[ROCm] Fix for the broken ROCm CSB.

[deven-amd]

The following commit breaks the --config=rocm build

f72695e

The above commit adds a couple of subtests that require support for the StatefulUnirformFullInt Op on the GPU. Currently ROCm does not support that Op on the GPU, which leads to those subtests failing.

The "fix" is to skip those subtests on the ROCm platform.

---

/cc @chsigg @whchung

[deven-amd]

@chsigg @tanzhenyu gentle ping

[tanzhenyu]

I don't understand why is this needed -- where is StatefulUniformFullInt used? I think the layer was only using stateless random ops.

[deven-amd]

The error message we get for the failure indicates that StatefulUniformFullInt is getting instantiated in the TF graph.

tensorflow.python.framework.errors_impl.NotFoundError: No registered 'StatefulUniformFullInt' OpKernel for 'GPU' devices compatible with node {{node StatefulUniformFullInt}}                                                                                                                                                                                                                                                        
        .  Registered:  device='XLA_GPU'; shape_dtype in [DT_FLOAT, DT_DOUBLE, DT_INT32, DT_UINT8, DT_INT16, ..., DT_UINT16, DT_COMPLEX128, DT_HALF, DT_UINT32, DT_UINT64]; dtype in [DT_INT32, DT_INT64, DT_UINT32, DT_UINT64]                                                                                                                                                                                                      
  device='XLA_CPU'; shape_dtype in [DT_FLOAT, DT_DOUBLE, DT_INT32, DT_UINT8, DT_INT16, ..., DT_UINT16, DT_COMPLEX128, DT_HALF, DT_UINT32, DT_UINT64]; dtype in [DT_INT32, DT_INT64, DT_UINT32, DT_UINT64]                                                                                                                                                                                                                            
  device='XLA_CPU_JIT'; shape_dtype in [DT_FLOAT, DT_DOUBLE, DT_INT32, DT_UINT8, DT_INT16, ..., DT_UINT16, DT_COMPLEX128, DT_HALF, DT_UINT32, DT_UINT64]; dtype in [DT_INT32, DT_INT64, DT_UINT32, DT_UINT64]                                                                                                                                                                                                                        
  device='CPU'; dtype in [DT_UINT64]                                                                                                                                                                                                                                                                                                                                                                                                 
  device='CPU'; dtype in [DT_UINT32]                                                                                                                                                                                                                                                                                                                                                                                                 
  device='CPU'; dtype in [DT_INT64]                                                                                                                                                                                                                                                                                                                                                                                                  
  device='CPU'; dtype in [DT_INT32]                                                                                                                                                                                                                                                                                                                                                                                                  
  device='XLA_GPU_JIT'; shape_dtype in [DT_FLOAT, DT_DOUBLE, DT_INT32, DT_UINT8, DT_INT16, ..., DT_UINT16, DT_COMPLEX128, DT_HALF, DT_UINT32, DT_UINT64]; dtype in [DT_INT32, DT_INT64, DT_UINT32, DT_UINT64]                                                                                                                                                                                                                        
 [Op:StatefulUniformFullInt] name: model_1/random_crop/stateful_uniform_full_int/                                                                                                                                                                                                                                                                                                                                                    

[deven-amd]

@chsigg @tanzhenyu gentle ping

[deven-amd]

@gbaned , anything we can do to help get this PR merged?