diff options
| author |  A. Unique TensorFlower <nobody@tensorflow.org> | 2016-05-26 16:27:02 -0800 |
|---|---|---|
| committer |  TensorFlower Gardener <gardener@tensorflow.org> | 2016-05-26 17:35:05 -0700 |
| commit | 57ab45bfc6f84b4cef1e225253d9b31191200a80 (patch) | |
| tree | d22dc978dcd9ad4b24d1814bdd509a2f01ace69c | |
| parent | 91f0b6c6a4d0f4f9259aac162b1fae5d4b24a0d1 (diff) | |
Adjust unit cost estimate for example_parsing_ops. The estimate was (mainly, I think) underestimating the cost of deserialization and dictionary lookup per dense key.
Improvements in wall time:
Run on rmlarsen3.mtv (12 X 3501 MHz CPUs); 2016-05-26T16:27:02.908401069-07:00
CPU: Intel Haswell with HyperThreading (6 cores) dL1:32KB dL2:256KB dL3:15MB
Benchmark Base (ns) New (ns) Improvement
------------------------------------------------------------------
BM_ParseDenseStringExample_128_10 555381 228527 +58.9%
BM_ParseDenseStringExample_128_100 5678223 1492853 +73.7%
BM_ParseDenseStringExample_128_1000 69209199 13411023 +80.6%
BM_ParseDenseStringExample_128_10000 285185983 270448314 +5.2%
Change: 123374081
| -rw-r--r-- | tensorflow/core/kernels/example_parsing_ops.cc | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/tensorflow/core/kernels/example_parsing_ops.cc b/tensorflow/core/kernels/example_parsing_ops.cc index 2dcf079dbf..8b56cae4ec 100644 --- a/tensorflow/core/kernels/example_parsing_ops.cc +++ b/tensorflow/core/kernels/example_parsing_ops.cc @@ -183,9 +183,9 @@ class ExampleParserOp : public OpKernel { auto worker_threads = *(ctx->device()->tensorflow_cpu_worker_threads()); // Estimate the cost of parsing each batch element. - int64 work_unit_size = 100 + 100 * num_sparse_; + int64 work_unit_size = 1000 + 100 * num_sparse_; for (int d = 0; d < num_dense_; ++d) { - work_unit_size += dense_shapes_[d].num_elements(); + work_unit_size += 100 + dense_shapes_[d].num_elements(); } mutex mu; |