Rollback of https://github.com/tensorflow/tensorflow/commit/8649852b75ed43fd62e1429086e4a8f5dd6d38ee.
The commit slowed down testGradientTensor4D and broke the Macos nightly build.
PiperOrigin-RevId: 396901985
Change-Id: I7d06790b814fc6dba9c32b6403dace4f63972fc5
diff --git a/tensorflow/core/kernels/redux_functor.h b/tensorflow/core/kernels/redux_functor.h
index 6681bec..e07fa53 100644
--- a/tensorflow/core/kernels/redux_functor.h
+++ b/tensorflow/core/kernels/redux_functor.h
@@ -230,6 +230,11 @@
input.template flat<InputT>().template cast<OutputT>().reshape(
output_dims);
return;
+ } else if (1 == inner_dim) {
+ // Equivalent to ReduceOuterDimensions.
+ const ReduceOuterDimensions<InputT, AccumT, OutputT, BinaryFunctor> redux;
+ redux(device, input_dims, input, output);
+ return;
}
// Compute block size along the outer dimension for efficiency.
diff --git a/tensorflow/python/kernel_tests/bias_op_base.py b/tensorflow/python/kernel_tests/bias_op_base.py
index b629123..a3b2ad5 100644
--- a/tensorflow/python/kernel_tests/bias_op_base.py
+++ b/tensorflow/python/kernel_tests/bias_op_base.py
@@ -254,7 +254,7 @@
self._testGradient(np_input, bias, dtype, data_format, use_gpu)
def testGradientTensor4D(self):
- for (data_format, use_gpu) in [("NHWC", False), ("NCHW", False)]:
+ for (data_format, use_gpu) in [("NHWC", False)]:
for dtype in (dtypes.float16, dtypes.float32, dtypes.float64):
np_input = np.arange(
1.0, 49.0,
@@ -273,13 +273,6 @@
self._testGradient(np_input,
np.random.rand(64).astype(dtype.as_numpy_dtype),
dtype, data_format, use_gpu)
- np_input = np.arange(
- 1.0, 129.0,
- dtype=dtype.as_numpy_dtype).reshape([4, 1, 1,
- 32]).astype(np.float32)
- self._testGradient(np_input,
- np.random.rand(32).astype(dtype.as_numpy_dtype),
- dtype, data_format, use_gpu)
def testGradientTensor5D(self):
for (data_format, use_gpu) in [("NHWC", False), ("NHWC", True),
