OpenCL Improvements (#7596)

* OpenCL improvements

Added Tile, Transpose and Range Ops double support for SYCL device.
Moved gpu_device_name() to test_util.py so now it can be used in force_gpu to pull either GPU or SYCL depending on what is available in the system.

* Improvements to the SYCL device support

 - Registration of Type Traits required for stride slice op
 - Registration of ConcatOffset, _ListToArray, _ArrayToList
   Pad, Reverse ( CPU ), ReverseV2 ( CPU ), Size, ExpandDims,
   Squeeze, StridedSlice, StridedSliceGrad, StridedSliceAssign,
   TileGrad, InvertPermutation, Transpose
 - Registration of Sycl kernels only for essential data types
 - Floor_div_real has been disabled for SYCL device
 - Device in control_flow_ops_py_test.py needed to be lower cased

* SYCL support improvements (#31)

* Improvements to the SYCL device support

This commit reduces number of failing tests when TensorFlow compiles
for OpenCL support.

 - Registration of Type Traits required for stride slice op
 - Registration of ConcatOffset, _ListToArray, _ArrayToList
   Pad, Reverse ( CPU ), ReverseV2 ( CPU ), Size, ExpandDims,
   Squeeze, StridedSlice, StridedSliceGrad, StridedSliceAssign,
   TileGrad, InvertPermutation, Transpose
 - Registration of Sycl kernels only for essential data types
 - Floor_div_real has been disabled for SYCL device
 - Device in control_flow_ops_py_test.py needed to be lower cased

* Fixes & Version bump (#33)

* Fix Unbuntu typo. (#38)

unbuntu -> ubuntu

* Add problem descriptions and solutions (#35)

* Add ComputeCpp lib folder to LD_LIBRARY_PATH

* Add ImportError problem + solution

If you get the error message "ImportError: libComputeCpp.so: cannot open shared
object file: No such file or directory", make sure you have added the
path to ComputeCpp's lib folder to your `LD_LIBRARY_PATH`.

* Add another ImportError problem + solution

If you get the error message "ImportError: cannot import name
'pywrap_tensorflow'" you may be standing in the TensorFlow directory.

* Improvements to the SYCL device support

* Registers FloorDiv, FloorMod and SoftMax Ops for SYCL device

* Workaround for 0 bytes allocation for SYCL device (#42)

* Sycl improvements (#44)

- Eigen version bump
 - Extends Cast and Cwise ops benchmark to cover Sycl device
 - Extends device_lib_test.py to cover Sycl device
 - Registers int32, string and ResourceHandler to run on host for
   Enter and RefEnter Sycl Ops
 - Enables RecudeMax op for Sycl since Eigen implementation is ready
 - Registers Less op for Sycl device

* Improved the formatting of the SYCL code

* Fixed compilation error.

* Made sure that using test sessions with force_gpu=True forces the
placement on a gpu device even if none is detected.

1 files changed, 15 insertions, 0 deletions

diff --git a/tensorflow/core/kernels/cwise_op_maximum.cc b/tensorflow/core/kernels/cwise_op_maximum.cc
index f93b5a8303..7311f25ec0 100644
--- a/tensorflow/core/kernels/cwise_op_maximum.cc
+++ b/tensorflow/core/kernels/cwise_op_maximum.cc
@@ -34,4 +34,19 @@ REGISTER_KERNEL_BUILDER(Name("Maximum")
                         BinaryOp<CPUDevice, functor::maximum<int32>>);
 #endif
 
+#ifdef TENSORFLOW_USE_SYCL
+REGISTER(BinaryOp, SYCL, "Maximum", functor::maximum, float);
+
+// A special GPU kernel for int32.
+// TODO(b/25387198): Also enable int32 in device memory. This kernel
+// registration requires all int32 inputs and outputs to be in host memory.
+REGISTER_KERNEL_BUILDER(Name("Maximum")
+                            .Device(DEVICE_SYCL)
+                            .HostMemory("x")
+                            .HostMemory("y")
+                            .HostMemory("z")
+                            .TypeConstraint<int32>("T"),
+                        BinaryOp<CPUDevice, functor::maximum<int32>>);
+#endif // TENSORFLOW_USE_SYCL
+
 }  // namespace tensorflow