Permalink
Browse files

Roll-forward of "Local Response Normalization GPU support via Stream …

…Executor."

Move AsDeviceMemory function into a StreamExecutorUtil class, not a GPUUtil one, since it's independent of GPUs.  Make lrn_op use the new version of that function.
Change: 127289319
  • Loading branch information...
rryan authored and tensorflower-gardener committed Jul 13, 2016
1 parent 8e69f98 commit 35df3ed43edabbc4ad1b2439bbc7de8917026d6e
View
@@ -249,6 +249,7 @@ tf_cuda_library(
"util/sparse/group_iterator.h",
"util/sparse/sparse_tensor.h",
"util/stat_summarizer.h",
+ "util/stream_executor_util.h",
"util/tensor_format.h",
"util/tensor_slice_reader.h",
"util/tensor_slice_reader_cache.h",
@@ -68,6 +68,9 @@ class GPUUtil {
// Map a Tensor as a DeviceMemory object wrapping the given typed
// buffer.
+ //
+ // NOTE: will be removed soon, see StreamExecutorUtil::AsDeviceMemory
+ // instead.
template <typename T>
static perftools::gputools::DeviceMemory<T> AsDeviceMemory(const Tensor& t) {
T* ptr = reinterpret_cast<T*>(const_cast<void*>(DMAHelper::base(&t)));
@@ -1332,7 +1332,7 @@ tf_kernel_libraries(
],
)
-tf_cc_test(
+tf_cuda_cc_test(
name = "lrn_op_test",
deps = [
":nn",
Oops, something went wrong.

0 comments on commit 35df3ed

Please sign in to comment.