diff options
author | Eugene Brevdo <ebrevdo@google.com> | 2017-06-29 15:33:13 -0700 |
---|---|---|
committer | TensorFlower Gardener <gardener@tensorflow.org> | 2017-06-29 15:37:15 -0700 |
commit | 8280e0ae9083a65b23608b34723f07e028a56dc8 (patch) | |
tree | 0f2df282cfd5cd712920e440cea88a093668cbf2 /third_party/cub.BUILD | |
parent | 4aa7c4d2330ce110b5be348144ee67143841272c (diff) |
GPU-enabled WhereOp using CUB.
* Import CUB.
* Add GPU-enabled async WhereOp.
* Added benchmarks.
* Added support for bool ResourceVariables on GPU.
Benchmark results on machine with single K40 tesla GPU:
Where on bool matrix shape [m x n] with p percentage values true below.
For small-medium sizes, running WhereOp on GPU is ~4-2x slower. For
realistic large problem sizes, it's 2-5x faster. This timing ignores
the time spent copying a tensor from GPU -> CPU and back from CPU -> GPU
when the WhereOp is between GPU computations (so the performance impact
should actually be better).
Benchmark: m_10_n_10_p_0.01_use_gpu_False wall_time: 9.01e-05 s Throughput: 0.00129 GB/s
Benchmark: m_10_n_10_p_0.01_use_gpu_True wall_time: 0.000187 s Throughput: 0.000621 GB/s
Benchmark: m_10_n_10_p_0.5_use_gpu_False wall_time: 9.3e-05 s Throughput: 0.00968 GB/s
Benchmark: m_10_n_10_p_0.5_use_gpu_True wall_time: 0.000252 s Throughput: 0.00357 GB/s
Benchmark: m_10_n_10_p_0.99_use_gpu_False wall_time: 0.000152 s Throughput: 0.0111 GB/s
Benchmark: m_10_n_10_p_0.99_use_gpu_True wall_time: 0.000245 s Throughput: 0.00687 GB/s
Benchmark: m_10_n_100_p_0.01_use_gpu_False wall_time: 9.3e-05 s Throughput: 0.0125 GB/s
Benchmark: m_10_n_100_p_0.01_use_gpu_True wall_time: 0.000253 s Throughput: 0.00458 GB/s
Benchmark: m_10_n_100_p_0.5_use_gpu_False wall_time: 9.8e-05 s Throughput: 0.0918 GB/s
Benchmark: m_10_n_100_p_0.5_use_gpu_True wall_time: 0.00026 s Throughput: 0.0346 GB/s
Benchmark: m_10_n_100_p_0.99_use_gpu_False wall_time: 0.000104 s Throughput: 0.162 GB/s
Benchmark: m_10_n_100_p_0.99_use_gpu_True wall_time: 0.000288 s Throughput: 0.0586 GB/s
Benchmark: m_10_n_1000_p_0.01_use_gpu_False wall_time: 0.000105 s Throughput: 0.111 GB/s
Benchmark: m_10_n_1000_p_0.01_use_gpu_True wall_time: 0.000283 s Throughput: 0.041 GB/s
Benchmark: m_10_n_1000_p_0.5_use_gpu_False wall_time: 0.000185 s Throughput: 0.486 GB/s
Benchmark: m_10_n_1000_p_0.5_use_gpu_True wall_time: 0.000335 s Throughput: 0.269 GB/s
Benchmark: m_10_n_1000_p_0.99_use_gpu_False wall_time: 0.000203 s Throughput: 0.83 GB/s
Benchmark: m_10_n_1000_p_0.99_use_gpu_True wall_time: 0.000346 s Throughput: 0.486 GB/s
Benchmark: m_10_n_10000_p_0.01_use_gpu_False wall_time: 0.00019 s Throughput: 0.609 GB/s
Benchmark: m_10_n_10000_p_0.01_use_gpu_True wall_time: 0.00028 s Throughput: 0.414 GB/s
Benchmark: m_10_n_10000_p_0.5_use_gpu_False wall_time: 0.00117 s Throughput: 0.771 GB/s
Benchmark: m_10_n_10000_p_0.5_use_gpu_True wall_time: 0.000426 s Throughput: 2.11 GB/s
Benchmark: m_10_n_10000_p_0.99_use_gpu_False wall_time: 0.0014 s Throughput: 1.2 GB/s
Benchmark: m_10_n_10000_p_0.99_use_gpu_True wall_time: 0.000482 s Throughput: 3.5 GB/s
Benchmark: m_10_n_100000_p_0.01_use_gpu_False wall_time: 0.00129 s Throughput: 0.899 GB/s
Benchmark: m_10_n_100000_p_0.01_use_gpu_True wall_time: 0.000336 s Throughput: 3.45 GB/s
Benchmark: m_10_n_100000_p_0.5_use_gpu_False wall_time: 0.0102 s Throughput: 0.885 GB/s
Benchmark: m_10_n_100000_p_0.5_use_gpu_True wall_time: 0.00136 s Throughput: 6.6 GB/s
Benchmark: m_10_n_100000_p_0.99_use_gpu_False wall_time: 0.0116 s Throughput: 1.45 GB/s
Benchmark: m_10_n_100000_p_0.99_use_gpu_True wall_time: 0.00233 s Throughput: 7.23 GB/s
Benchmark: m_10_n_1000000_p_0.01_use_gpu_False wall_time: 0.0111 s Throughput: 1.04 GB/s
Benchmark: m_10_n_1000000_p_0.01_use_gpu_True wall_time: 0.00109 s Throughput: 10.6 GB/s
Benchmark: m_10_n_1000000_p_0.5_use_gpu_False wall_time: 0.0895 s Throughput: 1.01 GB/s
Benchmark: m_10_n_1000000_p_0.5_use_gpu_True wall_time: 0.0103 s Throughput: 8.7 GB/s
Benchmark: m_10_n_1000000_p_0.99_use_gpu_False wall_time: 0.107 s Throughput: 1.58 GB/s
Benchmark: m_10_n_1000000_p_0.99_use_gpu_True wall_time: 0.0201 s Throughput: 8.39 GB/s
PiperOrigin-RevId: 160582709
Diffstat (limited to 'third_party/cub.BUILD')
-rw-r--r-- | third_party/cub.BUILD | 26 |
1 files changed, 26 insertions, 0 deletions
diff --git a/third_party/cub.BUILD b/third_party/cub.BUILD new file mode 100644 index 0000000000..29159c9dad --- /dev/null +++ b/third_party/cub.BUILD @@ -0,0 +1,26 @@ +# Description: CUB library which is a set of primitives for GPU programming. + +package( + default_visibility = ["//visibility:public"], +) + +licenses(["notice"]) # BSD + +exports_files(["LICENSE.TXT"]) + +load("@local_config_cuda//cuda:build_defs.bzl", "cuda_default_copts", "if_cuda") + +filegroup( + name = "cub_header_files", + srcs = glob([ + "cub/**", + ]), +) + +cc_library( + name = "cub", + hdrs = if_cuda([":cub_header_files"]), + deps = [ + "@local_config_cuda//cuda:cuda_headers", + ], +) |