fp16 support for BiasAdd. Includes support for atomic adds for Eigen::half,

although beware, they are going to be very slow.

Also:

  - Remove the testGradientBias() test, since it was a fp64-only duplication of
    tests we already had in other subtests.
  - Extend gradient microbenchmarks to measure NCHW and NHWC, not just the
    default in a two-dimensional tensor.
  - Fix cuda_builtin::__ldg() definition; seemingly calling ::__ldg() on gcudacc
    returns only zero, not the __ldg() function we defined a few lines further
    up.
Change: 121970776

0 files changed, 0 insertions, 0 deletions