963e564d91f6544c7a49ac80a618729eb7b6a36e |
|
10-Aug-2017 |
Gao, Xiang <qasdfgtyuiop@gmail.com> |
Ops and kernels for reduce_slice_ops (#9063) * init impl done * improve doc, shape inference * add and pass shape inference * check bound * small fix * add all types of reduction * add cpu kernels * write python tests * make it able to build independent on GPU * fix problems and pass all tests * add to py * remove redundant code * improve doc * change ops signature * add Cuda{2D,3D}LaunchConfig that max occupancy * support axis * remove default val, check input<=0 * modify tests to net api * fix some compilation err * fix Const in flat_inner_outer_dims * pass build * fix shape test * fix names in macros * fix test * specify reduceop by macro * clean code * further simplify code * misc fixes * add max size check * fix typo * fix typo * tests, docs, and related changes * build the test * buildify * partially support vec indices * vec indices finish, not tested * pass cpu tests * fix gpu functor * fix code style * update doc * fix code style * fix code style * sync cuda_kernel_config change * buildify * typename->class in template template typename in template template (See N4051: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4051.html) is supported in gcc starting at 5(See: https://gcc.gnu.org/projects/cxx-status.html). So using typename instead of class will cause the compilation fails at gcc-4. * move to contrib * some fixes * some fixes * some fixes * buildify * build cleanup * add cmake
/external/tensorflow/tensorflow/contrib/reduce_slice_ops/ops/reduce_slice_ops_test.cc
|