diff options
| author |  A. Unique TensorFlower <nobody@tensorflow.org> | 2016-03-11 09:08:22 -0800 |
|---|---|---|
| committer |  TensorFlower Gardener <gardener@tensorflow.org> | 2016-03-11 11:42:13 -0800 |
| commit | fc53648e0c0f5110bfab75a02b9dc75260b913d3 (patch) | |
| tree | caf4cf00eb9f29a22958461feb58a6628ccb94a7 /tensorflow/core/kernels/resize_bicubic_op.cc | |
| parent | 625d62abef3d64ab9025cf248719b76aea3b7fc9 (diff) | |
Use faster floating point rounding and floor primitives in a couple of image processing kernels. Specifically
1) round(x) -> lrintf(x)
2) int i = static_cast<int>(floor(x)) -> int i = x; for x >=0.
1) lrintf() is much faster than round() as it compiles down to a single instruction. Thanks to sesse@ for the tip.
2) Casting float to int is truncating.
This change speeds up bicubic resizer by ~7%. There is no benchmark for simple_distorted_bounding_box, but I verified that the change does not cause any regression on imagenet.
Benchmark Base (ns) New (ns) Improvement
------------------------------------------------------------------
BM_ResizeBicubic_8_32_3 111319 103691 +6.9%
BM_ResizeBicubic_8_128_3 1513173 1415275 +6.5%
BM_ResizeBicubic_8_512_3 23232055 21353694 +8.1%
BM_ResizeBicubic_8_1024_3 91817334 86537924 +5.7%
BM_ResizeBicubic_16_32_3 199235 184428 +7.4%
BM_ResizeBicubic_16_128_3 2946248 2744626 +6.8%
BM_ResizeBicubic_16_512_3 45957878 42832137 +6.8%
BM_ResizeBicubic_16_1024_3 184548577 170523376 +7.6%
BM_ResizeBicubic_32_32_3 432991 400042 +7.6%
BM_ResizeBicubic_32_128_3 5767107 5395839 +6.4%
BM_ResizeBicubic_32_512_3 94064648 86132128 +8.4%
BM_ResizeBicubic_32_1024_3 370615017 341913896 +7.7%
Change: 116978161
Diffstat (limited to 'tensorflow/core/kernels/resize_bicubic_op.cc')
| -rw-r--r-- | tensorflow/core/kernels/resize_bicubic_op.cc | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/tensorflow/core/kernels/resize_bicubic_op.cc b/tensorflow/core/kernels/resize_bicubic_op.cc index 370786f453..f81383984b 100644 --- a/tensorflow/core/kernels/resize_bicubic_op.cc +++ b/tensorflow/core/kernels/resize_bicubic_op.cc @@ -62,9 +62,9 @@ inline int64 Bound(int64 val, int64 limit) { inline void GetWeightsAndIndices(float scale, int64 out_loc, int64 limit, std::array<float, 4>* weights, std::array<int64, 4>* indices) { - const int64 in_loc = floor(scale * out_loc); + const int64 in_loc = scale * out_loc; const float delta = scale * out_loc - in_loc; - const int64 offset = round(delta * kTableSize); + const int64 offset = lrintf(delta * kTableSize); const float* coeffs_tab = GetCoeffsTable(); *weights = {{coeffs_tab[offset * 2 + 1], coeffs_tab[offset * 2], coeffs_tab[(kTableSize - offset) * 2], |