Closed jwlawson closed 7 years ago
Atomic support in SYCL is not designed in a way that plays nicely with Tensorflow and Eigen. Here we provide a new implementation for MaxPoolGrad and MaxPool3DGrad which does not rely on atomics, and so avoids any such problems.
Atomic support in SYCL is not designed in a way that plays nicely with Tensorflow and Eigen. Here we provide a new implementation for MaxPoolGrad and MaxPool3DGrad which does not rely on atomics, and so avoids any such problems.