experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Fix Mac OSX build linker error.

Browse Source 
ImportError: dlopen(/Volumes/BuildData/tmpfs/tmp/bazel/d1b2600cd78e76a92812a06683f5de10/execroot/org_tensorflow/bazel-out/host/bin/tensorflow/python/keras/api/create_tensorflow.python_api_1_keras_python_api_gen.runfiles/org_tensorflow/tensorflow/python/_pywrap_tensorflow_internal.so, 6): Symbol not found: __ZN10tensorflow14CSRMatMulCPUOpINSt3__17complexIdEEE10kMaxShardsE
PiperOrigin-RevId: 272567253
This commit is contained in:
Anudhyan Boral 2019-10-02 18:42:09 -07:00 committed by TensorFlower Gardener
parent 44c25c8870
commit b9d938352a
1 changed files with 14 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
tensorflow/core/kernels/sparse/mat_mul_op.cc
View File

@ -43,6 +43,20 @@ limitations under the License.
namespace tensorflow {
// TODO(anudhyan): These constants may be tuned based on the performance of
// 'benchmark_sparse_matrix_mat_vec_mul'. We would like to find constants
// which work across hardware platforms for typical matrix sizes. It should be
// possible to observe at least 30-50% improvement as we increase the number
// of threads by 1. If not, then it may we worth increasing kMaxShards and
// kNumShardsPerThread. However, once we have too many shards, latency may be
// dominated by per-shard overhead.
//
// Maximum number of shards into which to divide the computation for each CSR
// Sparse Matrix instance.
static constexpr int32 kMaxShards = 20;
// Number of shards allocated to each thread.
static constexpr int32 kNumShardsPerThread = 3;
typedef Eigen::ThreadPoolDevice CPUDevice;
typedef Eigen::GpuDevice GPUDevice;
@ -147,16 +161,6 @@ class CSRMatMulCPUOp : public CSRMatMulOp<CPUDevice, T> {
using MatrixMap = Eigen::Map<Matrix>;
public:
// TODO(anudhyan): These constants may be tuned based on the performance of
// 'benchmark_sparse_matrix_mat_vec_mul'. We would like to find constants
// which work across hardware platforms for typical matrix sizes. It should be
// possible to observe at least 30-50% improvement as we increase the number
// of threads by 1. If not, then it may we worth increasing kMaxShards and
// kNumShardsPerThread. However, once we have too many shards, latency may be
// dominated by per-shard overhead.
static constexpr int32 kMaxShards = 20;
static constexpr int32 kNumShardsPerThread = 3;
explicit CSRMatMulCPUOp(OpKernelConstruction* c)
: CSRMatMulOp<CPUDevice, T>(c) {}