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

1 files changed, 6 insertions, 2 deletions

diff --git a/tensorflow/core/platform/test_benchmark.h b/tensorflow/core/platform/test_benchmark.h
index 4c90cf7f85..3536548f00 100644
--- a/tensorflow/core/platform/test_benchmark.h
+++ b/tensorflow/core/platform/test_benchmark.h
@@ -17,6 +17,7 @@ limitations under the License.
 #ifndef TENSORFLOW_PLATFORM_TEST_BENCHMARK_H_
 #define TENSORFLOW_PLATFORM_TEST_BENCHMARK_H_
 
+#include <utility>
 #include <vector>
 #include "tensorflow/core/platform/macros.h"
 #include "tensorflow/core/platform/platform.h"
@@ -48,20 +49,23 @@ class Benchmark {
  public:
   Benchmark(const char* name, void (*fn)(int));
   Benchmark(const char* name, void (*fn)(int, int));
+  Benchmark(const char* name, void (*fn)(int, int, int));
 
   Benchmark* Arg(int x);
+  Benchmark* ArgPair(int x, int y);
   Benchmark* Range(int lo, int hi);
   static void Run(const char* pattern);
 
  private:
   string name_;
   int num_args_;
-  std::vector<int> args_;
+  std::vector<std::pair<int, int>> args_;
   void (*fn0_)(int) = nullptr;
   void (*fn1_)(int, int) = nullptr;
+  void (*fn2_)(int, int, int) = nullptr;
 
   void Register();
-  void Run(int arg, int* run_count, double* run_seconds);
+  void Run(int arg1, int arg2, int* run_count, double* run_seconds);
 };
 #endif