зеркало из https://github.com/microsoft/LightGBM.git
[ci] prevent trailing whitespace, ensure files end with newline (#6373)
This commit is contained in:
James Lamb
GitHub
Родитель
6a1ec44437
Коммит
631e0a2a7b
|
|
@ -56,7 +56,7 @@ cp --remove-destination /usr/lib/llvm-${CLANG_VERSION}/bin/* /usr/bin/
|
|||
# per https://www.stats.ox.ac.uk/pub/bdr/Rconfig/r-devel-linux-x86_64-fedora-clang
|
||||
#
|
||||
# clang was built to use libc++: for a version built to default to libstdc++
|
||||
# (as shipped by Fedora/Debian/Ubuntu), add -stdlib=libc++ to CXX
|
||||
# (as shipped by Fedora/Debian/Ubuntu), add -stdlib=libc++ to CXX
|
||||
# and install the libcxx-devel/libc++-dev package.
|
||||
mkdir -p "${HOME}/.R"
|
||||
|
||||
|
|
|
|||
|
|
@ -1,3 +1,4 @@
|
|||
# exclude files which are auto-generated by build tools
|
||||
exclude: |
|
||||
(?x)^(
|
||||
build|
|
||||
|
|
@ -5,8 +6,17 @@ exclude: |
|
|||
lightgbm-python|
|
||||
lightgbm_r|
|
||||
)$
|
||||
|R-package/configure$
|
||||
|R-package/inst/Makevars$
|
||||
|R-package/inst/Makevars.win$
|
||||
|R-package/man/.*Rd$
|
||||
|
||||
repos:
|
||||
- repo: https://github.com/pre-commit/pre-commit-hooks
|
||||
rev: v4.5.0
|
||||
hooks:
|
||||
- id: end-of-file-fixer
|
||||
- id: trailing-whitespace
|
||||
- repo: https://github.com/pycqa/isort
|
||||
rev: 5.13.2
|
||||
hooks:
|
||||
|
|
|
|||
|
|
@ -1,2 +1,2 @@
|
|||
YEAR: 2016
|
||||
COPYRIGHT HOLDER: Microsoft Corporation
|
||||
COPYRIGHT HOLDER: Microsoft Corporation
|
||||
|
|
|
|||
|
|
@ -14,7 +14,7 @@ warning was not fixed within 14 days.
|
|||
```text
|
||||
/usr/local/clang-trunk/bin/../include/c++/v1/__fwd/string_view.h:22:41:
|
||||
warning: 'char_traits<fmt::detail::char8_type>' is deprecated:
|
||||
char_traits<T> for T not equal to char, wchar_t, char8_t, char16_t or char32_t is non-standard and is provided for a temporary period.
|
||||
char_traits<T> for T not equal to char, wchar_t, char8_t, char16_t or char32_t is non-standard and is provided for a temporary period.
|
||||
It will be removed in LLVM 19, so please migrate off of it. [-Wdeprecated-declarations]
|
||||
```
|
||||
|
||||
|
|
|
|||
|
|
@ -14,7 +14,7 @@ Instead, please report them to the Microsoft Security Response Center (MSRC) at
|
|||
|
||||
If you prefer to submit without logging in, send email to [secure@microsoft.com](mailto:secure@microsoft.com). If possible, encrypt your message with our PGP key; please download it from the [Microsoft Security Response Center PGP Key page](https://aka.ms/opensource/security/pgpkey).
|
||||
|
||||
You should receive a response within 24 hours. If for some reason you do not, please follow up via email to ensure we received your original message. Additional information can be found at [microsoft.com/msrc](https://aka.ms/opensource/security/msrc).
|
||||
You should receive a response within 24 hours. If for some reason you do not, please follow up via email to ensure we received your original message. Additional information can be found at [microsoft.com/msrc](https://aka.ms/opensource/security/msrc).
|
||||
|
||||
Please include the requested information listed below (as much as you can provide) to help us better understand the nature and scope of the possible issue:
|
||||
|
||||
|
|
|
|||
|
|
@ -4,7 +4,7 @@
|
|||
# Prepare a source distribution of the R package
|
||||
# to be submitted to CRAN.
|
||||
#
|
||||
# [arguments]
|
||||
# [arguments]
|
||||
#
|
||||
# --r-executable Customize the R executable used by `R CMD build`.
|
||||
# Useful if building the R package in an environment with
|
||||
|
|
|
|||
|
|
@ -113,8 +113,8 @@ Unlike a categorical feature, however, ``positions`` are used to adjust the targ
|
|||
The position file corresponds with training data file line by line, and has one position per line. And if the name of training data file is ``train.txt``, the position file should be named as ``train.txt.position`` and placed in the same folder as the data file.
|
||||
In this case, LightGBM will load the position file automatically if it exists. The positions can also be specified through the ``Dataset`` constructor when using Python API. If the positions are specified in both approaches, the ``.position`` file will be ignored.
|
||||
|
||||
Currently, implemented is an approach to model position bias by using an idea of Generalized Additive Models (`GAM <https://en.wikipedia.org/wiki/Generalized_additive_model>`_) to linearly decompose the document score ``s`` into the sum of a relevance component ``f`` and a positional component ``g``: ``s(x, pos) = f(x) + g(pos)`` where the former component depends on the original query-document features and the latter depends on the position of an item.
|
||||
During the training, the compound scoring function ``s(x, pos)`` is fit with a standard ranking algorithm (e.g., LambdaMART) which boils down to jointly learning the relevance component ``f(x)`` (it is later returned as an unbiased model) and the position factors ``g(pos)`` that help better explain the observed (biased) labels.
|
||||
Similar score decomposition ideas have previously been applied for classification & pointwise ranking tasks with assumptions of binary labels and binary relevance (a.k.a. "two-tower" models, refer to the papers: `Towards Disentangling Relevance and Bias in Unbiased Learning to Rank <https://arxiv.org/abs/2212.13937>`_, `PAL: a position-bias aware learning framework for CTR prediction in live recommender systems <https://dl.acm.org/doi/10.1145/3298689.3347033>`_, `A General Framework for Debiasing in CTR Prediction <https://arxiv.org/abs/2112.02767>`_).
|
||||
In LightGBM, we adapt this idea to general pairwise Lerarning-to-Rank with arbitrary ordinal relevance labels.
|
||||
Currently, implemented is an approach to model position bias by using an idea of Generalized Additive Models (`GAM <https://en.wikipedia.org/wiki/Generalized_additive_model>`_) to linearly decompose the document score ``s`` into the sum of a relevance component ``f`` and a positional component ``g``: ``s(x, pos) = f(x) + g(pos)`` where the former component depends on the original query-document features and the latter depends on the position of an item.
|
||||
During the training, the compound scoring function ``s(x, pos)`` is fit with a standard ranking algorithm (e.g., LambdaMART) which boils down to jointly learning the relevance component ``f(x)`` (it is later returned as an unbiased model) and the position factors ``g(pos)`` that help better explain the observed (biased) labels.
|
||||
Similar score decomposition ideas have previously been applied for classification & pointwise ranking tasks with assumptions of binary labels and binary relevance (a.k.a. "two-tower" models, refer to the papers: `Towards Disentangling Relevance and Bias in Unbiased Learning to Rank <https://arxiv.org/abs/2212.13937>`_, `PAL: a position-bias aware learning framework for CTR prediction in live recommender systems <https://dl.acm.org/doi/10.1145/3298689.3347033>`_, `A General Framework for Debiasing in CTR Prediction <https://arxiv.org/abs/2112.02767>`_).
|
||||
In LightGBM, we adapt this idea to general pairwise Lerarning-to-Rank with arbitrary ordinal relevance labels.
|
||||
Besides, GAMs have been used in the context of explainable ML (`Accurate Intelligible Models with Pairwise Interactions <https://www.cs.cornell.edu/~yinlou/papers/lou-kdd13.pdf>`_) to linearly decompose the contribution of each feature (and possibly their pairwise interactions) to the overall score, for subsequent analysis and interpretation of their effects in the trained models.
|
||||
|
|
|
|||
|
|
@ -21,7 +21,7 @@ LightGBM uses histogram-based algorithms\ `[4, 5, 6] <#references>`__, which buc
|
|||
- To get one leaf's histograms in a binary tree, use the histogram subtraction of its parent and its neighbor
|
||||
|
||||
- So it needs to construct histograms for only one leaf (with smaller ``#data`` than its neighbor). It then can get histograms of its neighbor by histogram subtraction with small cost (``O(#bins)``)
|
||||
|
||||
|
||||
- **Reduce memory usage**
|
||||
|
||||
- Replaces continuous values with discrete bins. If ``#bins`` is small, can use small data type, e.g. uint8\_t, to store training data
|
||||
|
|
|
|||
|
|
@ -107,7 +107,7 @@ Example of using GPU (``gpu_platform_id = 0`` and ``gpu_device_id = 0`` in our s
|
|||
[LightGBM] [Info] 40 dense feature groups (0.12 MB) transferred to GPU in 0.004211 secs. 76 sparse feature groups.
|
||||
[LightGBM] [Info] No further splits with positive gain, best gain: -inf
|
||||
[LightGBM] [Info] Trained a tree with leaves=16 and depth=8
|
||||
[1]: test's rmse:1.10643e-17
|
||||
[1]: test's rmse:1.10643e-17
|
||||
[LightGBM] [Info] No further splits with positive gain, best gain: -inf
|
||||
[LightGBM] [Info] Trained a tree with leaves=7 and depth=5
|
||||
[2]: test's rmse:0
|
||||
|
|
@ -145,11 +145,11 @@ Example of using CPU (``gpu_platform_id = 0``, ``gpu_device_id = 1``). The GPU d
|
|||
[LightGBM] [Info] 40 dense feature groups (0.12 MB) transferred to GPU in 0.004540 secs. 76 sparse feature groups.
|
||||
[LightGBM] [Info] No further splits with positive gain, best gain: -inf
|
||||
[LightGBM] [Info] Trained a tree with leaves=16 and depth=8
|
||||
[1]: test's rmse:1.10643e-17
|
||||
[1]: test's rmse:1.10643e-17
|
||||
[LightGBM] [Info] No further splits with positive gain, best gain: -inf
|
||||
[LightGBM] [Info] Trained a tree with leaves=7 and depth=5
|
||||
[2]: test's rmse:0
|
||||
|
||||
|
||||
|
||||
Known issues:
|
||||
|
||||
|
|
|
|||
|
|
@ -61,7 +61,7 @@ Now we are ready to checkout LightGBM and compile it with GPU support:
|
|||
cd LightGBM
|
||||
mkdir build
|
||||
cd build
|
||||
cmake -DUSE_GPU=1 ..
|
||||
cmake -DUSE_GPU=1 ..
|
||||
# if you have installed NVIDIA CUDA to a customized location, you should specify paths to OpenCL headers and library like the following:
|
||||
# cmake -DUSE_GPU=1 -DOpenCL_LIBRARY=/usr/local/cuda/lib64/libOpenCL.so -DOpenCL_INCLUDE_DIR=/usr/local/cuda/include/ ..
|
||||
make -j$(nproc)
|
||||
|
|
|
|||
|
|
@ -75,7 +75,7 @@ The list includes the commits where the major feature added is considered workin
|
|||
* 22/06/2017 [Microsoft/LightGBM@d862b3e](https://github.com/microsoft/LightGBM/pull/642): CIntegration: Travis OSX Support (Pull Request 642)
|
||||
* 20/06/2017 [Microsoft/LightGBM@80c641c](https://github.com/microsoft/LightGBM/pull/635): Release: Python pip package (Pull Request 635)
|
||||
* 18/06/2017 [Microsoft/LightGBM@4d2aa84](https://github.com/microsoft/LightGBM/pull/634): CIntegration: AppVeyor Support (Pull Request 634)
|
||||
* 06/06/2017 [Microsoft/LightGBM@2c9ce59](https://github.com/microsoft/LightGBM/pull/592): Release: R-package version 0.2 (Pull Request 592)
|
||||
* 06/06/2017 [Microsoft/LightGBM@2c9ce59](https://github.com/microsoft/LightGBM/pull/592): Release: R-package version 0.2 (Pull Request 592)
|
||||
* 05/06/2017 [Microsoft/LightGBM@f98d75f](https://github.com/microsoft/LightGBM/pull/584): Feature: Use custom compiler for R-package (Pull Request 584)
|
||||
* 29/05/2017 [Microsoft/LightGBM@993bbd5](https://github.com/microsoft/LightGBM/pull/559): Parameter: Early Stopping for predictions (Pull Request 559)
|
||||
* 26/05/2017 [Microsoft/LightGBM@3abff37](https://github.com/microsoft/LightGBM/commit/3abff370bb353293e4a03e516111dd02785fbd97): Feature: Parameter to disable missing values (Commit)
|
||||
|
|
|
|||
|
|
@ -17,4 +17,4 @@ help:
|
|||
# Catch-all target: route all unknown targets to Sphinx using the new
|
||||
# "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS).
|
||||
%: Makefile
|
||||
@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
|
||||
@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
|
||||
|
|
|
|||
|
|
@ -1 +1 @@
|
|||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="118" height="20"><linearGradient id="b" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="a"><rect width="118" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#a)"><path fill="#555" d="M0 0h55v20H0z"/><path fill="#4c1" d="M55 0h63v20H55z"/><path fill="url(#b)" d="M0 0h118v20H0z"/></g><g fill="#fff" text-anchor="middle" font-family="DejaVu Sans,Verdana,Geneva,sans-serif" font-size="110"> <text x="285" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="450">artifacts</text><text x="285" y="140" transform="scale(.1)" textLength="450">artifacts</text><text x="855" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="530">download</text><text x="855" y="140" transform="scale(.1)" textLength="530">download</text></g> </svg>
|
||||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="118" height="20"><linearGradient id="b" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="a"><rect width="118" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#a)"><path fill="#555" d="M0 0h55v20H0z"/><path fill="#4c1" d="M55 0h63v20H55z"/><path fill="url(#b)" d="M0 0h118v20H0z"/></g><g fill="#fff" text-anchor="middle" font-family="DejaVu Sans,Verdana,Geneva,sans-serif" font-size="110"> <text x="285" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="450">artifacts</text><text x="285" y="140" transform="scale(.1)" textLength="450">artifacts</text><text x="855" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="530">download</text><text x="855" y="140" transform="scale(.1)" textLength="530">download</text></g> </svg>
|
||||
|
|
|
|||
|
До Ширина: | Высота: | Размер: 964 B После Ширина: | Высота: | Размер: 965 B |
|
|
@ -1 +1 @@
|
|||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="122" height="20"><linearGradient id="b" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="a"><rect width="122" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#a)"><path fill="#555" d="M0 0h55v20H0z"/><path fill="#9f9f9f" d="M55 0h67v20H55z"/><path fill="url(#b)" d="M0 0h122v20H0z"/></g><g fill="#fff" text-anchor="middle" font-family="DejaVu Sans,Verdana,Geneva,sans-serif" font-size="110"> <text x="285" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="450">artifacts</text><text x="285" y="140" transform="scale(.1)" textLength="450">artifacts</text><text x="875" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="570">fetching...</text><text x="875" y="140" transform="scale(.1)" textLength="570">fetching...</text></g> </svg>
|
||||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="122" height="20"><linearGradient id="b" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="a"><rect width="122" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#a)"><path fill="#555" d="M0 0h55v20H0z"/><path fill="#9f9f9f" d="M55 0h67v20H55z"/><path fill="url(#b)" d="M0 0h122v20H0z"/></g><g fill="#fff" text-anchor="middle" font-family="DejaVu Sans,Verdana,Geneva,sans-serif" font-size="110"> <text x="285" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="450">artifacts</text><text x="285" y="140" transform="scale(.1)" textLength="450">artifacts</text><text x="875" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="570">fetching...</text><text x="875" y="140" transform="scale(.1)" textLength="570">fetching...</text></g> </svg>
|
||||
|
|
|
|||
|
До Ширина: | Высота: | Размер: 973 B После Ширина: | Высота: | Размер: 974 B |
|
|
@ -1 +1 @@
|
|||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="302" height="20"><linearGradient id="b" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="a"><rect width="302" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#a)"><path fill="#555" d="M0 0h55v20H0z"/><path fill="#9f9f9f" d="M55 0h247v20H55z"/><path fill="url(#b)" d="M0 0h302v20H0z"/></g><g fill="#fff" text-anchor="middle" font-family="DejaVu Sans,Verdana,Geneva,sans-serif" font-size="110"> <text x="285" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="450">artifacts</text><text x="285" y="140" transform="scale(.1)" textLength="450">artifacts</text><text x="1775" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="2370">link is available only on Read the Docs site</text><text x="1775" y="140" transform="scale(.1)" textLength="2370">link is available only on Read the Docs site</text></g> </svg>
|
||||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="302" height="20"><linearGradient id="b" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="a"><rect width="302" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#a)"><path fill="#555" d="M0 0h55v20H0z"/><path fill="#9f9f9f" d="M55 0h247v20H55z"/><path fill="url(#b)" d="M0 0h302v20H0z"/></g><g fill="#fff" text-anchor="middle" font-family="DejaVu Sans,Verdana,Geneva,sans-serif" font-size="110"> <text x="285" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="450">artifacts</text><text x="285" y="140" transform="scale(.1)" textLength="450">artifacts</text><text x="1775" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="2370">link is available only on Read the Docs site</text><text x="1775" y="140" transform="scale(.1)" textLength="2370">link is available only on Read the Docs site</text></g> </svg>
|
||||
|
|
|
|||
|
До Ширина: | Высота: | Размер: 1.0 KiB После Ширина: | Высота: | Размер: 1.0 KiB |
Различия файлов скрыты, потому что одна или несколько строк слишком длинны
|
До Ширина: | Высота: | Размер: 87 KiB После Ширина: | Высота: | Размер: 87 KiB |
Различия файлов скрыты, потому что одна или несколько строк слишком длинны
|
До Ширина: | Высота: | Размер: 59 KiB После Ширина: | Высота: | Размер: 59 KiB |
|
|
@ -34,7 +34,7 @@ Machine Learning Challenge Winning Solutions
|
|||
| 1st | [Google Analytics Customer Revenue Prediction](https://www.kaggle.com/c/ga-customer-revenue-prediction) | [link](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/82614#latest-482575) | 2019.3 |
|
||||
| 1st | [VSB Power Line Fault Detection](https://www.kaggle.com/c/vsb-power-line-fault-detection) | [link](https://www.kaggle.com/c/vsb-power-line-fault-detection/discussion/87038#latest-521846) | 2019.3 |
|
||||
| 5th | [Elo Merchant Category Recommendation](https://www.kaggle.com/c/elo-merchant-category-recommendation) | [link](https://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/82314#latest-525737) | 2019.2 |
|
||||
| 2nd | [PLAsTiCC Astronomical Classification](https://www.kaggle.com/c/PLAsTiCC-2018) | [link](https://www.kaggle.com/c/PLAsTiCC-2018/discussion/75059#latest-462457) | 2018.12 |
|
||||
| 2nd | [PLAsTiCC Astronomical Classification](https://www.kaggle.com/c/PLAsTiCC-2018) | [link](https://www.kaggle.com/c/PLAsTiCC-2018/discussion/75059#latest-462457) | 2018.12 |
|
||||
| 1st | [Google Research Doodle Recognition Challenge](https://www.kaggle.com/c/quickdraw-doodle-recognition) | [link](https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/73738#latest-550028) | 2018.12 |
|
||||
| 1st | [Home Credit Group Home Credit Default Risk](https://www.kaggle.com/c/home-credit-default-risk) | [link](https://www.kaggle.com/c/home-credit-default-risk/discussion/64480#latest-514514) | 2018.8 |
|
||||
| 2nd | [Home Credit Group Home Credit Default Risk](https://www.kaggle.com/c/home-credit-default-risk) | [link](https://www.kaggle.com/c/home-credit-default-risk/discussion/64722#latest-394948) | 2018.8 |
|
||||
|
|
@ -45,11 +45,11 @@ Machine Learning Challenge Winning Solutions
|
|||
| 2nd | [Avito Demand Prediction Challenge](https://www.kaggle.com/c/avito-demand-prediction) | [link](https://www.kaggle.com/c/avito-demand-prediction/discussion/59871#latest-470807) | 2018.6 |
|
||||
| 3rd | [Avito Demand Prediction Challenge](https://www.kaggle.com/c/avito-demand-prediction) | [link](https://www.kaggle.com/c/avito-demand-prediction/discussion/59885#latest-364403) | 2018.6 |
|
||||
| 1st | [TalkingData AdTracking Fraud Detection Challenge](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection) | [link](https://www.kaggle.com/c/talkingdata-adtracking-fraud-detection/discussion/56475)| 2018.5 |
|
||||
| 1st | [DonorsChoose.org Application Screening](https://www.kaggle.com/c/donorschoose-application-screening)| [link](https://www.kaggle.com/shadowwarrior/1st-place-solution/notebook) | 2018.4 |
|
||||
| 1st | [DonorsChoose.org Application Screening](https://www.kaggle.com/c/donorschoose-application-screening)| [link](https://www.kaggle.com/shadowwarrior/1st-place-solution/notebook) | 2018.4 |
|
||||
| 1st | [Toxic Comment Classification Challenge](https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge)| [link](https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/52557) | 2018.3 |
|
||||
| 1st | [Mercari Price Suggestion Challenge](https://www.kaggle.com/c/mercari-price-suggestion-challenge) | [link](https://www.kaggle.com/c/mercari-price-suggestion-challenge/discussion/50256) | 2018.2 |
|
||||
| 1st | [IEEE's Signal Processing Society, Camera Model Identification](https://www.kaggle.com/c/sp-society-camera-model-identification)| [link](https://www.kaggle.com/c/sp-society-camera-model-identification/discussion/49367) | 2018.2 |
|
||||
| 1st | [Recruit Restaurant Visitor Forecasting](https://www.kaggle.com/c/recruit-restaurant-visitor-forecasting) | [link](https://www.kaggle.com/pureheart/1st-place-lgb-model-public-0-470-private-0-502/comments) | 2018.2|
|
||||
| 1st | [IEEE's Signal Processing Society, Camera Model Identification](https://www.kaggle.com/c/sp-society-camera-model-identification)| [link](https://www.kaggle.com/c/sp-society-camera-model-identification/discussion/49367) | 2018.2 |
|
||||
| 1st | [Recruit Restaurant Visitor Forecasting](https://www.kaggle.com/c/recruit-restaurant-visitor-forecasting) | [link](https://www.kaggle.com/pureheart/1st-place-lgb-model-public-0-470-private-0-502/comments) | 2018.2|
|
||||
| 1st | [WSDM CUP 2018 - KKBox's Music Recommendation Challenge](https://www.kaggle.com/c/kkbox-music-recommendation-challenge) | [link](https://www.kaggle.com/c/kkbox-music-recommendation-challenge/discussion/45942) | 2017.12 |
|
||||
| 1st | [Porto Seguro’s Safe Driver Prediction](https://www.kaggle.com/c/porto-seguro-safe-driver-prediction) | [link](https://www.kaggle.com/c/porto-seguro-safe-driver-prediction/discussion/44629) |2017.11 |
|
||||
| 1st | [Quora Question Pairs](https://www.kaggle.com/c/quora-question-pairs) | [link](https://www.kaggle.com/c/quora-question-pairs/discussion/34355) | 2017.6 |
|
||||
|
|
|
|||
|
|
@ -12,10 +12,10 @@ boosting_type = gbdt
|
|||
objective = binary
|
||||
|
||||
# eval metrics, support multi metric, delimited by ',' , support following metrics
|
||||
# l1
|
||||
# l1
|
||||
# l2 , default metric for regression
|
||||
# ndcg , default metric for lambdarank
|
||||
# auc
|
||||
# auc
|
||||
# binary_logloss , default metric for binary
|
||||
# binary_error
|
||||
metric = binary_logloss,auc
|
||||
|
|
@ -29,7 +29,7 @@ is_training_metric = true
|
|||
# column in data to use as label
|
||||
label_column = 0
|
||||
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
max_bin = 255
|
||||
|
||||
# training data
|
||||
|
|
@ -39,7 +39,7 @@ data = binary.train
|
|||
|
||||
# validation data, support multi validation data, separated by ','
|
||||
# if existing weight file, should name to "binary.test.weight"
|
||||
# alias: valid, test, test_data,
|
||||
# alias: valid, test, test_data,
|
||||
valid_data = binary.test
|
||||
|
||||
# number of trees(iterations), alias: num_tree, num_iteration, num_iterations, num_round, num_rounds
|
||||
|
|
@ -59,10 +59,10 @@ num_leaves = 63
|
|||
# alias: tree
|
||||
tree_learner = serial
|
||||
|
||||
# number of threads for multi-threading. One thread will use each CPU. The default is the CPU count.
|
||||
# number of threads for multi-threading. One thread will use each CPU. The default is the CPU count.
|
||||
# num_threads = 8
|
||||
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# alias: sub_feature
|
||||
feature_fraction = 0.8
|
||||
|
||||
|
|
|
|||
|
|
@ -14,10 +14,10 @@ objective = binary
|
|||
linear_tree = true
|
||||
|
||||
# eval metrics, support multi metric, delimited by ',' , support following metrics
|
||||
# l1
|
||||
# l1
|
||||
# l2 , default metric for regression
|
||||
# ndcg , default metric for lambdarank
|
||||
# auc
|
||||
# auc
|
||||
# binary_logloss , default metric for binary
|
||||
# binary_error
|
||||
metric = binary_logloss,auc
|
||||
|
|
@ -28,7 +28,7 @@ metric_freq = 1
|
|||
# true if need output metric for training data, alias: tranining_metric, train_metric
|
||||
is_training_metric = true
|
||||
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
max_bin = 255
|
||||
|
||||
# training data
|
||||
|
|
@ -38,7 +38,7 @@ data = binary.train
|
|||
|
||||
# validation data, support multi validation data, separated by ','
|
||||
# if existing weight file, should name to "binary.test.weight"
|
||||
# alias: valid, test, test_data,
|
||||
# alias: valid, test, test_data,
|
||||
valid_data = binary.test
|
||||
|
||||
# number of trees(iterations), alias: num_tree, num_iteration, num_iterations, num_round, num_rounds
|
||||
|
|
@ -58,10 +58,10 @@ num_leaves = 63
|
|||
# alias: tree
|
||||
tree_learner = serial
|
||||
|
||||
# number of threads for multi-threading. One thread will use each CPU. The default is set to CPU count.
|
||||
# number of threads for multi-threading. One thread will use each CPU. The default is set to CPU count.
|
||||
# num_threads = 8
|
||||
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# alias: sub_feature
|
||||
feature_fraction = 0.8
|
||||
|
||||
|
|
|
|||
|
|
@ -29,5 +29,5 @@ Run the following command in this folder:
|
|||
Data Format
|
||||
-----------
|
||||
|
||||
To learn more about the query format used in this example, check out the
|
||||
To learn more about the query format used in this example, check out the
|
||||
[query data format](https://lightgbm.readthedocs.io/en/latest/Parameters.html#query-data).
|
||||
|
|
|
|||
|
|
@ -12,10 +12,10 @@ boosting_type = gbdt
|
|||
objective = lambdarank
|
||||
|
||||
# eval metrics, support multi metric, delimited by ',' , support following metrics
|
||||
# l1
|
||||
# l1
|
||||
# l2 , default metric for regression
|
||||
# ndcg , default metric for lambdarank
|
||||
# auc
|
||||
# auc
|
||||
# binary_logloss , default metric for binary
|
||||
# binary_error
|
||||
metric = ndcg
|
||||
|
|
@ -32,7 +32,7 @@ is_training_metric = true
|
|||
# column in data to use as label
|
||||
label_column = 0
|
||||
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
max_bin = 255
|
||||
|
||||
# training data
|
||||
|
|
@ -44,7 +44,7 @@ data = rank.train
|
|||
# validation data, support multi validation data, separated by ','
|
||||
# if existing weight file, should name to "rank.test.weight"
|
||||
# if existing query file, should name to "rank.test.query"
|
||||
# alias: valid, test, test_data,
|
||||
# alias: valid, test, test_data,
|
||||
valid_data = rank.test
|
||||
|
||||
# number of trees(iterations), alias: num_tree, num_iteration, num_iterations, num_round, num_rounds
|
||||
|
|
@ -64,10 +64,10 @@ num_leaves = 31
|
|||
# alias: tree
|
||||
tree_learner = serial
|
||||
|
||||
# number of threads for multi-threading. One thread will use one CPU, defalut is setted to #cpu.
|
||||
# number of threads for multi-threading. One thread will use one CPU, defalut is setted to #cpu.
|
||||
# num_threads = 8
|
||||
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# alias: sub_feature
|
||||
feature_fraction = 1.0
|
||||
|
||||
|
|
|
|||
|
|
@ -13,10 +13,10 @@ boosting_type = gbdt
|
|||
objective = multiclass
|
||||
|
||||
# eval metrics, support multi metric, delimited by ',' , support following metrics
|
||||
# l1
|
||||
# l1
|
||||
# l2 , default metric for regression
|
||||
# ndcg , default metric for lambdarank
|
||||
# auc
|
||||
# auc
|
||||
# binary_logloss , default metric for binary
|
||||
# binary_error
|
||||
# multi_logloss
|
||||
|
|
@ -44,7 +44,7 @@ is_training_metric = true
|
|||
# column in data to use as label
|
||||
label_column = 0
|
||||
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
max_bin = 255
|
||||
|
||||
# training data
|
||||
|
|
|
|||
|
|
@ -12,10 +12,10 @@ boosting_type = gbdt
|
|||
objective = binary
|
||||
|
||||
# eval metrics, support multi metric, delimite by ',' , support following metrics
|
||||
# l1
|
||||
# l1
|
||||
# l2 , default metric for regression
|
||||
# ndcg , default metric for lambdarank
|
||||
# auc
|
||||
# auc
|
||||
# binary_logloss , default metric for binary
|
||||
# binary_error
|
||||
metric = binary_logloss,auc
|
||||
|
|
@ -29,7 +29,7 @@ is_training_metric = true
|
|||
# column in data to use as label
|
||||
label_column = 0
|
||||
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
max_bin = 255
|
||||
|
||||
# training data
|
||||
|
|
@ -39,7 +39,7 @@ data = binary.train
|
|||
|
||||
# validation data, support multi validation data, separated by ','
|
||||
# if existing weight file, should name to "binary.test.weight"
|
||||
# alias: valid, test, test_data,
|
||||
# alias: valid, test, test_data,
|
||||
valid_data = binary.test
|
||||
|
||||
# number of trees(iterations), alias: num_tree, num_iteration, num_iterations, num_round, num_rounds
|
||||
|
|
@ -62,7 +62,7 @@ tree_learner = feature
|
|||
# number of threads for multi-threading. One thread will use each CPU. The default is the CPU count.
|
||||
# num_threads = 8
|
||||
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# alias: sub_feature
|
||||
feature_fraction = 0.8
|
||||
|
||||
|
|
|
|||
|
|
@ -23,11 +23,11 @@ Examples include:
|
|||
- [simple_example.py](https://github.com/microsoft/LightGBM/blob/master/examples/python-guide/simple_example.py)
|
||||
- Construct Dataset
|
||||
- Basic train and predict
|
||||
- Eval during training
|
||||
- Eval during training
|
||||
- Early stopping
|
||||
- Save model to file
|
||||
- [sklearn_example.py](https://github.com/microsoft/LightGBM/blob/master/examples/python-guide/sklearn_example.py)
|
||||
- Create data for learning with sklearn interface
|
||||
- Create data for learning with sklearn interface
|
||||
- Basic train and predict with sklearn interface
|
||||
- Feature importances with sklearn interface
|
||||
- Self-defined eval metric with sklearn interface
|
||||
|
|
|
|||
|
|
@ -12,10 +12,10 @@ boosting_type = gbdt
|
|||
objective = regression
|
||||
|
||||
# eval metrics, support multi metric, delimite by ',' , support following metrics
|
||||
# l1
|
||||
# l1
|
||||
# l2 , default metric for regression
|
||||
# ndcg , default metric for lambdarank
|
||||
# auc
|
||||
# auc
|
||||
# binary_logloss , default metric for binary
|
||||
# binary_error
|
||||
metric = l2
|
||||
|
|
@ -29,7 +29,7 @@ is_training_metric = true
|
|||
# column in data to use as label
|
||||
label_column = 0
|
||||
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
max_bin = 255
|
||||
|
||||
# forced bin thresholds
|
||||
|
|
@ -42,7 +42,7 @@ data = regression.train
|
|||
|
||||
# validation data, support multi validation data, separated by ','
|
||||
# if exsting weight file, should name to "regression.test.weight"
|
||||
# alias: valid, test, test_data,
|
||||
# alias: valid, test, test_data,
|
||||
valid_data = regression.test
|
||||
|
||||
# number of trees(iterations), alias: num_tree, num_iteration, num_iterations, num_round, num_rounds
|
||||
|
|
@ -62,10 +62,10 @@ num_leaves = 31
|
|||
# alias: tree
|
||||
tree_learner = serial
|
||||
|
||||
# number of threads for multi-threading. One thread will use one CPU, default is setted to #cpu.
|
||||
# number of threads for multi-threading. One thread will use one CPU, default is setted to #cpu.
|
||||
# num_threads = 8
|
||||
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# alias: sub_feature
|
||||
feature_fraction = 0.9
|
||||
|
||||
|
|
|
|||
|
|
@ -29,5 +29,5 @@ Run the following command in this folder:
|
|||
Data Format
|
||||
-----------
|
||||
|
||||
To learn more about the query format used in this example, check out the
|
||||
To learn more about the query format used in this example, check out the
|
||||
[query data format](https://lightgbm.readthedocs.io/en/latest/Parameters.html#query-data).
|
||||
|
|
|
|||
|
|
@ -12,10 +12,10 @@ boosting_type = gbdt
|
|||
objective = rank_xendcg
|
||||
|
||||
# eval metrics, support multi metric, delimite by ',' , support following metrics
|
||||
# l1
|
||||
# l1
|
||||
# l2 , default metric for regression
|
||||
# ndcg , default metric for lambdarank
|
||||
# auc
|
||||
# auc
|
||||
# binary_logloss , default metric for binary
|
||||
# binary_error
|
||||
metric = ndcg
|
||||
|
|
@ -32,7 +32,7 @@ is_training_metric = true
|
|||
# column in data to use as label
|
||||
label_column = 0
|
||||
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
# number of bins for feature bucket, 255 is a recommend setting, it can save memories, and also has good accuracy.
|
||||
max_bin = 255
|
||||
|
||||
# training data
|
||||
|
|
@ -44,7 +44,7 @@ data = rank.train
|
|||
# validation data, support multi validation data, separated by ','
|
||||
# if existing weight file, should name to "rank.test.weight"
|
||||
# if existing query file, should name to "rank.test.query"
|
||||
# alias: valid, test, test_data,
|
||||
# alias: valid, test, test_data,
|
||||
valid_data = rank.test
|
||||
|
||||
# number of trees(iterations), alias: num_tree, num_iteration, num_iterations, num_round, num_rounds
|
||||
|
|
@ -68,7 +68,7 @@ tree_learner = serial
|
|||
num_threads = 1
|
||||
objective_seed = 1025
|
||||
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# feature sub-sample, will random select 80% feature to train on each iteration
|
||||
# alias: sub_feature
|
||||
feature_fraction = 1.0
|
||||
|
||||
|
|
|
|||
|
|
@ -144,7 +144,7 @@ class BinMapper {
|
|||
|
||||
/*!
|
||||
* \brief Maximum categorical value
|
||||
* \return Maximum categorical value for categorical features, 0 for numerical features
|
||||
* \return Maximum categorical value for categorical features, 0 for numerical features
|
||||
*/
|
||||
inline int MaxCatValue() const {
|
||||
if (bin_2_categorical_.size() == 0) {
|
||||
|
|
|
|||
|
|
@ -128,7 +128,7 @@ class Network {
|
|||
const ReduceFunction& reducer);
|
||||
|
||||
/*!
|
||||
* \brief Performing all_gather by using Bruck algorithm.
|
||||
* \brief Performing all_gather by using Bruck algorithm.
|
||||
Communication times is O(log(n)), and communication cost is O(send_size * number_machine)
|
||||
* It can be used when all nodes have same input size.
|
||||
* \param input Input data
|
||||
|
|
@ -138,7 +138,7 @@ class Network {
|
|||
static void Allgather(char* input, comm_size_t send_size, char* output);
|
||||
|
||||
/*!
|
||||
* \brief Performing all_gather by using Bruck algorithm.
|
||||
* \brief Performing all_gather by using Bruck algorithm.
|
||||
Communication times is O(log(n)), and communication cost is O(all_size)
|
||||
* It can be used when nodes have different input size.
|
||||
* \param input Input data
|
||||
|
|
@ -150,7 +150,7 @@ class Network {
|
|||
static void Allgather(char* input, const comm_size_t* block_start, const comm_size_t* block_len, char* output, comm_size_t all_size);
|
||||
|
||||
/*!
|
||||
* \brief Perform reduce scatter by using recursive halving algorithm.
|
||||
* \brief Perform reduce scatter by using recursive halving algorithm.
|
||||
Communication times is O(log(n)), and communication cost is O(input_size)
|
||||
* \param input Input data
|
||||
* \param input_size The size of input data
|
||||
|
|
|
|||
|
|
@ -1232,7 +1232,7 @@ struct __TToStringHelper<T, true, true> {
|
|||
* Converts an array to a string with with values separated by the space character.
|
||||
* This method replaces Common's ``ArrayToString`` and ``ArrayToStringFast`` functionality
|
||||
* and is locale-independent.
|
||||
*
|
||||
*
|
||||
* \note If ``high_precision_output`` is set to true,
|
||||
* floating point values are output with more digits of precision.
|
||||
*/
|
||||
|
|
|
|||
|
|
@ -1,4 +1,4 @@
|
|||
PMML Generator
|
||||
PMML Generator
|
||||
==============
|
||||
|
||||
The old Python convert script is removed due to it cannot support the new format of categorical features.
|
||||
|
|
|
|||
|
|
@ -107,7 +107,7 @@ class SingleRowPredictorInner {
|
|||
|
||||
/*!
|
||||
* \brief Object to store resources meant for single-row Fast Predict methods.
|
||||
*
|
||||
*
|
||||
* For legacy reasons this is called `FastConfig` in the public C API.
|
||||
*
|
||||
* Meant to be used by the *Fast* predict methods only.
|
||||
|
|
|
|||
|
|
@ -25,36 +25,36 @@ typedef unsigned char uchar;
|
|||
template<typename T>
|
||||
__device__ double as_double(const T t) {
|
||||
static_assert(sizeof(T) == sizeof(double), "size mismatch");
|
||||
double d;
|
||||
memcpy(&d, &t, sizeof(T));
|
||||
double d;
|
||||
memcpy(&d, &t, sizeof(T));
|
||||
return d;
|
||||
}
|
||||
template<typename T>
|
||||
__device__ unsigned long long as_ulong_ulong(const T t) {
|
||||
static_assert(sizeof(T) == sizeof(unsigned long long), "size mismatch");
|
||||
unsigned long long u;
|
||||
memcpy(&u, &t, sizeof(T));
|
||||
unsigned long long u;
|
||||
memcpy(&u, &t, sizeof(T));
|
||||
return u;
|
||||
}
|
||||
template<typename T>
|
||||
__device__ float as_float(const T t) {
|
||||
static_assert(sizeof(T) == sizeof(float), "size mismatch");
|
||||
float f;
|
||||
memcpy(&f, &t, sizeof(T));
|
||||
float f;
|
||||
memcpy(&f, &t, sizeof(T));
|
||||
return f;
|
||||
}
|
||||
template<typename T>
|
||||
__device__ unsigned int as_uint(const T t) {
|
||||
static_assert(sizeof(T) == sizeof(unsigned int), "size_mismatch");
|
||||
unsigned int u;
|
||||
memcpy(&u, &t, sizeof(T));
|
||||
unsigned int u;
|
||||
memcpy(&u, &t, sizeof(T));
|
||||
return u;
|
||||
}
|
||||
template<typename T>
|
||||
__device__ uchar4 as_uchar4(const T t) {
|
||||
static_assert(sizeof(T) == sizeof(uchar4), "size mismatch");
|
||||
uchar4 u;
|
||||
memcpy(&u, &t, sizeof(T));
|
||||
uchar4 u;
|
||||
memcpy(&u, &t, sizeof(T));
|
||||
return u;
|
||||
}
|
||||
|
||||
|
|
@ -158,4 +158,3 @@ DECLARE(histogram256);
|
|||
} // namespace LightGBM
|
||||
|
||||
#endif // LIGHTGBM_TREELEARNER_KERNELS_HISTOGRAM_16_64_256_HU_
|
||||
|
||||
|
|
|
|||
|
|
@ -38,7 +38,7 @@ class LeafSplits {
|
|||
}
|
||||
|
||||
/*!
|
||||
* \brief Init split on current leaf on partial data.
|
||||
* \brief Init split on current leaf on partial data.
|
||||
* \param leaf Index of current leaf
|
||||
* \param data_partition current data partition
|
||||
* \param sum_gradients
|
||||
|
|
@ -54,7 +54,7 @@ class LeafSplits {
|
|||
}
|
||||
|
||||
/*!
|
||||
* \brief Init split on current leaf on partial data.
|
||||
* \brief Init split on current leaf on partial data.
|
||||
* \param leaf Index of current leaf
|
||||
* \param data_partition current data partition
|
||||
* \param sum_gradients
|
||||
|
|
|
|||
|
|
@ -73,12 +73,12 @@ typedef uint acc_int_type;
|
|||
// local memory size in bytes
|
||||
#define LOCAL_MEM_SIZE (DWORD_FEATURES * (sizeof(uint) + 2 * sizeof(acc_type)) * NUM_BINS * NUM_BANKS)
|
||||
|
||||
// unroll the atomic operation for a few times. Takes more code space,
|
||||
// unroll the atomic operation for a few times. Takes more code space,
|
||||
// but compiler can generate better code for faster atomics.
|
||||
#define UNROLL_ATOMIC 1
|
||||
|
||||
// Options passed by compiler at run time:
|
||||
// IGNORE_INDICES will be set when the kernel does not
|
||||
// IGNORE_INDICES will be set when the kernel does not
|
||||
// #define IGNORE_INDICES
|
||||
// #define POWER_FEATURE_WORKGROUPS 10
|
||||
|
||||
|
|
@ -161,7 +161,7 @@ R""()
|
|||
// this function will be called by histogram16
|
||||
// we have one sub-histogram of one feature in registers, and need to read others
|
||||
void within_kernel_reduction16x8(uchar8 feature_mask,
|
||||
__global const acc_type* restrict feature4_sub_hist,
|
||||
__global const acc_type* restrict feature4_sub_hist,
|
||||
const uint skip_id,
|
||||
acc_type stat_val,
|
||||
const ushort num_sub_hist,
|
||||
|
|
@ -173,7 +173,7 @@ void within_kernel_reduction16x8(uchar8 feature_mask,
|
|||
uchar is_hessian_first = (ltid >> LOG2_DWORD_FEATURES) & 1; // hessian or gradient
|
||||
ushort bin_id = ltid >> (LOG2_DWORD_FEATURES + 1); // range 0 - 16
|
||||
ushort i;
|
||||
#if POWER_FEATURE_WORKGROUPS != 0
|
||||
#if POWER_FEATURE_WORKGROUPS != 0
|
||||
// if there is only 1 work group, no need to do the reduction
|
||||
// add all sub-histograms for 4 features
|
||||
__global const acc_type* restrict p = feature4_sub_hist + ltid;
|
||||
|
|
@ -185,7 +185,7 @@ void within_kernel_reduction16x8(uchar8 feature_mask,
|
|||
// skip the counters we already have
|
||||
p += 2 * DWORD_FEATURES * NUM_BINS;
|
||||
for (i = i + 1; i < num_sub_hist; ++i) {
|
||||
stat_val += *p;
|
||||
stat_val += *p;
|
||||
p += NUM_BINS * DWORD_FEATURES * 2;
|
||||
}
|
||||
#endif
|
||||
|
|
@ -208,12 +208,12 @@ R""()
|
|||
|
||||
__attribute__((reqd_work_group_size(LOCAL_SIZE_0, 1, 1)))
|
||||
#if USE_CONSTANT_BUF == 1
|
||||
__kernel void histogram16(__global const uchar4* restrict feature_data_base,
|
||||
__kernel void histogram16(__global const uchar4* restrict feature_data_base,
|
||||
__constant const uchar8* restrict feature_masks __attribute__((max_constant_size(65536))),
|
||||
const data_size_t feature_size,
|
||||
__constant const data_size_t* restrict data_indices __attribute__((max_constant_size(65536))),
|
||||
const data_size_t num_data,
|
||||
__constant const score_t* restrict ordered_gradients __attribute__((max_constant_size(65536))),
|
||||
__constant const data_size_t* restrict data_indices __attribute__((max_constant_size(65536))),
|
||||
const data_size_t num_data,
|
||||
__constant const score_t* restrict ordered_gradients __attribute__((max_constant_size(65536))),
|
||||
#if CONST_HESSIAN == 0
|
||||
__constant const score_t* restrict ordered_hessians __attribute__((max_constant_size(65536))),
|
||||
#else
|
||||
|
|
@ -223,18 +223,18 @@ __kernel void histogram16(__global const uchar4* restrict feature_data_base,
|
|||
__global volatile int * sync_counters,
|
||||
__global acc_type* restrict hist_buf_base) {
|
||||
#else
|
||||
__kernel void histogram16(__global const uchar4* feature_data_base,
|
||||
__kernel void histogram16(__global const uchar4* feature_data_base,
|
||||
__constant const uchar8* restrict feature_masks __attribute__((max_constant_size(65536))),
|
||||
const data_size_t feature_size,
|
||||
__global const data_size_t* data_indices,
|
||||
const data_size_t num_data,
|
||||
__global const score_t* ordered_gradients,
|
||||
__global const data_size_t* data_indices,
|
||||
const data_size_t num_data,
|
||||
__global const score_t* ordered_gradients,
|
||||
#if CONST_HESSIAN == 0
|
||||
__global const score_t* ordered_hessians,
|
||||
#else
|
||||
const score_t const_hessian,
|
||||
#endif
|
||||
__global char* restrict output_buf,
|
||||
__global char* restrict output_buf,
|
||||
__global volatile int * sync_counters,
|
||||
__global acc_type* restrict hist_buf_base) {
|
||||
#endif
|
||||
|
|
@ -260,38 +260,38 @@ __kernel void histogram16(__global const uchar4* feature_data_base,
|
|||
// there are 8 banks (sub-histograms) used by 256 threads total 8 KB
|
||||
/* memory layout of gh_hist:
|
||||
-----------------------------------------------------------------------------------------------
|
||||
bk0_g_f0_bin0 bk0_g_f1_bin0 bk0_g_f2_bin0 bk0_g_f3_bin0 bk0_g_f4_bin0 bk0_g_f5_bin0 bk0_g_f6_bin0 bk0_g_f7_bin0
|
||||
bk0_g_f0_bin0 bk0_g_f1_bin0 bk0_g_f2_bin0 bk0_g_f3_bin0 bk0_g_f4_bin0 bk0_g_f5_bin0 bk0_g_f6_bin0 bk0_g_f7_bin0
|
||||
bk0_h_f0_bin0 bk0_h_f1_bin0 bk0_h_f2_bin0 bk0_h_f3_bin0 bk0_h_f4_bin0 bk0_h_f5_bin0 bk0_h_f6_bin0 bk0_h_f7_bin0
|
||||
bk1_g_f0_bin0 bk1_g_f1_bin0 bk1_g_f2_bin0 bk1_g_f3_bin0 bk1_g_f4_bin0 bk1_g_f5_bin0 bk1_g_f6_bin0 bk1_g_f7_bin0
|
||||
bk1_g_f0_bin0 bk1_g_f1_bin0 bk1_g_f2_bin0 bk1_g_f3_bin0 bk1_g_f4_bin0 bk1_g_f5_bin0 bk1_g_f6_bin0 bk1_g_f7_bin0
|
||||
bk1_h_f0_bin0 bk1_h_f1_bin0 bk1_h_f2_bin0 bk1_h_f3_bin0 bk1_h_f4_bin0 bk1_h_f5_bin0 bk1_h_f6_bin0 bk1_h_f7_bin0
|
||||
bk2_g_f0_bin0 bk2_g_f1_bin0 bk2_g_f2_bin0 bk2_g_f3_bin0 bk2_g_f4_bin0 bk2_g_f5_bin0 bk2_g_f6_bin0 bk2_g_f7_bin0
|
||||
bk2_g_f0_bin0 bk2_g_f1_bin0 bk2_g_f2_bin0 bk2_g_f3_bin0 bk2_g_f4_bin0 bk2_g_f5_bin0 bk2_g_f6_bin0 bk2_g_f7_bin0
|
||||
bk2_h_f0_bin0 bk2_h_f1_bin0 bk2_h_f2_bin0 bk2_h_f3_bin0 bk2_h_f4_bin0 bk2_h_f5_bin0 bk2_h_f6_bin0 bk2_h_f7_bin0
|
||||
bk3_g_f0_bin0 bk3_g_f1_bin0 bk3_g_f2_bin0 bk3_g_f3_bin0 bk3_g_f4_bin0 bk3_g_f5_bin0 bk3_g_f6_bin0 bk3_g_f7_bin0
|
||||
bk3_g_f0_bin0 bk3_g_f1_bin0 bk3_g_f2_bin0 bk3_g_f3_bin0 bk3_g_f4_bin0 bk3_g_f5_bin0 bk3_g_f6_bin0 bk3_g_f7_bin0
|
||||
bk3_h_f0_bin0 bk3_h_f1_bin0 bk3_h_f2_bin0 bk3_h_f3_bin0 bk3_h_f4_bin0 bk3_h_f5_bin0 bk3_h_f6_bin0 bk3_h_f7_bin0
|
||||
bk4_g_f0_bin0 bk4_g_f1_bin0 bk4_g_f2_bin0 bk4_g_f3_bin0 bk4_g_f4_bin0 bk4_g_f5_bin0 bk4_g_f6_bin0 bk4_g_f7_bin0
|
||||
bk4_g_f0_bin0 bk4_g_f1_bin0 bk4_g_f2_bin0 bk4_g_f3_bin0 bk4_g_f4_bin0 bk4_g_f5_bin0 bk4_g_f6_bin0 bk4_g_f7_bin0
|
||||
bk4_h_f0_bin0 bk4_h_f1_bin0 bk4_h_f2_bin0 bk4_h_f3_bin0 bk4_h_f4_bin0 bk4_h_f5_bin0 bk4_h_f6_bin0 bk4_h_f7_bin0
|
||||
bk5_g_f0_bin0 bk5_g_f1_bin0 bk5_g_f2_bin0 bk5_g_f3_bin0 bk5_g_f4_bin0 bk5_g_f5_bin0 bk5_g_f6_bin0 bk5_g_f7_bin0
|
||||
bk5_g_f0_bin0 bk5_g_f1_bin0 bk5_g_f2_bin0 bk5_g_f3_bin0 bk5_g_f4_bin0 bk5_g_f5_bin0 bk5_g_f6_bin0 bk5_g_f7_bin0
|
||||
bk5_h_f0_bin0 bk5_h_f1_bin0 bk5_h_f2_bin0 bk5_h_f3_bin0 bk5_h_f4_bin0 bk5_h_f5_bin0 bk5_h_f6_bin0 bk5_h_f7_bin0
|
||||
bk6_g_f0_bin0 bk6_g_f1_bin0 bk6_g_f2_bin0 bk6_g_f3_bin0 bk6_g_f4_bin0 bk6_g_f5_bin0 bk6_g_f6_bin0 bk6_g_f7_bin0
|
||||
bk6_g_f0_bin0 bk6_g_f1_bin0 bk6_g_f2_bin0 bk6_g_f3_bin0 bk6_g_f4_bin0 bk6_g_f5_bin0 bk6_g_f6_bin0 bk6_g_f7_bin0
|
||||
bk6_h_f0_bin0 bk6_h_f1_bin0 bk6_h_f2_bin0 bk6_h_f3_bin0 bk6_h_f4_bin0 bk6_h_f5_bin0 bk6_h_f6_bin0 bk6_h_f7_bin0
|
||||
bk7_g_f0_bin0 bk7_g_f1_bin0 bk7_g_f2_bin0 bk7_g_f3_bin0 bk7_g_f4_bin0 bk7_g_f5_bin0 bk7_g_f6_bin0 bk7_g_f7_bin0
|
||||
bk7_g_f0_bin0 bk7_g_f1_bin0 bk7_g_f2_bin0 bk7_g_f3_bin0 bk7_g_f4_bin0 bk7_g_f5_bin0 bk7_g_f6_bin0 bk7_g_f7_bin0
|
||||
bk7_h_f0_bin0 bk7_h_f1_bin0 bk7_h_f2_bin0 bk7_h_f3_bin0 bk7_h_f4_bin0 bk7_h_f5_bin0 bk7_h_f6_bin0 bk7_h_f7_bin0
|
||||
...
|
||||
bk0_g_f0_bin16 bk0_g_f1_bin16 bk0_g_f2_bin16 bk0_g_f3_bin16 bk0_g_f4_bin16 bk0_g_f5_bin16 bk0_g_f6_bin16 bk0_g_f7_bin16
|
||||
bk0_g_f0_bin16 bk0_g_f1_bin16 bk0_g_f2_bin16 bk0_g_f3_bin16 bk0_g_f4_bin16 bk0_g_f5_bin16 bk0_g_f6_bin16 bk0_g_f7_bin16
|
||||
bk0_h_f0_bin16 bk0_h_f1_bin16 bk0_h_f2_bin16 bk0_h_f3_bin16 bk0_h_f4_bin16 bk0_h_f5_bin16 bk0_h_f6_bin16 bk0_h_f7_bin16
|
||||
bk1_g_f0_bin16 bk1_g_f1_bin16 bk1_g_f2_bin16 bk1_g_f3_bin16 bk1_g_f4_bin16 bk1_g_f5_bin16 bk1_g_f6_bin16 bk1_g_f7_bin16
|
||||
bk1_g_f0_bin16 bk1_g_f1_bin16 bk1_g_f2_bin16 bk1_g_f3_bin16 bk1_g_f4_bin16 bk1_g_f5_bin16 bk1_g_f6_bin16 bk1_g_f7_bin16
|
||||
bk1_h_f0_bin16 bk1_h_f1_bin16 bk1_h_f2_bin16 bk1_h_f3_bin16 bk1_h_f4_bin16 bk1_h_f5_bin16 bk1_h_f6_bin16 bk1_h_f7_bin16
|
||||
bk2_g_f0_bin16 bk2_g_f1_bin16 bk2_g_f2_bin16 bk2_g_f3_bin16 bk2_g_f4_bin16 bk2_g_f5_bin16 bk2_g_f6_bin16 bk2_g_f7_bin16
|
||||
bk2_g_f0_bin16 bk2_g_f1_bin16 bk2_g_f2_bin16 bk2_g_f3_bin16 bk2_g_f4_bin16 bk2_g_f5_bin16 bk2_g_f6_bin16 bk2_g_f7_bin16
|
||||
bk2_h_f0_bin16 bk2_h_f1_bin16 bk2_h_f2_bin16 bk2_h_f3_bin16 bk2_h_f4_bin16 bk2_h_f5_bin16 bk2_h_f6_bin16 bk2_h_f7_bin16
|
||||
bk3_g_f0_bin16 bk3_g_f1_bin16 bk3_g_f2_bin16 bk3_g_f3_bin16 bk3_g_f4_bin16 bk3_g_f5_bin16 bk3_g_f6_bin16 bk3_g_f7_bin16
|
||||
bk3_g_f0_bin16 bk3_g_f1_bin16 bk3_g_f2_bin16 bk3_g_f3_bin16 bk3_g_f4_bin16 bk3_g_f5_bin16 bk3_g_f6_bin16 bk3_g_f7_bin16
|
||||
bk3_h_f0_bin16 bk3_h_f1_bin16 bk3_h_f2_bin16 bk3_h_f3_bin16 bk3_h_f4_bin16 bk3_h_f5_bin16 bk3_h_f6_bin16 bk3_h_f7_bin16
|
||||
bk4_g_f0_bin16 bk4_g_f1_bin16 bk4_g_f2_bin16 bk4_g_f3_bin16 bk4_g_f4_bin16 bk4_g_f5_bin16 bk4_g_f6_bin16 bk4_g_f7_bin16
|
||||
bk4_g_f0_bin16 bk4_g_f1_bin16 bk4_g_f2_bin16 bk4_g_f3_bin16 bk4_g_f4_bin16 bk4_g_f5_bin16 bk4_g_f6_bin16 bk4_g_f7_bin16
|
||||
bk4_h_f0_bin16 bk4_h_f1_bin16 bk4_h_f2_bin16 bk4_h_f3_bin16 bk4_h_f4_bin16 bk4_h_f5_bin16 bk4_h_f6_bin16 bk4_h_f7_bin16
|
||||
bk5_g_f0_bin16 bk5_g_f1_bin16 bk5_g_f2_bin16 bk5_g_f3_bin16 bk5_g_f4_bin16 bk5_g_f5_bin16 bk5_g_f6_bin16 bk5_g_f7_bin16
|
||||
bk5_g_f0_bin16 bk5_g_f1_bin16 bk5_g_f2_bin16 bk5_g_f3_bin16 bk5_g_f4_bin16 bk5_g_f5_bin16 bk5_g_f6_bin16 bk5_g_f7_bin16
|
||||
bk5_h_f0_bin16 bk5_h_f1_bin16 bk5_h_f2_bin16 bk5_h_f3_bin16 bk5_h_f4_bin16 bk5_h_f5_bin16 bk5_h_f6_bin16 bk5_h_f7_bin16
|
||||
bk6_g_f0_bin16 bk6_g_f1_bin16 bk6_g_f2_bin16 bk6_g_f3_bin16 bk6_g_f4_bin16 bk6_g_f5_bin16 bk6_g_f6_bin16 bk6_g_f7_bin16
|
||||
bk6_g_f0_bin16 bk6_g_f1_bin16 bk6_g_f2_bin16 bk6_g_f3_bin16 bk6_g_f4_bin16 bk6_g_f5_bin16 bk6_g_f6_bin16 bk6_g_f7_bin16
|
||||
bk6_h_f0_bin16 bk6_h_f1_bin16 bk6_h_f2_bin16 bk6_h_f3_bin16 bk6_h_f4_bin16 bk6_h_f5_bin16 bk6_h_f6_bin16 bk6_h_f7_bin16
|
||||
bk7_g_f0_bin16 bk7_g_f1_bin16 bk7_g_f2_bin16 bk7_g_f3_bin16 bk7_g_f4_bin16 bk7_g_f5_bin16 bk7_g_f6_bin16 bk7_g_f7_bin16
|
||||
bk7_g_f0_bin16 bk7_g_f1_bin16 bk7_g_f2_bin16 bk7_g_f3_bin16 bk7_g_f4_bin16 bk7_g_f5_bin16 bk7_g_f6_bin16 bk7_g_f7_bin16
|
||||
bk7_h_f0_bin16 bk7_h_f1_bin16 bk7_h_f2_bin16 bk7_h_f3_bin16 bk7_h_f4_bin16 bk7_h_f5_bin16 bk7_h_f6_bin16 bk7_h_f7_bin16
|
||||
-----------------------------------------------------------------------------------------------
|
||||
*/
|
||||
|
|
@ -333,7 +333,7 @@ __kernel void histogram16(__global const uchar4* feature_data_base,
|
|||
uchar is_hessian_first = (ltid >> LOG2_DWORD_FEATURES) & 1;
|
||||
// thread 0-15 write result to bank0, 16-31 to bank1, 32-47 to bank2, 48-63 to bank3, etc
|
||||
ushort bank = (ltid >> (LOG2_DWORD_FEATURES + 1)) & BANK_MASK;
|
||||
|
||||
|
||||
ushort group_feature = group_id >> POWER_FEATURE_WORKGROUPS;
|
||||
// each 2^POWER_FEATURE_WORKGROUPS workgroups process on one feature (compile-time constant)
|
||||
// feature_size is the number of examples per feature
|
||||
|
|
@ -615,12 +615,12 @@ R""()
|
|||
)""
|
||||
R""()
|
||||
*/
|
||||
|
||||
|
||||
#if ENABLE_ALL_FEATURES == 0
|
||||
// restore feature_mask
|
||||
feature_mask = feature_masks[group_feature];
|
||||
#endif
|
||||
|
||||
|
||||
// now reduce the 4 banks of subhistograms into 1
|
||||
acc_type stat_val = 0.0f;
|
||||
uint cnt_val = 0;
|
||||
|
|
@ -644,7 +644,7 @@ R""()
|
|||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
|
||||
// now thread 0 - 7 holds feature 0 - 7's gradient for bin 0 and counter bin 0
|
||||
// now thread 8 - 15 holds feature 0 - 7's hessian for bin 0 and counter bin 1
|
||||
// now thread 16- 23 holds feature 0 - 7's gradient for bin 1 and counter bin 2
|
||||
|
|
@ -664,7 +664,7 @@ R""()
|
|||
// thread 8 - 15 read counters stored by thread 0 - 7
|
||||
// thread 24- 31 read counters stored by thread 8 - 15
|
||||
// thread 40- 47 read counters stored by thread 16- 23, etc
|
||||
stat_val = const_hessian *
|
||||
stat_val = const_hessian *
|
||||
cnt_hist[((ltid - DWORD_FEATURES) >> (LOG2_DWORD_FEATURES + 1)) * DWORD_FEATURES + (ltid & DWORD_FEATURES_MASK)];
|
||||
}
|
||||
else {
|
||||
|
|
@ -688,12 +688,12 @@ R""()
|
|||
h_f0_bin1 h_f1_bin1 h_f2_bin1 h_f3_bin1 h_f4_bin1 h_f5_bin1 h_f6_bin1 h_f7_bin1
|
||||
...
|
||||
...
|
||||
g_f0_bin16 g_f1_bin16 g_f2_bin16 g_f3_bin16 g_f4_bin16 g_f5_bin16 g_f6_bin16 g_f7_bin16
|
||||
h_f0_bin16 h_f1_bin16 h_f2_bin16 h_f3_bin16 h_f4_bin16 h_f5_bin16 h_f6_bin16 h_f7_bin16
|
||||
g_f0_bin16 g_f1_bin16 g_f2_bin16 g_f3_bin16 g_f4_bin16 g_f5_bin16 g_f6_bin16 g_f7_bin16
|
||||
h_f0_bin16 h_f1_bin16 h_f2_bin16 h_f3_bin16 h_f4_bin16 h_f5_bin16 h_f6_bin16 h_f7_bin16
|
||||
c_f0_bin0 c_f1_bin0 c_f2_bin0 c_f3_bin0 c_f4_bin0 c_f5_bin0 c_f6_bin0 c_f7_bin0
|
||||
c_f0_bin1 c_f1_bin1 c_f2_bin1 c_f3_bin1 c_f4_bin1 c_f5_bin1 c_f6_bin1 c_f7_bin1
|
||||
...
|
||||
c_f0_bin16 c_f1_bin16 c_f2_bin16 c_f3_bin16 c_f4_bin16 c_f5_bin16 c_f6_bin16 c_f7_bin16
|
||||
c_f0_bin16 c_f1_bin16 c_f2_bin16 c_f3_bin16 c_f4_bin16 c_f5_bin16 c_f6_bin16 c_f7_bin16
|
||||
*/
|
||||
// if there is only one workgroup processing this feature4, don't even need to write
|
||||
uint feature4_id = (group_id >> POWER_FEATURE_WORKGROUPS);
|
||||
|
|
@ -704,7 +704,7 @@ R""()
|
|||
output[0 * DWORD_FEATURES * NUM_BINS + ltid] = stat_val;
|
||||
barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);
|
||||
mem_fence(CLK_GLOBAL_MEM_FENCE);
|
||||
// To avoid the cost of an extra reducing kernel, we have to deal with some
|
||||
// To avoid the cost of an extra reducing kernel, we have to deal with some
|
||||
// gray area in OpenCL. We want the last work group that process this feature to
|
||||
// make the final reduction, and other threads will just quit.
|
||||
// This requires that the results written by other workgroups available to the
|
||||
|
|
@ -750,13 +750,13 @@ R""()
|
|||
#endif
|
||||
// locate our feature4's block in output memory
|
||||
uint output_offset = (feature4_id << POWER_FEATURE_WORKGROUPS);
|
||||
__global acc_type const * restrict feature4_subhists =
|
||||
__global acc_type const * restrict feature4_subhists =
|
||||
(__global acc_type *)output_buf + output_offset * DWORD_FEATURES * 2 * NUM_BINS;
|
||||
// skip reading the data already in local memory
|
||||
uint skip_id = group_id ^ output_offset;
|
||||
// locate output histogram location for this feature4
|
||||
__global acc_type* restrict hist_buf = hist_buf_base + feature4_id * DWORD_FEATURES * 2 * NUM_BINS;
|
||||
within_kernel_reduction16x8(feature_mask, feature4_subhists, skip_id, stat_val,
|
||||
within_kernel_reduction16x8(feature_mask, feature4_subhists, skip_id, stat_val,
|
||||
1 << POWER_FEATURE_WORKGROUPS, hist_buf, (__local acc_type *)shared_array);
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -47,12 +47,12 @@ typedef uint acc_int_type;
|
|||
#endif
|
||||
#define LOCAL_MEM_SIZE (4 * (sizeof(uint) + 2 * sizeof(acc_type)) * NUM_BINS)
|
||||
|
||||
// unroll the atomic operation for a few times. Takes more code space,
|
||||
// unroll the atomic operation for a few times. Takes more code space,
|
||||
// but compiler can generate better code for faster atomics.
|
||||
#define UNROLL_ATOMIC 1
|
||||
|
||||
// Options passed by compiler at run time:
|
||||
// IGNORE_INDICES will be set when the kernel does not
|
||||
// IGNORE_INDICES will be set when the kernel does not
|
||||
// #define IGNORE_INDICES
|
||||
// #define POWER_FEATURE_WORKGROUPS 10
|
||||
|
||||
|
|
@ -137,7 +137,7 @@ R""()
|
|||
// this function will be called by histogram256
|
||||
// we have one sub-histogram of one feature in local memory, and need to read others
|
||||
void within_kernel_reduction256x4(uchar4 feature_mask,
|
||||
__global const acc_type* restrict feature4_sub_hist,
|
||||
__global const acc_type* restrict feature4_sub_hist,
|
||||
const uint skip_id,
|
||||
const uint old_val_f0_cont_bin0,
|
||||
const ushort num_sub_hist,
|
||||
|
|
@ -314,12 +314,12 @@ R""()
|
|||
*/
|
||||
__attribute__((reqd_work_group_size(LOCAL_SIZE_0, 1, 1)))
|
||||
#if USE_CONSTANT_BUF == 1
|
||||
__kernel void histogram256(__global const uchar4* restrict feature_data_base,
|
||||
__kernel void histogram256(__global const uchar4* restrict feature_data_base,
|
||||
__constant const uchar4* restrict feature_masks __attribute__((max_constant_size(65536))),
|
||||
const data_size_t feature_size,
|
||||
__constant const data_size_t* restrict data_indices __attribute__((max_constant_size(65536))),
|
||||
const data_size_t num_data,
|
||||
__constant const score_t* restrict ordered_gradients __attribute__((max_constant_size(65536))),
|
||||
__constant const data_size_t* restrict data_indices __attribute__((max_constant_size(65536))),
|
||||
const data_size_t num_data,
|
||||
__constant const score_t* restrict ordered_gradients __attribute__((max_constant_size(65536))),
|
||||
#if CONST_HESSIAN == 0
|
||||
__constant const score_t* restrict ordered_hessians __attribute__((max_constant_size(65536))),
|
||||
#else
|
||||
|
|
@ -329,18 +329,18 @@ __kernel void histogram256(__global const uchar4* restrict feature_data_base,
|
|||
__global volatile int * sync_counters,
|
||||
__global acc_type* restrict hist_buf_base) {
|
||||
#else
|
||||
__kernel void histogram256(__global const uchar4* feature_data_base,
|
||||
__kernel void histogram256(__global const uchar4* feature_data_base,
|
||||
__constant const uchar4* restrict feature_masks __attribute__((max_constant_size(65536))),
|
||||
const data_size_t feature_size,
|
||||
__global const data_size_t* data_indices,
|
||||
const data_size_t num_data,
|
||||
__global const score_t* ordered_gradients,
|
||||
__global const data_size_t* data_indices,
|
||||
const data_size_t num_data,
|
||||
__global const score_t* ordered_gradients,
|
||||
#if CONST_HESSIAN == 0
|
||||
__global const score_t* ordered_hessians,
|
||||
#else
|
||||
const score_t const_hessian,
|
||||
#endif
|
||||
__global char* restrict output_buf,
|
||||
__global char* restrict output_buf,
|
||||
__global volatile int * sync_counters,
|
||||
__global acc_type* restrict hist_buf_base) {
|
||||
#endif
|
||||
|
|
@ -363,20 +363,20 @@ __kernel void histogram256(__global const uchar4* feature_data_base,
|
|||
// gradient/hessian histograms
|
||||
// assume this starts at 32 * 4 = 128-byte boundary
|
||||
// total size: 2 * 4 * 256 * size_of(float) = 8 KB
|
||||
// organization: each feature/grad/hessian is at a different bank,
|
||||
// organization: each feature/grad/hessian is at a different bank,
|
||||
// as independent of the feature value as possible
|
||||
__local acc_type * gh_hist = (__local acc_type *)shared_array;
|
||||
// counter histogram
|
||||
// total size: 4 * 256 * size_of(uint) = 4 KB
|
||||
#if CONST_HESSIAN == 1
|
||||
__local uint * cnt_hist = (__local uint *)(gh_hist + 2 * 4 * NUM_BINS);
|
||||
#endif
|
||||
#endif
|
||||
|
||||
// thread 0, 1, 2, 3 compute histograms for gradients first
|
||||
// thread 4, 5, 6, 7 compute histograms for Hessians first
|
||||
// etc.
|
||||
uchar is_hessian_first = (ltid >> 2) & 1;
|
||||
|
||||
|
||||
ushort group_feature = group_id >> POWER_FEATURE_WORKGROUPS;
|
||||
// each 2^POWER_FEATURE_WORKGROUPS workgroups process on one feature (compile-time constant)
|
||||
// feature_size is the number of examples per feature
|
||||
|
|
@ -725,7 +725,7 @@ R""()
|
|||
}
|
||||
barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);
|
||||
mem_fence(CLK_GLOBAL_MEM_FENCE);
|
||||
// To avoid the cost of an extra reducing kernel, we have to deal with some
|
||||
// To avoid the cost of an extra reducing kernel, we have to deal with some
|
||||
// gray area in OpenCL. We want the last work group that process this feature to
|
||||
// make the final reduction, and other threads will just quit.
|
||||
// This requires that the results written by other workgroups available to the
|
||||
|
|
@ -773,15 +773,15 @@ R""()
|
|||
#endif
|
||||
// locate our feature4's block in output memory
|
||||
uint output_offset = (feature4_id << POWER_FEATURE_WORKGROUPS);
|
||||
__global acc_type const * restrict feature4_subhists =
|
||||
__global acc_type const * restrict feature4_subhists =
|
||||
(__global acc_type *)output_buf + output_offset * 4 * 2 * NUM_BINS;
|
||||
// skip reading the data already in local memory
|
||||
uint skip_id = group_id ^ output_offset;
|
||||
// locate output histogram location for this feature4
|
||||
__global acc_type* restrict hist_buf = hist_buf_base + feature4_id * 4 * 2 * NUM_BINS;
|
||||
within_kernel_reduction256x4(feature_mask, feature4_subhists, skip_id, old_val, 1 << POWER_FEATURE_WORKGROUPS,
|
||||
within_kernel_reduction256x4(feature_mask, feature4_subhists, skip_id, old_val, 1 << POWER_FEATURE_WORKGROUPS,
|
||||
hist_buf, (__local acc_type *)shared_array);
|
||||
// if (ltid == 0)
|
||||
// if (ltid == 0)
|
||||
// printf("workgroup %d reduction done, %g %g %g %g %g %g %g %g\n", group_id, hist_buf[0], hist_buf[3*NUM_BINS], hist_buf[2*3*NUM_BINS], hist_buf[3*3*NUM_BINS], hist_buf[1], hist_buf[3*NUM_BINS+1], hist_buf[2*3*NUM_BINS+1], hist_buf[3*3*NUM_BINS+1]);
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -65,12 +65,12 @@ typedef uint acc_int_type;
|
|||
// local memory size in bytes
|
||||
#define LOCAL_MEM_SIZE (4 * (sizeof(uint) + 2 * sizeof(acc_type)) * NUM_BINS * NUM_BANKS)
|
||||
|
||||
// unroll the atomic operation for a few times. Takes more code space,
|
||||
// unroll the atomic operation for a few times. Takes more code space,
|
||||
// but compiler can generate better code for faster atomics.
|
||||
#define UNROLL_ATOMIC 1
|
||||
|
||||
// Options passed by compiler at run time:
|
||||
// IGNORE_INDICES will be set when the kernel does not
|
||||
// IGNORE_INDICES will be set when the kernel does not
|
||||
// #define IGNORE_INDICES
|
||||
// #define POWER_FEATURE_WORKGROUPS 10
|
||||
|
||||
|
|
@ -155,7 +155,7 @@ R""()
|
|||
// this function will be called by histogram64
|
||||
// we have one sub-histogram of one feature in registers, and need to read others
|
||||
void within_kernel_reduction64x4(uchar4 feature_mask,
|
||||
__global const acc_type* restrict feature4_sub_hist,
|
||||
__global const acc_type* restrict feature4_sub_hist,
|
||||
const uint skip_id,
|
||||
acc_type g_val, acc_type h_val,
|
||||
const ushort num_sub_hist,
|
||||
|
|
@ -166,7 +166,7 @@ void within_kernel_reduction64x4(uchar4 feature_mask,
|
|||
ushort feature_id = ltid & 3; // range 0 - 4
|
||||
const ushort bin_id = ltid >> 2; // range 0 - 63W
|
||||
ushort i;
|
||||
#if POWER_FEATURE_WORKGROUPS != 0
|
||||
#if POWER_FEATURE_WORKGROUPS != 0
|
||||
// if there is only 1 work group, no need to do the reduction
|
||||
// add all sub-histograms for 4 features
|
||||
__global const acc_type* restrict p = feature4_sub_hist + ltid;
|
||||
|
|
@ -212,12 +212,12 @@ R""()
|
|||
*/
|
||||
__attribute__((reqd_work_group_size(LOCAL_SIZE_0, 1, 1)))
|
||||
#if USE_CONSTANT_BUF == 1
|
||||
__kernel void histogram64(__global const uchar4* restrict feature_data_base,
|
||||
__kernel void histogram64(__global const uchar4* restrict feature_data_base,
|
||||
__constant const uchar4* restrict feature_masks __attribute__((max_constant_size(65536))),
|
||||
const data_size_t feature_size,
|
||||
__constant const data_size_t* restrict data_indices __attribute__((max_constant_size(65536))),
|
||||
const data_size_t num_data,
|
||||
__constant const score_t* restrict ordered_gradients __attribute__((max_constant_size(65536))),
|
||||
__constant const data_size_t* restrict data_indices __attribute__((max_constant_size(65536))),
|
||||
const data_size_t num_data,
|
||||
__constant const score_t* restrict ordered_gradients __attribute__((max_constant_size(65536))),
|
||||
#if CONST_HESSIAN == 0
|
||||
__constant const score_t* restrict ordered_hessians __attribute__((max_constant_size(65536))),
|
||||
#else
|
||||
|
|
@ -227,18 +227,18 @@ __kernel void histogram64(__global const uchar4* restrict feature_data_base,
|
|||
__global volatile int * sync_counters,
|
||||
__global acc_type* restrict hist_buf_base) {
|
||||
#else
|
||||
__kernel void histogram64(__global const uchar4* feature_data_base,
|
||||
__kernel void histogram64(__global const uchar4* feature_data_base,
|
||||
__constant const uchar4* restrict feature_masks __attribute__((max_constant_size(65536))),
|
||||
const data_size_t feature_size,
|
||||
__global const data_size_t* data_indices,
|
||||
const data_size_t num_data,
|
||||
__global const score_t* ordered_gradients,
|
||||
__global const data_size_t* data_indices,
|
||||
const data_size_t num_data,
|
||||
__global const score_t* ordered_gradients,
|
||||
#if CONST_HESSIAN == 0
|
||||
__global const score_t* ordered_hessians,
|
||||
#else
|
||||
const score_t const_hessian,
|
||||
#endif
|
||||
__global char* restrict output_buf,
|
||||
__global char* restrict output_buf,
|
||||
__global volatile int * sync_counters,
|
||||
__global acc_type* restrict hist_buf_base) {
|
||||
#endif
|
||||
|
|
@ -313,7 +313,7 @@ __kernel void histogram64(__global const uchar4* feature_data_base,
|
|||
uchar is_hessian_first = (ltid >> 2) & 1;
|
||||
// thread 0-7 write result to bank0, 8-15 to bank1, 16-23 to bank2, 24-31 to bank3
|
||||
ushort bank = (ltid >> 3) & BANK_MASK;
|
||||
|
||||
|
||||
ushort group_feature = group_id >> POWER_FEATURE_WORKGROUPS;
|
||||
// each 2^POWER_FEATURE_WORKGROUPS workgroups process on one feature (compile-time constant)
|
||||
// feature_size is the number of examples per feature
|
||||
|
|
@ -582,7 +582,7 @@ R""()
|
|||
atomic_local_add_f(gh_hist + addr2, s0_stat2);
|
||||
#endif
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
|
||||
/* Makes MSVC happy with long string literal
|
||||
)""
|
||||
R""()
|
||||
|
|
@ -591,7 +591,7 @@ R""()
|
|||
// restore feature_mask
|
||||
feature_mask = feature_masks[group_feature];
|
||||
#endif
|
||||
|
||||
|
||||
// now reduce the 4 banks of subhistograms into 1
|
||||
/* memory layout of gh_hist:
|
||||
-----------------------------------------------------------------------------------------------
|
||||
|
|
@ -680,7 +680,7 @@ R""()
|
|||
output[1 * 4 * NUM_BINS + ltid] = h_val;
|
||||
barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);
|
||||
mem_fence(CLK_GLOBAL_MEM_FENCE);
|
||||
// To avoid the cost of an extra reducing kernel, we have to deal with some
|
||||
// To avoid the cost of an extra reducing kernel, we have to deal with some
|
||||
// gray area in OpenCL. We want the last work group that process this feature to
|
||||
// make the final reduction, and other threads will just quit.
|
||||
// This requires that the results written by other workgroups available to the
|
||||
|
|
@ -726,13 +726,13 @@ R""()
|
|||
#endif
|
||||
// locate our feature4's block in output memory
|
||||
uint output_offset = (feature4_id << POWER_FEATURE_WORKGROUPS);
|
||||
__global acc_type const * restrict feature4_subhists =
|
||||
__global acc_type const * restrict feature4_subhists =
|
||||
(__global acc_type *)output_buf + output_offset * 4 * 2 * NUM_BINS;
|
||||
// skip reading the data already in local memory
|
||||
uint skip_id = group_id ^ output_offset;
|
||||
// locate output histogram location for this feature4
|
||||
__global acc_type* restrict hist_buf = hist_buf_base + feature4_id * 4 * 2 * NUM_BINS;
|
||||
within_kernel_reduction64x4(feature_mask, feature4_subhists, skip_id, g_val, h_val,
|
||||
within_kernel_reduction64x4(feature_mask, feature4_subhists, skip_id, g_val, h_val,
|
||||
1 << POWER_FEATURE_WORKGROUPS, hist_buf, (__local acc_type *)shared_array);
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -6,7 +6,7 @@
|
|||
*/
|
||||
/**
|
||||
* This wraps the StringArray.hpp class for SWIG usage,
|
||||
* adding the basic C-style wrappers needed to make it
|
||||
* adding the basic C-style wrappers needed to make it
|
||||
* usable for the users of the low-level lightgbmJNI API.
|
||||
*/
|
||||
|
||||
|
|
@ -23,7 +23,7 @@
|
|||
|
||||
/**
|
||||
* @brief Creates a new StringArray and returns its handle.
|
||||
*
|
||||
*
|
||||
* @param num_strings number of strings to store.
|
||||
* @param string_size the maximum number of characters that can be stored in each string.
|
||||
* @return StringArrayHandle or nullptr in case of allocation failure.
|
||||
|
|
@ -38,7 +38,7 @@
|
|||
|
||||
/**
|
||||
* @brief Free the StringArray object.
|
||||
*
|
||||
*
|
||||
* @param handle StringArray handle.
|
||||
*/
|
||||
void StringArrayHandle_free(StringArrayHandle handle)
|
||||
|
|
@ -49,7 +49,7 @@
|
|||
/**
|
||||
* @brief Return the raw pointer to the array of strings.
|
||||
* Wrapped in Java into String[] automatically.
|
||||
*
|
||||
*
|
||||
* @param handle StringArray handle.
|
||||
* @return Raw pointer to the string array which `various.i` maps to String[].
|
||||
*/
|
||||
|
|
@ -60,7 +60,7 @@
|
|||
|
||||
/**
|
||||
* For the end user to extract a specific string from the StringArray object.
|
||||
*
|
||||
*
|
||||
* @param handle StringArray handle.
|
||||
* @param index index of the string to retrieve from the array.
|
||||
* @return raw pointer to string at index, or nullptr if out of bounds.
|
||||
|
|
@ -72,7 +72,7 @@
|
|||
|
||||
/**
|
||||
* @brief Replaces one string of the array at index with the new content.
|
||||
*
|
||||
*
|
||||
* @param handle StringArray handle.
|
||||
* @param index Index of the string to replace
|
||||
* @param new_content The content to replace
|
||||
|
|
@ -85,7 +85,7 @@
|
|||
|
||||
/**
|
||||
* @brief Retrieve the number of strings in the StringArray.
|
||||
*
|
||||
*
|
||||
* @param handle StringArray handle.
|
||||
* @return number of strings that the array stores.
|
||||
*/
|
||||
|
|
|
|||
|
|
@ -6,11 +6,11 @@
|
|||
* This SWIG interface extension provides support to
|
||||
* the pointer manipulation methods present in the standard
|
||||
* SWIG wrappers, but with support for larger arrays.
|
||||
*
|
||||
*
|
||||
* SWIG provides this in https://github.com/swig/swig/blob/master/Lib/carrays.i
|
||||
* but the standard methods only provide arrays with up to
|
||||
* max(int32_t) elements.
|
||||
*
|
||||
*
|
||||
* The `long_array_functions` wrappers extend this
|
||||
* to arrays of size max(int64_t) instead of max(int32_t).
|
||||
*/
|
||||
|
|
@ -103,7 +103,7 @@ void delete_##NAME(TYPE *ary);
|
|||
TYPE NAME##_getitem(TYPE *ary, int64_t index);
|
||||
void NAME##_setitem(TYPE *ary, int64_t index, TYPE value);
|
||||
|
||||
%enddef
|
||||
%enddef
|
||||
|
||||
/* Custom template for arrays of pointers */
|
||||
%define %ptr_array_functions(TYPE,NAME)
|
||||
|
|
|
|||
|
|
@ -277,7 +277,7 @@ namespace LightGBM {
|
|||
/*!
|
||||
* Pushes data from 1 thread into a Dataset based on thread_id and nrows.
|
||||
* e.g. with 100 rows, thread 0 will push rows 0-49, and thread 2 will push rows 50-99.
|
||||
* Note that rows are still pushed in microbatches within their range.
|
||||
* Note that rows are still pushed in microbatches within their range.
|
||||
*/
|
||||
void TestUtils::PushSparseBatch(DatasetHandle dataset_handle,
|
||||
int32_t nrows,
|
||||
|
|
|
|||
|
|
@ -351,4 +351,4 @@
|
|||
<Filter>src\utils</Filter>
|
||||
</ClCompile>
|
||||
</ItemGroup>
|
||||
</Project>
|
||||
</Project>
|
||||
|
|
|
|||
Загрузка…
Ссылка в новой задаче