Jeffrey Mak

Supports both NCHW and NHWC formats. CUDA also provides kernels for small images, however the corresponding SYCL kernels are not written as the file is getting quite long.

Supports both NHWC and NCHW formats, although it depends on Conv2D when depth is 1.

The failure has to do with SYCL having a lower float precision than CPU.

File: depthwise_conv_grad_op.cc This operation is performed with one atomic add operation in CUDA, so it cannot be added in SYCL for now.

File: resize_nearest_neighbor_op.cc This operation is performed with one atomic add operation in CUDA, so it cannot be added in SYCL for now.

file: resize_bilinear_op.cc This operation is performed with four atomic add's in CUDA, so it cannot be added in SYCL for now.