w12 - 4.f * w22) / 144.f;
@@ -384,7 +407,7 @@
     out45 = (w02 - 2.f * w12 + 4.f * w22) / 24.f;
 
     // Row 5
-    float out50, out51, out52, out53, out54, out55;
+    DATA_TYPE out50, out51, out52, out53, out54, out55;
     out50 = (w20) / 4.f;
     out51 = (-w20 - w21 - w22) / 6.f;
     out52 = (-w20 + w21 - w22) / 6.f;
@@ -397,48 +420,48 @@
     int y0 = get_global_id(0); // idx channel
 
     // Get output address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(float) + y0 * dst_stride_y;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(DATA_TYPE) + y0 * dst_stride_y;
 
     // Store the values across the channels
     // 36 channels for 3x3 kernels
     // 6  channels for 3x1 or 1x3 kernels
-    *(__global float *)(dst_addr + 0 * dst_stride_z) = out00;
-    *(__global float *)(dst_addr + 1 * dst_stride_z) = out01;
-    *(__global float *)(dst_addr + 2 * dst_stride_z) = out02;
-    *(__global float *)(dst_addr + 3 * dst_stride_z) = out03;
-    *(__global float *)(dst_addr + 4 * dst_stride_z) = out04;
-    *(__global float *)(dst_addr + 5 * dst_stride_z) = out05;
+    *(__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z) = out00;
+    *(__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z) = out01;
+    *(__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z) = out02;
+    *(__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z) = out03;
+    *(__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z) = out04;
+    *(__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z) = out05;
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    *(__global float *)(dst_addr + 6 * dst_stride_z)  = out10;
-    *(__global float *)(dst_addr + 7 * dst_stride_z)  = out11;
-    *(__global float *)(dst_addr + 8 * dst_stride_z)  = out12;
-    *(__global float *)(dst_addr + 9 * dst_stride_z)  = out13;
-    *(__global float *)(dst_addr + 10 * dst_stride_z) = out14;
-    *(__global float *)(dst_addr + 11 * dst_stride_z) = out15;
-    *(__global float *)(dst_addr + 12 * dst_stride_z) = out20;
-    *(__global float *)(dst_addr + 13 * dst_stride_z) = out21;
-    *(__global float *)(dst_addr + 14 * dst_stride_z) = out22;
-    *(__global float *)(dst_addr + 15 * dst_stride_z) = out23;
-    *(__global float *)(dst_addr + 16 * dst_stride_z) = out24;
-    *(__global float *)(dst_addr + 17 * dst_stride_z) = out25;
-    *(__global float *)(dst_addr + 18 * dst_stride_z) = out30;
-    *(__global float *)(dst_addr + 19 * dst_stride_z) = out31;
-    *(__global float *)(dst_addr + 20 * dst_stride_z) = out32;
-    *(__global float *)(dst_addr + 21 * dst_stride_z) = out33;
-    *(__global float *)(dst_addr + 22 * dst_stride_z) = out34;
-    *(__global float *)(dst_addr + 23 * dst_stride_z) = out35;
-    *(__global float *)(dst_addr + 24 * dst_stride_z) = out40;
-    *(__global float *)(dst_addr + 25 * dst_stride_z) = out41;
-    *(__global float *)(dst_addr + 26 * dst_stride_z) = out42;
-    *(__global float *)(dst_addr + 27 * dst_stride_z) = out43;
-    *(__global float *)(dst_addr + 28 * dst_stride_z) = out44;
-    *(__global float *)(dst_addr + 29 * dst_stride_z) = out45;
-    *(__global float *)(dst_addr + 30 * dst_stride_z) = out50;
-    *(__global float *)(dst_addr + 31 * dst_stride_z) = out51;
-    *(__global float *)(dst_addr + 32 * dst_stride_z) = out52;
-    *(__global float *)(dst_addr + 33 * dst_stride_z) = out53;
-    *(__global float *)(dst_addr + 34 * dst_stride_z) = out54;
-    *(__global float *)(dst_addr + 35 * dst_stride_z) = out55;
+    *(__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z)  = out10;
+    *(__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z)  = out11;
+    *(__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z)  = out12;
+    *(__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z)  = out13;
+    *(__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z) = out14;
+    *(__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z) = out15;
+    *(__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z) = out20;
+    *(__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z) = out21;
+    *(__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z) = out22;
+    *(__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z) = out23;
+    *(__global DATA_TYPE *)(dst_addr + 16 * dst_stride_z) = out24;
+    *(__global DATA_TYPE *)(dst_addr + 17 * dst_stride_z) = out25;
+    *(__global DATA_TYPE *)(dst_addr + 18 * dst_stride_z) = out30;
+    *(__global DATA_TYPE *)(dst_addr + 19 * dst_stride_z) = out31;
+    *(__global DATA_TYPE *)(dst_addr + 20 * dst_stride_z) = out32;
+    *(__global DATA_TYPE *)(dst_addr + 21 * dst_stride_z) = out33;
+    *(__global DATA_TYPE *)(dst_addr + 22 * dst_stride_z) = out34;
+    *(__global DATA_TYPE *)(dst_addr + 23 * dst_stride_z) = out35;
+    *(__global DATA_TYPE *)(dst_addr + 24 * dst_stride_z) = out40;
+    *(__global DATA_TYPE *)(dst_addr + 25 * dst_stride_z) = out41;
+    *(__global DATA_TYPE *)(dst_addr + 26 * dst_stride_z) = out42;
+    *(__global DATA_TYPE *)(dst_addr + 27 * dst_stride_z) = out43;
+    *(__global DATA_TYPE *)(dst_addr + 28 * dst_stride_z) = out44;
+    *(__global DATA_TYPE *)(dst_addr + 29 * dst_stride_z) = out45;
+    *(__global DATA_TYPE *)(dst_addr + 30 * dst_stride_z) = out50;
+    *(__global DATA_TYPE *)(dst_addr + 31 * dst_stride_z) = out51;
+    *(__global DATA_TYPE *)(dst_addr + 32 * dst_stride_z) = out52;
+    *(__global DATA_TYPE *)(dst_addr + 33 * dst_stride_z) = out53;
+    *(__global DATA_TYPE *)(dst_addr + 34 * dst_stride_z) = out54;
+    *(__global DATA_TYPE *)(dst_addr + 35 * dst_stride_z) = out55;
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 }
 
@@ -448,8 +471,9 @@
  *
  * @note If this kernel is used to perform Winograd filter transform 5x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd filter transform 1x5, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -478,177 +502,192 @@
 
     // Load the values from the input tensor
 #if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
-    float4 w00 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float  w01 = *((__global float *)(src_addr + 0 * src_stride_y) + 4);
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    w00           = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    DATA_TYPE w01 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_y) + 4);
 #elif defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float4 w00 = (float4)(*((__global float *)(src_addr + 0 * src_stride_y)),
-                          *((__global float *)(src_addr + 1 * src_stride_y)),
-                          *((__global float *)(src_addr + 2 * src_stride_y)),
-                          *((__global float *)(src_addr + 3 * src_stride_y)));
-    float w01 = *((__global float *)(src_addr + 4 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    w00 = (VEC_DATA_TYPE(DATA_TYPE, 4))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)),
+                                        *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)),
+                                        *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)),
+                                        *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y)));
+    DATA_TYPE w01 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_y));
 #else  // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float4 w00 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float  w01 = *((__global float *)(src_addr + 0 * src_stride_y) + 4);
-    float4 w10 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float  w11 = *((__global float *)(src_addr + 1 * src_stride_y) + 4);
-    float4 w20 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
-    float  w21 = *((__global float *)(src_addr + 2 * src_stride_y) + 4);
-    float4 w30 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
-    float  w31 = *((__global float *)(src_addr + 3 * src_stride_y) + 4);
-    float4 w40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y));
-    float  w41 = *((__global float *)(src_addr + 4 * src_stride_y) + 4);
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    w00           = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    DATA_TYPE w01 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_y) + 4);
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    w10           = vload4(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    DATA_TYPE w11 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y) + 4);
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    w20           = vload4(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
+    DATA_TYPE w21 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y) + 4);
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    w30           = vload4(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y));
+    DATA_TYPE w31 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y) + 4);
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    w40           = vload4(0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y));
+    DATA_TYPE w41 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_y) + 4);
 #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
 
     // Transform the input tile
 
     // Row 0
-    float8 out0 = 0.0f;
-    out0.s0     = w00.s0;
-    out0.s1     = -2.f * (w00.s0 + w00.s1 + w00.s2 + w00.s3 + w01) / 9.f;
-    out0.s2     = -2.f * (w00.s0 - w00.s1 + w00.s2 - w00.s3 + w01) / 9.f;
-    out0.s3     = (w00.s0 + 2.f * w00.s1 + 4.f * w00.s2 + 8.f * w00.s3 + 16.f * w01) / 90.f;
-    out0.s4     = (w00.s0 - 2.f * w00.s1 + 4.f * w00.s2 - 8.f * w00.s3 + 16.f * w01) / 90.f;
-    out0.s5     = (16.f * w00.s0 + 8.f * w00.s1 + 4.f * w00.s2 + 2.f * w00.s3 + w01) / 180.f;
-    out0.s6     = (16.f * w00.s0 - 8.f * w00.s1 + 4.f * w00.s2 - 2.f * w00.s3 + w01) / 180.f;
-    out0.s7     = w01;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out0    = 0.0f;
+    out0.s0 = w00.s0;
+    out0.s1 = -2.f * (w00.s0 + w00.s1 + w00.s2 + w00.s3 + w01) / 9.f;
+    out0.s2 = -2.f * (w00.s0 - w00.s1 + w00.s2 - w00.s3 + w01) / 9.f;
+    out0.s3 = (w00.s0 + 2.f * w00.s1 + 4.f * w00.s2 + 8.f * w00.s3 + 16.f * w01) / 90.f;
+    out0.s4 = (w00.s0 - 2.f * w00.s1 + 4.f * w00.s2 - 8.f * w00.s3 + 16.f * w01) / 90.f;
+    out0.s5 = (16.f * w00.s0 + 8.f * w00.s1 + 4.f * w00.s2 + 2.f * w00.s3 + w01) / 180.f;
+    out0.s6 = (16.f * w00.s0 - 8.f * w00.s1 + 4.f * w00.s2 - 2.f * w00.s3 + w01) / 180.f;
+    out0.s7 = w01;
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
     // Row 1
-    float8 out1 = 0.0f;
-    out1.s0     = -2.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) / 9.f;
-    out1.s1     = 4.f * ((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) +
-                         (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 81.f;
-    out1.s2     = 4.f * ((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) -
-                         (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 81.f;
-    out1.s3     = -((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + 2.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) + 8.f *
-                    (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + 16.f * (w01 + w11 + w21 + w31 + w41)) / 405.f;
-    out1.s4     = -((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - 2.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) - 8.f *
-                    (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + 16.f * (w01 + w11 + w21 + w31 + w41)) / 405.f;
-    out1.s5     = -(16.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + 8.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) + 2.f *
-                    (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 810.f;
-    out1.s6     = -(16.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - 8.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) - 2.f *
-                    (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 810.f;
-    out1.s7     = -2.f * (w01 + w11 + w21 + w31 + w41) / 9.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out1    = 0.0f;
+    out1.s0 = -2.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) / 9.f;
+    out1.s1 = 4.f * ((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) +
+                     (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 81.f;
+    out1.s2 = 4.f * ((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) -
+                     (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 81.f;
+    out1.s3 = -((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + 2.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) + 8.f *
+                (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + 16.f * (w01 + w11 + w21 + w31 + w41)) / 405.f;
+    out1.s4 = -((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - 2.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) - 8.f *
+                (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + 16.f * (w01 + w11 + w21 + w31 + w41)) / 405.f;
+    out1.s5 = -(16.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + 8.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) + 2.f *
+                (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 810.f;
+    out1.s6 = -(16.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - 8.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) - 2.f *
+                (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 810.f;
+    out1.s7 = -2.f * (w01 + w11 + w21 + w31 + w41) / 9.f;
 
     // Row 2
-    float8 out2 = 0.0f;
-    out2.s0     = -2.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) / 9.f;
-    out2.s1     = 4.f * ((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) +
-                         (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 81.f;
-    out2.s2     = 4.f * ((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) -
-                         (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 81.f;
-    out2.s3     = -((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + 2.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) + 8.f *
-                    (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + 16.f * (w01 - w11 + w21 - w31 + w41)) / 405.f;
-    out2.s4     = -((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - 2.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) - 8.f *
-                    (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + 16.f * (w01 - w11 + w21 - w31 + w41)) / 405.f;
-    out2.s5     = -(16.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + 8.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) + 2.f *
-                    (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 810.f;
-    out2.s6     = -(16.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - 8.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) - 2.f *
-                    (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 810.f;
-    out2.s7     = -2.f * (w01 - w11 + w21 - w31 + w41) / 9.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out2    = 0.0f;
+    out2.s0 = -2.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) / 9.f;
+    out2.s1 = 4.f * ((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) +
+                     (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 81.f;
+    out2.s2 = 4.f * ((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) -
+                     (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 81.f;
+    out2.s3 = -((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + 2.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) + 8.f *
+                (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + 16.f * (w01 - w11 + w21 - w31 + w41)) / 405.f;
+    out2.s4 = -((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - 2.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) - 8.f *
+                (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + 16.f * (w01 - w11 + w21 - w31 + w41)) / 405.f;
+    out2.s5 = -(16.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + 8.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) + 2.f *
+                (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 810.f;
+    out2.s6 = -(16.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - 8.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) - 2.f *
+                (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 810.f;
+    out2.s7 = -2.f * (w01 - w11 + w21 - w31 + w41) / 9.f;
 
     // Row 3
-    float8 out3 = 0.0f;
-    out3.s0     = (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) / 90.f;
-    out3.s1     = -((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) +
-                    (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
-                    (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 405.f;
-    out3.s2     = -((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) +
-                    (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
-                    (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 405.f;
-    out3.s3     = ((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + 2.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
-                   (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + 8.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
-                   (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 8100.f;
-    out3.s4     = ((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - 2.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
-                   (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - 8.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
-                   (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 8100.f;
-    out3.s5     = (16.f * (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + 8.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
-                   (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + 2.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
-                   (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 16200.f;
-    out3.s6     = (16.f * (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - 8.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
-                   (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - 2.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
-                   (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 16200.f;
-    out3.s7     = (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) / 90.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out3    = 0.0f;
+    out3.s0 = (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) / 90.f;
+    out3.s1 = -((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) +
+                (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
+                (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 405.f;
+    out3.s2 = -((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) +
+                (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
+                (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 405.f;
+    out3.s3 = ((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + 2.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+               (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + 8.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
+               (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 8100.f;
+    out3.s4 = ((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - 2.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+               (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - 8.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
+               (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 8100.f;
+    out3.s5 = (16.f * (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + 8.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+               (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + 2.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
+               (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 16200.f;
+    out3.s6 = (16.f * (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - 8.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+               (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - 2.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) +
+               (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 16200.f;
+    out3.s7 = (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) / 90.f;
 
     // Row 4
-    float8 out4 = 0.0f;
-    out4.s0     = (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) / 90.f;
-    out4.s1     = -((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) +
-                    (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
-                    (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 405.f;
-    out4.s2     = -((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) +
-                    (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
-                    (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 405.f;
-    out4.s3     = ((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + 2.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
-                   (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + 8.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
-                   (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 8100.f;
-    out4.s4     = ((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - 2.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
-                   (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - 8.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
-                   (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 8100.f;
-    out4.s5     = (16.f * (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + 8.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
-                   (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + 2.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
-                   (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 16200.f;
-    out4.s6     = (16.f * (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - 8.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
-                   (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - 2.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
-                   (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 16200.f;
-    out4.s7     = (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) / 90.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out4    = 0.0f;
+    out4.s0 = (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) / 90.f;
+    out4.s1 = -((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) +
+                (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
+                (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 405.f;
+    out4.s2 = -((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) +
+                (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
+                (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 405.f;
+    out4.s3 = ((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + 2.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+               (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + 8.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
+               (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 8100.f;
+    out4.s4 = ((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - 2.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+               (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - 8.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + 16.f *
+               (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 8100.f;
+    out4.s5 = (16.f * (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + 8.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+               (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + 2.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
+               (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 16200.f;
+    out4.s6 = (16.f * (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - 8.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f *
+               (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - 2.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) +
+               (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 16200.f;
+    out4.s7 = (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) / 90.f;
 
     // Row 5
-    float8 out5 = 0.0f;
-    out5.s0     = (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) / 180.f;
-    out5.s1     = -((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) +
-                    (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
-                    (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 810.f;
-    out5.s2     = -((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) +
-                    (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
-                    (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 810.f;
-    out5.s3     = ((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + 2.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f *
-                   (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + 8.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + 16.f *
-                   (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 16200.f;
-    out5.s4     = ((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - 2.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f *
-                   (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - 8.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + 16.f *
-                   (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 16200.f;
-    out5.s5     = (16.f * (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + 8.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f *
-                   (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + 2.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
-                   (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 32400.f;
-    out5.s6     = (16.f * (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - 8.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f *
-                   (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - 2.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
-                   (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 32400.f;
-    out5.s7     = (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) / 180.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out5    = 0.0f;
+    out5.s0 = (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) / 180.f;
+    out5.s1 = -((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) +
+                (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
+                (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 810.f;
+    out5.s2 = -((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) +
+                (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
+                (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 810.f;
+    out5.s3 = ((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + 2.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f *
+               (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + 8.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + 16.f *
+               (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 16200.f;
+    out5.s4 = ((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - 2.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f *
+               (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - 8.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + 16.f *
+               (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 16200.f;
+    out5.s5 = (16.f * (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + 8.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f *
+               (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + 2.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
+               (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 32400.f;
+    out5.s6 = (16.f * (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - 8.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f *
+               (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - 2.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) +
+               (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 32400.f;
+    out5.s7 = (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) / 180.f;
 
     // Row 6
-    float8 out6 = 0.0f;
-    out6.s0     = (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) / 180.f;
-    out6.s1     = -((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) +
-                    (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
-                    (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 810.f;
-    out6.s2     = -((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) +
-                    (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
-                    (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 810.f;
-    out6.s3     = ((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + 2.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f *
-                   (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + 8.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + 16.f *
-                   (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 16200.f;
-    out6.s4     = ((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - 2.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f *
-                   (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - 8.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + 16.f *
-                   (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 16200.f;
-    out6.s5     = (16.f * (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + 8.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f *
-                   (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + 2.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
-                   (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 32400.f;
-    out6.s6     = (16.f * (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - 8.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f *
-                   (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - 2.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
-                   (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 32400.f;
-    out6.s7     = (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) / 180.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out6    = 0.0f;
+    out6.s0 = (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) / 180.f;
+    out6.s1 = -((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) +
+                (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
+                (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 810.f;
+    out6.s2 = -((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) +
+                (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
+                (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 810.f;
+    out6.s3 = ((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + 2.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f *
+               (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + 8.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + 16.f *
+               (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 16200.f;
+    out6.s4 = ((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - 2.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f *
+               (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - 8.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + 16.f *
+               (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 16200.f;
+    out6.s5 = (16.f * (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + 8.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f *
+               (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + 2.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
+               (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 32400.f;
+    out6.s6 = (16.f * (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - 8.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f *
+               (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - 2.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) +
+               (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 32400.f;
+    out6.s7 = (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) / 180.f;
 
     // Row 7
-    float8 out7 = 0.0f;
-    out7.s0     = w40.s0;
-    out7.s1     = -2.f * (w40.s0 + w40.s1 + w40.s2 + w40.s3 + w41) / 9.f;
-    out7.s2     = -2.f * (w40.s0 - w40.s1 + w40.s2 - w40.s3 + w41) / 9.f;
-    out7.s3     = (w40.s0 + 2.f * w40.s1 + 4.f * w40.s2 + 8.f * w40.s3 + 16.f * w41) / 90.f;
-    out7.s4     = (w40.s0 - 2.f * w40.s1 + 4.f * w40.s2 - 8.f * w40.s3 + 16.f * w41) / 90.f;
-    out7.s5     = (16.f * w40.s0 + 8.f * w40.s1 + 4.f * w40.s2 + 2.f * w40.s3 + w41) / 180.f;
-    out7.s6     = (16.f * w40.s0 - 8.f * w40.s1 + 4.f * w40.s2 - 2.f * w40.s3 + w41) / 180.f;
-    out7.s7     = w41;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out7    = 0.0f;
+    out7.s0 = w40.s0;
+    out7.s1 = -2.f * (w40.s0 + w40.s1 + w40.s2 + w40.s3 + w41) / 9.f;
+    out7.s2 = -2.f * (w40.s0 - w40.s1 + w40.s2 - w40.s3 + w41) / 9.f;
+    out7.s3 = (w40.s0 + 2.f * w40.s1 + 4.f * w40.s2 + 8.f * w40.s3 + 16.f * w41) / 90.f;
+    out7.s4 = (w40.s0 - 2.f * w40.s1 + 4.f * w40.s2 - 8.f * w40.s3 + 16.f * w41) / 90.f;
+    out7.s5 = (16.f * w40.s0 + 8.f * w40.s1 + 4.f * w40.s2 + 2.f * w40.s3 + w41) / 180.f;
+    out7.s6 = (16.f * w40.s0 - 8.f * w40.s1 + 4.f * w40.s2 - 2.f * w40.s3 + w41) / 180.f;
+    out7.s7 = w41;
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 
     int z  = get_global_id(2);
@@ -656,75 +695,75 @@
     int y0 = z % SRC_DIM_Z; // idx channel
 
     // Get output address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(float) + y0 * dst_stride_y;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(DATA_TYPE) + y0 * dst_stride_y;
 
     // Store the values across the channels
-    *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0;
-    *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1;
-    *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2;
-    *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3;
-    *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4;
-    *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5;
-    *(__global float *)(dst_addr + 6 * dst_stride_z) = out0.s6;
-    *(__global float *)(dst_addr + 7 * dst_stride_z) = out0.s7;
+    *(__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z) = out0.s0;
+    *(__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z) = out0.s1;
+    *(__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z) = out0.s2;
+    *(__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z) = out0.s3;
+    *(__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z) = out0.s4;
+    *(__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z) = out0.s5;
+    *(__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z) = out0.s6;
+    *(__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z) = out0.s7;
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    *(__global float *)(dst_addr + 8 * dst_stride_z)  = out1.s0;
-    *(__global float *)(dst_addr + 9 * dst_stride_z)  = out1.s1;
-    *(__global float *)(dst_addr + 10 * dst_stride_z) = out1.s2;
-    *(__global float *)(dst_addr + 11 * dst_stride_z) = out1.s3;
-    *(__global float *)(dst_addr + 12 * dst_stride_z) = out1.s4;
-    *(__global float *)(dst_addr + 13 * dst_stride_z) = out1.s5;
-    *(__global float *)(dst_addr + 14 * dst_stride_z) = out1.s6;
-    *(__global float *)(dst_addr + 15 * dst_stride_z) = out1.s7;
-    *(__global float *)(dst_addr + 16 * dst_stride_z) = out2.s0;
-    *(__global float *)(dst_addr + 17 * dst_stride_z) = out2.s1;
-    *(__global float *)(dst_addr + 18 * dst_stride_z) = out2.s2;
-    *(__global float *)(dst_addr + 19 * dst_stride_z) = out2.s3;
-    *(__global float *)(dst_addr + 20 * dst_stride_z) = out2.s4;
-    *(__global float *)(dst_addr + 21 * dst_stride_z) = out2.s5;
-    *(__global float *)(dst_addr + 22 * dst_stride_z) = out2.s6;
-    *(__global float *)(dst_addr + 23 * dst_stride_z) = out2.s7;
-    *(__global float *)(dst_addr + 24 * dst_stride_z) = out3.s0;
-    *(__global float *)(dst_addr + 25 * dst_stride_z) = out3.s1;
-    *(__global float *)(dst_addr + 26 * dst_stride_z) = out3.s2;
-    *(__global float *)(dst_addr + 27 * dst_stride_z) = out3.s3;
-    *(__global float *)(dst_addr + 28 * dst_stride_z) = out3.s4;
-    *(__global float *)(dst_addr + 29 * dst_stride_z) = out3.s5;
-    *(__global float *)(dst_addr + 30 * dst_stride_z) = out3.s6;
-    *(__global float *)(dst_addr + 31 * dst_stride_z) = out3.s7;
-    *(__global float *)(dst_addr + 32 * dst_stride_z) = out4.s0;
-    *(__global float *)(dst_addr + 33 * dst_stride_z) = out4.s1;
-    *(__global float *)(dst_addr + 34 * dst_stride_z) = out4.s2;
-    *(__global float *)(dst_addr + 35 * dst_stride_z) = out4.s3;
-    *(__global float *)(dst_addr + 36 * dst_stride_z) = out4.s4;
-    *(__global float *)(dst_addr + 37 * dst_stride_z) = out4.s5;
-    *(__global float *)(dst_addr + 38 * dst_stride_z) = out4.s6;
-    *(__global float *)(dst_addr + 39 * dst_stride_z) = out4.s7;
-    *(__global float *)(dst_addr + 40 * dst_stride_z) = out5.s0;
-    *(__global float *)(dst_addr + 41 * dst_stride_z) = out5.s1;
-    *(__global float *)(dst_addr + 42 * dst_stride_z) = out5.s2;
-    *(__global float *)(dst_addr + 43 * dst_stride_z) = out5.s3;
-    *(__global float *)(dst_addr + 44 * dst_stride_z) = out5.s4;
-    *(__global float *)(dst_addr + 45 * dst_stride_z) = out5.s5;
-    *(__global float *)(dst_addr + 46 * dst_stride_z) = out5.s6;
-    *(__global float *)(dst_addr + 47 * dst_stride_z) = out5.s7;
-    *(__global float *)(dst_addr + 48 * dst_stride_z) = out6.s0;
-    *(__global float *)(dst_addr + 49 * dst_stride_z) = out6.s1;
-    *(__global float *)(dst_addr + 50 * dst_stride_z) = out6.s2;
-    *(__global float *)(dst_addr + 51 * dst_stride_z) = out6.s3;
-    *(__global float *)(dst_addr + 52 * dst_stride_z) = out6.s4;
-    *(__global float *)(dst_addr + 53 * dst_stride_z) = out6.s5;
-    *(__global float *)(dst_addr + 54 * dst_stride_z) = out6.s6;
-    *(__global float *)(dst_addr + 55 * dst_stride_z) = out6.s7;
-    *(__global float *)(dst_addr + 56 * dst_stride_z) = out7.s0;
-    *(__global float *)(dst_addr + 57 * dst_stride_z) = out7.s1;
-    *(__global float *)(dst_addr + 58 * dst_stride_z) = out7.s2;
-    *(__global float *)(dst_addr + 59 * dst_stride_z) = out7.s3;
-    *(__global float *)(dst_addr + 60 * dst_stride_z) = out7.s4;
-    *(__global float *)(dst_addr + 61 * dst_stride_z) = out7.s5;
-    *(__global float *)(dst_addr + 62 * dst_stride_z) = out7.s6;
-    *(__global float *)(dst_addr + 63 * dst_stride_z) = out7.s7;
+    *(__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z)  = out1.s0;
+    *(__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z)  = out1.s1;
+    *(__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z) = out1.s2;
+    *(__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z) = out1.s3;
+    *(__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z) = out1.s4;
+    *(__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z) = out1.s5;
+    *(__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z) = out1.s6;
+    *(__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z) = out1.s7;
+    *(__global DATA_TYPE *)(dst_addr + 16 * dst_stride_z) = out2.s0;
+    *(__global DATA_TYPE *)(dst_addr + 17 * dst_stride_z) = out2.s1;
+    *(__global DATA_TYPE *)(dst_addr + 18 * dst_stride_z) = out2.s2;
+    *(__global DATA_TYPE *)(dst_addr + 19 * dst_stride_z) = out2.s3;
+    *(__global DATA_TYPE *)(dst_addr + 20 * dst_stride_z) = out2.s4;
+    *(__global DATA_TYPE *)(dst_addr + 21 * dst_stride_z) = out2.s5;
+    *(__global DATA_TYPE *)(dst_addr + 22 * dst_stride_z) = out2.s6;
+    *(__global DATA_TYPE *)(dst_addr + 23 * dst_stride_z) = out2.s7;
+    *(__global DATA_TYPE *)(dst_addr + 24 * dst_stride_z) = out3.s0;
+    *(__global DATA_TYPE *)(dst_addr + 25 * dst_stride_z) = out3.s1;
+    *(__global DATA_TYPE *)(dst_addr + 26 * dst_stride_z) = out3.s2;
+    *(__global DATA_TYPE *)(dst_addr + 27 * dst_stride_z) = out3.s3;
+    *(__global DATA_TYPE *)(dst_addr + 28 * dst_stride_z) = out3.s4;
+    *(__global DATA_TYPE *)(dst_addr + 29 * dst_stride_z) = out3.s5;
+    *(__global DATA_TYPE *)(dst_addr + 30 * dst_stride_z) = out3.s6;
+    *(__global DATA_TYPE *)(dst_addr + 31 * dst_stride_z) = out3.s7;
+    *(__global DATA_TYPE *)(dst_addr + 32 * dst_stride_z) = out4.s0;
+    *(__global DATA_TYPE *)(dst_addr + 33 * dst_stride_z) = out4.s1;
+    *(__global DATA_TYPE *)(dst_addr + 34 * dst_stride_z) = out4.s2;
+    *(__global DATA_TYPE *)(dst_addr + 35 * dst_stride_z) = out4.s3;
+    *(__global DATA_TYPE *)(dst_addr + 36 * dst_stride_z) = out4.s4;
+    *(__global DATA_TYPE *)(dst_addr + 37 * dst_stride_z) = out4.s5;
+    *(__global DATA_TYPE *)(dst_addr + 38 * dst_stride_z) = out4.s6;
+    *(__global DATA_TYPE *)(dst_addr + 39 * dst_stride_z) = out4.s7;
+    *(__global DATA_TYPE *)(dst_addr + 40 * dst_stride_z) = out5.s0;
+    *(__global DATA_TYPE *)(dst_addr + 41 * dst_stride_z) = out5.s1;
+    *(__global DATA_TYPE *)(dst_addr + 42 * dst_stride_z) = out5.s2;
+    *(__global DATA_TYPE *)(dst_addr + 43 * dst_stride_z) = out5.s3;
+    *(__global DATA_TYPE *)(dst_addr + 44 * dst_stride_z) = out5.s4;
+    *(__global DATA_TYPE *)(dst_addr + 45 * dst_stride_z) = out5.s5;
+    *(__global DATA_TYPE *)(dst_addr + 46 * dst_stride_z) = out5.s6;
+    *(__global DATA_TYPE *)(dst_addr + 47 * dst_stride_z) = out5.s7;
+    *(__global DATA_TYPE *)(dst_addr + 48 * dst_stride_z) = out6.s0;
+    *(__global DATA_TYPE *)(dst_addr + 49 * dst_stride_z) = out6.s1;
+    *(__global DATA_TYPE *)(dst_addr + 50 * dst_stride_z) = out6.s2;
+    *(__global DATA_TYPE *)(dst_addr + 51 * dst_stride_z) = out6.s3;
+    *(__global DATA_TYPE *)(dst_addr + 52 * dst_stride_z) = out6.s4;
+    *(__global DATA_TYPE *)(dst_addr + 53 * dst_stride_z) = out6.s5;
+    *(__global DATA_TYPE *)(dst_addr + 54 * dst_stride_z) = out6.s6;
+    *(__global DATA_TYPE *)(dst_addr + 55 * dst_stride_z) = out6.s7;
+    *(__global DATA_TYPE *)(dst_addr + 56 * dst_stride_z) = out7.s0;
+    *(__global DATA_TYPE *)(dst_addr + 57 * dst_stride_z) = out7.s1;
+    *(__global DATA_TYPE *)(dst_addr + 58 * dst_stride_z) = out7.s2;
+    *(__global DATA_TYPE *)(dst_addr + 59 * dst_stride_z) = out7.s3;
+    *(__global DATA_TYPE *)(dst_addr + 60 * dst_stride_z) = out7.s4;
+    *(__global DATA_TYPE *)(dst_addr + 61 * dst_stride_z) = out7.s5;
+    *(__global DATA_TYPE *)(dst_addr + 62 * dst_stride_z) = out7.s6;
+    *(__global DATA_TYPE *)(dst_addr + 63 * dst_stride_z) = out7.s7;
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 }
 
@@ -733,8 +772,9 @@
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note If this kernel is used to perform Winograd filter transform 5x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd filter transform 1x5, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -759,242 +799,250 @@
 {
     Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, SRC_DIM_Z);
 
-    const __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + get_global_id(0) * sizeof(float) + get_global_id(1) * src_step_y + get_global_id(2) * src_step_w;
+    const __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + get_global_id(0) * sizeof(DATA_TYPE) + get_global_id(1) * src_step_y + get_global_id(2) * src_step_w;
 
 #if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
     // Load the values from the input tensor
-    float w00 = *((__global float *)(src_addr + 0 * src_stride_z));
-    float w01 = *((__global float *)(src_addr + 1 * src_stride_z));
-    float w02 = *((__global float *)(src_addr + 2 * src_stride_z));
-    float w03 = *((__global float *)(src_addr + 3 * src_stride_z));
-    float w04 = *((__global float *)(src_addr + 4 * src_stride_z));
+    DATA_TYPE w00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z));
+    DATA_TYPE w01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z));
+    DATA_TYPE w02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z));
+    DATA_TYPE w03 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z));
+    DATA_TYPE w04 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z));
 #else  // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
     // Load the values from the input tensor
-    float w00 = *((__global float *)(src_addr + 0 * src_stride_y));
-    float w01 = *((__global float *)(src_addr + 1 * src_stride_y));
-    float w02 = *((__global float *)(src_addr + 2 * src_stride_y));
-    float w03 = *((__global float *)(src_addr + 3 * src_stride_y));
-    float w04 = *((__global float *)(src_addr + 4 * src_stride_y));
+    DATA_TYPE w00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    DATA_TYPE w01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    DATA_TYPE w02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
+    DATA_TYPE w03 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y));
+    DATA_TYPE w04 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_y));
 #endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    float w10 = *((__global float *)(src_addr + 1 * src_stride_z + 0 * src_stride_y));
-    float w11 = *((__global float *)(src_addr + 1 * src_stride_z + 1 * src_stride_y));
-    float w12 = *((__global float *)(src_addr + 1 * src_stride_z + 2 * src_stride_y));
-    float w13 = *((__global float *)(src_addr + 1 * src_stride_z + 3 * src_stride_y));
-    float w14 = *((__global float *)(src_addr + 1 * src_stride_z + 4 * src_stride_y));
-    float w20 = *((__global float *)(src_addr + 2 * src_stride_z + 0 * src_stride_y));
-    float w21 = *((__global float *)(src_addr + 2 * src_stride_z + 1 * src_stride_y));
-    float w22 = *((__global float *)(src_addr + 2 * src_stride_z + 2 * src_stride_y));
-    float w23 = *((__global float *)(src_addr + 2 * src_stride_z + 3 * src_stride_y));
-    float w24 = *((__global float *)(src_addr + 2 * src_stride_z + 4 * src_stride_y));
-    float w30 = *((__global float *)(src_addr + 3 * src_stride_z + 0 * src_stride_y));
-    float w31 = *((__global float *)(src_addr + 3 * src_stride_z + 1 * src_stride_y));
-    float w32 = *((__global float *)(src_addr + 3 * src_stride_z + 2 * src_stride_y));
-    float w33 = *((__global float *)(src_addr + 3 * src_stride_z + 3 * src_stride_y));
-    float w34 = *((__global float *)(src_addr + 3 * src_stride_z + 4 * src_stride_y));
-    float w40 = *((__global float *)(src_addr + 4 * src_stride_z + 0 * src_stride_y));
-    float w41 = *((__global float *)(src_addr + 4 * src_stride_z + 1 * src_stride_y));
-    float w42 = *((__global float *)(src_addr + 4 * src_stride_z + 2 * src_stride_y));
-    float w43 = *((__global float *)(src_addr + 4 * src_stride_z + 3 * src_stride_y));
-    float w44 = *((__global float *)(src_addr + 4 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w10 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w11 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w12 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w13 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w14 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w20 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w21 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w22 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w23 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w24 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w30 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w31 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w32 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w33 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w34 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z + 4 * src_stride_y));
+    DATA_TYPE w40 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 0 * src_stride_y));
+    DATA_TYPE w41 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 1 * src_stride_y));
+    DATA_TYPE w42 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 2 * src_stride_y));
+    DATA_TYPE w43 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 3 * src_stride_y));
+    DATA_TYPE w44 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z + 4 * src_stride_y));
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 
     // Row 0
-    float8 out0 = 0.0f;
-    out0.s0     = w00;
-    out0.s1     = -2.f * (w00 + w01 + w02 + w03 + w04) / 9.f;
-    out0.s2     = -2.f * (w00 - w01 + w02 - w03 + w04) / 9.f;
-    out0.s3     = (w00 + 2.f * w01 + 4.f * w02 + 8.f * w03 + 16.f * w04) / 90.f;
-    out0.s4     = (w00 - 2.f * w01 + 4.f * w02 - 8.f * w03 + 16.f * w04) / 90.f;
-    out0.s5     = (16.f * w00 + 8.f * w01 + 4.f * w02 + 2.f * w03 + w04) / 180.f;
-    out0.s6     = (16.f * w00 - 8.f * w01 + 4.f * w02 - 2.f * w03 + w04) / 180.f;
-    out0.s7     = w04;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out0    = 0.0f;
+    out0.s0 = w00;
+    out0.s1 = -2.f * (w00 + w01 + w02 + w03 + w04) / 9.f;
+    out0.s2 = -2.f * (w00 - w01 + w02 - w03 + w04) / 9.f;
+    out0.s3 = (w00 + 2.f * w01 + 4.f * w02 + 8.f * w03 + 16.f * w04) / 90.f;
+    out0.s4 = (w00 - 2.f * w01 + 4.f * w02 - 8.f * w03 + 16.f * w04) / 90.f;
+    out0.s5 = (16.f * w00 + 8.f * w01 + 4.f * w02 + 2.f * w03 + w04) / 180.f;
+    out0.s6 = (16.f * w00 - 8.f * w01 + 4.f * w02 - 2.f * w03 + w04) / 180.f;
+    out0.s7 = w04;
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
     // Row 1
-    float8 out1 = 0.0f;
-    out1.s0     = -2.f * (w00 + w10 + w20 + w30 + w40) / 9.f;
-    out1.s1     = 4.f * ((w00 + w10 + w20 + w30 + w40) + (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) + (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f;
-    out1.s2     = 4.f * ((w00 + w10 + w20 + w30 + w40) - (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) - (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f;
-    out1.s3     = -((w00 + w10 + w20 + w30 + w40) + 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f *
-                    (w04 + w14 + w24 + w34 + w44)) / 405.f;
-    out1.s4     = -((w00 + w10 + w20 + w30 + w40) - 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f *
-                    (w04 + w14 + w24 + w34 + w44)) / 405.f;
-    out1.s5     = -(16.f * (w00 + w10 + w20 + w30 + w40) + 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 2.f * (w03 + w13 + w23 + w33 + w43) +
-                    (w04 + w14 + w24 + w34 + w44)) / 810.f;
-    out1.s6     = -(16.f * (w00 + w10 + w20 + w30 + w40) - 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 2.f * (w03 + w13 + w23 + w33 + w43) +
-                    (w04 + w14 + w24 + w34 + w44)) / 810.f;
-    out1.s7     = -2.f * (w04 + w14 + w24 + w34 + w44) / 9.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out1    = 0.0f;
+    out1.s0 = -2.f * (w00 + w10 + w20 + w30 + w40) / 9.f;
+    out1.s1 = 4.f * ((w00 + w10 + w20 + w30 + w40) + (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) + (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f;
+    out1.s2 = 4.f * ((w00 + w10 + w20 + w30 + w40) - (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) - (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f;
+    out1.s3 = -((w00 + w10 + w20 + w30 + w40) + 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f *
+                (w04 + w14 + w24 + w34 + w44)) / 405.f;
+    out1.s4 = -((w00 + w10 + w20 + w30 + w40) - 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f *
+                (w04 + w14 + w24 + w34 + w44)) / 405.f;
+    out1.s5 = -(16.f * (w00 + w10 + w20 + w30 + w40) + 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 2.f * (w03 + w13 + w23 + w33 + w43) +
+                (w04 + w14 + w24 + w34 + w44)) / 810.f;
+    out1.s6 = -(16.f * (w00 + w10 + w20 + w30 + w40) - 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 2.f * (w03 + w13 + w23 + w33 + w43) +
+                (w04 + w14 + w24 + w34 + w44)) / 810.f;
+    out1.s7 = -2.f * (w04 + w14 + w24 + w34 + w44) / 9.f;
 
     // Row 2
-    float8 out2 = 0.0f;
-    out2.s0     = -2.f * (w00 - w10 + w20 - w30 + w40) / 9.f;
-    out2.s1     = 4.f * ((w00 - w10 + w20 - w30 + w40) + (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) + (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f;
-    out2.s2     = 4.f * ((w00 - w10 + w20 - w30 + w40) - (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) - (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f;
-    out2.s3     = -((w00 - w10 + w20 - w30 + w40) + 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f *
-                    (w04 - w14 + w24 - w34 + w44)) / 405.f;
-    out2.s4     = -((w00 - w10 + w20 - w30 + w40) - 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f *
-                    (w04 - w14 + w24 - w34 + w44)) / 405.f;
-    out2.s5     = -(16.f * (w00 - w10 + w20 - w30 + w40) + 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 2.f * (w03 - w13 + w23 - w33 + w43) +
-                    (w04 - w14 + w24 - w34 + w44)) / 810.f;
-    out2.s6     = -(16.f * (w00 - w10 + w20 - w30 + w40) - 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 2.f * (w03 - w13 + w23 - w33 + w43) +
-                    (w04 - w14 + w24 - w34 + w44)) / 810.f;
-    out2.s7     = -2.f * (w04 - w14 + w24 - w34 + w44) / 9.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out2    = 0.0f;
+    out2.s0 = -2.f * (w00 - w10 + w20 - w30 + w40) / 9.f;
+    out2.s1 = 4.f * ((w00 - w10 + w20 - w30 + w40) + (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) + (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f;
+    out2.s2 = 4.f * ((w00 - w10 + w20 - w30 + w40) - (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) - (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f;
+    out2.s3 = -((w00 - w10 + w20 - w30 + w40) + 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f *
+                (w04 - w14 + w24 - w34 + w44)) / 405.f;
+    out2.s4 = -((w00 - w10 + w20 - w30 + w40) - 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f *
+                (w04 - w14 + w24 - w34 + w44)) / 405.f;
+    out2.s5 = -(16.f * (w00 - w10 + w20 - w30 + w40) + 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 2.f * (w03 - w13 + w23 - w33 + w43) +
+                (w04 - w14 + w24 - w34 + w44)) / 810.f;
+    out2.s6 = -(16.f * (w00 - w10 + w20 - w30 + w40) - 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 2.f * (w03 - w13 + w23 - w33 + w43) +
+                (w04 - w14 + w24 - w34 + w44)) / 810.f;
+    out2.s7 = -2.f * (w04 - w14 + w24 - w34 + w44) / 9.f;
 
     // Row 3
-    float8 out3 = 0.0f;
-    out3.s0     = (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) / 90.f;
-    out3.s1     = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) +
-                    (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f;
-    out3.s2     = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) -
-                    (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f;
-    out3.s3     = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f *
-                   (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 8.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f;
-    out3.s4     = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f *
-                   (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 8.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f;
-    out3.s5     = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f *
-                   (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f;
-    out3.s6     = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f *
-                   (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f;
-    out3.s7     = (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44) / 90.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out3    = 0.0f;
+    out3.s0 = (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) / 90.f;
+    out3.s1 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) +
+                (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f;
+    out3.s2 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) -
+                (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f;
+    out3.s3 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 8.f
+               * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f;
+    out3.s4 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 8.f
+               * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f;
+    out3.s5 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f *
+               (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f;
+    out3.s6 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f *
+               (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f;
+    out3.s7 = (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44) / 90.f;
 
     // Row 4
-    float8 out4 = 0.0f;
-    out4.s0     = (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) / 90.f;
-    out4.s1     = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) +
-                    (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f;
-    out4.s2     = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) -
-                    (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f;
-    out4.s3     = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f *
-                   (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 8.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f;
-    out4.s4     = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f *
-                   (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 8.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f;
-    out4.s5     = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f *
-                   (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f;
-    out4.s6     = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f *
-                   (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f;
-    out4.s7     = (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44) / 90.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out4    = 0.0f;
+    out4.s0 = (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) / 90.f;
+    out4.s1 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) +
+                (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f;
+    out4.s2 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) -
+                (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f;
+    out4.s3 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 8.f
+               * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f;
+    out4.s4 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 8.f
+               * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f;
+    out4.s5 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f *
+               (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f;
+    out4.s6 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f *
+               (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f;
+    out4.s7 = (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44) / 90.f;
 
     // Row 5
-    float8 out5 = 0.0f;
-    out5.s0     = (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) / 180.f;
-    out5.s1     = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) +
-                    (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f;
-    out5.s2     = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) -
-                    (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f;
-    out5.s3     = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f *
-                   (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 8.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f;
-    out5.s4     = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f *
-                   (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 8.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f;
-    out5.s5     = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f *
-                   (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f;
-    out5.s6     = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f *
-                   (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f;
-    out5.s7     = (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44) / 180.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out5    = 0.0f;
+    out5.s0 = (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) / 180.f;
+    out5.s1 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) +
+                (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f;
+    out5.s2 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) -
+                (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f;
+    out5.s3 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 8.f
+               * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f;
+    out5.s4 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 8.f
+               * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f;
+    out5.s5 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f *
+               (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f;
+    out5.s6 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f *
+               (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f;
+    out5.s7 = (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44) / 180.f;
 
     // Row 6
-    float8 out6 = 0.0f;
-    out6.s0     = (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) / 180.f;
-    out6.s1     = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) +
-                    (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f;
-    out6.s2     = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) -
-                    (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f;
-    out6.s3     = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f *
-                   (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 8.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f;
-    out6.s4     = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f *
-                   (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 8.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f;
-    out6.s5     = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f *
-                   (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f;
-    out6.s6     = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f *
-                   (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f;
-    out6.s7     = (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44) / 180.f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out6    = 0.0f;
+    out6.s0 = (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) / 180.f;
+    out6.s1 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) +
+                (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f;
+    out6.s2 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) -
+                (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f;
+    out6.s3 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 8.f
+               * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f;
+    out6.s4 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 8.f
+               * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f;
+    out6.s5 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f *
+               (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f;
+    out6.s6 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f *
+               (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f;
+    out6.s7 = (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44) / 180.f;
 
     // Row 7
-    float8 out7 = 0.0f;
-    out7.s0     = w40;
-    out7.s1     = -2.f * (w40 + w41 + w42 + w43 + w44) / 9.f;
-    out7.s2     = -2.f * (w40 - w41 + w42 - w43 + w44) / 9.f;
-    out7.s3     = (w40 + 2.f * w41 + 4.f * w42 + 8.f * w43 + 16.f * w44) / 90.f;
-    out7.s4     = (w40 - 2.f * w41 + 4.f * w42 - 8.f * w43 + 16.f * w44) / 90.f;
-    out7.s5     = (16.f * w40 + 8.f * w41 + 4.f * w42 + 2.f * w43 + w44) / 180.f;
-    out7.s6     = (16.f * w40 - 8.f * w41 + 4.f * w42 - 2.f * w43 + w44) / 180.f;
-    out7.s7     = w44;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out7    = 0.0f;
+    out7.s0 = w40;
+    out7.s1 = -2.f * (w40 + w41 + w42 + w43 + w44) / 9.f;
+    out7.s2 = -2.f * (w40 - w41 + w42 - w43 + w44) / 9.f;
+    out7.s3 = (w40 + 2.f * w41 + 4.f * w42 + 8.f * w43 + 16.f * w44) / 90.f;
+    out7.s4 = (w40 - 2.f * w41 + 4.f * w42 - 8.f * w43 + 16.f * w44) / 90.f;
+    out7.s5 = (16.f * w40 + 8.f * w41 + 4.f * w42 + 2.f * w43 + w44) / 180.f;
+    out7.s6 = (16.f * w40 - 8.f * w41 + 4.f * w42 - 2.f * w43 + w44) / 180.f;
+    out7.s7 = w44;
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 
     int x0 = get_global_id(2); // idx filter
     int y0 = get_global_id(0); // idx channel
 
     // Get output address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(float) + y0 * dst_stride_y;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(DATA_TYPE) + y0 * dst_stride_y;
 
     // Store the values across the channels
-    *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0;
-    *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1;
-    *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2;
-    *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3;
-    *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4;
-    *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5;
-    *(__global float *)(dst_addr + 6 * dst_stride_z) = out0.s6;
-    *(__global float *)(dst_addr + 7 * dst_stride_z) = out0.s7;
+    *(__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z) = out0.s0;
+    *(__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z) = out0.s1;
+    *(__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z) = out0.s2;
+    *(__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z) = out0.s3;
+    *(__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z) = out0.s4;
+    *(__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z) = out0.s5;
+    *(__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z) = out0.s6;
+    *(__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z) = out0.s7;
 
 #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
-    *(__global float *)(dst_addr + 8 * dst_stride_z)  = out1.s0;
-    *(__global float *)(dst_addr + 9 * dst_stride_z)  = out1.s1;
-    *(__global float *)(dst_addr + 10 * dst_stride_z) = out1.s2;
-    *(__global float *)(dst_addr + 11 * dst_stride_z) = out1.s3;
-    *(__global float *)(dst_addr + 12 * dst_stride_z) = out1.s4;
-    *(__global float *)(dst_addr + 13 * dst_stride_z) = out1.s5;
-    *(__global float *)(dst_addr + 14 * dst_stride_z) = out1.s6;
-    *(__global float *)(dst_addr + 15 * dst_stride_z) = out1.s7;
-    *(__global float *)(dst_addr + 16 * dst_stride_z) = out2.s0;
-    *(__global float *)(dst_addr + 17 * dst_stride_z) = out2.s1;
-    *(__global float *)(dst_addr + 18 * dst_stride_z) = out2.s2;
-    *(__global float *)(dst_addr + 19 * dst_stride_z) = out2.s3;
-    *(__global float *)(dst_addr + 20 * dst_stride_z) = out2.s4;
-    *(__global float *)(dst_addr + 21 * dst_stride_z) = out2.s5;
-    *(__global float *)(dst_addr + 22 * dst_stride_z) = out2.s6;
-    *(__global float *)(dst_addr + 23 * dst_stride_z) = out2.s7;
-    *(__global float *)(dst_addr + 24 * dst_stride_z) = out3.s0;
-    *(__global float *)(dst_addr + 25 * dst_stride_z) = out3.s1;
-    *(__global float *)(dst_addr + 26 * dst_stride_z) = out3.s2;
-    *(__global float *)(dst_addr + 27 * dst_stride_z) = out3.s3;
-    *(__global float *)(dst_addr + 28 * dst_stride_z) = out3.s4;
-    *(__global float *)(dst_addr + 29 * dst_stride_z) = out3.s5;
-    *(__global float *)(dst_addr + 30 * dst_stride_z) = out3.s6;
-    *(__global float *)(dst_addr + 31 * dst_stride_z) = out3.s7;
-    *(__global float *)(dst_addr + 32 * dst_stride_z) = out4.s0;
-    *(__global float *)(dst_addr + 33 * dst_stride_z) = out4.s1;
-    *(__global float *)(dst_addr + 34 * dst_stride_z) = out4.s2;
-    *(__global float *)(dst_addr + 35 * dst_stride_z) = out4.s3;
-    *(__global float *)(dst_addr + 36 * dst_stride_z) = out4.s4;
-    *(__global float *)(dst_addr + 37 * dst_stride_z) = out4.s5;
-    *(__global float *)(dst_addr + 38 * dst_stride_z) = out4.s6;
-    *(__global float *)(dst_addr + 39 * dst_stride_z) = out4.s7;
-    *(__global float *)(dst_addr + 40 * dst_stride_z) = out5.s0;
-    *(__global float *)(dst_addr + 41 * dst_stride_z) = out5.s1;
-    *(__global float *)(dst_addr + 42 * dst_stride_z) = out5.s2;
-    *(__global float *)(dst_addr + 43 * dst_stride_z) = out5.s3;
-    *(__global float *)(dst_addr + 44 * dst_stride_z) = out5.s4;
-    *(__global float *)(dst_addr + 45 * dst_stride_z) = out5.s5;
-    *(__global float *)(dst_addr + 46 * dst_stride_z) = out5.s6;
-    *(__global float *)(dst_addr + 47 * dst_stride_z) = out5.s7;
-    *(__global float *)(dst_addr + 48 * dst_stride_z) = out6.s0;
-    *(__global float *)(dst_addr + 49 * dst_stride_z) = out6.s1;
-    *(__global float *)(dst_addr + 50 * dst_stride_z) = out6.s2;
-    *(__global float *)(dst_addr + 51 * dst_stride_z) = out6.s3;
-    *(__global float *)(dst_addr + 52 * dst_stride_z) = out6.s4;
-    *(__global float *)(dst_addr + 53 * dst_stride_z) = out6.s5;
-    *(__global float *)(dst_addr + 54 * dst_stride_z) = out6.s6;
-    *(__global float *)(dst_addr + 55 * dst_stride_z) = out6.s7;
-    *(__global float *)(dst_addr + 56 * dst_stride_z) = out7.s0;
-    *(__global float *)(dst_addr + 57 * dst_stride_z) = out7.s1;
-    *(__global float *)(dst_addr + 58 * dst_stride_z) = out7.s2;
-    *(__global float *)(dst_addr + 59 * dst_stride_z) = out7.s3;
-    *(__global float *)(dst_addr + 60 * dst_stride_z) = out7.s4;
-    *(__global float *)(dst_addr + 61 * dst_stride_z) = out7.s5;
-    *(__global float *)(dst_addr + 62 * dst_stride_z) = out7.s6;
-    *(__global float *)(dst_addr + 63 * dst_stride_z) = out7.s7;
+    *(__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z)  = out1.s0;
+    *(__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z)  = out1.s1;
+    *(__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z) = out1.s2;
+    *(__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z) = out1.s3;
+    *(__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z) = out1.s4;
+    *(__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z) = out1.s5;
+    *(__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z) = out1.s6;
+    *(__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z) = out1.s7;
+    *(__global DATA_TYPE *)(dst_addr + 16 * dst_stride_z) = out2.s0;
+    *(__global DATA_TYPE *)(dst_addr + 17 * dst_stride_z) = out2.s1;
+    *(__global DATA_TYPE *)(dst_addr + 18 * dst_stride_z) = out2.s2;
+    *(__global DATA_TYPE *)(dst_addr + 19 * dst_stride_z) = out2.s3;
+    *(__global DATA_TYPE *)(dst_addr + 20 * dst_stride_z) = out2.s4;
+    *(__global DATA_TYPE *)(dst_addr + 21 * dst_stride_z) = out2.s5;
+    *(__global DATA_TYPE *)(dst_addr + 22 * dst_stride_z) = out2.s6;
+    *(__global DATA_TYPE *)(dst_addr + 23 * dst_stride_z) = out2.s7;
+    *(__global DATA_TYPE *)(dst_addr + 24 * dst_stride_z) = out3.s0;
+    *(__global DATA_TYPE *)(dst_addr + 25 * dst_stride_z) = out3.s1;
+    *(__global DATA_TYPE *)(dst_addr + 26 * dst_stride_z) = out3.s2;
+    *(__global DATA_TYPE *)(dst_addr + 27 * dst_stride_z) = out3.s3;
+    *(__global DATA_TYPE *)(dst_addr + 28 * dst_stride_z) = out3.s4;
+    *(__global DATA_TYPE *)(dst_addr + 29 * dst_stride_z) = out3.s5;
+    *(__global DATA_TYPE *)(dst_addr + 30 * dst_stride_z) = out3.s6;
+    *(__global DATA_TYPE *)(dst_addr + 31 * dst_stride_z) = out3.s7;
+    *(__global DATA_TYPE *)(dst_addr + 32 * dst_stride_z) = out4.s0;
+    *(__global DATA_TYPE *)(dst_addr + 33 * dst_stride_z) = out4.s1;
+    *(__global DATA_TYPE *)(dst_addr + 34 * dst_stride_z) = out4.s2;
+    *(__global DATA_TYPE *)(dst_addr + 35 * dst_stride_z) = out4.s3;
+    *(__global DATA_TYPE *)(dst_addr + 36 * dst_stride_z) = out4.s4;
+    *(__global DATA_TYPE *)(dst_addr + 37 * dst_stride_z) = out4.s5;
+    *(__global DATA_TYPE *)(dst_addr + 38 * dst_stride_z) = out4.s6;
+    *(__global DATA_TYPE *)(dst_addr + 39 * dst_stride_z) = out4.s7;
+    *(__global DATA_TYPE *)(dst_addr + 40 * dst_stride_z) = out5.s0;
+    *(__global DATA_TYPE *)(dst_addr + 41 * dst_stride_z) = out5.s1;
+    *(__global DATA_TYPE *)(dst_addr + 42 * dst_stride_z) = out5.s2;
+    *(__global DATA_TYPE *)(dst_addr + 43 * dst_stride_z) = out5.s3;
+    *(__global DATA_TYPE *)(dst_addr + 44 * dst_stride_z) = out5.s4;
+    *(__global DATA_TYPE *)(dst_addr + 45 * dst_stride_z) = out5.s5;
+    *(__global DATA_TYPE *)(dst_addr + 46 * dst_stride_z) = out5.s6;
+    *(__global DATA_TYPE *)(dst_addr + 47 * dst_stride_z) = out5.s7;
+    *(__global DATA_TYPE *)(dst_addr + 48 * dst_stride_z) = out6.s0;
+    *(__global DATA_TYPE *)(dst_addr + 49 * dst_stride_z) = out6.s1;
+    *(__global DATA_TYPE *)(dst_addr + 50 * dst_stride_z) = out6.s2;
+    *(__global DATA_TYPE *)(dst_addr + 51 * dst_stride_z) = out6.s3;
+    *(__global DATA_TYPE *)(dst_addr + 52 * dst_stride_z) = out6.s4;
+    *(__global DATA_TYPE *)(dst_addr + 53 * dst_stride_z) = out6.s5;
+    *(__global DATA_TYPE *)(dst_addr + 54 * dst_stride_z) = out6.s6;
+    *(__global DATA_TYPE *)(dst_addr + 55 * dst_stride_z) = out6.s7;
+    *(__global DATA_TYPE *)(dst_addr + 56 * dst_stride_z) = out7.s0;
+    *(__global DATA_TYPE *)(dst_addr + 57 * dst_stride_z) = out7.s1;
+    *(__global DATA_TYPE *)(dst_addr + 58 * dst_stride_z) = out7.s2;
+    *(__global DATA_TYPE *)(dst_addr + 59 * dst_stride_z) = out7.s3;
+    *(__global DATA_TYPE *)(dst_addr + 60 * dst_stride_z) = out7.s4;
+    *(__global DATA_TYPE *)(dst_addr + 61 * dst_stride_z) = out7.s5;
+    *(__global DATA_TYPE *)(dst_addr + 62 * dst_stride_z) = out7.s6;
+    *(__global DATA_TYPE *)(dst_addr + 63 * dst_stride_z) = out7.s7;
 #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL)
 }
 #endif // defined(SRC_DIM_Z)
@@ -1004,8 +1052,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1052,8 +1101,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1100,8 +1150,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1148,8 +1199,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1196,8 +1248,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1246,8 +1299,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1294,8 +1348,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1342,8 +1397,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1390,8 +1446,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1438,8 +1495,9 @@
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
diff --git a/src/core/CL/cl_kernels/winograd_input_transform.cl b/src/core/CL/cl_kernels/winograd_input_transform.cl
index da18e4a..205e416 100644
--- a/src/core/CL/cl_kernels/winograd_input_transform.cl
+++ b/src/core/CL/cl_kernels/winograd_input_transform.cl
@@ -52,8 +52,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2
  * @note If this kernel is used to perform Winograd input transform 3x1, -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd input transform 1x3, -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -79,76 +80,86 @@
     int z = get_global_id(2);
 
     // Compute input address
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(float) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z;
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z;
 
-    src_addr = src_addr - ((int)PAD_LEFT * sizeof(float)) - ((int)PAD_TOP * src_stride_y);
+    src_addr = src_addr - ((int)PAD_LEFT * sizeof(DATA_TYPE)) - ((int)PAD_TOP * src_stride_y);
 
 #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL)
-    float4 in_row0 = vload4(0, (__global float *)(src_addr));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row0 = vload4(0, (__global DATA_TYPE *)(src_addr));
 #elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
-    float4 in_row0 = (float4)(*((__global float *)(src_addr + 0 * src_stride_y)),
-                              *((__global float *)(src_addr + 1 * src_stride_y)),
-                              *((__global float *)(src_addr + 2 * src_stride_y)),
-                              *((__global float *)(src_addr + 3 * src_stride_y)));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)),
+                                            *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)),
+                                            *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)),
+                                            *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y)));
 #else                                            // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    float4       in_row0 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float4       in_row1 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float4       in_row2 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
-    float4       in_row3 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row0 = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row1 = vload4(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row2 = vload4(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row3 = vload4(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y));
 #endif                                           // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
-    float4 tmp0 = in_row0;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp0 = in_row0;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
     tmp0 -= in_row2;
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
-    float out00 = tmp0.s0 - tmp0.s2;
-    float out01 = tmp0.s1 + tmp0.s2;
-    float out02 = tmp0.s2 - tmp0.s1;
-    float out03 = tmp0.s1 - tmp0.s3;
+    DATA_TYPE out00 = tmp0.s0 - tmp0.s2;
+    DATA_TYPE out01 = tmp0.s1 + tmp0.s2;
+    DATA_TYPE out02 = tmp0.s2 - tmp0.s1;
+    DATA_TYPE out03 = tmp0.s1 - tmp0.s3;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    float4 tmp1 = in_row1 + in_row2;
-    float4 tmp2 = in_row2 - in_row1;
-    float4 tmp3 = in_row1 - in_row3;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp1 = in_row1 + in_row2;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp2 = in_row2 - in_row1;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp3 = in_row1 - in_row3;
 
-    float out10 = tmp1.s0 - tmp1.s2;
-    float out11 = tmp1.s1 + tmp1.s2;
-    float out12 = tmp1.s2 - tmp1.s1;
-    float out13 = tmp1.s1 - tmp1.s3;
+    DATA_TYPE out10 = tmp1.s0 - tmp1.s2;
+    DATA_TYPE out11 = tmp1.s1 + tmp1.s2;
+    DATA_TYPE out12 = tmp1.s2 - tmp1.s1;
+    DATA_TYPE out13 = tmp1.s1 - tmp1.s3;
 
-    float out20 = tmp2.s0 - tmp2.s2;
-    float out21 = tmp2.s1 + tmp2.s2;
-    float out22 = tmp2.s2 - tmp2.s1;
-    float out23 = tmp2.s1 - tmp2.s3;
+    DATA_TYPE out20 = tmp2.s0 - tmp2.s2;
+    DATA_TYPE out21 = tmp2.s1 + tmp2.s2;
+    DATA_TYPE out22 = tmp2.s2 - tmp2.s1;
+    DATA_TYPE out23 = tmp2.s1 - tmp2.s3;
 
-    float out30 = tmp3.s0 - tmp3.s2;
-    float out31 = tmp3.s1 + tmp3.s2;
-    float out32 = tmp3.s2 - tmp3.s1;
-    float out33 = tmp3.s1 - tmp3.s3;
+    DATA_TYPE out30 = tmp3.s0 - tmp3.s2;
+    DATA_TYPE out31 = tmp3.s1 + tmp3.s2;
+    DATA_TYPE out32 = tmp3.s2 - tmp3.s1;
+    DATA_TYPE out33 = tmp3.s1 - tmp3.s3;
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(float) + (x + y * (int)NUM_TILES_X) * dst_stride_y;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y;
 
-    *((__global float *)(dst_addr + 0 * dst_stride_z)) = out00; // in_row0.s0; out00;
-    *((__global float *)(dst_addr + 1 * dst_stride_z)) = out01; // in_row0.s1; out01;
-    *((__global float *)(dst_addr + 2 * dst_stride_z)) = out02; // in_row0.s2; out02;
-    *((__global float *)(dst_addr + 3 * dst_stride_z)) = out03; // in_row0.s3; out03;
+    *((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z)) = out00; // in_row0.s0; out00;
+    *((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z)) = out01; // in_row0.s1; out01;
+    *((__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z)) = out02; // in_row0.s2; out02;
+    *((__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z)) = out03; // in_row0.s3; out03;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    *((__global float *)(dst_addr + 4 * dst_stride_z))  = out10;
-    *((__global float *)(dst_addr + 5 * dst_stride_z))  = out11;
-    *((__global float *)(dst_addr + 6 * dst_stride_z))  = out12;
-    *((__global float *)(dst_addr + 7 * dst_stride_z))  = out13;
-    *((__global float *)(dst_addr + 8 * dst_stride_z))  = out20;
-    *((__global float *)(dst_addr + 9 * dst_stride_z))  = out21;
-    *((__global float *)(dst_addr + 10 * dst_stride_z)) = out22;
-    *((__global float *)(dst_addr + 11 * dst_stride_z)) = out23;
-    *((__global float *)(dst_addr + 12 * dst_stride_z)) = out30;
-    *((__global float *)(dst_addr + 13 * dst_stride_z)) = out31;
-    *((__global float *)(dst_addr + 14 * dst_stride_z)) = out32;
-    *((__global float *)(dst_addr + 15 * dst_stride_z)) = out33;
+    *((__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z))  = out10;
+    *((__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z))  = out11;
+    *((__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z))  = out12;
+    *((__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z))  = out13;
+    *((__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z))  = out20;
+    *((__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z))  = out21;
+    *((__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z)) = out22;
+    *((__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z)) = out23;
+    *((__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z)) = out30;
+    *((__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z)) = out31;
+    *((__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z)) = out32;
+    *((__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z)) = out33;
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 }
 
@@ -160,8 +171,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2
  * @note If this kernel is used to perform Winograd input transform 3x1, -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd input transform 1x3, -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -187,97 +199,133 @@
     int z = get_global_id(2) * 2;
 
     // Compute input address
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(float) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z;
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z;
 
-    src_addr = src_addr - ((int)PAD_LEFT * sizeof(float)) - ((int)PAD_TOP * src_stride_y);
+    src_addr = src_addr - ((int)PAD_LEFT * sizeof(DATA_TYPE)) - ((int)PAD_TOP * src_stride_y);
 
 #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL)
-    float4 in_row0 = vload4(0, (__global float *)(src_addr));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row0 = vload4(0, (__global DATA_TYPE *)(src_addr));
 #elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
-    float4 in_row0 = (float4)(*((__global float *)(src_addr + 0 * src_stride_y)),
-                              *((__global float *)(src_addr + 1 * src_stride_y)),
-                              *((__global float *)(src_addr + 2 * src_stride_y)),
-                              *((__global float *)(src_addr + 3 * src_stride_y)));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row0 = (VEC_DATA_TYPE(DATA_TYPE, 4))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)),
+                                            *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)),
+                                            *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)),
+                                            *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y)));
 #else                                            // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    float4       in_row0 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float4       in_row1 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float4       in_row2 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
-    float4       in_row3 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row0 = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row1 = vload4(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row2 = vload4(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row3 = vload4(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y));
 #endif                                           // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
     src_addr += src_stride_z;
 #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL)
-    float4 in_row4 = vload4(0, (__global float *)(src_addr));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row4 = vload4(0, (__global DATA_TYPE *)(src_addr));
 #elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
-    float4 in_row4 = (float4)(*((__global float *)(src_addr + 0 * src_stride_y)),
-                              *((__global float *)(src_addr + 1 * src_stride_y)),
-                              *((__global float *)(src_addr + 2 * src_stride_y)),
-                              *((__global float *)(src_addr + 3 * src_stride_y)));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row4 = (VEC_DATA_TYPE(DATA_TYPE, 4))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)),
+                                            *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)),
+                                            *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)),
+                                            *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y)));
 #else                                            // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    float4       in_row4 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float4       in_row5 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float4       in_row6 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
-    float4       in_row7 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row4 = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row5 = vload4(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row6 = vload4(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    in_row7 = vload4(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y));
 #endif                                           // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
-    float4 tmp0 = in_row0;
-    float4 tmp4 = in_row4;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp0 = in_row0;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp4 = in_row4;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
     tmp0 -= in_row2;
     tmp4 -= in_row6;
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
-    float2 out00 = (float2)(tmp0.s0 - tmp0.s2, tmp4.s0 - tmp4.s2);
-    float2 out01 = (float2)(tmp0.s1 + tmp0.s2, tmp4.s1 + tmp4.s2);
-    float2 out02 = (float2)(tmp0.s2 - tmp0.s1, tmp4.s2 - tmp4.s1);
-    float2 out03 = (float2)(tmp0.s1 - tmp0.s3, tmp4.s1 - tmp4.s3);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out00 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp0.s0 - tmp0.s2, tmp4.s0 - tmp4.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out01 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp0.s1 + tmp0.s2, tmp4.s1 + tmp4.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out02 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp0.s2 - tmp0.s1, tmp4.s2 - tmp4.s1);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out03 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp0.s1 - tmp0.s3, tmp4.s1 - tmp4.s3);
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    float4 tmp1 = in_row1 + in_row2;
-    float4 tmp2 = in_row2 - in_row1;
-    float4 tmp3 = in_row1 - in_row3;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp1 = in_row1 + in_row2;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp2 = in_row2 - in_row1;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp3 = in_row1 - in_row3;
 
-    float4 tmp5 = in_row5 + in_row6;
-    float4 tmp6 = in_row6 - in_row5;
-    float4 tmp7 = in_row5 - in_row7;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp5 = in_row5 + in_row6;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp6 = in_row6 - in_row5;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp7 = in_row5 - in_row7;
 
-    float2 out10 = (float2)(tmp1.s0 - tmp1.s2, tmp5.s0 - tmp5.s2);
-    float2 out11 = (float2)(tmp1.s1 + tmp1.s2, tmp5.s1 + tmp5.s2);
-    float2 out12 = (float2)(tmp1.s2 - tmp1.s1, tmp5.s2 - tmp5.s1);
-    float2 out13 = (float2)(tmp1.s1 - tmp1.s3, tmp5.s1 - tmp5.s3);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out10 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp1.s0 - tmp1.s2, tmp5.s0 - tmp5.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out11 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp1.s1 + tmp1.s2, tmp5.s1 + tmp5.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out12 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp1.s2 - tmp1.s1, tmp5.s2 - tmp5.s1);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out13 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp1.s1 - tmp1.s3, tmp5.s1 - tmp5.s3);
 
-    float2 out20 = (float2)(tmp2.s0 - tmp2.s2, tmp6.s0 - tmp6.s2);
-    float2 out21 = (float2)(tmp2.s1 + tmp2.s2, tmp6.s1 + tmp6.s2);
-    float2 out22 = (float2)(tmp2.s2 - tmp2.s1, tmp6.s2 - tmp6.s1);
-    float2 out23 = (float2)(tmp2.s1 - tmp2.s3, tmp6.s1 - tmp6.s3);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out20 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp2.s0 - tmp2.s2, tmp6.s0 - tmp6.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out21 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp2.s1 + tmp2.s2, tmp6.s1 + tmp6.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out22 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp2.s2 - tmp2.s1, tmp6.s2 - tmp6.s1);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out23 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp2.s1 - tmp2.s3, tmp6.s1 - tmp6.s3);
 
-    float2 out30 = (float2)(tmp3.s0 - tmp3.s2, tmp7.s0 - tmp7.s2);
-    float2 out31 = (float2)(tmp3.s1 + tmp3.s2, tmp7.s1 + tmp7.s2);
-    float2 out32 = (float2)(tmp3.s2 - tmp3.s1, tmp7.s2 - tmp7.s1);
-    float2 out33 = (float2)(tmp3.s1 - tmp3.s3, tmp7.s1 - tmp7.s3);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out30 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp3.s0 - tmp3.s2, tmp7.s0 - tmp7.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out31 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp3.s1 + tmp3.s2, tmp7.s1 + tmp7.s2);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out32 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp3.s2 - tmp3.s1, tmp7.s2 - tmp7.s1);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    out33 = (VEC_DATA_TYPE(DATA_TYPE, 2))(tmp3.s1 - tmp3.s3, tmp7.s1 - tmp7.s3);
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(float) + (x + y * (int)NUM_TILES_X) * dst_stride_y;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y;
 
-    vstore2(out00, 0, (__global float *)(dst_addr + 0 * dst_stride_z));
-    vstore2(out01, 0, (__global float *)(dst_addr + 1 * dst_stride_z));
-    vstore2(out02, 0, (__global float *)(dst_addr + 2 * dst_stride_z));
-    vstore2(out03, 0, (__global float *)(dst_addr + 3 * dst_stride_z));
+    vstore2(out00, 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z));
+    vstore2(out01, 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z));
+    vstore2(out02, 0, (__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z));
+    vstore2(out03, 0, (__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z));
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    vstore2(out10, 0, (__global float *)(dst_addr + 4 * dst_stride_z));
-    vstore2(out11, 0, (__global float *)(dst_addr + 5 * dst_stride_z));
-    vstore2(out12, 0, (__global float *)(dst_addr + 6 * dst_stride_z));
-    vstore2(out13, 0, (__global float *)(dst_addr + 7 * dst_stride_z));
-    vstore2(out20, 0, (__global float *)(dst_addr + 8 * dst_stride_z));
-    vstore2(out21, 0, (__global float *)(dst_addr + 9 * dst_stride_z));
-    vstore2(out22, 0, (__global float *)(dst_addr + 10 * dst_stride_z));
-    vstore2(out23, 0, (__global float *)(dst_addr + 11 * dst_stride_z));
-    vstore2(out30, 0, (__global float *)(dst_addr + 12 * dst_stride_z));
-    vstore2(out31, 0, (__global float *)(dst_addr + 13 * dst_stride_z));
-    vstore2(out32, 0, (__global float *)(dst_addr + 14 * dst_stride_z));
-    vstore2(out33, 0, (__global float *)(dst_addr + 15 * dst_stride_z));
+    vstore2(out10, 0, (__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z));
+    vstore2(out11, 0, (__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z));
+    vstore2(out12, 0, (__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z));
+    vstore2(out13, 0, (__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z));
+    vstore2(out20, 0, (__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z));
+    vstore2(out21, 0, (__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z));
+    vstore2(out22, 0, (__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z));
+    vstore2(out23, 0, (__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z));
+    vstore2(out30, 0, (__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z));
+    vstore2(out31, 0, (__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z));
+    vstore2(out32, 0, (__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z));
+    vstore2(out33, 0, (__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z));
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 }
 
@@ -289,8 +337,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2
  * @note If this kernel is used to perform Winograd input transform 3x1, -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd input transform 1x3, -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -316,30 +365,34 @@
     int z = get_global_id(2);
 
     // Compute input address
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(float) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z;
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z;
 
-    src_addr = src_addr - ((int)PAD_LEFT * sizeof(float)) - ((int)PAD_TOP * src_stride_y);
+    src_addr = src_addr - ((int)PAD_LEFT * sizeof(DATA_TYPE)) - ((int)PAD_TOP * src_stride_y);
 
 #if defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
     // Row0
-    float4 d00 = (float4)(*((__global float *)(src_addr + 0 * src_stride_y)),
-                          *((__global float *)(src_addr + 1 * src_stride_y)),
-                          *((__global float *)(src_addr + 2 * src_stride_y)),
-                          *((__global float *)(src_addr + 3 * src_stride_y)));
-    float2 d01 = (float2)(*((__global float *)(src_addr + 4 * src_stride_y)),
-                          *((__global float *)(src_addr + 5 * src_stride_y)));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    d00 = (VEC_DATA_TYPE(DATA_TYPE, 4))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)),
+                                        *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)),
+                                        *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)),
+                                        *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y)));
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    d01 = (VEC_DATA_TYPE(DATA_TYPE, 2))(*((__global DATA_TYPE *)(src_addr + 4 * src_stride_y)),
+                                        *((__global DATA_TYPE *)(src_addr + 5 * src_stride_y)));
 #else  // defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
     // Row0
-    float4       d00     = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
-    float2       d01     = vload2(2, (__global float *)(src_addr + 0 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    d00 = vload4(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    d01 = vload2(2, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
 #endif // defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
-    float out0 = 0.0f;
-    float out1 = 0.0f;
-    float out2 = 0.0f;
-    float out3 = 0.0f;
-    float out4 = 0.0f;
-    float out5 = 0.0f;
+    DATA_TYPE out0 = 0.0f;
+    DATA_TYPE out1 = 0.0f;
+    DATA_TYPE out2 = 0.0f;
+    DATA_TYPE out3 = 0.0f;
+    DATA_TYPE out4 = 0.0f;
+    DATA_TYPE out5 = 0.0f;
 
     // Channels [0, 5]: [out00, out01, out02, out03, out04, out05]
     out0 += 16.0f * d00.s0 - 20.0f * d00.s2 + 4.0f * d01.s0;
@@ -351,16 +404,18 @@
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
     // Row4
-    float4 d40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y));
-    float2 d41 = vload2(2, (__global float *)(src_addr + 4 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    d40 = vload4(0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    d41 = vload2(2, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y));
 
     // k0, k1, k2, k3, k4, k5 are common terms for row0, row1, row2, row3 and row4
-    float k0 = d41.s0;
-    float k1 = d41.s0;
-    float k2 = d41.s0;
-    float k3 = d41.s0;
-    float k4 = d41.s0;
-    float k5 = 0.0f;
+    DATA_TYPE k0 = d41.s0;
+    DATA_TYPE k1 = d41.s0;
+    DATA_TYPE k2 = d41.s0;
+    DATA_TYPE k3 = d41.s0;
+    DATA_TYPE k4 = d41.s0;
+    DATA_TYPE k5 = 0.0f;
 
     k0 += 4.0f * d40.s0 - 5.0f * d40.s2;
     k1 += -4.0f * d40.s1 - 4.0f * d40.s2 + d40.s3;
@@ -377,8 +432,10 @@
     out5 += k5;
 
     // Row2
-    float4 d20 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
-    float2 d21 = vload2(2, (__global float *)(src_addr + 2 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    d20 = vload4(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    d21 = vload2(2, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
 
     out0 += -20.0f * d20.s0 + 25.0f * d20.s2 - 5.0f * d21.s0;
     out1 += +20.0f * d20.s1 + 20.0f * d20.s2 - 5.0f * d20.s3 - 5.0f * d21.s0;
@@ -389,9 +446,9 @@
 #endif // #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
     // Compute destination address
-    __global float *dst_addr = (__global float *)(dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(float) + (x + y * (int)NUM_TILES_X) * dst_stride_y);
+    __global DATA_TYPE *dst_addr = (__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y);
 
-    uint dst_plane_stride = dst_stride_z / sizeof(float);
+    uint dst_plane_stride = dst_stride_z / sizeof(DATA_TYPE);
 
     *(dst_addr) = out0;
     dst_addr += dst_plane_stride;
@@ -407,69 +464,73 @@
     dst_addr += dst_plane_stride;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    float out6  = k0;
-    float out7  = k1;
-    float out8  = k2;
-    float out9  = k3;
-    float out10 = k4;
-    float out11 = k5;
-    float out12 = k0;
-    float out13 = k1;
-    float out14 = k2;
-    float out15 = k3;
-    float out16 = k4;
-    float out17 = k5;
-    float out18 = k0;
-    float out19 = k1;
-    float out20 = k2;
-    float out21 = k3;
-    float out22 = k4;
-    float out23 = k5;
-    float out24 = k0;
-    float out25 = k1;
-    float out26 = k2;
-    float out27 = k3;
-    float out28 = k4;
-    float out29 = k5;
+    DATA_TYPE out6  = k0;
+    DATA_TYPE out7  = k1;
+    DATA_TYPE out8  = k2;
+    DATA_TYPE out9  = k3;
+    DATA_TYPE out10 = k4;
+    DATA_TYPE out11 = k5;
+    DATA_TYPE out12 = k0;
+    DATA_TYPE out13 = k1;
+    DATA_TYPE out14 = k2;
+    DATA_TYPE out15 = k3;
+    DATA_TYPE out16 = k4;
+    DATA_TYPE out17 = k5;
+    DATA_TYPE out18 = k0;
+    DATA_TYPE out19 = k1;
+    DATA_TYPE out20 = k2;
+    DATA_TYPE out21 = k3;
+    DATA_TYPE out22 = k4;
+    DATA_TYPE out23 = k5;
+    DATA_TYPE out24 = k0;
+    DATA_TYPE out25 = k1;
+    DATA_TYPE out26 = k2;
+    DATA_TYPE out27 = k3;
+    DATA_TYPE out28 = k4;
+    DATA_TYPE out29 = k5;
 
     // Row1
-    float4 d10 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
-    float2 d11 = vload2(2, (__global float *)(src_addr + 1 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    d10 = vload4(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    d11 = vload2(2, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
 
     // Row3
-    float4 d30 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
-    float2 d31 = vload2(2, (__global float *)(src_addr + 3 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    d30 = vload4(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    d31 = vload2(2, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y));
 
     // Compute common parts for the channels between [6, 29]
     // Channels [6, 11]:  [out10, out11, out12, out13, out14, out15]
     // Channels [12, 17]: [out20, out21, out22, out23, out24, out25]
-    float part0  = -16.0f * d20.s0 + 20.0f * d20.s2 - 4.0f * d21.s0;
-    float part1  = 16.0f * d10.s0 - 20.0f * d10.s2 + 4.0f * d11.s0 - 4.0f * d30.s0 + 5.0f * d30.s2 - d31.s0;
-    float part2  = 16.0f * d20.s2 - 4.0f * d21.s0;
-    float part3  = 16.0f * d20.s1 - 4.0f * d20.s3;
-    float part4  = 16.0f * d10.s2 - 4.0f * d11.s0 - 4.0f * d30.s2 + d31.s0;
-    float part5  = 16.0f * d10.s1 - 4.0f * d10.s3 - 4.0f * d30.s1 + d30.s3;
-    float part6  = 4.0f * d20.s2 - 4.0f * d21.s0;
-    float part7  = 8.0f * d10.s1 - 8.0f * d10.s3 - 2.0f * d30.s1 + 2.0f * d30.s3;
-    float part8  = 4.0f * d10.s2 - 4.0f * d11.s0 - d30.s2 + d31.s0;
-    float part9  = 8.0f * d20.s1 - 8.0f * d20.s3;
-    float part10 = -16.0f * d20.s1 + 20.0f * d20.s3 - 4.0f * d21.s1;
-    float part11 = -16.0f * d10.s1 + 20.0f * d10.s3 - 4.0f * d11.s1 + 4.0f * d30.s1 - 5.0f * d30.s3 + d31.s1;
+    DATA_TYPE part0  = -16.0f * d20.s0 + 20.0f * d20.s2 - 4.0f * d21.s0;
+    DATA_TYPE part1  = 16.0f * d10.s0 - 20.0f * d10.s2 + 4.0f * d11.s0 - 4.0f * d30.s0 + 5.0f * d30.s2 - d31.s0;
+    DATA_TYPE part2  = 16.0f * d20.s2 - 4.0f * d21.s0;
+    DATA_TYPE part3  = 16.0f * d20.s1 - 4.0f * d20.s3;
+    DATA_TYPE part4  = 16.0f * d10.s2 - 4.0f * d11.s0 - 4.0f * d30.s2 + d31.s0;
+    DATA_TYPE part5  = 16.0f * d10.s1 - 4.0f * d10.s3 - 4.0f * d30.s1 + d30.s3;
+    DATA_TYPE part6  = 4.0f * d20.s2 - 4.0f * d21.s0;
+    DATA_TYPE part7  = 8.0f * d10.s1 - 8.0f * d10.s3 - 2.0f * d30.s1 + 2.0f * d30.s3;
+    DATA_TYPE part8  = 4.0f * d10.s2 - 4.0f * d11.s0 - d30.s2 + d31.s0;
+    DATA_TYPE part9  = 8.0f * d20.s1 - 8.0f * d20.s3;
+    DATA_TYPE part10 = -16.0f * d20.s1 + 20.0f * d20.s3 - 4.0f * d21.s1;
+    DATA_TYPE part11 = -16.0f * d10.s1 + 20.0f * d10.s3 - 4.0f * d11.s1 + 4.0f * d30.s1 - 5.0f * d30.s3 + d31.s1;
 
     // Channels [18, 23]: [out30, out31, out32, out33, out34, out35]
     // Channels [24, 29]: [out40, out41, out42, out43, out44, out45]
-    float part12 = 8.0f * d10.s0 - 10.0f * d10.s2 + 2.0f * d11.s0 - 8.0f * d30.s0 + 10.0f * d30.s2 - 2.0f * d31.s0;
-    float part13 = part0 * 0.25f; // -4.0f * d20.s0 + 5.0f * d20.s2 - d21.s0
-    float part14 = part2 * 0.25f; // 4.0f * d20.s2 - d21.s0
-    float part15 = 8.0f * d10.s1 - 2.0f * d10.s3 - 8.0f * d30.s1 + 2.0f * d30.s3;
-    float part16 = 8.0f * d10.s2 - 2.0f * d11.s0 - 8.0f * d30.s2 + 2.0f * d31.s0;
-    float part17 = part3 * 0.25f; // 4.0f * d20.s1 - d20.s3
-    float part18 = part6 * 0.25f; // d20.s2 - d21.s0
-    float part19 = 4.0f * d10.s1 - 4.0f * d10.s3 - 4.0f * d30.s1 + 4.0f * d30.s3;
-    float part20 = 2.0f * d10.s2 - 2.0f * d11.s0 - 2.0f * d30.s2 + 2.0f * d31.s0;
-    float part21 = part9 * 0.25f;                                                 // 2.0f * (d20.s1 - d20.s3)
-    float part22 = part10 * 0.25f;                                                // - 4.0f * d20.s1 + 5.0f * d20.s3 - d21.s1
-    float part23 = part11 * 0.5f + 6.0f * d30.s1 - 7.5f * d30.s3 + 1.5f * d31.s1; // - 8.0f * d10.s1 + 10.0f * d10.s3 - 2.0f * d11.s1 + 8.0f * d30.s1 - 10.0f * d30.s3 + 2.0f * d31.s1;
+    DATA_TYPE part12 = 8.0f * d10.s0 - 10.0f * d10.s2 + 2.0f * d11.s0 - 8.0f * d30.s0 + 10.0f * d30.s2 - 2.0f * d31.s0;
+    DATA_TYPE part13 = part0 * 0.25f; // -4.0f * d20.s0 + 5.0f * d20.s2 - d21.s0
+    DATA_TYPE part14 = part2 * 0.25f; // 4.0f * d20.s2 - d21.s0
+    DATA_TYPE part15 = 8.0f * d10.s1 - 2.0f * d10.s3 - 8.0f * d30.s1 + 2.0f * d30.s3;
+    DATA_TYPE part16 = 8.0f * d10.s2 - 2.0f * d11.s0 - 8.0f * d30.s2 + 2.0f * d31.s0;
+    DATA_TYPE part17 = part3 * 0.25f; // 4.0f * d20.s1 - d20.s3
+    DATA_TYPE part18 = part6 * 0.25f; // d20.s2 - d21.s0
+    DATA_TYPE part19 = 4.0f * d10.s1 - 4.0f * d10.s3 - 4.0f * d30.s1 + 4.0f * d30.s3;
+    DATA_TYPE part20 = 2.0f * d10.s2 - 2.0f * d11.s0 - 2.0f * d30.s2 + 2.0f * d31.s0;
+    DATA_TYPE part21 = part9 * 0.25f;                                                 // 2.0f * (d20.s1 - d20.s3)
+    DATA_TYPE part22 = part10 * 0.25f;                                                // - 4.0f * d20.s1 + 5.0f * d20.s3 - d21.s1
+    DATA_TYPE part23 = part11 * 0.5f + 6.0f * d30.s1 - 7.5f * d30.s3 + 1.5f * d31.s1; // - 8.0f * d10.s1 + 10.0f * d10.s3 - 2.0f * d11.s1 + 8.0f * d30.s1 - 10.0f * d30.s3 + 2.0f * d31.s1;
 
     out6 += part0 - part1;
     out12 += part0 + part1;
@@ -548,8 +609,10 @@
     dst_addr += dst_plane_stride;
 
     // Row5
-    float4 d50 = vload4(0, (__global float *)(src_addr + 5 * src_stride_y));
-    float2 d51 = vload2(2, (__global float *)(src_addr + 5 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    d50 = vload4(0, (__global DATA_TYPE *)(src_addr + 5 * src_stride_y));
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    d51 = vload2(2, (__global DATA_TYPE *)(src_addr + 5 * src_stride_y));
 
     // Channels [30, 35]
     out0 = 16.0f * d10.s0 - 20.0f * d10.s2 - 20.0f * d30.s0 + 25.0f * d30.s2 + 4.0f * d50.s0 - 5.0f * d50.s2 + d51.s0 + 4.0f * d11.s0 - 5.0f * d31.s0;
@@ -585,8 +648,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note If this kernel is used to perform Winograd input transform 3x1, -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd input transform 1x3, -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -611,7 +675,7 @@
     int y = get_global_id(1);
     int z = get_global_id(2);
 
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(float);
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE);
 
     // Clamp coordinates. This clamp is valid for all rows
     int4 y_coord0 = (int4)(y * OUTPUT_TILE_W) + (int4)(0, 1, 2, 3) - (int4)PAD_LEFT;
@@ -637,19 +701,19 @@
     // Clamp z coordinate
     z_coord = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    float d40 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d41 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    float d42 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    float d43 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    float d44 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d45 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d40 = *(__global DATA_TYPE *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d41 = *(__global DATA_TYPE *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d42 = *(__global DATA_TYPE *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d43 = *(__global DATA_TYPE *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d44 = *(__global DATA_TYPE *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d45 = *(__global DATA_TYPE *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
 
-    float k0 = d44;
-    float k1 = d44;
-    float k2 = d44;
-    float k3 = d44;
-    float k4 = d44;
-    float k5 = (float)0.0f;
+    DATA_TYPE k0 = d44;
+    DATA_TYPE k1 = d44;
+    DATA_TYPE k2 = d44;
+    DATA_TYPE k3 = d44;
+    DATA_TYPE k4 = d44;
+    DATA_TYPE k5 = (DATA_TYPE)0.0f;
 
     k0 += 4.0f * d40 - 5.0f * d42;
     k1 += -4.0f * d41 - 4.0f * d42 + d43;
@@ -674,12 +738,12 @@
     valid_y1 = y_coord1;
 #endif // if PAD_TOP == 0, we cannot read out of bound
 
-    float d00 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d01 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    float d02 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    float d03 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    float d04 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d05 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d00 = *(__global DATA_TYPE *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d01 = *(__global DATA_TYPE *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d02 = *(__global DATA_TYPE *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d03 = *(__global DATA_TYPE *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d04 = *(__global DATA_TYPE *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d05 = *(__global DATA_TYPE *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
 #else  // !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
     int4 z_coords0 = (int4)(z * OUTPUT_TILE_H) + (int4)(0, 1, 2, 3) - (int4)PAD_TOP;
     int2 z_coords1 = (int2)(z * OUTPUT_TILE_H) + (int2)(4, 5) - (int2)PAD_TOP;
@@ -692,20 +756,20 @@
     z_coords0 = clamp((int4)z_coords0, (int4)0, (int4)((int)SRC_DIM_2 - 1));
     z_coords1 = clamp((int2)z_coords1, (int2)0, (int2)((int)SRC_DIM_2 - 1));
 
-    float d00 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coords0.s0 * src_stride_z);
-    float d01 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coords0.s1 * src_stride_z);
-    float d02 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coords0.s2 * src_stride_z);
-    float d03 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coords0.s3 * src_stride_z);
-    float d04 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coords1.s0 * src_stride_z);
-    float d05 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coords1.s1 * src_stride_z);
+    DATA_TYPE d00 = *(__global DATA_TYPE *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coords0.s0 * src_stride_z);
+    DATA_TYPE d01 = *(__global DATA_TYPE *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coords0.s1 * src_stride_z);
+    DATA_TYPE d02 = *(__global DATA_TYPE *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coords0.s2 * src_stride_z);
+    DATA_TYPE d03 = *(__global DATA_TYPE *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coords0.s3 * src_stride_z);
+    DATA_TYPE d04 = *(__global DATA_TYPE *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coords1.s0 * src_stride_z);
+    DATA_TYPE d05 = *(__global DATA_TYPE *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coords1.s1 * src_stride_z);
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
-    float out0 = 16.0f * d00 - 20.0f * d02 + 4.0f * d04;
-    float out1 = -16.0f * d01 - 16.0f * d02 + 4.0f * d03 + 4.0f * d04;
-    float out2 = 16.0f * d01 - 16.0f * d02 - 4.0f * d03 + 4.0f * d04;
-    float out3 = -8.0f * d01 - 4.0f * d02 + 8.0f * d03 + 4.0f * d04;
-    float out4 = 8.0f * d01 - 4.0f * d02 - 8.0f * d03 + 4.0f * d04;
-    float out5 = 16.0f * d01 - 20.0f * d03 + 4.0f * d05;
+    DATA_TYPE out0 = 16.0f * d00 - 20.0f * d02 + 4.0f * d04;
+    DATA_TYPE out1 = -16.0f * d01 - 16.0f * d02 + 4.0f * d03 + 4.0f * d04;
+    DATA_TYPE out2 = 16.0f * d01 - 16.0f * d02 - 4.0f * d03 + 4.0f * d04;
+    DATA_TYPE out3 = -8.0f * d01 - 4.0f * d02 + 8.0f * d03 + 4.0f * d04;
+    DATA_TYPE out4 = 8.0f * d01 - 4.0f * d02 - 8.0f * d03 + 4.0f * d04;
+    DATA_TYPE out5 = 16.0f * d01 - 20.0f * d03 + 4.0f * d05;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
     // Row2
@@ -716,12 +780,12 @@
     valid_y1 = select(valid_y1, (int2)SRC_DIM_1, (int2)z_coord >= (int)SRC_DIM_2);
     z_coord  = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    float d20 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d21 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    float d22 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    float d23 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    float d24 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d25 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d20 = *(__global DATA_TYPE *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d21 = *(__global DATA_TYPE *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d22 = *(__global DATA_TYPE *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d23 = *(__global DATA_TYPE *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d24 = *(__global DATA_TYPE *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d25 = *(__global DATA_TYPE *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
 
     out0 += k0;
     out1 += k1;
@@ -729,30 +793,30 @@
     out3 += k3;
     out4 += k4;
     out5 += k5;
-    float out6  = k0;
-    float out7  = k1;
-    float out8  = k2;
-    float out9  = k3;
-    float out10 = k4;
-    float out11 = k5;
-    float out12 = k0;
-    float out13 = k1;
-    float out14 = k2;
-    float out15 = k3;
-    float out16 = k4;
-    float out17 = k5;
-    float out18 = k0;
-    float out19 = k1;
-    float out20 = k2;
-    float out21 = k3;
-    float out22 = k4;
-    float out23 = k5;
-    float out24 = k0;
-    float out25 = k1;
-    float out26 = k2;
-    float out27 = k3;
-    float out28 = k4;
-    float out29 = k5;
+    DATA_TYPE out6  = k0;
+    DATA_TYPE out7  = k1;
+    DATA_TYPE out8  = k2;
+    DATA_TYPE out9  = k3;
+    DATA_TYPE out10 = k4;
+    DATA_TYPE out11 = k5;
+    DATA_TYPE out12 = k0;
+    DATA_TYPE out13 = k1;
+    DATA_TYPE out14 = k2;
+    DATA_TYPE out15 = k3;
+    DATA_TYPE out16 = k4;
+    DATA_TYPE out17 = k5;
+    DATA_TYPE out18 = k0;
+    DATA_TYPE out19 = k1;
+    DATA_TYPE out20 = k2;
+    DATA_TYPE out21 = k3;
+    DATA_TYPE out22 = k4;
+    DATA_TYPE out23 = k5;
+    DATA_TYPE out24 = k0;
+    DATA_TYPE out25 = k1;
+    DATA_TYPE out26 = k2;
+    DATA_TYPE out27 = k3;
+    DATA_TYPE out28 = k4;
+    DATA_TYPE out29 = k5;
 
     // Channels [0, 5]: [out00, out01, out02, out03, out04, out05]
     out0 += -20.0f * d20 + 25.0f * d22 - 5.0f * d24;
@@ -764,20 +828,20 @@
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
     // Compute destination address
-    __global float *dst_addr         = (__global float *)(dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(float) + (y + z * (int)NUM_TILES_X) * dst_stride_y);
-    uint            dst_plane_stride = dst_stride_z / sizeof(float);
+    __global DATA_TYPE *dst_addr                              = (__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) + (y + z * (int)NUM_TILES_X) * dst_stride_y);
+    uint                                     dst_plane_stride = dst_stride_z / sizeof(DATA_TYPE);
 
-    *((__global float *)dst_addr) = out0;
+    *((__global DATA_TYPE *)dst_addr) = out0;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out1;
+    *((__global DATA_TYPE *)dst_addr) = out1;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out2;
+    *((__global DATA_TYPE *)dst_addr) = out2;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out3;
+    *((__global DATA_TYPE *)dst_addr) = out3;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out4;
+    *((__global DATA_TYPE *)dst_addr) = out4;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out5;
+    *((__global DATA_TYPE *)dst_addr) = out5;
     dst_addr += dst_plane_stride;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
@@ -787,12 +851,12 @@
     valid_y0 = y_coord0;
     valid_y1 = y_coord1;
 
-    float d10 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d11 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    float d12 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    float d13 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    float d14 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d15 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d10 = *(__global DATA_TYPE *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d11 = *(__global DATA_TYPE *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d12 = *(__global DATA_TYPE *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d13 = *(__global DATA_TYPE *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d14 = *(__global DATA_TYPE *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d15 = *(__global DATA_TYPE *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Row3
     z_coord  = (z * (int)OUTPUT_TILE_H) - (int)PAD_TOP + 3;
@@ -803,43 +867,43 @@
     z_coord  = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
     z_coord  = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    float d30 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d31 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    float d32 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    float d33 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    float d34 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d35 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d30 = *(__global DATA_TYPE *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d31 = *(__global DATA_TYPE *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d32 = *(__global DATA_TYPE *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d33 = *(__global DATA_TYPE *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d34 = *(__global DATA_TYPE *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d35 = *(__global DATA_TYPE *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Compute common parts for the channels between [6, 29]
     // Channels [6, 11]:  [out10, out11, out12, out13, out14, out15]
     // Channels [12, 17]: [out20, out21, out22, out23, out24, out25]
-    float part0  = -16.0f * d20 + 20.0f * d22 - 4.0f * d24;
-    float part1  = 16.0f * d10 - 20.0f * d12 + 4.0f * d14 - 4.0f * d30 + 5.0f * d32 - d34;
-    float part2  = 16.0f * d22 - 4.0f * d24;
-    float part3  = 16.0f * d21 - 4.0f * d23;
-    float part4  = 16.0f * d12 - 4.0f * d14 - 4.0f * d32 + d34;
-    float part5  = 16.0f * d11 - 4.0f * d13 - 4.0f * d31 + d33;
-    float part6  = 4.0f * d22 - 4.0f * d24;
-    float part7  = 8.0f * d11 - 8.0f * d13 - 2.0f * d31 + 2.0f * d33;
-    float part8  = 4.0f * d12 - 4.0f * d14 - d32 + d34;
-    float part9  = 8.0f * d21 - 8.0f * d23;
-    float part10 = -16.0f * d21 + 20.0f * d23 - 4.0f * d25;
-    float part11 = -16.0f * d11 + 20.0f * d13 - 4.0f * d15 + 4.0f * d31 - 5.0f * d33 + d35;
+    DATA_TYPE part0  = -16.0f * d20 + 20.0f * d22 - 4.0f * d24;
+    DATA_TYPE part1  = 16.0f * d10 - 20.0f * d12 + 4.0f * d14 - 4.0f * d30 + 5.0f * d32 - d34;
+    DATA_TYPE part2  = 16.0f * d22 - 4.0f * d24;
+    DATA_TYPE part3  = 16.0f * d21 - 4.0f * d23;
+    DATA_TYPE part4  = 16.0f * d12 - 4.0f * d14 - 4.0f * d32 + d34;
+    DATA_TYPE part5  = 16.0f * d11 - 4.0f * d13 - 4.0f * d31 + d33;
+    DATA_TYPE part6  = 4.0f * d22 - 4.0f * d24;
+    DATA_TYPE part7  = 8.0f * d11 - 8.0f * d13 - 2.0f * d31 + 2.0f * d33;
+    DATA_TYPE part8  = 4.0f * d12 - 4.0f * d14 - d32 + d34;
+    DATA_TYPE part9  = 8.0f * d21 - 8.0f * d23;
+    DATA_TYPE part10 = -16.0f * d21 + 20.0f * d23 - 4.0f * d25;
+    DATA_TYPE part11 = -16.0f * d11 + 20.0f * d13 - 4.0f * d15 + 4.0f * d31 - 5.0f * d33 + d35;
 
     // Channels [18, 23]: [out30, out31, out32, out33, out34, out35]
     // Channels [24, 29]: [out40, out41, out42, out43, out44, out45]
-    float part12 = 8.0f * d10 - 10.0f * d12 + 2.0f * d14 - 8.0f * d30 + 10.0f * d32 - 2.0f * d34;
-    float part13 = part0 * 0.25f; // -4.0f * d20 + 5.0f * d22 - d24
-    float part14 = part2 * 0.25f; // 4.0f * d22 - d24
-    float part15 = 8.0f * d11 - 2.0f * d13 - 8.0f * d31 + 2.0f * d33;
-    float part16 = 8.0f * d12 - 2.0f * d14 - 8.0f * d32 + 2.0f * d34;
-    float part17 = part3 * 0.25f; // 4.0f * d21 - d23
-    float part18 = part6 * 0.25f; // d22 - d24
-    float part19 = 4.0f * d11 - 4.0f * d13 - 4.0f * d31 + 4.0f * d33;
-    float part20 = 2.0f * d12 - 2.0f * d14 - 2.0f * d32 + 2.0f * d34;
-    float part21 = part9 * 0.25f;                                        // 2.0f * (d21 - d23)
-    float part22 = part10 * 0.25f;                                       // - 4.0f * d21 + 5.0f * d23 - d25
-    float part23 = part11 * 0.5f + 6.0f * d31 - 7.5f * d33 + 1.5f * d35; // - 8.0f * d11 + 10.0f * d13 - 2.0f * d15 + 8.0f * d31 - 10.0f * d33 + 2.0f * d35;
+    DATA_TYPE part12 = 8.0f * d10 - 10.0f * d12 + 2.0f * d14 - 8.0f * d30 + 10.0f * d32 - 2.0f * d34;
+    DATA_TYPE part13 = part0 * 0.25f; // -4.0f * d20 + 5.0f * d22 - d24
+    DATA_TYPE part14 = part2 * 0.25f; // 4.0f * d22 - d24
+    DATA_TYPE part15 = 8.0f * d11 - 2.0f * d13 - 8.0f * d31 + 2.0f * d33;
+    DATA_TYPE part16 = 8.0f * d12 - 2.0f * d14 - 8.0f * d32 + 2.0f * d34;
+    DATA_TYPE part17 = part3 * 0.25f; // 4.0f * d21 - d23
+    DATA_TYPE part18 = part6 * 0.25f; // d22 - d24
+    DATA_TYPE part19 = 4.0f * d11 - 4.0f * d13 - 4.0f * d31 + 4.0f * d33;
+    DATA_TYPE part20 = 2.0f * d12 - 2.0f * d14 - 2.0f * d32 + 2.0f * d34;
+    DATA_TYPE part21 = part9 * 0.25f;                                        // 2.0f * (d21 - d23)
+    DATA_TYPE part22 = part10 * 0.25f;                                       // - 4.0f * d21 + 5.0f * d23 - d25
+    DATA_TYPE part23 = part11 * 0.5f + 6.0f * d31 - 7.5f * d33 + 1.5f * d35; // - 8.0f * d11 + 10.0f * d13 - 2.0f * d15 + 8.0f * d31 - 10.0f * d33 + 2.0f * d35;
 
     out6 += part0 - part1;
     out12 += part0 + part1;
@@ -867,54 +931,54 @@
     out23 += part22 + part23;
     out29 += part22 - part23;
 
-    *((__global float *)dst_addr) = out6;
+    *((__global DATA_TYPE *)dst_addr) = out6;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out7;
+    *((__global DATA_TYPE *)dst_addr) = out7;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out8;
+    *((__global DATA_TYPE *)dst_addr) = out8;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out9;
+    *((__global DATA_TYPE *)dst_addr) = out9;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out10;
+    *((__global DATA_TYPE *)dst_addr) = out10;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out11;
+    *((__global DATA_TYPE *)dst_addr) = out11;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out12;
+    *((__global DATA_TYPE *)dst_addr) = out12;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out13;
+    *((__global DATA_TYPE *)dst_addr) = out13;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out14;
+    *((__global DATA_TYPE *)dst_addr) = out14;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out15;
+    *((__global DATA_TYPE *)dst_addr) = out15;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out16;
+    *((__global DATA_TYPE *)dst_addr) = out16;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out17;
+    *((__global DATA_TYPE *)dst_addr) = out17;
     dst_addr += dst_plane_stride;
 
-    *((__global float *)dst_addr) = out18;
+    *((__global DATA_TYPE *)dst_addr) = out18;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out19;
+    *((__global DATA_TYPE *)dst_addr) = out19;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out20;
+    *((__global DATA_TYPE *)dst_addr) = out20;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out21;
+    *((__global DATA_TYPE *)dst_addr) = out21;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out22;
+    *((__global DATA_TYPE *)dst_addr) = out22;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out23;
+    *((__global DATA_TYPE *)dst_addr) = out23;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out24;
+    *((__global DATA_TYPE *)dst_addr) = out24;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out25;
+    *((__global DATA_TYPE *)dst_addr) = out25;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out26;
+    *((__global DATA_TYPE *)dst_addr) = out26;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out27;
+    *((__global DATA_TYPE *)dst_addr) = out27;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out28;
+    *((__global DATA_TYPE *)dst_addr) = out28;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out29;
+    *((__global DATA_TYPE *)dst_addr) = out29;
     dst_addr += dst_plane_stride;
 
     // Row5
@@ -926,12 +990,12 @@
     z_coord  = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
     z_coord  = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    float d50 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d51 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    float d52 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    float d53 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    float d54 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    float d55 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d50 = *(__global DATA_TYPE *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d51 = *(__global DATA_TYPE *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d52 = *(__global DATA_TYPE *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d53 = *(__global DATA_TYPE *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d54 = *(__global DATA_TYPE *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    DATA_TYPE d55 = *(__global DATA_TYPE *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Channels [30, 35]
     out0 = 16.0f * d10 - 20.0f * d12 - 20.0f * d30 + 25.0f * d32 + 4.0f * d50 - 5.0f * d52 + d54 + 4.0f * d14 - 5.0f * d34;
@@ -941,17 +1005,17 @@
     out4 = 8.0f * d11 - 4.0f * d12 - 8.0f * d13 - 10.0f * d31 + 5.0f * d32 + 10.0f * d33 + 2.0f * d51 - 2.0f * d53 - d52 + d54 + 4.0f * d14 - 5.0f * d34;
     out5 = 16.0f * d11 - 20.0f * d13 + 4.0f * d15 - 20.0f * d31 + 25.0f * d33 - 5.0f * d35 + 4.0f * d51 - 5.0f * d53 + d55;
 
-    *((__global float *)dst_addr) = out0;
+    *((__global DATA_TYPE *)dst_addr) = out0;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out1;
+    *((__global DATA_TYPE *)dst_addr) = out1;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out2;
+    *((__global DATA_TYPE *)dst_addr) = out2;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out3;
+    *((__global DATA_TYPE *)dst_addr) = out3;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out4;
+    *((__global DATA_TYPE *)dst_addr) = out4;
     dst_addr += dst_plane_stride;
-    *((__global float *)dst_addr) = out5;
+    *((__global DATA_TYPE *)dst_addr) = out5;
     dst_addr += dst_plane_stride;
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 }
@@ -966,8 +1030,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note If this kernel is used to perform Winograd input transform 5x1, -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd input transform 1x5, -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -993,7 +1058,7 @@
     int z = get_global_id(2);
 
     // Compute input address
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(float);
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE);
 
 #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL)
     // Clamp coordinates. This clamp is valid for all rows
@@ -1005,21 +1070,25 @@
     int z_coord = z * OUTPUT_TILE_H;
 
     // Load the input tile
-    float8 in_row0;
-    in_row0.s0 = *(__global float *)(src_addr + y_coord.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s1 = *(__global float *)(src_addr + y_coord.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s2 = *(__global float *)(src_addr + y_coord.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s3 = *(__global float *)(src_addr + y_coord.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s4 = *(__global float *)(src_addr + y_coord.s4 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s5 = *(__global float *)(src_addr + y_coord.s5 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s6 = *(__global float *)(src_addr + y_coord.s6 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s7 = *(__global float *)(src_addr + y_coord.s7 * (int)src_stride_y + z_coord * src_stride_z);
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    in_row0;
+    in_row0.s0 = *(__global DATA_TYPE *)(src_addr + y_coord.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s1 = *(__global DATA_TYPE *)(src_addr + y_coord.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s2 = *(__global DATA_TYPE *)(src_addr + y_coord.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s3 = *(__global DATA_TYPE *)(src_addr + y_coord.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s4 = *(__global DATA_TYPE *)(src_addr + y_coord.s4 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s5 = *(__global DATA_TYPE *)(src_addr + y_coord.s5 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s6 = *(__global DATA_TYPE *)(src_addr + y_coord.s6 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s7 = *(__global DATA_TYPE *)(src_addr + y_coord.s7 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Calculate common factors for intermediate tensor
-    float8 comm_fact0 = 0.0f;
-    float8 tmp0       = in_row0;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    comm_fact0 = 0.0f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    tmp0 = in_row0;
 
-    float8 out0 = (float8)0.0f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out0 = (VEC_DATA_TYPE(DATA_TYPE, 8))0.0f;
 
     OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0);
 
@@ -1035,25 +1104,30 @@
     z_coord      = clamp(z_coord, (int8)0, (int8)SRC_DIM_2 - 1);                 // Clamp z coordinate
 
     // Load the input tile
-    float8 in_row0;
-    in_row0.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord.s0 * src_stride_z);
-    in_row0.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord.s1 * src_stride_z);
-    in_row0.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord.s2 * src_stride_z);
-    in_row0.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord.s3 * src_stride_z);
-    in_row0.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord.s4 * src_stride_z);
-    in_row0.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord.s5 * src_stride_z);
-    in_row0.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord.s6 * src_stride_z);
-    in_row0.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord.s7 * src_stride_z);
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    in_row0;
+    in_row0.s0 = *(__global DATA_TYPE *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord.s0 * src_stride_z);
+    in_row0.s1 = *(__global DATA_TYPE *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord.s1 * src_stride_z);
+    in_row0.s2 = *(__global DATA_TYPE *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord.s2 * src_stride_z);
+    in_row0.s3 = *(__global DATA_TYPE *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord.s3 * src_stride_z);
+    in_row0.s4 = *(__global DATA_TYPE *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord.s4 * src_stride_z);
+    in_row0.s5 = *(__global DATA_TYPE *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord.s5 * src_stride_z);
+    in_row0.s6 = *(__global DATA_TYPE *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord.s6 * src_stride_z);
+    in_row0.s7 = *(__global DATA_TYPE *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord.s7 * src_stride_z);
 
     // Calculate common factors for intermediate tensor
-    float8 comm_fact0 = 0.0f;
-    float8 tmp0       = in_row0;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    comm_fact0 = 0.0f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    tmp0 = in_row0;
 
-    float8 out0 = (float8)0.0f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out0 = (VEC_DATA_TYPE(DATA_TYPE, 8))0.0f;
 
     OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0);
 #else                                            // defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL)
-    float8 in_row0, in_row1, in_row2, in_row3, in_row4, in_row5, in_row6, in_row7;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    in_row0, in_row1, in_row2, in_row3, in_row4, in_row5, in_row6, in_row7;
 
     // Clamp coordinates. This clamp is valid for all rows
     int8 y_coord = (int8)(y * OUTPUT_TILE_W) + (int8)(0, 1, 2, 3, 4, 5, 6, 7) - (int8)PAD_LEFT;
@@ -1066,14 +1140,14 @@
     z_coord      = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);                             // Clamp z coordinate
 
     // Load the input tile
-    in_row0.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row0.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s0 = *(__global DATA_TYPE *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s1 = *(__global DATA_TYPE *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s2 = *(__global DATA_TYPE *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s3 = *(__global DATA_TYPE *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s4 = *(__global DATA_TYPE *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s5 = *(__global DATA_TYPE *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s6 = *(__global DATA_TYPE *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row0.s7 = *(__global DATA_TYPE *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Row1
     z_coord = (z * (int)OUTPUT_TILE_H) - (int)PAD_TOP + 1;
@@ -1081,14 +1155,14 @@
     valid_y = select(valid_y, (int8)SRC_DIM_1, (int8)z_coord >= (int)SRC_DIM_2);
     z_coord = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    in_row1.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row1.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row1.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row1.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row1.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row1.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row1.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row1.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row1.s0 = *(__global DATA_TYPE *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row1.s1 = *(__global DATA_TYPE *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row1.s2 = *(__global DATA_TYPE *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row1.s3 = *(__global DATA_TYPE *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row1.s4 = *(__global DATA_TYPE *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row1.s5 = *(__global DATA_TYPE *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row1.s6 = *(__global DATA_TYPE *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row1.s7 = *(__global DATA_TYPE *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Row2
     z_coord = (z * (int)OUTPUT_TILE_H) - (int)PAD_TOP + 2;
@@ -1096,14 +1170,14 @@
     valid_y = select(valid_y, (int8)SRC_DIM_1, (int8)z_coord >= (int)SRC_DIM_2);
     z_coord = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    in_row2.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row2.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row2.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row2.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row2.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row2.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row2.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row2.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row2.s0 = *(__global DATA_TYPE *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row2.s1 = *(__global DATA_TYPE *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row2.s2 = *(__global DATA_TYPE *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row2.s3 = *(__global DATA_TYPE *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row2.s4 = *(__global DATA_TYPE *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row2.s5 = *(__global DATA_TYPE *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row2.s6 = *(__global DATA_TYPE *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row2.s7 = *(__global DATA_TYPE *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Row3
     z_coord = (z * (int)OUTPUT_TILE_H) - (int)PAD_TOP + 3;
@@ -1111,14 +1185,14 @@
     valid_y = select(valid_y, (int8)SRC_DIM_1, (int8)z_coord >= (int)SRC_DIM_2);
     z_coord = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    in_row3.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row3.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row3.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row3.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row3.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row3.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row3.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row3.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row3.s0 = *(__global DATA_TYPE *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row3.s1 = *(__global DATA_TYPE *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row3.s2 = *(__global DATA_TYPE *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row3.s3 = *(__global DATA_TYPE *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row3.s4 = *(__global DATA_TYPE *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row3.s5 = *(__global DATA_TYPE *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row3.s6 = *(__global DATA_TYPE *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row3.s7 = *(__global DATA_TYPE *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Row4
     z_coord = (z * (int)OUTPUT_TILE_H) - (int)PAD_TOP + 4;
@@ -1126,14 +1200,14 @@
     valid_y = select(valid_y, (int8)SRC_DIM_1, (int8)z_coord >= (int)SRC_DIM_2);
     z_coord = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    in_row4.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row4.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row4.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row4.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row4.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row4.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row4.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row4.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row4.s0 = *(__global DATA_TYPE *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row4.s1 = *(__global DATA_TYPE *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row4.s2 = *(__global DATA_TYPE *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row4.s3 = *(__global DATA_TYPE *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row4.s4 = *(__global DATA_TYPE *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row4.s5 = *(__global DATA_TYPE *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row4.s6 = *(__global DATA_TYPE *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row4.s7 = *(__global DATA_TYPE *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Row5
     z_coord = (z * (int)OUTPUT_TILE_H) - (int)PAD_TOP + 5;
@@ -1141,14 +1215,14 @@
     valid_y = select(valid_y, (int8)SRC_DIM_1, (int8)z_coord >= (int)SRC_DIM_2);
     z_coord = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    in_row5.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row5.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row5.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row5.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row5.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row5.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row5.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row5.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row5.s0 = *(__global DATA_TYPE *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row5.s1 = *(__global DATA_TYPE *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row5.s2 = *(__global DATA_TYPE *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row5.s3 = *(__global DATA_TYPE *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row5.s4 = *(__global DATA_TYPE *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row5.s5 = *(__global DATA_TYPE *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row5.s6 = *(__global DATA_TYPE *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row5.s7 = *(__global DATA_TYPE *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Row6
     z_coord = (z * (int)OUTPUT_TILE_H) - (int)PAD_TOP + 6;
@@ -1156,14 +1230,14 @@
     valid_y = select(valid_y, (int8)SRC_DIM_1, (int8)z_coord >= (int)SRC_DIM_2);
     z_coord = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    in_row6.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row6.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row6.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row6.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row6.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row6.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row6.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row6.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row6.s0 = *(__global DATA_TYPE *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row6.s1 = *(__global DATA_TYPE *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row6.s2 = *(__global DATA_TYPE *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row6.s3 = *(__global DATA_TYPE *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row6.s4 = *(__global DATA_TYPE *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row6.s5 = *(__global DATA_TYPE *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row6.s6 = *(__global DATA_TYPE *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row6.s7 = *(__global DATA_TYPE *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
 
     // Row7
     z_coord = (z * (int)OUTPUT_TILE_H) - (int)PAD_TOP + 7;
@@ -1171,39 +1245,43 @@
     valid_y = select(valid_y, (int8)SRC_DIM_1, (int8)z_coord >= (int)SRC_DIM_2);
     z_coord = clamp(z_coord, 0, (int)SRC_DIM_2 - 1);
 
-    in_row7.s0 = *(__global float *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row7.s1 = *(__global float *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row7.s2 = *(__global float *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row7.s3 = *(__global float *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row7.s4 = *(__global float *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row7.s5 = *(__global float *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row7.s6 = *(__global float *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
-    in_row7.s7 = *(__global float *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row7.s0 = *(__global DATA_TYPE *)(src_addr + valid_y.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row7.s1 = *(__global DATA_TYPE *)(src_addr + valid_y.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row7.s2 = *(__global DATA_TYPE *)(src_addr + valid_y.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row7.s3 = *(__global DATA_TYPE *)(src_addr + valid_y.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row7.s4 = *(__global DATA_TYPE *)(src_addr + valid_y.s4 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row7.s5 = *(__global DATA_TYPE *)(src_addr + valid_y.s5 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row7.s6 = *(__global DATA_TYPE *)(src_addr + valid_y.s6 * (int)src_stride_y + z_coord * src_stride_z);
+    in_row7.s7 = *(__global DATA_TYPE *)(src_addr + valid_y.s7 * (int)src_stride_y + z_coord * src_stride_z);
 
-    float8 comm_fact0 = in_row2 + in_row6 - 4.25f * in_row4;
-    float8 comm_fact1 = in_row1 + in_row5 - 4.25f * in_row3;
-    float8 comm_fact2 = 0.25f * in_row2 - 1.25f * in_row4 + in_row6;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    comm_fact0 = in_row2 + in_row6 - (DATA_TYPE)4.25f * in_row4;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    comm_fact1 = in_row1 + in_row5 - (DATA_TYPE)4.25f * in_row3;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    comm_fact2 = (DATA_TYPE)0.25f * in_row2 - (DATA_TYPE)1.25f * in_row4 + in_row6;
 
     // Calculate intermediate tensor and reuse common factor vectors
-    const float8 tmp0 = in_row0 - in_row6 + 5.25f * in_row4 - 5.25f * in_row2;
-    const float8 tmp1 = comm_fact0 + comm_fact1;
-    const float8 tmp2 = comm_fact0 - comm_fact1;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp0 = in_row0 - in_row6 + (DATA_TYPE)5.25f * in_row4 - (DATA_TYPE)5.25f * in_row2;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp1 = comm_fact0 + comm_fact1;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp2 = comm_fact0 - comm_fact1;
 
-    comm_fact0 = 2.5f * in_row3;
-    comm_fact1 = 0.5f * in_row1 - comm_fact0 + 2.f * in_row5;
+    comm_fact0 = (DATA_TYPE)2.5f * in_row3;
+    comm_fact1 = (DATA_TYPE)0.5f * in_row1 - comm_fact0 + (DATA_TYPE)2.f * in_row5;
 
-    const float8 tmp3 = comm_fact1 + comm_fact2;
-    const float8 tmp4 = comm_fact2 - comm_fact1;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp3 = comm_fact1 + comm_fact2;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp4 = comm_fact2 - comm_fact1;
 
-    comm_fact1 = 2.f * in_row1 - comm_fact0 + 0.5f * in_row5;
-    comm_fact2 = 4.f * in_row2 - 5.f * in_row4 + in_row6;
+    comm_fact1 = (DATA_TYPE)2.f * in_row1 - comm_fact0 + (DATA_TYPE)0.5f * in_row5;
+    comm_fact2 = (DATA_TYPE)4.f * in_row2 - (DATA_TYPE)5.f * in_row4 + in_row6;
 
-    const float8 tmp5 = comm_fact1 + comm_fact2;
-    const float8 tmp6 = comm_fact2 - comm_fact1;
-    const float8 tmp7 = in_row7 - in_row1 + 5.25f * in_row3 - 5.25f * in_row5;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp5 = comm_fact1 + comm_fact2;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp6 = comm_fact2 - comm_fact1;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp7 = in_row7 - in_row1 + (DATA_TYPE)5.25f * in_row3 - (DATA_TYPE)5.25f * in_row5;
 
     // Calculate output rows (reuse comm_fact0 vector)
-    float8 out0, out1, out2, out3, out4, out5, out6, out7;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out0, out1, out2, out3, out4, out5, out6, out7;
     OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0);
     OUTPUT_ROW_4x4_5x5(out1, tmp1, comm_fact0);
     OUTPUT_ROW_4x4_5x5(out2, tmp2, comm_fact0);
@@ -1215,74 +1293,74 @@
 #endif                                           // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
     // Store values across the channels
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(float) + (y + z * (int)NUM_TILES_X) * dst_stride_y;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) + (y + z * (int)NUM_TILES_X) * dst_stride_y;
 
-    *((__global float *)(dst_addr + 0 * dst_stride_z)) = out0.s0;
-    *((__global float *)(dst_addr + 1 * dst_stride_z)) = out0.s1;
-    *((__global float *)(dst_addr + 2 * dst_stride_z)) = out0.s2;
-    *((__global float *)(dst_addr + 3 * dst_stride_z)) = out0.s3;
-    *((__global float *)(dst_addr + 4 * dst_stride_z)) = out0.s4;
-    *((__global float *)(dst_addr + 5 * dst_stride_z)) = out0.s5;
-    *((__global float *)(dst_addr + 6 * dst_stride_z)) = out0.s6;
-    *((__global float *)(dst_addr + 7 * dst_stride_z)) = out0.s7;
+    *((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z)) = out0.s0;
+    *((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z)) = out0.s1;
+    *((__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z)) = out0.s2;
+    *((__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z)) = out0.s3;
+    *((__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z)) = out0.s4;
+    *((__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z)) = out0.s5;
+    *((__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z)) = out0.s6;
+    *((__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z)) = out0.s7;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    *((__global float *)(dst_addr + 8 * dst_stride_z))  = out1.s0;
-    *((__global float *)(dst_addr + 9 * dst_stride_z))  = out1.s1;
-    *((__global float *)(dst_addr + 10 * dst_stride_z)) = out1.s2;
-    *((__global float *)(dst_addr + 11 * dst_stride_z)) = out1.s3;
-    *((__global float *)(dst_addr + 12 * dst_stride_z)) = out1.s4;
-    *((__global float *)(dst_addr + 13 * dst_stride_z)) = out1.s5;
-    *((__global float *)(dst_addr + 14 * dst_stride_z)) = out1.s6;
-    *((__global float *)(dst_addr + 15 * dst_stride_z)) = out1.s7;
-    *((__global float *)(dst_addr + 16 * dst_stride_z)) = out2.s0;
-    *((__global float *)(dst_addr + 17 * dst_stride_z)) = out2.s1;
-    *((__global float *)(dst_addr + 18 * dst_stride_z)) = out2.s2;
-    *((__global float *)(dst_addr + 19 * dst_stride_z)) = out2.s3;
-    *((__global float *)(dst_addr + 20 * dst_stride_z)) = out2.s4;
-    *((__global float *)(dst_addr + 21 * dst_stride_z)) = out2.s5;
-    *((__global float *)(dst_addr + 22 * dst_stride_z)) = out2.s6;
-    *((__global float *)(dst_addr + 23 * dst_stride_z)) = out2.s7;
-    *((__global float *)(dst_addr + 24 * dst_stride_z)) = out3.s0;
-    *((__global float *)(dst_addr + 25 * dst_stride_z)) = out3.s1;
-    *((__global float *)(dst_addr + 26 * dst_stride_z)) = out3.s2;
-    *((__global float *)(dst_addr + 27 * dst_stride_z)) = out3.s3;
-    *((__global float *)(dst_addr + 28 * dst_stride_z)) = out3.s4;
-    *((__global float *)(dst_addr + 29 * dst_stride_z)) = out3.s5;
-    *((__global float *)(dst_addr + 30 * dst_stride_z)) = out3.s6;
-    *((__global float *)(dst_addr + 31 * dst_stride_z)) = out3.s7;
-    *((__global float *)(dst_addr + 32 * dst_stride_z)) = out4.s0;
-    *((__global float *)(dst_addr + 33 * dst_stride_z)) = out4.s1;
-    *((__global float *)(dst_addr + 34 * dst_stride_z)) = out4.s2;
-    *((__global float *)(dst_addr + 35 * dst_stride_z)) = out4.s3;
-    *((__global float *)(dst_addr + 36 * dst_stride_z)) = out4.s4;
-    *((__global float *)(dst_addr + 37 * dst_stride_z)) = out4.s5;
-    *((__global float *)(dst_addr + 38 * dst_stride_z)) = out4.s6;
-    *((__global float *)(dst_addr + 39 * dst_stride_z)) = out4.s7;
-    *((__global float *)(dst_addr + 40 * dst_stride_z)) = out5.s0;
-    *((__global float *)(dst_addr + 41 * dst_stride_z)) = out5.s1;
-    *((__global float *)(dst_addr + 42 * dst_stride_z)) = out5.s2;
-    *((__global float *)(dst_addr + 43 * dst_stride_z)) = out5.s3;
-    *((__global float *)(dst_addr + 44 * dst_stride_z)) = out5.s4;
-    *((__global float *)(dst_addr + 45 * dst_stride_z)) = out5.s5;
-    *((__global float *)(dst_addr + 46 * dst_stride_z)) = out5.s6;
-    *((__global float *)(dst_addr + 47 * dst_stride_z)) = out5.s7;
-    *((__global float *)(dst_addr + 48 * dst_stride_z)) = out6.s0;
-    *((__global float *)(dst_addr + 49 * dst_stride_z)) = out6.s1;
-    *((__global float *)(dst_addr + 50 * dst_stride_z)) = out6.s2;
-    *((__global float *)(dst_addr + 51 * dst_stride_z)) = out6.s3;
-    *((__global float *)(dst_addr + 52 * dst_stride_z)) = out6.s4;
-    *((__global float *)(dst_addr + 53 * dst_stride_z)) = out6.s5;
-    *((__global float *)(dst_addr + 54 * dst_stride_z)) = out6.s6;
-    *((__global float *)(dst_addr + 55 * dst_stride_z)) = out6.s7;
-    *((__global float *)(dst_addr + 56 * dst_stride_z)) = out7.s0;
-    *((__global float *)(dst_addr + 57 * dst_stride_z)) = out7.s1;
-    *((__global float *)(dst_addr + 58 * dst_stride_z)) = out7.s2;
-    *((__global float *)(dst_addr + 59 * dst_stride_z)) = out7.s3;
-    *((__global float *)(dst_addr + 60 * dst_stride_z)) = out7.s4;
-    *((__global float *)(dst_addr + 61 * dst_stride_z)) = out7.s5;
-    *((__global float *)(dst_addr + 62 * dst_stride_z)) = out7.s6;
-    *((__global float *)(dst_addr + 63 * dst_stride_z)) = out7.s7;
+    *((__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z))  = out1.s0;
+    *((__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z))  = out1.s1;
+    *((__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z)) = out1.s2;
+    *((__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z)) = out1.s3;
+    *((__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z)) = out1.s4;
+    *((__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z)) = out1.s5;
+    *((__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z)) = out1.s6;
+    *((__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z)) = out1.s7;
+    *((__global DATA_TYPE *)(dst_addr + 16 * dst_stride_z)) = out2.s0;
+    *((__global DATA_TYPE *)(dst_addr + 17 * dst_stride_z)) = out2.s1;
+    *((__global DATA_TYPE *)(dst_addr + 18 * dst_stride_z)) = out2.s2;
+    *((__global DATA_TYPE *)(dst_addr + 19 * dst_stride_z)) = out2.s3;
+    *((__global DATA_TYPE *)(dst_addr + 20 * dst_stride_z)) = out2.s4;
+    *((__global DATA_TYPE *)(dst_addr + 21 * dst_stride_z)) = out2.s5;
+    *((__global DATA_TYPE *)(dst_addr + 22 * dst_stride_z)) = out2.s6;
+    *((__global DATA_TYPE *)(dst_addr + 23 * dst_stride_z)) = out2.s7;
+    *((__global DATA_TYPE *)(dst_addr + 24 * dst_stride_z)) = out3.s0;
+    *((__global DATA_TYPE *)(dst_addr + 25 * dst_stride_z)) = out3.s1;
+    *((__global DATA_TYPE *)(dst_addr + 26 * dst_stride_z)) = out3.s2;
+    *((__global DATA_TYPE *)(dst_addr + 27 * dst_stride_z)) = out3.s3;
+    *((__global DATA_TYPE *)(dst_addr + 28 * dst_stride_z)) = out3.s4;
+    *((__global DATA_TYPE *)(dst_addr + 29 * dst_stride_z)) = out3.s5;
+    *((__global DATA_TYPE *)(dst_addr + 30 * dst_stride_z)) = out3.s6;
+    *((__global DATA_TYPE *)(dst_addr + 31 * dst_stride_z)) = out3.s7;
+    *((__global DATA_TYPE *)(dst_addr + 32 * dst_stride_z)) = out4.s0;
+    *((__global DATA_TYPE *)(dst_addr + 33 * dst_stride_z)) = out4.s1;
+    *((__global DATA_TYPE *)(dst_addr + 34 * dst_stride_z)) = out4.s2;
+    *((__global DATA_TYPE *)(dst_addr + 35 * dst_stride_z)) = out4.s3;
+    *((__global DATA_TYPE *)(dst_addr + 36 * dst_stride_z)) = out4.s4;
+    *((__global DATA_TYPE *)(dst_addr + 37 * dst_stride_z)) = out4.s5;
+    *((__global DATA_TYPE *)(dst_addr + 38 * dst_stride_z)) = out4.s6;
+    *((__global DATA_TYPE *)(dst_addr + 39 * dst_stride_z)) = out4.s7;
+    *((__global DATA_TYPE *)(dst_addr + 40 * dst_stride_z)) = out5.s0;
+    *((__global DATA_TYPE *)(dst_addr + 41 * dst_stride_z)) = out5.s1;
+    *((__global DATA_TYPE *)(dst_addr + 42 * dst_stride_z)) = out5.s2;
+    *((__global DATA_TYPE *)(dst_addr + 43 * dst_stride_z)) = out5.s3;
+    *((__global DATA_TYPE *)(dst_addr + 44 * dst_stride_z)) = out5.s4;
+    *((__global DATA_TYPE *)(dst_addr + 45 * dst_stride_z)) = out5.s5;
+    *((__global DATA_TYPE *)(dst_addr + 46 * dst_stride_z)) = out5.s6;
+    *((__global DATA_TYPE *)(dst_addr + 47 * dst_stride_z)) = out5.s7;
+    *((__global DATA_TYPE *)(dst_addr + 48 * dst_stride_z)) = out6.s0;
+    *((__global DATA_TYPE *)(dst_addr + 49 * dst_stride_z)) = out6.s1;
+    *((__global DATA_TYPE *)(dst_addr + 50 * dst_stride_z)) = out6.s2;
+    *((__global DATA_TYPE *)(dst_addr + 51 * dst_stride_z)) = out6.s3;
+    *((__global DATA_TYPE *)(dst_addr + 52 * dst_stride_z)) = out6.s4;
+    *((__global DATA_TYPE *)(dst_addr + 53 * dst_stride_z)) = out6.s5;
+    *((__global DATA_TYPE *)(dst_addr + 54 * dst_stride_z)) = out6.s6;
+    *((__global DATA_TYPE *)(dst_addr + 55 * dst_stride_z)) = out6.s7;
+    *((__global DATA_TYPE *)(dst_addr + 56 * dst_stride_z)) = out7.s0;
+    *((__global DATA_TYPE *)(dst_addr + 57 * dst_stride_z)) = out7.s1;
+    *((__global DATA_TYPE *)(dst_addr + 58 * dst_stride_z)) = out7.s2;
+    *((__global DATA_TYPE *)(dst_addr + 59 * dst_stride_z)) = out7.s3;
+    *((__global DATA_TYPE *)(dst_addr + 60 * dst_stride_z)) = out7.s4;
+    *((__global DATA_TYPE *)(dst_addr + 61 * dst_stride_z)) = out7.s5;
+    *((__global DATA_TYPE *)(dst_addr + 62 * dst_stride_z)) = out7.s6;
+    *((__global DATA_TYPE *)(dst_addr + 63 * dst_stride_z)) = out7.s7;
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 }
 #endif // defined(SRC_DIM_1) && defined(SRC_DIM_2)
@@ -1295,8 +1373,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2
  * @note If this kernel is used to perform Winograd input transform 5x1, -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd input transform 1x5, -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1322,68 +1401,74 @@
     int z = get_global_id(2);
 
     // Compute input address
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(float) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z;
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * OUTPUT_TILE_W * sizeof(DATA_TYPE) + y * OUTPUT_TILE_H * src_stride_y + z * src_stride_z;
 
-    src_addr = src_addr - ((int)PAD_LEFT * sizeof(float)) - ((int)PAD_TOP * src_stride_y);
+    src_addr = src_addr - ((int)PAD_LEFT * sizeof(DATA_TYPE)) - ((int)PAD_TOP * src_stride_y);
 
     // Load input tile
 #if defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL)
-    const float8 in_row0 = vload8(0, (__global float *)(src_addr));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row0 = vload8(0, (__global DATA_TYPE *)(src_addr));
 #elif defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL) // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL)
-    const float8 in_row0 = (float8)(*((__global float *)(src_addr + 0 * src_stride_y)),
-                                    *((__global float *)(src_addr + 1 * src_stride_y)),
-                                    *((__global float *)(src_addr + 2 * src_stride_y)),
-                                    *((__global float *)(src_addr + 3 * src_stride_y)),
-                                    *((__global float *)(src_addr + 4 * src_stride_y)),
-                                    *((__global float *)(src_addr + 5 * src_stride_y)),
-                                    *((__global float *)(src_addr + 6 * src_stride_y)),
-                                    *((__global float *)(src_addr + 7 * src_stride_y)));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row0 = (VEC_DATA_TYPE(DATA_TYPE, 8))(*((__global DATA_TYPE *)(src_addr + 0 * src_stride_y)),
+                                                                              *((__global DATA_TYPE *)(src_addr + 1 * src_stride_y)),
+                                                                              *((__global DATA_TYPE *)(src_addr + 2 * src_stride_y)),
+                                                                              *((__global DATA_TYPE *)(src_addr + 3 * src_stride_y)),
+                                                                              *((__global DATA_TYPE *)(src_addr + 4 * src_stride_y)),
+                                                                              *((__global DATA_TYPE *)(src_addr + 5 * src_stride_y)),
+                                                                              *((__global DATA_TYPE *)(src_addr + 6 * src_stride_y)),
+                                                                              *((__global DATA_TYPE *)(src_addr + 7 * src_stride_y)));
 #else                                            // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    const float8 in_row0 = vload8(0, (__global float *)(src_addr + 0 * src_stride_y));
-    const float8 in_row1 = vload8(0, (__global float *)(src_addr + 1 * src_stride_y));
-    const float8 in_row2 = vload8(0, (__global float *)(src_addr + 2 * src_stride_y));
-    const float8 in_row3 = vload8(0, (__global float *)(src_addr + 3 * src_stride_y));
-    const float8 in_row4 = vload8(0, (__global float *)(src_addr + 4 * src_stride_y));
-    const float8 in_row5 = vload8(0, (__global float *)(src_addr + 5 * src_stride_y));
-    const float8 in_row6 = vload8(0, (__global float *)(src_addr + 6 * src_stride_y));
-    const float8 in_row7 = vload8(0, (__global float *)(src_addr + 7 * src_stride_y));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row0 = vload8(0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row1 = vload8(0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row2 = vload8(0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row3 = vload8(0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row4 = vload8(0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row5 = vload8(0, (__global DATA_TYPE *)(src_addr + 5 * src_stride_y));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row6 = vload8(0, (__global DATA_TYPE *)(src_addr + 6 * src_stride_y));
+    const VEC_DATA_TYPE(DATA_TYPE, 8) in_row7 = vload8(0, (__global DATA_TYPE *)(src_addr + 7 * src_stride_y));
 #endif                                           // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
     // Calculate common factors for intermediate tensor
-    float8 tmp0       = in_row0;
-    float8 comm_fact0 = 0.0f;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    tmp0 = in_row0;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    comm_fact0 = 0.0f;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    comm_fact0 += in_row2 + in_row6 - 4.25f * in_row4;
-    tmp0 += -in_row6 + 5.25f * in_row4 - 5.25f * in_row2;
+    comm_fact0 += in_row2 + in_row6 - (DATA_TYPE)4.25 * in_row4;
+    tmp0 += -in_row6 + (DATA_TYPE)5.25 * in_row4 - (DATA_TYPE)5.25 * in_row2;
 
-    float8 comm_fact1 = in_row1 + in_row5 - 4.25f * in_row3;
-    float8 comm_fact2 = 0.25f * in_row2 - 1.25f * in_row4 + in_row6;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    comm_fact1 = in_row1 + in_row5 - (DATA_TYPE)4.25 * in_row3;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    comm_fact2 = (DATA_TYPE)0.25 * in_row2 - (DATA_TYPE)1.25 * in_row4 + in_row6;
 
-    const float8 tmp1 = comm_fact0 + comm_fact1;
-    const float8 tmp2 = comm_fact0 - comm_fact1;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp1 = comm_fact0 + comm_fact1;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp2 = comm_fact0 - comm_fact1;
 
-    comm_fact0 = 2.5f * in_row3;
-    comm_fact1 = 0.5f * in_row1 - comm_fact0 + 2.f * in_row5;
+    comm_fact0 = (DATA_TYPE)2.5 * in_row3;
+    comm_fact1 = (DATA_TYPE)0.5 * in_row1 - comm_fact0 + (DATA_TYPE)2.0 * in_row5;
 
-    const float8 tmp3 = comm_fact1 + comm_fact2;
-    const float8 tmp4 = comm_fact2 - comm_fact1;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp3 = comm_fact1 + comm_fact2;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp4 = comm_fact2 - comm_fact1;
 
-    comm_fact1 = 2.f * in_row1 - comm_fact0 + 0.5f * in_row5;
-    comm_fact2 = 4.f * in_row2 - 5.f * in_row4 + in_row6;
+    comm_fact1 = (DATA_TYPE)2.0 * in_row1 - comm_fact0 + (DATA_TYPE)0.5 * in_row5;
+    comm_fact2 = (DATA_TYPE)4.0 * in_row2 - (DATA_TYPE)5.0 * in_row4 + in_row6;
 
-    const float8 tmp5 = comm_fact1 + comm_fact2;
-    const float8 tmp6 = comm_fact2 - comm_fact1;
-    const float8 tmp7 = in_row7 - in_row1 + 5.25f * in_row3 - 5.25f * in_row5;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp5 = comm_fact1 + comm_fact2;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp6 = comm_fact2 - comm_fact1;
+    const VEC_DATA_TYPE(DATA_TYPE, 8) tmp7 = in_row7 - in_row1 + (DATA_TYPE)5.25 * in_row3 - (DATA_TYPE)5.25 * in_row5;
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
     // Calculate output rows (reuse comm_fact0 vector)
-    float8 out0;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out0;
 
     OUTPUT_ROW_4x4_5x5(out0, tmp0, comm_fact0);
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    float8 out1, out2, out3, out4, out5, out6, out7;
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    out1, out2, out3, out4, out5, out6, out7;
 
     OUTPUT_ROW_4x4_5x5(out1, tmp1, comm_fact0);
     OUTPUT_ROW_4x4_5x5(out2, tmp2, comm_fact0);
@@ -1395,74 +1480,74 @@
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 
     // Store values across the channels
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(float) + (x + y * (int)NUM_TILES_X) * dst_stride_y;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * sizeof(DATA_TYPE) + (x + y * (int)NUM_TILES_X) * dst_stride_y;
 
-    *((__global float *)(dst_addr + 0 * dst_stride_z)) = out0.s0;
-    *((__global float *)(dst_addr + 1 * dst_stride_z)) = out0.s1;
-    *((__global float *)(dst_addr + 2 * dst_stride_z)) = out0.s2;
-    *((__global float *)(dst_addr + 3 * dst_stride_z)) = out0.s3;
-    *((__global float *)(dst_addr + 4 * dst_stride_z)) = out0.s4;
-    *((__global float *)(dst_addr + 5 * dst_stride_z)) = out0.s5;
-    *((__global float *)(dst_addr + 6 * dst_stride_z)) = out0.s6;
-    *((__global float *)(dst_addr + 7 * dst_stride_z)) = out0.s7;
+    *((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_z)) = out0.s0;
+    *((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_z)) = out0.s1;
+    *((__global DATA_TYPE *)(dst_addr + 2 * dst_stride_z)) = out0.s2;
+    *((__global DATA_TYPE *)(dst_addr + 3 * dst_stride_z)) = out0.s3;
+    *((__global DATA_TYPE *)(dst_addr + 4 * dst_stride_z)) = out0.s4;
+    *((__global DATA_TYPE *)(dst_addr + 5 * dst_stride_z)) = out0.s5;
+    *((__global DATA_TYPE *)(dst_addr + 6 * dst_stride_z)) = out0.s6;
+    *((__global DATA_TYPE *)(dst_addr + 7 * dst_stride_z)) = out0.s7;
 
 #if !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
-    *((__global float *)(dst_addr + 8 * dst_stride_z))  = out1.s0;
-    *((__global float *)(dst_addr + 9 * dst_stride_z))  = out1.s1;
-    *((__global float *)(dst_addr + 10 * dst_stride_z)) = out1.s2;
-    *((__global float *)(dst_addr + 11 * dst_stride_z)) = out1.s3;
-    *((__global float *)(dst_addr + 12 * dst_stride_z)) = out1.s4;
-    *((__global float *)(dst_addr + 13 * dst_stride_z)) = out1.s5;
-    *((__global float *)(dst_addr + 14 * dst_stride_z)) = out1.s6;
-    *((__global float *)(dst_addr + 15 * dst_stride_z)) = out1.s7;
-    *((__global float *)(dst_addr + 16 * dst_stride_z)) = out2.s0;
-    *((__global float *)(dst_addr + 17 * dst_stride_z)) = out2.s1;
-    *((__global float *)(dst_addr + 18 * dst_stride_z)) = out2.s2;
-    *((__global float *)(dst_addr + 19 * dst_stride_z)) = out2.s3;
-    *((__global float *)(dst_addr + 20 * dst_stride_z)) = out2.s4;
-    *((__global float *)(dst_addr + 21 * dst_stride_z)) = out2.s5;
-    *((__global float *)(dst_addr + 22 * dst_stride_z)) = out2.s6;
-    *((__global float *)(dst_addr + 23 * dst_stride_z)) = out2.s7;
-    *((__global float *)(dst_addr + 24 * dst_stride_z)) = out3.s0;
-    *((__global float *)(dst_addr + 25 * dst_stride_z)) = out3.s1;
-    *((__global float *)(dst_addr + 26 * dst_stride_z)) = out3.s2;
-    *((__global float *)(dst_addr + 27 * dst_stride_z)) = out3.s3;
-    *((__global float *)(dst_addr + 28 * dst_stride_z)) = out3.s4;
-    *((__global float *)(dst_addr + 29 * dst_stride_z)) = out3.s5;
-    *((__global float *)(dst_addr + 30 * dst_stride_z)) = out3.s6;
-    *((__global float *)(dst_addr + 31 * dst_stride_z)) = out3.s7;
-    *((__global float *)(dst_addr + 32 * dst_stride_z)) = out4.s0;
-    *((__global float *)(dst_addr + 33 * dst_stride_z)) = out4.s1;
-    *((__global float *)(dst_addr + 34 * dst_stride_z)) = out4.s2;
-    *((__global float *)(dst_addr + 35 * dst_stride_z)) = out4.s3;
-    *((__global float *)(dst_addr + 36 * dst_stride_z)) = out4.s4;
-    *((__global float *)(dst_addr + 37 * dst_stride_z)) = out4.s5;
-    *((__global float *)(dst_addr + 38 * dst_stride_z)) = out4.s6;
-    *((__global float *)(dst_addr + 39 * dst_stride_z)) = out4.s7;
-    *((__global float *)(dst_addr + 40 * dst_stride_z)) = out5.s0;
-    *((__global float *)(dst_addr + 41 * dst_stride_z)) = out5.s1;
-    *((__global float *)(dst_addr + 42 * dst_stride_z)) = out5.s2;
-    *((__global float *)(dst_addr + 43 * dst_stride_z)) = out5.s3;
-    *((__global float *)(dst_addr + 44 * dst_stride_z)) = out5.s4;
-    *((__global float *)(dst_addr + 45 * dst_stride_z)) = out5.s5;
-    *((__global float *)(dst_addr + 46 * dst_stride_z)) = out5.s6;
-    *((__global float *)(dst_addr + 47 * dst_stride_z)) = out5.s7;
-    *((__global float *)(dst_addr + 48 * dst_stride_z)) = out6.s0;
-    *((__global float *)(dst_addr + 49 * dst_stride_z)) = out6.s1;
-    *((__global float *)(dst_addr + 50 * dst_stride_z)) = out6.s2;
-    *((__global float *)(dst_addr + 51 * dst_stride_z)) = out6.s3;
-    *((__global float *)(dst_addr + 52 * dst_stride_z)) = out6.s4;
-    *((__global float *)(dst_addr + 53 * dst_stride_z)) = out6.s5;
-    *((__global float *)(dst_addr + 54 * dst_stride_z)) = out6.s6;
-    *((__global float *)(dst_addr + 55 * dst_stride_z)) = out6.s7;
-    *((__global float *)(dst_addr + 56 * dst_stride_z)) = out7.s0;
-    *((__global float *)(dst_addr + 57 * dst_stride_z)) = out7.s1;
-    *((__global float *)(dst_addr + 58 * dst_stride_z)) = out7.s2;
-    *((__global float *)(dst_addr + 59 * dst_stride_z)) = out7.s3;
-    *((__global float *)(dst_addr + 60 * dst_stride_z)) = out7.s4;
-    *((__global float *)(dst_addr + 61 * dst_stride_z)) = out7.s5;
-    *((__global float *)(dst_addr + 62 * dst_stride_z)) = out7.s6;
-    *((__global float *)(dst_addr + 63 * dst_stride_z)) = out7.s7;
+    *((__global DATA_TYPE *)(dst_addr + 8 * dst_stride_z))  = out1.s0;
+    *((__global DATA_TYPE *)(dst_addr + 9 * dst_stride_z))  = out1.s1;
+    *((__global DATA_TYPE *)(dst_addr + 10 * dst_stride_z)) = out1.s2;
+    *((__global DATA_TYPE *)(dst_addr + 11 * dst_stride_z)) = out1.s3;
+    *((__global DATA_TYPE *)(dst_addr + 12 * dst_stride_z)) = out1.s4;
+    *((__global DATA_TYPE *)(dst_addr + 13 * dst_stride_z)) = out1.s5;
+    *((__global DATA_TYPE *)(dst_addr + 14 * dst_stride_z)) = out1.s6;
+    *((__global DATA_TYPE *)(dst_addr + 15 * dst_stride_z)) = out1.s7;
+    *((__global DATA_TYPE *)(dst_addr + 16 * dst_stride_z)) = out2.s0;
+    *((__global DATA_TYPE *)(dst_addr + 17 * dst_stride_z)) = out2.s1;
+    *((__global DATA_TYPE *)(dst_addr + 18 * dst_stride_z)) = out2.s2;
+    *((__global DATA_TYPE *)(dst_addr + 19 * dst_stride_z)) = out2.s3;
+    *((__global DATA_TYPE *)(dst_addr + 20 * dst_stride_z)) = out2.s4;
+    *((__global DATA_TYPE *)(dst_addr + 21 * dst_stride_z)) = out2.s5;
+    *((__global DATA_TYPE *)(dst_addr + 22 * dst_stride_z)) = out2.s6;
+    *((__global DATA_TYPE *)(dst_addr + 23 * dst_stride_z)) = out2.s7;
+    *((__global DATA_TYPE *)(dst_addr + 24 * dst_stride_z)) = out3.s0;
+    *((__global DATA_TYPE *)(dst_addr + 25 * dst_stride_z)) = out3.s1;
+    *((__global DATA_TYPE *)(dst_addr + 26 * dst_stride_z)) = out3.s2;
+    *((__global DATA_TYPE *)(dst_addr + 27 * dst_stride_z)) = out3.s3;
+    *((__global DATA_TYPE *)(dst_addr + 28 * dst_stride_z)) = out3.s4;
+    *((__global DATA_TYPE *)(dst_addr + 29 * dst_stride_z)) = out3.s5;
+    *((__global DATA_TYPE *)(dst_addr + 30 * dst_stride_z)) = out3.s6;
+    *((__global DATA_TYPE *)(dst_addr + 31 * dst_stride_z)) = out3.s7;
+    *((__global DATA_TYPE *)(dst_addr + 32 * dst_stride_z)) = out4.s0;
+    *((__global DATA_TYPE *)(dst_addr + 33 * dst_stride_z)) = out4.s1;
+    *((__global DATA_TYPE *)(dst_addr + 34 * dst_stride_z)) = out4.s2;
+    *((__global DATA_TYPE *)(dst_addr + 35 * dst_stride_z)) = out4.s3;
+    *((__global DATA_TYPE *)(dst_addr + 36 * dst_stride_z)) = out4.s4;
+    *((__global DATA_TYPE *)(dst_addr + 37 * dst_stride_z)) = out4.s5;
+    *((__global DATA_TYPE *)(dst_addr + 38 * dst_stride_z)) = out4.s6;
+    *((__global DATA_TYPE *)(dst_addr + 39 * dst_stride_z)) = out4.s7;
+    *((__global DATA_TYPE *)(dst_addr + 40 * dst_stride_z)) = out5.s0;
+    *((__global DATA_TYPE *)(dst_addr + 41 * dst_stride_z)) = out5.s1;
+    *((__global DATA_TYPE *)(dst_addr + 42 * dst_stride_z)) = out5.s2;
+    *((__global DATA_TYPE *)(dst_addr + 43 * dst_stride_z)) = out5.s3;
+    *((__global DATA_TYPE *)(dst_addr + 44 * dst_stride_z)) = out5.s4;
+    *((__global DATA_TYPE *)(dst_addr + 45 * dst_stride_z)) = out5.s5;
+    *((__global DATA_TYPE *)(dst_addr + 46 * dst_stride_z)) = out5.s6;
+    *((__global DATA_TYPE *)(dst_addr + 47 * dst_stride_z)) = out5.s7;
+    *((__global DATA_TYPE *)(dst_addr + 48 * dst_stride_z)) = out6.s0;
+    *((__global DATA_TYPE *)(dst_addr + 49 * dst_stride_z)) = out6.s1;
+    *((__global DATA_TYPE *)(dst_addr + 50 * dst_stride_z)) = out6.s2;
+    *((__global DATA_TYPE *)(dst_addr + 51 * dst_stride_z)) = out6.s3;
+    *((__global DATA_TYPE *)(dst_addr + 52 * dst_stride_z)) = out6.s4;
+    *((__global DATA_TYPE *)(dst_addr + 53 * dst_stride_z)) = out6.s5;
+    *((__global DATA_TYPE *)(dst_addr + 54 * dst_stride_z)) = out6.s6;
+    *((__global DATA_TYPE *)(dst_addr + 55 * dst_stride_z)) = out6.s7;
+    *((__global DATA_TYPE *)(dst_addr + 56 * dst_stride_z)) = out7.s0;
+    *((__global DATA_TYPE *)(dst_addr + 57 * dst_stride_z)) = out7.s1;
+    *((__global DATA_TYPE *)(dst_addr + 58 * dst_stride_z)) = out7.s2;
+    *((__global DATA_TYPE *)(dst_addr + 59 * dst_stride_z)) = out7.s3;
+    *((__global DATA_TYPE *)(dst_addr + 60 * dst_stride_z)) = out7.s4;
+    *((__global DATA_TYPE *)(dst_addr + 61 * dst_stride_z)) = out7.s5;
+    *((__global DATA_TYPE *)(dst_addr + 62 * dst_stride_z)) = out7.s6;
+    *((__global DATA_TYPE *)(dst_addr + 63 * dst_stride_z)) = out7.s7;
 #endif // !defined(WINOGRAD_INPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_INPUT_TRANSFORM_VERTICAL)
 }
 
@@ -1474,8 +1559,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=2
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1521,8 +1607,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=2
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1568,8 +1655,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1615,8 +1703,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=2
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2
  * @note -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1665,8 +1754,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1714,8 +1804,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1764,8 +1855,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2
  * @note -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1811,8 +1903,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2
  * @note -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1858,8 +1951,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1905,8 +1999,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -1955,8 +2050,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
@@ -2004,8 +2100,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note -DWINOGRAD_INPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32/F16
  * @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)
diff --git a/src/core/CL/cl_kernels/winograd_output_transform.cl b/src/core/CL/cl_kernels/winograd_output_transform.cl
index a1e7b3e..9be51f2 100644
--- a/src/core/CL/cl_kernels/winograd_output_transform.cl
+++ b/src/core/CL/cl_kernels/winograd_output_transform.cl
@@ -31,8 +31,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2
  * @note If this kernel is used to perform Winograd output transform 3x1, -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd output transform 1x3, -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -64,49 +65,49 @@
     const __global uchar *src_addr = tensor3D_offset(&src, 0, 0, 0);
 
     // Load the values across the 16 or 4 channels to compose the 4x4 or 4x1 tile
-    float d00 = *((__global float *)(src_addr + 0 * src_stride_z));
-    float d01 = *((__global float *)(src_addr + 1 * src_stride_z));
-    float d02 = *((__global float *)(src_addr + 2 * src_stride_z));
-    float d03 = *((__global float *)(src_addr + 3 * src_stride_z));
+    DATA_TYPE d00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z));
+    DATA_TYPE d01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z));
+    DATA_TYPE d02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z));
+    DATA_TYPE d03 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z));
 
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
     // Compute the 2x1 or 1x2 output tile
     // out00 = d00 + d01 + d02
     // out01 = d01 - d02 - d03
 
-    float out00 = d00 + d01 + d02;
-    float out01 = d01 - d02 - d03;
+    DATA_TYPE out00 = d00 + d01 + d02;
+    DATA_TYPE out01 = d01 - d02 - d03;
 #else  // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    float d10 = *((__global float *)(src_addr + 4 * src_stride_z));
-    float d11 = *((__global float *)(src_addr + 5 * src_stride_z));
-    float d12 = *((__global float *)(src_addr + 6 * src_stride_z));
-    float d13 = *((__global float *)(src_addr + 7 * src_stride_z));
+    DATA_TYPE d10 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z));
+    DATA_TYPE d11 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z));
+    DATA_TYPE d12 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z));
+    DATA_TYPE d13 = *((__global DATA_TYPE *)(src_addr + 7 * src_stride_z));
 
-    float d20 = *((__global float *)(src_addr + 8 * src_stride_z));
-    float d21 = *((__global float *)(src_addr + 9 * src_stride_z));
-    float d22 = *((__global float *)(src_addr + 10 * src_stride_z));
-    float d23 = *((__global float *)(src_addr + 11 * src_stride_z));
+    DATA_TYPE d20 = *((__global DATA_TYPE *)(src_addr + 8 * src_stride_z));
+    DATA_TYPE d21 = *((__global DATA_TYPE *)(src_addr + 9 * src_stride_z));
+    DATA_TYPE d22 = *((__global DATA_TYPE *)(src_addr + 10 * src_stride_z));
+    DATA_TYPE d23 = *((__global DATA_TYPE *)(src_addr + 11 * src_stride_z));
 
-    float d30 = *((__global float *)(src_addr + 12 * src_stride_z));
-    float d31 = *((__global float *)(src_addr + 13 * src_stride_z));
-    float d32 = *((__global float *)(src_addr + 14 * src_stride_z));
-    float d33 = *((__global float *)(src_addr + 15 * src_stride_z));
+    DATA_TYPE d30 = *((__global DATA_TYPE *)(src_addr + 12 * src_stride_z));
+    DATA_TYPE d31 = *((__global DATA_TYPE *)(src_addr + 13 * src_stride_z));
+    DATA_TYPE d32 = *((__global DATA_TYPE *)(src_addr + 14 * src_stride_z));
+    DATA_TYPE d33 = *((__global DATA_TYPE *)(src_addr + 15 * src_stride_z));
 
     // Compute the 2x2 output tile
-    float k0 = d01 + d11 + d21;
-    float k1 = d02 + d12 + d22;
-    float k2 = d11 - d21 - d31;
-    float k3 = d12 - d22 - d32;
+    DATA_TYPE k0 = d01 + d11 + d21;
+    DATA_TYPE k1 = d02 + d12 + d22;
+    DATA_TYPE k2 = d11 - d21 - d31;
+    DATA_TYPE k3 = d12 - d22 - d32;
 
     // out00 = d00 + d10 + d20 + d01 + d11 + d21 + d02 + d12 + d22
     // out01 = d01 + d11 + d21 - (d02 + d12 + d22) - (d03 + d13 + d23)
     // out10 = d10 - d20 - d30 + (d11 - d21 - d31) + (d12 - d22 - d32)
     // out11 = d11 - d21 - d31 - (d12 - d22 - d32) - (d13 - d23 - d33)
 
-    float out00 = d10;
-    float out01 = -d13;
-    float out10 = d10;
-    float out11 = -d13;
+    DATA_TYPE out00 = d10;
+    DATA_TYPE out01 = -d13;
+    DATA_TYPE out10 = d10;
+    DATA_TYPE out11 = -d13;
 
     out00 += d00 + d20 + k0 + k1;
     out01 += k0 - k1 - (d03 + d23);
@@ -123,31 +124,31 @@
     // Add bias
     Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
 
-    float b = (float) * ((__global float *)(vector_offset(&bias, z_out)));
+    DATA_TYPE b = (DATA_TYPE) * ((__global DATA_TYPE *)(vector_offset(&bias, z_out)));
 
-    out00 += (float)b;
-    out01 += (float)b;
+    out00 += (DATA_TYPE)b;
+    out01 += (DATA_TYPE)b;
 #endif // defined(HAS_BIAS)
 
     // Get output address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z;
 
     // Store the output tile
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    *((__global float *)(dst_addr + 0 * dst_stride_y)) = out00;
-    *((__global float *)(dst_addr + 1 * dst_stride_y)) = out01;
+    *((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y)) = out00;
+    *((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y)) = out01;
 #else  // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    vstore2((float2)(out00, out01), 0, (__global float *)(dst_addr + 0 * dst_stride_y));
+    vstore2((VEC_DATA_TYPE(DATA_TYPE, 2))(out00, out01), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y));
 #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
 #if !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 #if defined(HAS_BIAS)
     // Add bias
-    out10 += (float)b;
-    out11 += (float)b;
+    out10 += (DATA_TYPE)b;
+    out11 += (DATA_TYPE)b;
 #endif // defined(HAS_BIAS)
 
-    vstore2((float2)(out10, out11), 0, (__global float *)(dst_addr + 1 * dst_stride_y));
+    vstore2((VEC_DATA_TYPE(DATA_TYPE, 2))(out10, out11), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y));
 #endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 }
 
@@ -158,8 +159,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note If this kernel is used to perform Winograd output transform 3x1, -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd output transform 1x3, -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -191,63 +193,63 @@
     const __global uchar *src_addr = tensor3D_offset(&src, 0, 0, 0);
 
     // Load the values across the channels to compose the 6x6 or 6x1 tile
-    float d00 = *((__global float *)(src_addr + 0 * src_stride_z));
-    float d01 = *((__global float *)(src_addr + 1 * src_stride_z));
-    float d02 = *((__global float *)(src_addr + 2 * src_stride_z));
-    float d03 = *((__global float *)(src_addr + 3 * src_stride_z));
-    float d04 = *((__global float *)(src_addr + 4 * src_stride_z));
-    float d05 = *((__global float *)(src_addr + 5 * src_stride_z));
+    DATA_TYPE d00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z));
+    DATA_TYPE d01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z));
+    DATA_TYPE d02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z));
+    DATA_TYPE d03 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z));
+    DATA_TYPE d04 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z));
+    DATA_TYPE d05 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z));
 
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
     // Compute out00, out01, out02 and out03
-    float out00 = d00 + d01 + d02 + d03 + d04;
-    float out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04;
-    float out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04;
-    float out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05;
+    DATA_TYPE out00 = d00 + d01 + d02 + d03 + d04;
+    DATA_TYPE out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04;
+    DATA_TYPE out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04;
+    DATA_TYPE out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05;
 #else  // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    float d10 = *((__global float *)(src_addr + 6 * src_stride_z));
-    float d11 = *((__global float *)(src_addr + 7 * src_stride_z));
-    float d12 = *((__global float *)(src_addr + 8 * src_stride_z));
-    float d13 = *((__global float *)(src_addr + 9 * src_stride_z));
-    float d14 = *((__global float *)(src_addr + 10 * src_stride_z));
-    float d15 = *((__global float *)(src_addr + 11 * src_stride_z));
+    DATA_TYPE d10 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z));
+    DATA_TYPE d11 = *((__global DATA_TYPE *)(src_addr + 7 * src_stride_z));
+    DATA_TYPE d12 = *((__global DATA_TYPE *)(src_addr + 8 * src_stride_z));
+    DATA_TYPE d13 = *((__global DATA_TYPE *)(src_addr + 9 * src_stride_z));
+    DATA_TYPE d14 = *((__global DATA_TYPE *)(src_addr + 10 * src_stride_z));
+    DATA_TYPE d15 = *((__global DATA_TYPE *)(src_addr + 11 * src_stride_z));
 
-    float d20 = *((__global float *)(src_addr + 12 * src_stride_z));
-    float d21 = *((__global float *)(src_addr + 13 * src_stride_z));
-    float d22 = *((__global float *)(src_addr + 14 * src_stride_z));
-    float d23 = *((__global float *)(src_addr + 15 * src_stride_z));
-    float d24 = *((__global float *)(src_addr + 16 * src_stride_z));
-    float d25 = *((__global float *)(src_addr + 17 * src_stride_z));
+    DATA_TYPE d20 = *((__global DATA_TYPE *)(src_addr + 12 * src_stride_z));
+    DATA_TYPE d21 = *((__global DATA_TYPE *)(src_addr + 13 * src_stride_z));
+    DATA_TYPE d22 = *((__global DATA_TYPE *)(src_addr + 14 * src_stride_z));
+    DATA_TYPE d23 = *((__global DATA_TYPE *)(src_addr + 15 * src_stride_z));
+    DATA_TYPE d24 = *((__global DATA_TYPE *)(src_addr + 16 * src_stride_z));
+    DATA_TYPE d25 = *((__global DATA_TYPE *)(src_addr + 17 * src_stride_z));
 
-    float d30 = *((__global float *)(src_addr + 18 * src_stride_z));
-    float d31 = *((__global float *)(src_addr + 19 * src_stride_z));
-    float d32 = *((__global float *)(src_addr + 20 * src_stride_z));
-    float d33 = *((__global float *)(src_addr + 21 * src_stride_z));
-    float d34 = *((__global float *)(src_addr + 22 * src_stride_z));
-    float d35 = *((__global float *)(src_addr + 23 * src_stride_z));
+    DATA_TYPE d30 = *((__global DATA_TYPE *)(src_addr + 18 * src_stride_z));
+    DATA_TYPE d31 = *((__global DATA_TYPE *)(src_addr + 19 * src_stride_z));
+    DATA_TYPE d32 = *((__global DATA_TYPE *)(src_addr + 20 * src_stride_z));
+    DATA_TYPE d33 = *((__global DATA_TYPE *)(src_addr + 21 * src_stride_z));
+    DATA_TYPE d34 = *((__global DATA_TYPE *)(src_addr + 22 * src_stride_z));
+    DATA_TYPE d35 = *((__global DATA_TYPE *)(src_addr + 23 * src_stride_z));
 
-    float d40 = *((__global float *)(src_addr + 24 * src_stride_z));
-    float d41 = *((__global float *)(src_addr + 25 * src_stride_z));
-    float d42 = *((__global float *)(src_addr + 26 * src_stride_z));
-    float d43 = *((__global float *)(src_addr + 27 * src_stride_z));
-    float d44 = *((__global float *)(src_addr + 28 * src_stride_z));
-    float d45 = *((__global float *)(src_addr + 29 * src_stride_z));
+    DATA_TYPE d40 = *((__global DATA_TYPE *)(src_addr + 24 * src_stride_z));
+    DATA_TYPE d41 = *((__global DATA_TYPE *)(src_addr + 25 * src_stride_z));
+    DATA_TYPE d42 = *((__global DATA_TYPE *)(src_addr + 26 * src_stride_z));
+    DATA_TYPE d43 = *((__global DATA_TYPE *)(src_addr + 27 * src_stride_z));
+    DATA_TYPE d44 = *((__global DATA_TYPE *)(src_addr + 28 * src_stride_z));
+    DATA_TYPE d45 = *((__global DATA_TYPE *)(src_addr + 29 * src_stride_z));
 
-    float d50 = *((__global float *)(src_addr + 30 * src_stride_z));
-    float d51 = *((__global float *)(src_addr + 31 * src_stride_z));
-    float d52 = *((__global float *)(src_addr + 32 * src_stride_z));
-    float d53 = *((__global float *)(src_addr + 33 * src_stride_z));
-    float d54 = *((__global float *)(src_addr + 34 * src_stride_z));
-    float d55 = *((__global float *)(src_addr + 35 * src_stride_z));
+    DATA_TYPE d50 = *((__global DATA_TYPE *)(src_addr + 30 * src_stride_z));
+    DATA_TYPE d51 = *((__global DATA_TYPE *)(src_addr + 31 * src_stride_z));
+    DATA_TYPE d52 = *((__global DATA_TYPE *)(src_addr + 32 * src_stride_z));
+    DATA_TYPE d53 = *((__global DATA_TYPE *)(src_addr + 33 * src_stride_z));
+    DATA_TYPE d54 = *((__global DATA_TYPE *)(src_addr + 34 * src_stride_z));
+    DATA_TYPE d55 = *((__global DATA_TYPE *)(src_addr + 35 * src_stride_z));
 
     // Compute out00, out01, out02 and out03
-    float out00 = d01 + d21 + d41 + d11 + d31;
-    float out01 = d01 + d21 + d41 + d11 + d31;
-    float out02 = d01 + d21 + d41 + d11 + d31;
-    float out03 = d01 + d21 + d41 + d11 + d31;
+    DATA_TYPE out00 = d01 + d21 + d41 + d11 + d31;
+    DATA_TYPE out01 = d01 + d21 + d41 + d11 + d31;
+    DATA_TYPE out02 = d01 + d21 + d41 + d11 + d31;
+    DATA_TYPE out03 = d01 + d21 + d41 + d11 + d31;
 
-    float k0 = d03 + d04 + d13 + d14 + d23 + d24 + d33 + d34 + d43 + d44;
-    float k1 = 2.0f * d03 - 2.0f * d04 + 2.0f * d13 - 2.0f * d14 + 2.0f * d23 - 2.0f * d24 + 2.0f * d33 - 2.0f * d34 + 2.0f * d43 - 2.0f * d44;
+    DATA_TYPE k0 = d03 + d04 + d13 + d14 + d23 + d24 + d33 + d34 + d43 + d44;
+    DATA_TYPE k1 = 2.0f * d03 - 2.0f * d04 + 2.0f * d13 - 2.0f * d14 + 2.0f * d23 - 2.0f * d24 + 2.0f * d33 - 2.0f * d34 + 2.0f * d43 - 2.0f * d44;
 
     out00 += k0 + d00 + d02 + d10 + d12 + d20 + d22 + d30 + d32 + d40 + d42;
     out01 += k1 - d02 - d12 - d22 - d32 - d42;
@@ -255,10 +257,10 @@
     out03 += 4.0f * k1 - d02 - d12 - d22 - d32 - d42 + d05 + d15 + d25 + d35 + d45;
 
     // Compute out10, out11, out12 and out13
-    float out10 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
-    float out11 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
-    float out12 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
-    float out13 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
+    DATA_TYPE out10 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
+    DATA_TYPE out11 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
+    DATA_TYPE out12 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
+    DATA_TYPE out13 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
 
     k0 = d13 + d14 - d23 - d24 + 2.0f * d33 + 2.0f * d34 - 2.0f * d43 - 2.0f * d44;
     k1 = 2.0f * d13 - 2.0f * d14 - 2.0f * d23 + 2.0f * d24 + 4.0f * d33 - 4.0f * d34 - 4.0f * d43 + 4.0f * d44;
@@ -269,10 +271,10 @@
     out13 += 4.0f * k1 - d12 + d15 + d22 - d25 - 2.0f * d32 + 2.0f * d35 + 2.0f * d42 - 2.0f * d45;
 
     // Compute out20, out21, out22 and out23
-    float out20 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
-    float out21 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
-    float out22 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
-    float out23 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
+    DATA_TYPE out20 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
+    DATA_TYPE out21 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
+    DATA_TYPE out22 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
+    DATA_TYPE out23 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
 
     k0 = d13 + d14 + d23 + d24 + 4.0f * d33 + 4.0f * d34 + 4.0f * d43 + 4.0f * d44;
     k1 = 2.0f * d13 - 2.0f * d14 + 2.0f * d23 - 2.0f * d24 + 8.0f * d33 - 8.0f * d34 + 8.0f * d43 - 8.0f * d44;
@@ -283,10 +285,10 @@
     out23 += 4.0f * k1 - d12 + d15 - d22 + d25 - 4.0f * d32 + 4.0f * d35 - 4.0f * d42 + 4.0f * d45;
 
     // Compute out30, out31, out32 and out33
-    float out30 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
-    float out31 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
-    float out32 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
-    float out33 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
+    DATA_TYPE out30 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
+    DATA_TYPE out31 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
+    DATA_TYPE out32 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
+    DATA_TYPE out33 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
 
     k0 = d13 + d14 - d23 - d24 + 8.0f * d33 + 8.0f * d34 - 8.0f * d43 - 8.0f * d44 + d53 + d54;
     k1 = 2.0f * d13 - 2.0f * d14 - 2.0f * d23 + 2.0f * d24 + 16.0f * d33 - 16.0f * d34 - 16.0f * d43 + 16.0f * d44 + 2.0f * d53 - 2.0f * d54;
@@ -306,48 +308,48 @@
     // Add bias
     Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
 
-    float b = (float) * ((__global float *)(vector_offset(&bias, z_out)));
+    DATA_TYPE b = (DATA_TYPE) * ((__global DATA_TYPE *)(vector_offset(&bias, z_out)));
 
-    out00 += (float)b;
-    out01 += (float)b;
-    out02 += (float)b;
-    out03 += (float)b;
+    out00 += (DATA_TYPE)b;
+    out01 += (DATA_TYPE)b;
+    out02 += (DATA_TYPE)b;
+    out03 += (DATA_TYPE)b;
 #endif // defined(HAS_BIAS)
 
     // Get output address
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z;
 
     // Store the output tile
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    *((__global float *)(dst_addr + 0 * dst_stride_y)) = out00;
-    *((__global float *)(dst_addr + 1 * dst_stride_y)) = out01;
-    *((__global float *)(dst_addr + 2 * dst_stride_y)) = out02;
-    *((__global float *)(dst_addr + 3 * dst_stride_y)) = out03;
+    *((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y)) = out00;
+    *((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y)) = out01;
+    *((__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y)) = out02;
+    *((__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y)) = out03;
 #else  // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    vstore4((float4)(out00, out01, out02, out03), 0, (__global float *)(dst_addr + 0 * dst_stride_y));
+    vstore4((VEC_DATA_TYPE(DATA_TYPE, 4))(out00, out01, out02, out03), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y));
 #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
 #if !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 #if defined(HAS_BIAS)
     // Add bias
-    out10 += (float)b;
-    out11 += (float)b;
-    out12 += (float)b;
-    out13 += (float)b;
+    out10 += (DATA_TYPE)b;
+    out11 += (DATA_TYPE)b;
+    out12 += (DATA_TYPE)b;
+    out13 += (DATA_TYPE)b;
 
-    out20 += (float)b;
-    out21 += (float)b;
-    out22 += (float)b;
-    out23 += (float)b;
+    out20 += (DATA_TYPE)b;
+    out21 += (DATA_TYPE)b;
+    out22 += (DATA_TYPE)b;
+    out23 += (DATA_TYPE)b;
 
-    out30 += (float)b;
-    out31 += (float)b;
-    out32 += (float)b;
-    out33 += (float)b;
+    out30 += (DATA_TYPE)b;
+    out31 += (DATA_TYPE)b;
+    out32 += (DATA_TYPE)b;
+    out33 += (DATA_TYPE)b;
 #endif // defined(HAS_BIAS)
-    vstore4((float4)(out10, out11, out12, out13), 0, (__global float *)(dst_addr + 1 * dst_stride_y));
-    vstore4((float4)(out20, out21, out22, out23), 0, (__global float *)(dst_addr + 2 * dst_stride_y));
-    vstore4((float4)(out30, out31, out32, out33), 0, (__global float *)(dst_addr + 3 * dst_stride_y));
+    vstore4((VEC_DATA_TYPE(DATA_TYPE, 4))(out10, out11, out12, out13), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y));
+    vstore4((VEC_DATA_TYPE(DATA_TYPE, 4))(out20, out21, out22, out23), 0, (__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y));
+    vstore4((VEC_DATA_TYPE(DATA_TYPE, 4))(out30, out31, out32, out33), 0, (__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y));
 #endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 }
 
@@ -358,8 +360,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note If this kernel is used to perform Winograd output transform 3x1, -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd output transform 1x3, -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -391,64 +394,64 @@
     const __global uchar *src_addr = tensor3D_offset(&src, 0, 0, 0);
 
     // Load the values across the 36 channels to compose the 6x6 or 6x1 tile
-    float d00 = *((__global float *)(src_addr + 0 * src_stride_z));
-    float d01 = *((__global float *)(src_addr + 1 * src_stride_z));
-    float d02 = *((__global float *)(src_addr + 2 * src_stride_z));
-    float d03 = *((__global float *)(src_addr + 3 * src_stride_z));
-    float d04 = *((__global float *)(src_addr + 4 * src_stride_z));
-    float d05 = *((__global float *)(src_addr + 5 * src_stride_z));
+    DATA_TYPE d00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z));
+    DATA_TYPE d01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z));
+    DATA_TYPE d02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z));
+    DATA_TYPE d03 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z));
+    DATA_TYPE d04 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z));
+    DATA_TYPE d05 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z));
 
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
     // Compute out00, out01, out02 and out03
-    float out00 = d00 + d01 + d02 + d03 + d04;
-    float out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04;
-    float out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04;
-    float out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05;
+    DATA_TYPE out00 = d00 + d01 + d02 + d03 + d04;
+    DATA_TYPE out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04;
+    DATA_TYPE out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04;
+    DATA_TYPE out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05;
 #else  // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
-    float d10 = *((__global float *)(src_addr + 6 * src_stride_z));
-    float d11 = *((__global float *)(src_addr + 7 * src_stride_z));
-    float d12 = *((__global float *)(src_addr + 8 * src_stride_z));
-    float d13 = *((__global float *)(src_addr + 9 * src_stride_z));
-    float d14 = *((__global float *)(src_addr + 10 * src_stride_z));
-    float d15 = *((__global float *)(src_addr + 11 * src_stride_z));
+    DATA_TYPE d10 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z));
+    DATA_TYPE d11 = *((__global DATA_TYPE *)(src_addr + 7 * src_stride_z));
+    DATA_TYPE d12 = *((__global DATA_TYPE *)(src_addr + 8 * src_stride_z));
+    DATA_TYPE d13 = *((__global DATA_TYPE *)(src_addr + 9 * src_stride_z));
+    DATA_TYPE d14 = *((__global DATA_TYPE *)(src_addr + 10 * src_stride_z));
+    DATA_TYPE d15 = *((__global DATA_TYPE *)(src_addr + 11 * src_stride_z));
 
-    float d20 = *((__global float *)(src_addr + 12 * src_stride_z));
-    float d21 = *((__global float *)(src_addr + 13 * src_stride_z));
-    float d22 = *((__global float *)(src_addr + 14 * src_stride_z));
-    float d23 = *((__global float *)(src_addr + 15 * src_stride_z));
-    float d24 = *((__global float *)(src_addr + 16 * src_stride_z));
-    float d25 = *((__global float *)(src_addr + 17 * src_stride_z));
+    DATA_TYPE d20 = *((__global DATA_TYPE *)(src_addr + 12 * src_stride_z));
+    DATA_TYPE d21 = *((__global DATA_TYPE *)(src_addr + 13 * src_stride_z));
+    DATA_TYPE d22 = *((__global DATA_TYPE *)(src_addr + 14 * src_stride_z));
+    DATA_TYPE d23 = *((__global DATA_TYPE *)(src_addr + 15 * src_stride_z));
+    DATA_TYPE d24 = *((__global DATA_TYPE *)(src_addr + 16 * src_stride_z));
+    DATA_TYPE d25 = *((__global DATA_TYPE *)(src_addr + 17 * src_stride_z));
 
-    float d30 = *((__global float *)(src_addr + 18 * src_stride_z));
-    float d31 = *((__global float *)(src_addr + 19 * src_stride_z));
-    float d32 = *((__global float *)(src_addr + 20 * src_stride_z));
-    float d33 = *((__global float *)(src_addr + 21 * src_stride_z));
-    float d34 = *((__global float *)(src_addr + 22 * src_stride_z));
-    float d35 = *((__global float *)(src_addr + 23 * src_stride_z));
+    DATA_TYPE d30 = *((__global DATA_TYPE *)(src_addr + 18 * src_stride_z));
+    DATA_TYPE d31 = *((__global DATA_TYPE *)(src_addr + 19 * src_stride_z));
+    DATA_TYPE d32 = *((__global DATA_TYPE *)(src_addr + 20 * src_stride_z));
+    DATA_TYPE d33 = *((__global DATA_TYPE *)(src_addr + 21 * src_stride_z));
+    DATA_TYPE d34 = *((__global DATA_TYPE *)(src_addr + 22 * src_stride_z));
+    DATA_TYPE d35 = *((__global DATA_TYPE *)(src_addr + 23 * src_stride_z));
 
-    float d40 = *((__global float *)(src_addr + 24 * src_stride_z));
-    float d41 = *((__global float *)(src_addr + 25 * src_stride_z));
-    float d42 = *((__global float *)(src_addr + 26 * src_stride_z));
-    float d43 = *((__global float *)(src_addr + 27 * src_stride_z));
-    float d44 = *((__global float *)(src_addr + 28 * src_stride_z));
-    float d45 = *((__global float *)(src_addr + 29 * src_stride_z));
+    DATA_TYPE d40 = *((__global DATA_TYPE *)(src_addr + 24 * src_stride_z));
+    DATA_TYPE d41 = *((__global DATA_TYPE *)(src_addr + 25 * src_stride_z));
+    DATA_TYPE d42 = *((__global DATA_TYPE *)(src_addr + 26 * src_stride_z));
+    DATA_TYPE d43 = *((__global DATA_TYPE *)(src_addr + 27 * src_stride_z));
+    DATA_TYPE d44 = *((__global DATA_TYPE *)(src_addr + 28 * src_stride_z));
+    DATA_TYPE d45 = *((__global DATA_TYPE *)(src_addr + 29 * src_stride_z));
 
-    float d50 = *((__global float *)(src_addr + 30 * src_stride_z));
-    float d51 = *((__global float *)(src_addr + 31 * src_stride_z));
-    float d52 = *((__global float *)(src_addr + 32 * src_stride_z));
-    float d53 = *((__global float *)(src_addr + 33 * src_stride_z));
-    float d54 = *((__global float *)(src_addr + 34 * src_stride_z));
-    float d55 = *((__global float *)(src_addr + 35 * src_stride_z));
+    DATA_TYPE d50 = *((__global DATA_TYPE *)(src_addr + 30 * src_stride_z));
+    DATA_TYPE d51 = *((__global DATA_TYPE *)(src_addr + 31 * src_stride_z));
+    DATA_TYPE d52 = *((__global DATA_TYPE *)(src_addr + 32 * src_stride_z));
+    DATA_TYPE d53 = *((__global DATA_TYPE *)(src_addr + 33 * src_stride_z));
+    DATA_TYPE d54 = *((__global DATA_TYPE *)(src_addr + 34 * src_stride_z));
+    DATA_TYPE d55 = *((__global DATA_TYPE *)(src_addr + 35 * src_stride_z));
 
     // Compute out00, out01, out02 and out03
-    float out00 = d01 + d21 + d41 + d11 + d31;
-    float out01 = d01 + d21 + d41 + d11 + d31;
-    float out02 = d01 + d21 + d41 + d11 + d31;
-    float out03 = d01 + d21 + d41 + d11 + d31;
+    DATA_TYPE out00 = d01 + d21 + d41 + d11 + d31;
+    DATA_TYPE out01 = d01 + d21 + d41 + d11 + d31;
+    DATA_TYPE out02 = d01 + d21 + d41 + d11 + d31;
+    DATA_TYPE out03 = d01 + d21 + d41 + d11 + d31;
 
-    float k0 = d03 + d04 + d13 + d14 + d23 + d24 + d33 + d34 + d43 + d44;
-    float k1 = 2.0f * d03 - 2.0f * d04 + 2.0f * d13 - 2.0f * d14 + 2.0f * d23 - 2.0f * d24 + 2.0f * d33 - 2.0f * d34 + 2.0f * d43 - 2.0f * d44;
+    DATA_TYPE k0 = d03 + d04 + d13 + d14 + d23 + d24 + d33 + d34 + d43 + d44;
+    DATA_TYPE k1 = 2.0f * d03 - 2.0f * d04 + 2.0f * d13 - 2.0f * d14 + 2.0f * d23 - 2.0f * d24 + 2.0f * d33 - 2.0f * d34 + 2.0f * d43 - 2.0f * d44;
 
     out00 += k0 + d00 + d02 + d10 + d12 + d20 + d22 + d30 + d32 + d40 + d42;
     out01 += k1 - d02 - d12 - d22 - d32 - d42;
@@ -456,10 +459,10 @@
     out03 += 4.0f * k1 - d02 - d12 - d22 - d32 - d42 + d05 + d15 + d25 + d35 + d45;
 
     // Compute out10, out11, out12 and out13
-    float out10 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
-    float out11 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
-    float out12 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
-    float out13 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
+    DATA_TYPE out10 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
+    DATA_TYPE out11 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
+    DATA_TYPE out12 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
+    DATA_TYPE out13 = d11 - d21 + 2.0f * d31 - 2.0f * d41;
 
     k0 = d13 + d14 - d23 - d24 + 2.0f * d33 + 2.0f * d34 - 2.0f * d43 - 2.0f * d44;
     k1 = 2.0f * d13 - 2.0f * d14 - 2.0f * d23 + 2.0f * d24 + 4.0f * d33 - 4.0f * d34 - 4.0f * d43 + 4.0f * d44;
@@ -470,10 +473,10 @@
     out13 += 4.0f * k1 - d12 + d15 + d22 - d25 - 2.0f * d32 + 2.0f * d35 + 2.0f * d42 - 2.0f * d45;
 
     // Compute out20, out21, out22 and out23
-    float out20 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
-    float out21 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
-    float out22 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
-    float out23 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
+    DATA_TYPE out20 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
+    DATA_TYPE out21 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
+    DATA_TYPE out22 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
+    DATA_TYPE out23 = d11 + d21 + 4.0f * d31 + 4.0f * d41;
 
     k0 = d13 + d14 + d23 + d24 + 4.0f * d33 + 4.0f * d34 + 4.0f * d43 + 4.0f * d44;
     k1 = 2.0f * d13 - 2.0f * d14 + 2.0f * d23 - 2.0f * d24 + 8.0f * d33 - 8.0f * d34 + 8.0f * d43 - 8.0f * d44;
@@ -484,10 +487,10 @@
     out23 += 4.0f * k1 - d12 + d15 - d22 + d25 - 4.0f * d32 + 4.0f * d35 - 4.0f * d42 + 4.0f * d45;
 
     // Compute out30, out31, out32 and out33
-    float out30 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
-    float out31 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
-    float out32 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
-    float out33 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
+    DATA_TYPE out30 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
+    DATA_TYPE out31 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
+    DATA_TYPE out32 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
+    DATA_TYPE out33 = d11 - d21 + 8.0f * d31 - 8.0f * d41 + d51;
 
     k0 = d13 + d14 - d23 - d24 + 8.0f * d33 + 8.0f * d34 - 8.0f * d43 - 8.0f * d44 + d53 + d54;
     k1 = 2.0f * d13 - 2.0f * d14 - 2.0f * d23 + 2.0f * d24 + 16.0f * d33 - 16.0f * d34 - 16.0f * d43 + 16.0f * d44 + 2.0f * d53 - 2.0f * d54;
@@ -507,72 +510,72 @@
     // Add bias
     Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
 
-    float b = (float) * ((__global float *)(vector_offset(&bias, x_out)));
+    DATA_TYPE b = (DATA_TYPE) * ((__global DATA_TYPE *)(vector_offset(&bias, x_out)));
 
-    out00 += (float)b;
-    out01 += (float)b;
-    out02 += (float)b;
-    out03 += (float)b;
+    out00 += (DATA_TYPE)b;
+    out01 += (DATA_TYPE)b;
+    out02 += (DATA_TYPE)b;
+    out03 += (DATA_TYPE)b;
 #if !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) & !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    out10 += (float)b;
-    out11 += (float)b;
-    out12 += (float)b;
-    out13 += (float)b;
+    out10 += (DATA_TYPE)b;
+    out11 += (DATA_TYPE)b;
+    out12 += (DATA_TYPE)b;
+    out13 += (DATA_TYPE)b;
 
-    out20 += (float)b;
-    out21 += (float)b;
-    out22 += (float)b;
-    out23 += (float)b;
+    out20 += (DATA_TYPE)b;
+    out21 += (DATA_TYPE)b;
+    out22 += (DATA_TYPE)b;
+    out23 += (DATA_TYPE)b;
 
-    out30 += (float)b;
-    out31 += (float)b;
-    out32 += (float)b;
-    out33 += (float)b;
+    out30 += (DATA_TYPE)b;
+    out31 += (DATA_TYPE)b;
+    out32 += (DATA_TYPE)b;
+    out33 += (DATA_TYPE)b;
 #endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) & !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
 #endif // defined(HAS_BIAS)
 
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z);
+    int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z);
     offset      = min(offset + (int4)(0, 1, 2, 3) * (int4)dst_stride_z, (int4)dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding).
 
     // Store the 1x4 output tile
-    *((__global float *)(dst_ptr + offset.s0)) = out00;
-    *((__global float *)(dst_ptr + offset.s1)) = out01;
-    *((__global float *)(dst_ptr + offset.s2)) = out02;
-    *((__global float *)(dst_ptr + offset.s3)) = out03;
+    *((__global DATA_TYPE *)(dst_ptr + offset.s0)) = out00;
+    *((__global DATA_TYPE *)(dst_ptr + offset.s1)) = out01;
+    *((__global DATA_TYPE *)(dst_ptr + offset.s2)) = out02;
+    *((__global DATA_TYPE *)(dst_ptr + offset.s3)) = out03;
 #elif defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL)
     // Store the 4x1 output tile
-    int offset = dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z;
+    int offset = dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z;
     int mult_y = min(dst_size - offset, 1);
 
-    *((__global float *)(dst_ptr + mult_y * 0 * dst_stride_y + offset)) = out00;
-    *((__global float *)(dst_ptr + mult_y * 1 * dst_stride_y + offset)) = out01;
-    *((__global float *)(dst_ptr + mult_y * 2 * dst_stride_y + offset)) = out02;
-    *((__global float *)(dst_ptr + mult_y * 3 * dst_stride_y + offset)) = out03;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y * 0 * dst_stride_y + offset)) = out00;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y * 1 * dst_stride_y + offset)) = out01;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y * 2 * dst_stride_y + offset)) = out02;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y * 3 * dst_stride_y + offset)) = out03;
 #else // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL)
     // Get output address
-    int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z);
+    int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z);
     offset      = min(offset + (int4)(0, 1, 2, 3) * (int4)dst_stride_z, (int4)dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding).
-    int4 mult_y = min((int4)dst_size - offset, (int4)1);                                       // If out of bound, we don't want to increase dst_stride_y, so we set the multiplier to 0. It will be 1 otherwise.
+    int4 mult_y = min((int4)dst_size - offset, (int4)1);                                 // If out of bound, we don't want to increase dst_stride_y, so we set the multiplier to 0. It will be 1 otherwise.
 
     // Store the 4x4 output tile
-    *((__global float *)(dst_ptr + mult_y.s0 * 0 * dst_stride_y + offset.s0)) = out00;
-    *((__global float *)(dst_ptr + mult_y.s0 * 1 * dst_stride_y + offset.s0)) = out01;
-    *((__global float *)(dst_ptr + mult_y.s0 * 2 * dst_stride_y + offset.s0)) = out02;
-    *((__global float *)(dst_ptr + mult_y.s0 * 3 * dst_stride_y + offset.s0)) = out03;
-    *((__global float *)(dst_ptr + mult_y.s1 * 0 * dst_stride_y + offset.s1)) = out10;
-    *((__global float *)(dst_ptr + mult_y.s1 * 1 * dst_stride_y + offset.s1)) = out11;
-    *((__global float *)(dst_ptr + mult_y.s1 * 2 * dst_stride_y + offset.s1)) = out12;
-    *((__global float *)(dst_ptr + mult_y.s1 * 3 * dst_stride_y + offset.s1)) = out13;
-    *((__global float *)(dst_ptr + mult_y.s2 * 0 * dst_stride_y + offset.s2)) = out20;
-    *((__global float *)(dst_ptr + mult_y.s2 * 1 * dst_stride_y + offset.s2)) = out21;
-    *((__global float *)(dst_ptr + mult_y.s2 * 2 * dst_stride_y + offset.s2)) = out22;
-    *((__global float *)(dst_ptr + mult_y.s2 * 3 * dst_stride_y + offset.s2)) = out23;
-    *((__global float *)(dst_ptr + mult_y.s3 * 0 * dst_stride_y + offset.s3)) = out30;
-    *((__global float *)(dst_ptr + mult_y.s3 * 1 * dst_stride_y + offset.s3)) = out31;
-    *((__global float *)(dst_ptr + mult_y.s3 * 2 * dst_stride_y + offset.s3)) = out32;
-    *((__global float *)(dst_ptr + mult_y.s3 * 3 * dst_stride_y + offset.s3)) = out33;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 0 * dst_stride_y + offset.s0)) = out00;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 1 * dst_stride_y + offset.s0)) = out01;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 2 * dst_stride_y + offset.s0)) = out02;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 3 * dst_stride_y + offset.s0)) = out03;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s1 * 0 * dst_stride_y + offset.s1)) = out10;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s1 * 1 * dst_stride_y + offset.s1)) = out11;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s1 * 2 * dst_stride_y + offset.s1)) = out12;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s1 * 3 * dst_stride_y + offset.s1)) = out13;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s2 * 0 * dst_stride_y + offset.s2)) = out20;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s2 * 1 * dst_stride_y + offset.s2)) = out21;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s2 * 2 * dst_stride_y + offset.s2)) = out22;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s2 * 3 * dst_stride_y + offset.s2)) = out23;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s3 * 0 * dst_stride_y + offset.s3)) = out30;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s3 * 1 * dst_stride_y + offset.s3)) = out31;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s3 * 2 * dst_stride_y + offset.s3)) = out32;
+    *((__global DATA_TYPE *)(dst_ptr + mult_y.s3 * 3 * dst_stride_y + offset.s3)) = out33;
 
 #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL)
 }
@@ -601,8 +604,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note If this kernel is used to perform Winograd output transform 3x1, -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd output transform 1x3, -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -639,114 +643,116 @@
     int y_out = (y_in / NUM_TILES_X) * OUTPUT_TILE_H;
     int z_out = get_global_id(0);
 
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z;
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z;
 
     // Load the values across the channels to compose the input tile
-    float d00 = *((__global float *)(src_addr + 0 * src_stride_z));
-    float d01 = *((__global float *)(src_addr + 1 * src_stride_z));
-    float d02 = *((__global float *)(src_addr + 2 * src_stride_z));
-    float d03 = *((__global float *)(src_addr + 3 * src_stride_z));
-    float d04 = *((__global float *)(src_addr + 4 * src_stride_z));
-    float d05 = *((__global float *)(src_addr + 5 * src_stride_z));
-    float d06 = *((__global float *)(src_addr + 6 * src_stride_z));
-    float d07 = *((__global float *)(src_addr + 7 * src_stride_z));
+    DATA_TYPE d00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z));
+    DATA_TYPE d01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z));
+    DATA_TYPE d02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z));
+    DATA_TYPE d03 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z));
+    DATA_TYPE d04 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z));
+    DATA_TYPE d05 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z));
+    DATA_TYPE d06 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z));
+    DATA_TYPE d07 = *((__global DATA_TYPE *)(src_addr + 7 * src_stride_z));
 
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
     // Compute out00, out01, out02 and out03
-    float out00 = d00 + d01 + d02 + d03 + d04 + 8.0f * d05 + 8.0f * d06;
-    float out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04 + 4.0f * d05 - 4.0f * d06;
-    float out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04 + 2.0f * d05 + 2.0f * d06;
-    float out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05 - d06 + d07;
+    DATA_TYPE out00 = d00 + d01 + d02 + d03 + d04 + 8.0f * d05 + 8.0f * d06;
+    DATA_TYPE out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04 + 4.0f * d05 - 4.0f * d06;
+    DATA_TYPE out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04 + 2.0f * d05 + 2.0f * d06;
+    DATA_TYPE out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05 - d06 + d07;
 
 #if defined(HAS_BIAS)
     // Add bias
     Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
 
-    float b = (float) * ((__global float *)(vector_offset(&bias, z_out)));
+    DATA_TYPE b = (DATA_TYPE) * ((__global DATA_TYPE *)(vector_offset(&bias, z_out)));
 
-    out00 += (float)b;
-    out01 += (float)b;
-    out02 += (float)b;
-    out03 += (float)b;
+    out00 += (DATA_TYPE)b;
+    out01 += (DATA_TYPE)b;
+    out02 += (DATA_TYPE)b;
+    out03 += (DATA_TYPE)b;
 #endif // defined(HAS_BIAS)
 
     // Store the output tile
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    *((__global float *)(dst_addr + 0 * dst_stride_y)) = out00;
-    *((__global float *)(dst_addr + 1 * dst_stride_y)) = out01;
-    *((__global float *)(dst_addr + 2 * dst_stride_y)) = out02;
-    *((__global float *)(dst_addr + 3 * dst_stride_y)) = out03;
+    *((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y)) = out00;
+    *((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y)) = out01;
+    *((__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y)) = out02;
+    *((__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y)) = out03;
 #else  // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    vstore4((float4)(out00, out01, out02, out03), 0, (__global float *)(dst_addr));
+    vstore4((VEC_DATA_TYPE(DATA_TYPE, 4))(out00, out01, out02, out03), 0, (__global DATA_TYPE *)(dst_addr));
 #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
 #else // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
-    float d10                                                           = *((__global float *)(src_addr + 8 * src_stride_z));
-    float d11                                                           = *((__global float *)(src_addr + 9 * src_stride_z));
-    float d12                                                           = *((__global float *)(src_addr + 10 * src_stride_z));
-    float d13                                                           = *((__global float *)(src_addr + 11 * src_stride_z));
-    float d14                                                           = *((__global float *)(src_addr + 12 * src_stride_z));
-    float d15                                                           = *((__global float *)(src_addr + 13 * src_stride_z));
-    float d16                                                           = *((__global float *)(src_addr + 14 * src_stride_z));
-    float d17                                                           = *((__global float *)(src_addr + 15 * src_stride_z));
+    DATA_TYPE d10                                                           = *((__global DATA_TYPE *)(src_addr + 8 * src_stride_z));
+    DATA_TYPE d11                                                           = *((__global DATA_TYPE *)(src_addr + 9 * src_stride_z));
+    DATA_TYPE d12                                                           = *((__global DATA_TYPE *)(src_addr + 10 * src_stride_z));
+    DATA_TYPE d13                                                           = *((__global DATA_TYPE *)(src_addr + 11 * src_stride_z));
+    DATA_TYPE d14                                                           = *((__global DATA_TYPE *)(src_addr + 12 * src_stride_z));
+    DATA_TYPE d15                                                           = *((__global DATA_TYPE *)(src_addr + 13 * src_stride_z));
+    DATA_TYPE d16                                                           = *((__global DATA_TYPE *)(src_addr + 14 * src_stride_z));
+    DATA_TYPE d17                                                           = *((__global DATA_TYPE *)(src_addr + 15 * src_stride_z));
 
-    float d20 = *((__global float *)(src_addr + 16 * src_stride_z));
-    float d21 = *((__global float *)(src_addr + 17 * src_stride_z));
-    float d22 = *((__global float *)(src_addr + 18 * src_stride_z));
-    float d23 = *((__global float *)(src_addr + 19 * src_stride_z));
-    float d24 = *((__global float *)(src_addr + 20 * src_stride_z));
-    float d25 = *((__global float *)(src_addr + 21 * src_stride_z));
-    float d26 = *((__global float *)(src_addr + 22 * src_stride_z));
-    float d27 = *((__global float *)(src_addr + 23 * src_stride_z));
+    DATA_TYPE d20 = *((__global DATA_TYPE *)(src_addr + 16 * src_stride_z));
+    DATA_TYPE d21 = *((__global DATA_TYPE *)(src_addr + 17 * src_stride_z));
+    DATA_TYPE d22 = *((__global DATA_TYPE *)(src_addr + 18 * src_stride_z));
+    DATA_TYPE d23 = *((__global DATA_TYPE *)(src_addr + 19 * src_stride_z));
+    DATA_TYPE d24 = *((__global DATA_TYPE *)(src_addr + 20 * src_stride_z));
+    DATA_TYPE d25 = *((__global DATA_TYPE *)(src_addr + 21 * src_stride_z));
+    DATA_TYPE d26 = *((__global DATA_TYPE *)(src_addr + 22 * src_stride_z));
+    DATA_TYPE d27 = *((__global DATA_TYPE *)(src_addr + 23 * src_stride_z));
 
-    float d30 = *((__global float *)(src_addr + 24 * src_stride_z));
-    float d31 = *((__global float *)(src_addr + 25 * src_stride_z));
-    float d32 = *((__global float *)(src_addr + 26 * src_stride_z));
-    float d33 = *((__global float *)(src_addr + 27 * src_stride_z));
-    float d34 = *((__global float *)(src_addr + 28 * src_stride_z));
-    float d35 = *((__global float *)(src_addr + 29 * src_stride_z));
-    float d36 = *((__global float *)(src_addr + 30 * src_stride_z));
-    float d37 = *((__global float *)(src_addr + 31 * src_stride_z));
+    DATA_TYPE d30 = *((__global DATA_TYPE *)(src_addr + 24 * src_stride_z));
+    DATA_TYPE d31 = *((__global DATA_TYPE *)(src_addr + 25 * src_stride_z));
+    DATA_TYPE d32 = *((__global DATA_TYPE *)(src_addr + 26 * src_stride_z));
+    DATA_TYPE d33 = *((__global DATA_TYPE *)(src_addr + 27 * src_stride_z));
+    DATA_TYPE d34 = *((__global DATA_TYPE *)(src_addr + 28 * src_stride_z));
+    DATA_TYPE d35 = *((__global DATA_TYPE *)(src_addr + 29 * src_stride_z));
+    DATA_TYPE d36 = *((__global DATA_TYPE *)(src_addr + 30 * src_stride_z));
+    DATA_TYPE d37 = *((__global DATA_TYPE *)(src_addr + 31 * src_stride_z));
 
-    float d40 = *((__global float *)(src_addr + 32 * src_stride_z));
-    float d41 = *((__global float *)(src_addr + 33 * src_stride_z));
-    float d42 = *((__global float *)(src_addr + 34 * src_stride_z));
-    float d43 = *((__global float *)(src_addr + 35 * src_stride_z));
-    float d44 = *((__global float *)(src_addr + 36 * src_stride_z));
-    float d45 = *((__global float *)(src_addr + 37 * src_stride_z));
-    float d46 = *((__global float *)(src_addr + 38 * src_stride_z));
-    float d47 = *((__global float *)(src_addr + 39 * src_stride_z));
+    DATA_TYPE d40 = *((__global DATA_TYPE *)(src_addr + 32 * src_stride_z));
+    DATA_TYPE d41 = *((__global DATA_TYPE *)(src_addr + 33 * src_stride_z));
+    DATA_TYPE d42 = *((__global DATA_TYPE *)(src_addr + 34 * src_stride_z));
+    DATA_TYPE d43 = *((__global DATA_TYPE *)(src_addr + 35 * src_stride_z));
+    DATA_TYPE d44 = *((__global DATA_TYPE *)(src_addr + 36 * src_stride_z));
+    DATA_TYPE d45 = *((__global DATA_TYPE *)(src_addr + 37 * src_stride_z));
+    DATA_TYPE d46 = *((__global DATA_TYPE *)(src_addr + 38 * src_stride_z));
+    DATA_TYPE d47 = *((__global DATA_TYPE *)(src_addr + 39 * src_stride_z));
 
-    float d50 = *((__global float *)(src_addr + 40 * src_stride_z));
-    float d51 = *((__global float *)(src_addr + 41 * src_stride_z));
-    float d52 = *((__global float *)(src_addr + 42 * src_stride_z));
-    float d53 = *((__global float *)(src_addr + 43 * src_stride_z));
-    float d54 = *((__global float *)(src_addr + 44 * src_stride_z));
-    float d55 = *((__global float *)(src_addr + 45 * src_stride_z));
-    float d56 = *((__global float *)(src_addr + 46 * src_stride_z));
-    float d57 = *((__global float *)(src_addr + 47 * src_stride_z));
+    DATA_TYPE d50 = *((__global DATA_TYPE *)(src_addr + 40 * src_stride_z));
+    DATA_TYPE d51 = *((__global DATA_TYPE *)(src_addr + 41 * src_stride_z));
+    DATA_TYPE d52 = *((__global DATA_TYPE *)(src_addr + 42 * src_stride_z));
+    DATA_TYPE d53 = *((__global DATA_TYPE *)(src_addr + 43 * src_stride_z));
+    DATA_TYPE d54 = *((__global DATA_TYPE *)(src_addr + 44 * src_stride_z));
+    DATA_TYPE d55 = *((__global DATA_TYPE *)(src_addr + 45 * src_stride_z));
+    DATA_TYPE d56 = *((__global DATA_TYPE *)(src_addr + 46 * src_stride_z));
+    DATA_TYPE d57 = *((__global DATA_TYPE *)(src_addr + 47 * src_stride_z));
 
-    float d60 = *((__global float *)(src_addr + 48 * src_stride_z));
-    float d61 = *((__global float *)(src_addr + 49 * src_stride_z));
-    float d62 = *((__global float *)(src_addr + 50 * src_stride_z));
-    float d63 = *((__global float *)(src_addr + 51 * src_stride_z));
-    float d64 = *((__global float *)(src_addr + 52 * src_stride_z));
-    float d65 = *((__global float *)(src_addr + 53 * src_stride_z));
-    float d66 = *((__global float *)(src_addr + 54 * src_stride_z));
-    float d67 = *((__global float *)(src_addr + 55 * src_stride_z));
+    DATA_TYPE d60 = *((__global DATA_TYPE *)(src_addr + 48 * src_stride_z));
+    DATA_TYPE d61 = *((__global DATA_TYPE *)(src_addr + 49 * src_stride_z));
+    DATA_TYPE d62 = *((__global DATA_TYPE *)(src_addr + 50 * src_stride_z));
+    DATA_TYPE d63 = *((__global DATA_TYPE *)(src_addr + 51 * src_stride_z));
+    DATA_TYPE d64 = *((__global DATA_TYPE *)(src_addr + 52 * src_stride_z));
+    DATA_TYPE d65 = *((__global DATA_TYPE *)(src_addr + 53 * src_stride_z));
+    DATA_TYPE d66 = *((__global DATA_TYPE *)(src_addr + 54 * src_stride_z));
+    DATA_TYPE d67 = *((__global DATA_TYPE *)(src_addr + 55 * src_stride_z));
 
-    float d70 = *((__global float *)(src_addr + 56 * src_stride_z));
-    float d71 = *((__global float *)(src_addr + 57 * src_stride_z));
-    float d72 = *((__global float *)(src_addr + 58 * src_stride_z));
-    float d73 = *((__global float *)(src_addr + 59 * src_stride_z));
-    float d74 = *((__global float *)(src_addr + 60 * src_stride_z));
-    float d75 = *((__global float *)(src_addr + 61 * src_stride_z));
-    float d76 = *((__global float *)(src_addr + 62 * src_stride_z));
-    float d77 = *((__global float *)(src_addr + 63 * src_stride_z));
+    DATA_TYPE d70 = *((__global DATA_TYPE *)(src_addr + 56 * src_stride_z));
+    DATA_TYPE d71 = *((__global DATA_TYPE *)(src_addr + 57 * src_stride_z));
+    DATA_TYPE d72 = *((__global DATA_TYPE *)(src_addr + 58 * src_stride_z));
+    DATA_TYPE d73 = *((__global DATA_TYPE *)(src_addr + 59 * src_stride_z));
+    DATA_TYPE d74 = *((__global DATA_TYPE *)(src_addr + 60 * src_stride_z));
+    DATA_TYPE d75 = *((__global DATA_TYPE *)(src_addr + 61 * src_stride_z));
+    DATA_TYPE d76 = *((__global DATA_TYPE *)(src_addr + 62 * src_stride_z));
+    DATA_TYPE d77 = *((__global DATA_TYPE *)(src_addr + 63 * src_stride_z));
 
     // Compute the 8x4 intermediate tensor
-    float4 comm_fact0, comm_fact1, comm_fact2;
-    float4 tmp_col0, tmp_col1, tmp_col2, tmp_col3, tmp_col4, tmp_col5, tmp_col6, tmp_col7;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    comm_fact0, comm_fact1, comm_fact2;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp_col0, tmp_col1, tmp_col2, tmp_col3, tmp_col4, tmp_col5, tmp_col6, tmp_col7;
 
     COMPUTE_TMP_COL(tmp_col0, d00, d10, d20, d30, d40, d50, d60, d70, comm_fact0);
     COMPUTE_TMP_COL(tmp_col1, d01, d11, d21, d31, d41, d51, d61, d71, comm_fact0);
@@ -762,33 +768,37 @@
     comm_fact1 = tmp_col3 + tmp_col4;
     comm_fact2 = tmp_col5 + tmp_col6;
 
-    float4 out_col0 = comm_fact0 + comm_fact1 + 8.f * comm_fact2 + tmp_col0;
-    float4 out_col2 = comm_fact0 + 4.f * comm_fact1 + 2.f * comm_fact2;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out_col0 = comm_fact0 + comm_fact1 + (DATA_TYPE)8.f * comm_fact2 + tmp_col0;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out_col2 = comm_fact0 + (DATA_TYPE)4.f * comm_fact1 + (DATA_TYPE)2.f * comm_fact2;
 
     comm_fact0 = tmp_col1 - tmp_col2;
     comm_fact1 = tmp_col3 - tmp_col4;
     comm_fact2 = tmp_col5 - tmp_col6;
 
-    float4 out_col1 = comm_fact0 + 2.f * comm_fact1 + 4.f * comm_fact2;
-    float4 out_col3 = comm_fact0 + 8.f * comm_fact1 + comm_fact2 + tmp_col7;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out_col1 = comm_fact0 + (DATA_TYPE)2.f * comm_fact1 + (DATA_TYPE)4.f * comm_fact2;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out_col3 = comm_fact0 + (DATA_TYPE)8.f * comm_fact1 + comm_fact2 + tmp_col7;
 
 #if defined(HAS_BIAS)
     // Add bias
     Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
 
-    float b = (float) * ((__global float *)(vector_offset(&bias, z_out)));
+    DATA_TYPE b = (DATA_TYPE) * ((__global DATA_TYPE *)(vector_offset(&bias, z_out)));
 
-    out_col0 += (float4)b;
-    out_col1 += (float4)b;
-    out_col2 += (float4)b;
-    out_col3 += (float4)b;
+    out_col0 += (VEC_DATA_TYPE(DATA_TYPE, 4))b;
+    out_col1 += (VEC_DATA_TYPE(DATA_TYPE, 4))b;
+    out_col2 += (VEC_DATA_TYPE(DATA_TYPE, 4))b;
+    out_col3 += (VEC_DATA_TYPE(DATA_TYPE, 4))b;
 #endif // defined(HAS_BIAS)
 
     // Store the output tile
-    vstore4((float4)(out_col0.s0, out_col1.s0, out_col2.s0, out_col3.s0), 0, (__global float *)(dst_addr + 0 * dst_stride_y));
-    vstore4((float4)(out_col0.s1, out_col1.s1, out_col2.s1, out_col3.s1), 0, (__global float *)(dst_addr + 1 * dst_stride_y));
-    vstore4((float4)(out_col0.s2, out_col1.s2, out_col2.s2, out_col3.s2), 0, (__global float *)(dst_addr + 2 * dst_stride_y));
-    vstore4((float4)(out_col0.s3, out_col1.s3, out_col2.s3, out_col3.s3), 0, (__global float *)(dst_addr + 3 * dst_stride_y));
+    vstore4((VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s0, out_col1.s0, out_col2.s0, out_col3.s0), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y));
+    vstore4((VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s1, out_col1.s1, out_col2.s1, out_col3.s1), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y));
+    vstore4((VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s2, out_col1.s2, out_col2.s2, out_col3.s2), 0, (__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y));
+    vstore4((VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s3, out_col1.s3, out_col2.s3, out_col3.s3), 0, (__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y));
 #endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 }
 
@@ -799,8 +809,9 @@
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note If this kernel is used to perform Winograd output transform 5x1, -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
  * @note If this kernel is used to perform Winograd output transform 1x5, -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -836,122 +847,124 @@
     int z_out = (y_in / NUM_TILES_X) * OUTPUT_TILE_H;
 
     // Load the values across the channels to compose the input tile
-    float d00 = *((__global float *)(src_addr + 0 * src_stride_z));
-    float d01 = *((__global float *)(src_addr + 1 * src_stride_z));
-    float d02 = *((__global float *)(src_addr + 2 * src_stride_z));
-    float d03 = *((__global float *)(src_addr + 3 * src_stride_z));
-    float d04 = *((__global float *)(src_addr + 4 * src_stride_z));
-    float d05 = *((__global float *)(src_addr + 5 * src_stride_z));
-    float d06 = *((__global float *)(src_addr + 6 * src_stride_z));
-    float d07 = *((__global float *)(src_addr + 7 * src_stride_z));
+    DATA_TYPE d00 = *((__global DATA_TYPE *)(src_addr + 0 * src_stride_z));
+    DATA_TYPE d01 = *((__global DATA_TYPE *)(src_addr + 1 * src_stride_z));
+    DATA_TYPE d02 = *((__global DATA_TYPE *)(src_addr + 2 * src_stride_z));
+    DATA_TYPE d03 = *((__global DATA_TYPE *)(src_addr + 3 * src_stride_z));
+    DATA_TYPE d04 = *((__global DATA_TYPE *)(src_addr + 4 * src_stride_z));
+    DATA_TYPE d05 = *((__global DATA_TYPE *)(src_addr + 5 * src_stride_z));
+    DATA_TYPE d06 = *((__global DATA_TYPE *)(src_addr + 6 * src_stride_z));
+    DATA_TYPE d07 = *((__global DATA_TYPE *)(src_addr + 7 * src_stride_z));
 
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
     // Compute out00, out01, out02 and out03
-    float out00 = d00 + d01 + d02 + d03 + d04 + 8.0f * d05 + 8.0f * d06;
-    float out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04 + 4.0f * d05 - 4.0f * d06;
-    float out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04 + 2.0f * d05 + 2.0f * d06;
-    float out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05 - d06 + d07;
+    DATA_TYPE out00 = d00 + d01 + d02 + d03 + d04 + 8.0f * d05 + 8.0f * d06;
+    DATA_TYPE out01 = d01 - d02 + 2.0f * d03 - 2.0f * d04 + 4.0f * d05 - 4.0f * d06;
+    DATA_TYPE out02 = d01 + d02 + 4.0f * d03 + 4.0f * d04 + 2.0f * d05 + 2.0f * d06;
+    DATA_TYPE out03 = d01 - d02 + 8.0f * d03 - 8.0f * d04 + d05 - d06 + d07;
 
 #if defined(HAS_BIAS)
     // Add bias
     Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
 
-    float b = (float) * ((__global float *)(vector_offset(&bias, x_out)));
+    DATA_TYPE b = (DATA_TYPE) * ((__global DATA_TYPE *)(vector_offset(&bias, x_out)));
 
-    out00 += (float)b;
-    out01 += (float)b;
-    out02 += (float)b;
-    out03 += (float)b;
+    out00 += (DATA_TYPE)b;
+    out01 += (DATA_TYPE)b;
+    out02 += (DATA_TYPE)b;
+    out03 += (DATA_TYPE)b;
 #endif // defined(HAS_BIAS)
 
     // Store the output tile
 #if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
     // Get output address
-    int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z);
+    int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z);
     offset      = min(offset + (int4)(0, 1, 2, 3) * (int4)dst_stride_z, (int4)dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding).
 
-    *(__global float *)(dst_ptr + offset.s0) = out00;
-    *(__global float *)(dst_ptr + offset.s1) = out01;
-    *(__global float *)(dst_ptr + offset.s2) = out02;
-    *(__global float *)(dst_ptr + offset.s3) = out03;
+    *(__global DATA_TYPE *)(dst_ptr + offset.s0) = out00;
+    *(__global DATA_TYPE *)(dst_ptr + offset.s1) = out01;
+    *(__global DATA_TYPE *)(dst_ptr + offset.s2) = out02;
+    *(__global DATA_TYPE *)(dst_ptr + offset.s3) = out03;
 #else  // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
     // Get output address
-    int offset = dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z;
+    int offset = dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z;
 
-    *(__global float *)(dst_ptr + 0 * dst_stride_y + offset) = out00;
-    *(__global float *)(dst_ptr + 1 * dst_stride_y + offset) = out01;
-    *(__global float *)(dst_ptr + 2 * dst_stride_y + offset) = out02;
-    *(__global float *)(dst_ptr + 3 * dst_stride_y + offset) = out03;
+    *(__global DATA_TYPE *)(dst_ptr + 0 * dst_stride_y + offset) = out00;
+    *(__global DATA_TYPE *)(dst_ptr + 1 * dst_stride_y + offset) = out01;
+    *(__global DATA_TYPE *)(dst_ptr + 2 * dst_stride_y + offset) = out02;
+    *(__global DATA_TYPE *)(dst_ptr + 3 * dst_stride_y + offset) = out03;
 #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
 #else // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 
-    float d10 = *((__global float *)(src_addr + 8 * src_stride_z));
-    float d11 = *((__global float *)(src_addr + 9 * src_stride_z));
-    float d12 = *((__global float *)(src_addr + 10 * src_stride_z));
-    float d13 = *((__global float *)(src_addr + 11 * src_stride_z));
-    float d14 = *((__global float *)(src_addr + 12 * src_stride_z));
-    float d15 = *((__global float *)(src_addr + 13 * src_stride_z));
-    float d16 = *((__global float *)(src_addr + 14 * src_stride_z));
-    float d17 = *((__global float *)(src_addr + 15 * src_stride_z));
+    DATA_TYPE d10 = *((__global DATA_TYPE *)(src_addr + 8 * src_stride_z));
+    DATA_TYPE d11 = *((__global DATA_TYPE *)(src_addr + 9 * src_stride_z));
+    DATA_TYPE d12 = *((__global DATA_TYPE *)(src_addr + 10 * src_stride_z));
+    DATA_TYPE d13 = *((__global DATA_TYPE *)(src_addr + 11 * src_stride_z));
+    DATA_TYPE d14 = *((__global DATA_TYPE *)(src_addr + 12 * src_stride_z));
+    DATA_TYPE d15 = *((__global DATA_TYPE *)(src_addr + 13 * src_stride_z));
+    DATA_TYPE d16 = *((__global DATA_TYPE *)(src_addr + 14 * src_stride_z));
+    DATA_TYPE d17 = *((__global DATA_TYPE *)(src_addr + 15 * src_stride_z));
 
-    float d20 = *((__global float *)(src_addr + 16 * src_stride_z));
-    float d21 = *((__global float *)(src_addr + 17 * src_stride_z));
-    float d22 = *((__global float *)(src_addr + 18 * src_stride_z));
-    float d23 = *((__global float *)(src_addr + 19 * src_stride_z));
-    float d24 = *((__global float *)(src_addr + 20 * src_stride_z));
-    float d25 = *((__global float *)(src_addr + 21 * src_stride_z));
-    float d26 = *((__global float *)(src_addr + 22 * src_stride_z));
-    float d27 = *((__global float *)(src_addr + 23 * src_stride_z));
+    DATA_TYPE d20 = *((__global DATA_TYPE *)(src_addr + 16 * src_stride_z));
+    DATA_TYPE d21 = *((__global DATA_TYPE *)(src_addr + 17 * src_stride_z));
+    DATA_TYPE d22 = *((__global DATA_TYPE *)(src_addr + 18 * src_stride_z));
+    DATA_TYPE d23 = *((__global DATA_TYPE *)(src_addr + 19 * src_stride_z));
+    DATA_TYPE d24 = *((__global DATA_TYPE *)(src_addr + 20 * src_stride_z));
+    DATA_TYPE d25 = *((__global DATA_TYPE *)(src_addr + 21 * src_stride_z));
+    DATA_TYPE d26 = *((__global DATA_TYPE *)(src_addr + 22 * src_stride_z));
+    DATA_TYPE d27 = *((__global DATA_TYPE *)(src_addr + 23 * src_stride_z));
 
-    float d30 = *((__global float *)(src_addr + 24 * src_stride_z));
-    float d31 = *((__global float *)(src_addr + 25 * src_stride_z));
-    float d32 = *((__global float *)(src_addr + 26 * src_stride_z));
-    float d33 = *((__global float *)(src_addr + 27 * src_stride_z));
-    float d34 = *((__global float *)(src_addr + 28 * src_stride_z));
-    float d35 = *((__global float *)(src_addr + 29 * src_stride_z));
-    float d36 = *((__global float *)(src_addr + 30 * src_stride_z));
-    float d37 = *((__global float *)(src_addr + 31 * src_stride_z));
+    DATA_TYPE d30 = *((__global DATA_TYPE *)(src_addr + 24 * src_stride_z));
+    DATA_TYPE d31 = *((__global DATA_TYPE *)(src_addr + 25 * src_stride_z));
+    DATA_TYPE d32 = *((__global DATA_TYPE *)(src_addr + 26 * src_stride_z));
+    DATA_TYPE d33 = *((__global DATA_TYPE *)(src_addr + 27 * src_stride_z));
+    DATA_TYPE d34 = *((__global DATA_TYPE *)(src_addr + 28 * src_stride_z));
+    DATA_TYPE d35 = *((__global DATA_TYPE *)(src_addr + 29 * src_stride_z));
+    DATA_TYPE d36 = *((__global DATA_TYPE *)(src_addr + 30 * src_stride_z));
+    DATA_TYPE d37 = *((__global DATA_TYPE *)(src_addr + 31 * src_stride_z));
 
-    float d40 = *((__global float *)(src_addr + 32 * src_stride_z));
-    float d41 = *((__global float *)(src_addr + 33 * src_stride_z));
-    float d42 = *((__global float *)(src_addr + 34 * src_stride_z));
-    float d43 = *((__global float *)(src_addr + 35 * src_stride_z));
-    float d44 = *((__global float *)(src_addr + 36 * src_stride_z));
-    float d45 = *((__global float *)(src_addr + 37 * src_stride_z));
-    float d46 = *((__global float *)(src_addr + 38 * src_stride_z));
-    float d47 = *((__global float *)(src_addr + 39 * src_stride_z));
+    DATA_TYPE d40 = *((__global DATA_TYPE *)(src_addr + 32 * src_stride_z));
+    DATA_TYPE d41 = *((__global DATA_TYPE *)(src_addr + 33 * src_stride_z));
+    DATA_TYPE d42 = *((__global DATA_TYPE *)(src_addr + 34 * src_stride_z));
+    DATA_TYPE d43 = *((__global DATA_TYPE *)(src_addr + 35 * src_stride_z));
+    DATA_TYPE d44 = *((__global DATA_TYPE *)(src_addr + 36 * src_stride_z));
+    DATA_TYPE d45 = *((__global DATA_TYPE *)(src_addr + 37 * src_stride_z));
+    DATA_TYPE d46 = *((__global DATA_TYPE *)(src_addr + 38 * src_stride_z));
+    DATA_TYPE d47 = *((__global DATA_TYPE *)(src_addr + 39 * src_stride_z));
 
-    float d50 = *((__global float *)(src_addr + 40 * src_stride_z));
-    float d51 = *((__global float *)(src_addr + 41 * src_stride_z));
-    float d52 = *((__global float *)(src_addr + 42 * src_stride_z));
-    float d53 = *((__global float *)(src_addr + 43 * src_stride_z));
-    float d54 = *((__global float *)(src_addr + 44 * src_stride_z));
-    float d55 = *((__global float *)(src_addr + 45 * src_stride_z));
-    float d56 = *((__global float *)(src_addr + 46 * src_stride_z));
-    float d57 = *((__global float *)(src_addr + 47 * src_stride_z));
+    DATA_TYPE d50 = *((__global DATA_TYPE *)(src_addr + 40 * src_stride_z));
+    DATA_TYPE d51 = *((__global DATA_TYPE *)(src_addr + 41 * src_stride_z));
+    DATA_TYPE d52 = *((__global DATA_TYPE *)(src_addr + 42 * src_stride_z));
+    DATA_TYPE d53 = *((__global DATA_TYPE *)(src_addr + 43 * src_stride_z));
+    DATA_TYPE d54 = *((__global DATA_TYPE *)(src_addr + 44 * src_stride_z));
+    DATA_TYPE d55 = *((__global DATA_TYPE *)(src_addr + 45 * src_stride_z));
+    DATA_TYPE d56 = *((__global DATA_TYPE *)(src_addr + 46 * src_stride_z));
+    DATA_TYPE d57 = *((__global DATA_TYPE *)(src_addr + 47 * src_stride_z));
 
-    float d60 = *((__global float *)(src_addr + 48 * src_stride_z));
-    float d61 = *((__global float *)(src_addr + 49 * src_stride_z));
-    float d62 = *((__global float *)(src_addr + 50 * src_stride_z));
-    float d63 = *((__global float *)(src_addr + 51 * src_stride_z));
-    float d64 = *((__global float *)(src_addr + 52 * src_stride_z));
-    float d65 = *((__global float *)(src_addr + 53 * src_stride_z));
-    float d66 = *((__global float *)(src_addr + 54 * src_stride_z));
-    float d67 = *((__global float *)(src_addr + 55 * src_stride_z));
+    DATA_TYPE d60 = *((__global DATA_TYPE *)(src_addr + 48 * src_stride_z));
+    DATA_TYPE d61 = *((__global DATA_TYPE *)(src_addr + 49 * src_stride_z));
+    DATA_TYPE d62 = *((__global DATA_TYPE *)(src_addr + 50 * src_stride_z));
+    DATA_TYPE d63 = *((__global DATA_TYPE *)(src_addr + 51 * src_stride_z));
+    DATA_TYPE d64 = *((__global DATA_TYPE *)(src_addr + 52 * src_stride_z));
+    DATA_TYPE d65 = *((__global DATA_TYPE *)(src_addr + 53 * src_stride_z));
+    DATA_TYPE d66 = *((__global DATA_TYPE *)(src_addr + 54 * src_stride_z));
+    DATA_TYPE d67 = *((__global DATA_TYPE *)(src_addr + 55 * src_stride_z));
 
-    float d70 = *((__global float *)(src_addr + 56 * src_stride_z));
-    float d71 = *((__global float *)(src_addr + 57 * src_stride_z));
-    float d72 = *((__global float *)(src_addr + 58 * src_stride_z));
-    float d73 = *((__global float *)(src_addr + 59 * src_stride_z));
-    float d74 = *((__global float *)(src_addr + 60 * src_stride_z));
-    float d75 = *((__global float *)(src_addr + 61 * src_stride_z));
-    float d76 = *((__global float *)(src_addr + 62 * src_stride_z));
-    float d77 = *((__global float *)(src_addr + 63 * src_stride_z));
+    DATA_TYPE d70 = *((__global DATA_TYPE *)(src_addr + 56 * src_stride_z));
+    DATA_TYPE d71 = *((__global DATA_TYPE *)(src_addr + 57 * src_stride_z));
+    DATA_TYPE d72 = *((__global DATA_TYPE *)(src_addr + 58 * src_stride_z));
+    DATA_TYPE d73 = *((__global DATA_TYPE *)(src_addr + 59 * src_stride_z));
+    DATA_TYPE d74 = *((__global DATA_TYPE *)(src_addr + 60 * src_stride_z));
+    DATA_TYPE d75 = *((__global DATA_TYPE *)(src_addr + 61 * src_stride_z));
+    DATA_TYPE d76 = *((__global DATA_TYPE *)(src_addr + 62 * src_stride_z));
+    DATA_TYPE d77 = *((__global DATA_TYPE *)(src_addr + 63 * src_stride_z));
 
     // Compute the 8x4 intermediate tensor
-    float4 comm_fact0, comm_fact1, comm_fact2;
-    float4 tmp_col0, tmp_col1, tmp_col2, tmp_col3, tmp_col4, tmp_col5, tmp_col6, tmp_col7;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    comm_fact0, comm_fact1, comm_fact2;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    tmp_col0, tmp_col1, tmp_col2, tmp_col3, tmp_col4, tmp_col5, tmp_col6, tmp_col7;
 
     COMPUTE_TMP_COL(tmp_col0, d00, d10, d20, d30, d40, d50, d60, d70, comm_fact0);
     COMPUTE_TMP_COL(tmp_col1, d01, d11, d21, d31, d41, d51, d61, d71, comm_fact0);
@@ -967,49 +980,53 @@
     comm_fact1 = tmp_col3 + tmp_col4;
     comm_fact2 = tmp_col5 + tmp_col6;
 
-    float4 out_col0 = comm_fact0 + comm_fact1 + 8.f * comm_fact2 + tmp_col0;
-    float4 out_col2 = comm_fact0 + 4.f * comm_fact1 + 2.f * comm_fact2;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out_col0 = comm_fact0 + comm_fact1 + (DATA_TYPE)8.f * comm_fact2 + tmp_col0;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out_col2 = comm_fact0 + (DATA_TYPE)4.f * comm_fact1 + (DATA_TYPE)2.f * comm_fact2;
 
     comm_fact0 = tmp_col1 - tmp_col2;
     comm_fact1 = tmp_col3 - tmp_col4;
     comm_fact2 = tmp_col5 - tmp_col6;
 
-    float4 out_col1 = comm_fact0 + 2.f * comm_fact1 + 4.f * comm_fact2;
-    float4 out_col3 = comm_fact0 + 8.f * comm_fact1 + comm_fact2 + tmp_col7;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out_col1 = comm_fact0 + (DATA_TYPE)2.f * comm_fact1 + (DATA_TYPE)4.f * comm_fact2;
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    out_col3 = comm_fact0 + (DATA_TYPE)8.f * comm_fact1 + comm_fact2 + tmp_col7;
 
 #if defined(HAS_BIAS)
     // Add bias
     Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
 
-    float b = (float) * ((__global float *)(vector_offset(&bias, x_out)));
+    DATA_TYPE b = (DATA_TYPE) * ((__global DATA_TYPE *)(vector_offset(&bias, x_out)));
 
-    out_col0 += (float4)b;
-    out_col1 += (float4)b;
-    out_col2 += (float4)b;
-    out_col3 += (float4)b;
+    out_col0 += (VEC_DATA_TYPE(DATA_TYPE, 4))b;
+    out_col1 += (VEC_DATA_TYPE(DATA_TYPE, 4))b;
+    out_col2 += (VEC_DATA_TYPE(DATA_TYPE, 4))b;
+    out_col3 += (VEC_DATA_TYPE(DATA_TYPE, 4))b;
 #endif // defined(HAS_BIAS)
     // Get output address
-    int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(float) + y_out * dst_stride_y + z_out * dst_stride_z);
+    int4 offset = (int4)(dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z);
     offset      = min(offset + (int4)(0, 1, 2, 3) * (int4)dst_stride_z, (int4)dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding).
     int4 mult_y = min((int4)dst_size - offset, (int4)1);                                 // If out of bound, we don't want to increase dst_stride_y, so we set the multiplier to 0. It will be 1 otherwise.
 
     // Store the output tile
-    *(__global float *)(dst_ptr + mult_y.s0 * 0 * (int)dst_stride_y + offset.s0) = out_col0.s0;
-    *(__global float *)(dst_ptr + mult_y.s0 * 1 * (int)dst_stride_y + offset.s0) = out_col1.s0;
-    *(__global float *)(dst_ptr + mult_y.s0 * 2 * (int)dst_stride_y + offset.s0) = out_col2.s0;
-    *(__global float *)(dst_ptr + mult_y.s0 * 3 * (int)dst_stride_y + offset.s0) = out_col3.s0;
-    *(__global float *)(dst_ptr + mult_y.s1 * 0 * (int)dst_stride_y + offset.s1) = out_col0.s1;
-    *(__global float *)(dst_ptr + mult_y.s1 * 1 * (int)dst_stride_y + offset.s1) = out_col1.s1;
-    *(__global float *)(dst_ptr + mult_y.s1 * 2 * (int)dst_stride_y + offset.s1) = out_col2.s1;
-    *(__global float *)(dst_ptr + mult_y.s1 * 3 * (int)dst_stride_y + offset.s1) = out_col3.s1;
-    *(__global float *)(dst_ptr + mult_y.s2 * 0 * (int)dst_stride_y + offset.s2) = out_col0.s2;
-    *(__global float *)(dst_ptr + mult_y.s2 * 1 * (int)dst_stride_y + offset.s2) = out_col1.s2;
-    *(__global float *)(dst_ptr + mult_y.s2 * 2 * (int)dst_stride_y + offset.s2) = out_col2.s2;
-    *(__global float *)(dst_ptr + mult_y.s2 * 3 * (int)dst_stride_y + offset.s2) = out_col3.s2;
-    *(__global float *)(dst_ptr + mult_y.s3 * 0 * (int)dst_stride_y + offset.s3) = out_col0.s3;
-    *(__global float *)(dst_ptr + mult_y.s3 * 1 * (int)dst_stride_y + offset.s3) = out_col1.s3;
-    *(__global float *)(dst_ptr + mult_y.s3 * 2 * (int)dst_stride_y + offset.s3) = out_col2.s3;
-    *(__global float *)(dst_ptr + mult_y.s3 * 3 * (int)dst_stride_y + offset.s3) = out_col3.s3;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 0 * (int)dst_stride_y + offset.s0) = out_col0.s0;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 1 * (int)dst_stride_y + offset.s0) = out_col1.s0;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 2 * (int)dst_stride_y + offset.s0) = out_col2.s0;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 3 * (int)dst_stride_y + offset.s0) = out_col3.s0;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s1 * 0 * (int)dst_stride_y + offset.s1) = out_col0.s1;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s1 * 1 * (int)dst_stride_y + offset.s1) = out_col1.s1;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s1 * 2 * (int)dst_stride_y + offset.s1) = out_col2.s1;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s1 * 3 * (int)dst_stride_y + offset.s1) = out_col3.s1;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s2 * 0 * (int)dst_stride_y + offset.s2) = out_col0.s2;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s2 * 1 * (int)dst_stride_y + offset.s2) = out_col1.s2;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s2 * 2 * (int)dst_stride_y + offset.s2) = out_col2.s2;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s2 * 3 * (int)dst_stride_y + offset.s2) = out_col3.s2;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s3 * 0 * (int)dst_stride_y + offset.s3) = out_col0.s3;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s3 * 1 * (int)dst_stride_y + offset.s3) = out_col1.s3;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s3 * 2 * (int)dst_stride_y + offset.s3) = out_col2.s3;
+    *(__global DATA_TYPE *)(dst_ptr + mult_y.s3 * 3 * (int)dst_stride_y + offset.s3) = out_col3.s3;
 #endif // defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) || defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
 }
 
@@ -1020,8 +1037,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=2
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1079,8 +1097,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1138,8 +1157,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1197,8 +1217,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1254,8 +1275,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=4
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=1
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1313,8 +1335,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=2
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1372,8 +1395,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1431,8 +1455,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1490,8 +1515,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
@@ -1547,8 +1573,9 @@
  * @note The width of the output tile must be passed at compile time using -DOUTPUT_TILE_W: e.g. -DOUTPUT_TILE_W=1
  * @note The height of the output tile must be passed at compile time using -DOUTPUT_TILE_H: e.g. -DOUTPUT_TILE_H=4
  * @note -DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL has to be passed at compile time
+ * @note The data type must be passed at compile time using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types: float/half.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32/F16
  * @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)
diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
index 818638c..55cc465 100644
--- a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
@@ -26,6 +26,7 @@
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/IAccessWindow.h"
@@ -45,7 +46,8 @@
 {
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
 
     const Size2D kernel_size      = winograd_info.kernel_size;
     const Size2D output_tile_size = winograd_info.output_tile_size;
@@ -109,9 +111,9 @@
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(input->info()->dimension(2)));
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL");
     build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_FILTER_TRANSFORM_VERTICAL");
-
     const Size2D kernel_size      = winograd_info.kernel_size;
     const Size2D output_tile_size = winograd_info.output_tile_size;
 
diff --git a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
index c4e472a..0406731 100644
--- a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
@@ -26,6 +26,7 @@
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/OpenCL.h"
 #include "arm_compute/core/Error.h"
@@ -41,7 +42,8 @@
 {
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
 
     const PadStrideInfo conv_info        = winograd_info.convolution_info;
     const Size2D        output_tile_size = winograd_info.output_tile_size;
@@ -159,6 +161,7 @@
     build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
     build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
     build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL");
     build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_INPUT_TRANSFORM_VERTICAL");
 
diff --git a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
index fa42596..75988c6 100644
--- a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
@@ -26,6 +26,7 @@
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/IAccessWindow.h"
@@ -47,7 +48,8 @@
 {
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
 
     ARM_COMPUTE_RETURN_ERROR_ON(output->data_layout() != winograd_info.output_data_layout);
 
@@ -162,6 +164,7 @@
     build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles.width));
     build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
     build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL");
     build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL");