Skip to content

GitLab

"csrc/gfx93/decode/sparse_fp8/components/config.h" did not exist on "082094b793fcc7452977d0a71a00e266a2e3061e"
Commit 3f1316d5 authored by traveller59's avatar traveller59
Browse files

initial release

parent a347176a
Show whitespace changes
Inline Side-by-side
Showing with 2484 additions and 2 deletions
.gitignore 0 → 100644
View file @ 3f1316d5
# Byte-compiled / optimized / DLL files
__pycache__/
*.py[cod]
*$py.class
# C extensions
*.so
*.o
*.out
# Distribution / packaging
.Python
build/
develop-eggs/
dist/
downloads/
eggs/
.eggs/
lib/
lib64/
parts/
sdist/
var/
wheels/
*.egg-info/
.installed.cfg
*.egg
MANIFEST
# PyInstaller
# Usually these files are written by a python script from a template
# before PyInstaller builds the exe, so as to inject date/other infos into it.
*.manifest
*.spec
# Installer logs
pip-log.txt
pip-delete-this-directory.txt
# Unit test / coverage reports
htmlcov/
.tox/
.coverage
.coverage.*
.cache
nosetests.xml
coverage.xml
*.cover
.hypothesis/
.pytest_cache/
# Translations
*.mo
*.pot
# Django stuff:
*.log
local_settings.py
db.sqlite3
# Flask stuff:
instance/
.webassets-cache
# Scrapy stuff:
.scrapy
# Sphinx documentation
docs/_build/
# PyBuilder
target/
# Jupyter Notebook
.ipynb_checkpoints
# pyenv
.python-version
# celery beat schedule file
celerybeat-schedule
# SageMath parsed files
*.sage.py
# Environments
.env
.venv
env/
venv/
ENV/
env.bak/
venv.bak/
# Spyder project settings
.spyderproject
.spyproject
# Rope project settings
.ropeproject
# mkdocs documentation
/site
# mypy
.mypy_cache/
.vscode
.gitmodules 0 → 100644
View file @ 3f1316d5
[submodule "third_party/pybind11"]
path = third_party/pybind11
url = https://github.com/pybind/pybind11.git
CMakeLists.txt 0 → 100644
View file @ 3f1316d5
cmake_minimum_required(VERSION 3.13 FATAL_ERROR)
project(SparseConv LANGUAGES CXX CUDA VERSION 1.0)
option(SPCONV_BuildTests "Build the unit tests when BUILD_TESTING is enabled." ON)
set(CMAKE_CXX_EXTENSIONS OFF) # avoid gnu++11 be added to CXX flags
set(CUDA_TOOLKIT_ROOT_DIR "${CMAKE_CUDA_COMPILER}")
get_filename_component(CUDA_TOOLKIT_ROOT_DIR "${CUDA_TOOLKIT_ROOT_DIR}" DIRECTORY)
get_filename_component(CUDA_TOOLKIT_ROOT_DIR "${CUDA_TOOLKIT_ROOT_DIR}" DIRECTORY)
if(WIN32) # true if windows (32 and 64 bit)
set(CUDA_LIB_PATH_HINTS "${CUDA_TOOLKIT_ROOT_DIR}/lib/x64")
add_compile_definitions(TV_WINDOWS)
else()
set(CUDA_LIB_PATH_HINTS "${CUDA_TOOLKIT_ROOT_DIR}/lib64")
endif()
find_library(CUDA_CUDART NAMES cudart HINTS ${CUDA_LIB_PATH_HINTS})
find_library(CUDA_CUBLAS NAMES cublas HINTS ${CUDA_LIB_PATH_HINTS})
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
add_compile_definitions(TV_DEBUG)
endif()
find_package(Torch REQUIRED)
add_definitions(-D_GLIBCXX_USE_CXX11_ABI=0)
add_compile_definitions(SPCONV_CUDA)
add_subdirectory(third_party/pybind11)
set(ALL_LIBS ${CUDA_CUDART} ${CUDA_CUBLAS} ${TORCH_LIBRARIES})
set(ALL_INCLUDE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
${PROJECT_SOURCE_DIR}/include)
add_subdirectory(src/spconv)
add_subdirectory(src/utils)
if (SPCONV_BuildTests)
include(CTest) #adds option BUILD_TESTING (default ON)
if(BUILD_TESTING)
enable_testing()
add_subdirectory(test)
endif()
endif()
LICENSE 0 → 100644
View file @ 3f1316d5
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright 2019 Yan Yan
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
\ No newline at end of file
README.md
View file @ 3f1316d5
# spconv
Spatial Sparse Convolution in PyTorch
# SpConv: PyTorch Spatially Sparse Convolution Library
This is a spatially sparse convolution library like [SparseConvNet](https://github.com/facebookresearch/SparseConvNet) but faster and easy to read. This library provide sparse convolution/transposed, submanifold convolution, inverse convolution and sparse maxpool.
If you need more kinds of spatial layers such as avg pool, please implement it by yourself, I don't have time to do this.
## Install
0. Use ```git clone xxx.git --recursive``` to clone this repo.
1. Download stable LibTorch from pytorch [website](https://pytorch.org/).
2. Set an environment variable ```LIBTORCH_ROOT``` to the root path of LibTorch.
3. Install boost headers to your system include path, you can use either ```sudo apt-get install libboostall-dev``` or download compressed files from boost official website and copy headers to include path.
4. Download cmake >= 3.13.2, then add cmake executables to PATH.
5. run ```python setup.py install```.
## Compare with SparseConvNet
### Features
* SparseConvNet's Sparse Convolution don't support padding and dilation, spconv support this.
* spconv only contains sparse convolutions, the batchnorm and activations can directly use layers from torch.nn, SparseConvNet contains lots of their own implementation of layers such as batchnorm and activations.
### Speed
* spconv is faster than SparseConvNet due to gpu indice generation and gather-gemm-scatter algorithm. SparseConvNet use hand-written gemm which is slow.
## Usage
### SparseConvTensor
```Python
features = # your features with shape [N, numPlanes]
indices = # your indices/coordinates with shape [N, ndim + 1], batch index must be put in indices[:, 0]
spatial_shape = # spatial shape of your sparse tensor.
batch_size = # batch size of your sparse tensor.
x = spconv.SparseConvTensor(features, indices, spatial_shape, batch_size)
x_dense_NHWC = x.dense() # convert sparse tensor to dense NCHW tensor.
print(x.sparity) # helper function to check sparity.
```
### Sparse Convolution
```Python
import spconv
from torch import nn
class ExampleNet(nn.Module):
def __init__(self, shape):
super().__init__()
self.net = spconv.SparseSequential(
spconv.SparseConv3d(32, 64, 3), # just like nn.Conv3d but don't support group and all([d > 1, s > 1])
nn.BatchNorm1d(64), # non-spatial layers can be used directly in SparseSequential.
nn.ReLU(),
spconv.SubMConv3d(64, 64, 3, indice_key="subm0"),
nn.BatchNorm1d(64),
nn.ReLU(),
# when use submanifold convolutions, their indices can be shared to save indices generation time.
spconv.SubMConv3d(64, 64, 3, indice_key="subm0"),
nn.BatchNorm1d(64),
nn.ReLU(),
spconv.SparseConvTranspose3d(64, 64, 3, 2),
nn.BatchNorm1d(64),
nn.ReLU(),
spconv.ToDense(), # convert spconv tensor to dense and convert it to NCHW format.
nn.Conv3d(64, 64, 3),
nn.BatchNorm1d(64),
nn.ReLU(),
)
self.shape = shape
def forward(self, features, coors, batch_size):
coors = coors.int() # unlike torch, this library only accept int coordinates.
x = spconv.SparseConvTensor(features, coors, self.shape, batch_size)
return self.net(x)# .dense()
```
### Inverse Convolution
Inverse sparse convolution means "inv" of sparse convolution. the output of inverse convolution contains same indices as input of sparse convolution.
Inverse convolution usually used in semantic segmentation.
```Python
class ExampleNet(nn.Module):
def __init__(self, shape):
super().__init__()
self.net = spconv.SparseSequential(
spconv.SparseConv3d(32, 64, 3, 2, indice_key="cp0"),
spconv.SparseInverseConv3d(64, 32, 3, 2,indice_key="cp0"),
)
self.shape = shape
def forward(self, features, coors, batch_size):
coors = coors.int()
x = spconv.SparseConvTensor(features, coors, self.shape, batch_size)
return self.net(x)
```
### Utility functions
* convert point cloud to voxel
```Python
voxel_generator = spconv.utils.VoxelGenerator(
voxel_size=[0.1, 0.1, 0.1],
point_cloud_range=[-50, -50, -3, 50, 50, 1],
max_num_points=30,
max_voxels=40000
)
points = # [N, 3+] tensor.
voxels, coords, num_points_per_voxel = voxel_generator.generate(points)
```
## Implementation Details
This implementation use gather-gemm-scatter framework to do sparse convolution.
## Projects using spconv:
* [second.pytorch](https://github.com/traveller59/second.pytorch): Point Cloud Object Detection in KITTI Dataset.
## Authors
* **Yan Yan** - *Initial work* - [traveller59](https://github.com/traveller59)
* **Bo Li** - *gpu indice generation idea* - [prclibo](https://github.com/prclibo)
## License
This project is licensed under the Apache license 2.0 License - see the [LICENSE.md](LICENSE.md) file for details
include/paramsgrid.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef PARAMS_GRID_H_
#define PARAMS_GRID_H_
#include <tuple>
#include <vector>
namespace detail {
template <class T> int getTotalSize(std::vector<T> arg) { return arg.size(); }
template <class T, class... TArgs>
int getTotalSize(std::vector<T> arg, std::vector<TArgs>... args) {
return arg.size() * getTotalSize(args...);
}
template <typename T> int getSize(std::vector<T> arg) { return arg.size(); }
template <int Idx, class TT, class T>
void assigner(TT &src, std::vector<int> counter, std::vector<T> &arg) {
std::get<Idx>(src) = arg[counter[Idx]];
}
template <int Idx, class TT, class T, class... TArgs>
void assigner(TT &src, std::vector<int> counter, std::vector<T> &arg,
std::vector<TArgs> &... args) {
std::get<Idx>(src) = arg[counter[Idx]];
assigner<Idx + 1>(src, counter, args...);
}
} // namespace detail
template <class... TArgs>
std::vector<std::tuple<TArgs...>> paramsGrid(std::vector<TArgs>... args) {
int length = detail::getTotalSize(args...);
std::vector<int> sizes = {detail::getSize(args)...};
int size = sizes.size();
std::vector<std::tuple<TArgs...>> params(length);
std::vector<int> counter(size);
for (int i = 0; i < length; ++i) {
detail::assigner<0>(params[i], counter, args...);
counter[size - 1] += 1;
for (int c = size - 1; c >= 0; --c) {
if (counter[c] == sizes[c] && c > 0) {
counter[c - 1] += 1;
counter[c] = 0;
}
}
}
return params;
}
#endif
\ No newline at end of file
include/prettyprint.h 0 → 100644
View file @ 3f1316d5
// Copyright Louis Delacroix 2010 - 2014.
// Distributed under the Boost Software License, Version 1.0.
// (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)
//
// A pretty printing library for C++
//
// Usage:
// Include this header, and operator<< will "just work".
#ifndef H_PRETTY_PRINT
#define H_PRETTY_PRINT
#include <cstddef>
#include <iterator>
#include <memory>
#include <ostream>
#include <set>
#include <tuple>
#include <type_traits>
#include <unordered_set>
#include <utility>
#include <valarray>
namespace pretty_print
{
namespace detail
{
// SFINAE type trait to detect whether T::const_iterator exists.
struct sfinae_base
{
using yes = char;
using no = yes[2];
};
template <typename T>
struct has_const_iterator : private sfinae_base
{
private:
template <typename C> static yes & test(typename C::const_iterator*);
template <typename C> static no & test(...);
public:
static const bool value = sizeof(test<T>(nullptr)) == sizeof(yes);
using type = T;
};
template <typename T>
struct has_begin_end : private sfinae_base
{
private:
template <typename C>
static yes & f(typename std::enable_if<
std::is_same<decltype(static_cast<typename C::const_iterator(C::*)() const>(&C::begin)),
typename C::const_iterator(C::*)() const>::value>::type *);
template <typename C> static no & f(...);
template <typename C>
static yes & g(typename std::enable_if<
std::is_same<decltype(static_cast<typename C::const_iterator(C::*)() const>(&C::end)),
typename C::const_iterator(C::*)() const>::value, void>::type*);
template <typename C> static no & g(...);
public:
static bool const beg_value = sizeof(f<T>(nullptr)) == sizeof(yes);
static bool const end_value = sizeof(g<T>(nullptr)) == sizeof(yes);
};
} // namespace detail
// Holds the delimiter values for a specific character type
template <typename TChar>
struct delimiters_values
{
using char_type = TChar;
const char_type * prefix;
const char_type * delimiter;
const char_type * postfix;
};
// Defines the delimiter values for a specific container and character type
template <typename T, typename TChar>
struct delimiters
{
using type = delimiters_values<TChar>;
static const type values;
};
// Functor to print containers. You can use this directly if you want
// to specificy a non-default delimiters type. The printing logic can
// be customized by specializing the nested template.
template <typename T,
typename TChar = char,
typename TCharTraits = ::std::char_traits<TChar>,
typename TDelimiters = delimiters<T, TChar>>
struct print_container_helper
{
using delimiters_type = TDelimiters;
using ostream_type = std::basic_ostream<TChar, TCharTraits>;
template <typename U>
struct printer
{
static void print_body(const U & c, ostream_type & stream)
{
using std::begin;
using std::end;
auto it = begin(c);
const auto the_end = end(c);
if (it != the_end)
{
for ( ; ; )
{
stream << *it;
if (++it == the_end) break;
if (delimiters_type::values.delimiter != NULL)
stream << delimiters_type::values.delimiter;
}
}
}
};
print_container_helper(const T & container)
: container_(container)
{ }
inline void operator()(ostream_type & stream) const
{
if (delimiters_type::values.prefix != NULL)
stream << delimiters_type::values.prefix;
printer<T>::print_body(container_, stream);
if (delimiters_type::values.postfix != NULL)
stream << delimiters_type::values.postfix;
}
private:
const T & container_;
};
// Specialization for pairs
template <typename T, typename TChar, typename TCharTraits, typename TDelimiters>
template <typename T1, typename T2>
struct print_container_helper<T, TChar, TCharTraits, TDelimiters>::printer<std::pair<T1, T2>>
{
using ostream_type = typename print_container_helper<T, TChar, TCharTraits, TDelimiters>::ostream_type;
static void print_body(const std::pair<T1, T2> & c, ostream_type & stream)
{
stream << c.first;
if (print_container_helper<T, TChar, TCharTraits, TDelimiters>::delimiters_type::values.delimiter != NULL)
stream << print_container_helper<T, TChar, TCharTraits, TDelimiters>::delimiters_type::values.delimiter;
stream << c.second;
}
};
// Specialization for tuples
template <typename T, typename TChar, typename TCharTraits, typename TDelimiters>
template <typename ...Args>
struct print_container_helper<T, TChar, TCharTraits, TDelimiters>::printer<std::tuple<Args...>>
{
using ostream_type = typename print_container_helper<T, TChar, TCharTraits, TDelimiters>::ostream_type;
using element_type = std::tuple<Args...>;
template <std::size_t I> struct Int { };
static void print_body(const element_type & c, ostream_type & stream)
{
tuple_print(c, stream, Int<0>());
}
static void tuple_print(const element_type &, ostream_type &, Int<sizeof...(Args)>)
{
}
static void tuple_print(const element_type & c, ostream_type & stream,
typename std::conditional<sizeof...(Args) != 0, Int<0>, std::nullptr_t>::type)
{
stream << std::get<0>(c);
tuple_print(c, stream, Int<1>());
}
template <std::size_t N>
static void tuple_print(const element_type & c, ostream_type & stream, Int<N>)
{
if (print_container_helper<T, TChar, TCharTraits, TDelimiters>::delimiters_type::values.delimiter != NULL)
stream << print_container_helper<T, TChar, TCharTraits, TDelimiters>::delimiters_type::values.delimiter;
stream << std::get<N>(c);
tuple_print(c, stream, Int<N + 1>());
}
};
// Prints a print_container_helper to the specified stream.
template<typename T, typename TChar, typename TCharTraits, typename TDelimiters>
inline std::basic_ostream<TChar, TCharTraits> & operator<<(
std::basic_ostream<TChar, TCharTraits> & stream,
const print_container_helper<T, TChar, TCharTraits, TDelimiters> & helper)
{
helper(stream);
return stream;
}
// Basic is_container template; specialize to derive from std::true_type for all desired container types
template <typename T>
struct is_container : public std::integral_constant<bool,
detail::has_const_iterator<T>::value &&
detail::has_begin_end<T>::beg_value &&
detail::has_begin_end<T>::end_value> { };
template <typename T, std::size_t N>
struct is_container<T[N]> : std::true_type { };
template <std::size_t N>
struct is_container<char[N]> : std::false_type { };
template <typename T>
struct is_container<std::valarray<T>> : std::true_type { };
template <typename T1, typename T2>
struct is_container<std::pair<T1, T2>> : std::true_type { };
template <typename ...Args>
struct is_container<std::tuple<Args...>> : std::true_type { };
// Default delimiters
template <typename T> struct delimiters<T, char> { static const delimiters_values<char> values; };
template <typename T> const delimiters_values<char> delimiters<T, char>::values = { "[", ", ", "]" };
template <typename T> struct delimiters<T, wchar_t> { static const delimiters_values<wchar_t> values; };
template <typename T> const delimiters_values<wchar_t> delimiters<T, wchar_t>::values = { L"[", L", ", L"]" };
// Delimiters for (multi)set and unordered_(multi)set
template <typename T, typename TComp, typename TAllocator>
struct delimiters< ::std::set<T, TComp, TAllocator>, char> { static const delimiters_values<char> values; };
template <typename T, typename TComp, typename TAllocator>
const delimiters_values<char> delimiters< ::std::set<T, TComp, TAllocator>, char>::values = { "{", ", ", "}" };
template <typename T, typename TComp, typename TAllocator>
struct delimiters< ::std::set<T, TComp, TAllocator>, wchar_t> { static const delimiters_values<wchar_t> values; };
template <typename T, typename TComp, typename TAllocator>
const delimiters_values<wchar_t> delimiters< ::std::set<T, TComp, TAllocator>, wchar_t>::values = { L"{", L", ", L"}" };
template <typename T, typename TComp, typename TAllocator>
struct delimiters< ::std::multiset<T, TComp, TAllocator>, char> { static const delimiters_values<char> values; };
template <typename T, typename TComp, typename TAllocator>
const delimiters_values<char> delimiters< ::std::multiset<T, TComp, TAllocator>, char>::values = { "{", ", ", "}" };
template <typename T, typename TComp, typename TAllocator>
struct delimiters< ::std::multiset<T, TComp, TAllocator>, wchar_t> { static const delimiters_values<wchar_t> values; };
template <typename T, typename TComp, typename TAllocator>
const delimiters_values<wchar_t> delimiters< ::std::multiset<T, TComp, TAllocator>, wchar_t>::values = { L"{", L", ", L"}" };
template <typename T, typename THash, typename TEqual, typename TAllocator>
struct delimiters< ::std::unordered_set<T, THash, TEqual, TAllocator>, char> { static const delimiters_values<char> values; };
template <typename T, typename THash, typename TEqual, typename TAllocator>
const delimiters_values<char> delimiters< ::std::unordered_set<T, THash, TEqual, TAllocator>, char>::values = { "{", ", ", "}" };
template <typename T, typename THash, typename TEqual, typename TAllocator>
struct delimiters< ::std::unordered_set<T, THash, TEqual, TAllocator>, wchar_t> { static const delimiters_values<wchar_t> values; };
template <typename T, typename THash, typename TEqual, typename TAllocator>
const delimiters_values<wchar_t> delimiters< ::std::unordered_set<T, THash, TEqual, TAllocator>, wchar_t>::values = { L"{", L", ", L"}" };
template <typename T, typename THash, typename TEqual, typename TAllocator>
struct delimiters< ::std::unordered_multiset<T, THash, TEqual, TAllocator>, char> { static const delimiters_values<char> values; };
template <typename T, typename THash, typename TEqual, typename TAllocator>
const delimiters_values<char> delimiters< ::std::unordered_multiset<T, THash, TEqual, TAllocator>, char>::values = { "{", ", ", "}" };
template <typename T, typename THash, typename TEqual, typename TAllocator>
struct delimiters< ::std::unordered_multiset<T, THash, TEqual, TAllocator>, wchar_t> { static const delimiters_values<wchar_t> values; };
template <typename T, typename THash, typename TEqual, typename TAllocator>
const delimiters_values<wchar_t> delimiters< ::std::unordered_multiset<T, THash, TEqual, TAllocator>, wchar_t>::values = { L"{", L", ", L"}" };
// Delimiters for pair and tuple
template <typename T1, typename T2> struct delimiters<std::pair<T1, T2>, char> { static const delimiters_values<char> values; };
template <typename T1, typename T2> const delimiters_values<char> delimiters<std::pair<T1, T2>, char>::values = { "(", ", ", ")" };
template <typename T1, typename T2> struct delimiters< ::std::pair<T1, T2>, wchar_t> { static const delimiters_values<wchar_t> values; };
template <typename T1, typename T2> const delimiters_values<wchar_t> delimiters< ::std::pair<T1, T2>, wchar_t>::values = { L"(", L", ", L")" };
template <typename ...Args> struct delimiters<std::tuple<Args...>, char> { static const delimiters_values<char> values; };
template <typename ...Args> const delimiters_values<char> delimiters<std::tuple<Args...>, char>::values = { "(", ", ", ")" };
template <typename ...Args> struct delimiters< ::std::tuple<Args...>, wchar_t> { static const delimiters_values<wchar_t> values; };
template <typename ...Args> const delimiters_values<wchar_t> delimiters< ::std::tuple<Args...>, wchar_t>::values = { L"(", L", ", L")" };
// Type-erasing helper class for easy use of custom delimiters.
// Requires TCharTraits = std::char_traits<TChar> and TChar = char or wchar_t, and MyDelims needs to be defined for TChar.
// Usage: "cout << pretty_print::custom_delims<MyDelims>(x)".
struct custom_delims_base
{
virtual ~custom_delims_base() { }
virtual std::ostream & stream(::std::ostream &) = 0;
virtual std::wostream & stream(::std::wostream &) = 0;
};
template <typename T, typename Delims>
struct custom_delims_wrapper : custom_delims_base
{
custom_delims_wrapper(const T & t_) : t(t_) { }
std::ostream & stream(std::ostream & s)
{
return s << print_container_helper<T, char, std::char_traits<char>, Delims>(t);
}
std::wostream & stream(std::wostream & s)
{
return s << print_container_helper<T, wchar_t, std::char_traits<wchar_t>, Delims>(t);
}
private:
const T & t;
};
template <typename Delims>
struct custom_delims
{
template <typename Container>
custom_delims(const Container & c) : base(new custom_delims_wrapper<Container, Delims>(c)) { }
std::unique_ptr<custom_delims_base> base;
};
template <typename TChar, typename TCharTraits, typename Delims>
inline std::basic_ostream<TChar, TCharTraits> & operator<<(std::basic_ostream<TChar, TCharTraits> & s, const custom_delims<Delims> & p)
{
return p.base->stream(s);
}
// A wrapper for a C-style array given as pointer-plus-size.
// Usage: std::cout << pretty_print_array(arr, n) << std::endl;
template<typename T>
struct array_wrapper_n
{
typedef const T * const_iterator;
typedef T value_type;
array_wrapper_n(const T * const a, size_t n) : _array(a), _n(n) { }
inline const_iterator begin() const { return _array; }
inline const_iterator end() const { return _array + _n; }
private:
const T * const _array;
size_t _n;
};
// A wrapper for hash-table based containers that offer local iterators to each bucket.
// Usage: std::cout << bucket_print(m, 4) << std::endl; (Prints bucket 5 of container m.)
template <typename T>
struct bucket_print_wrapper
{
typedef typename T::const_local_iterator const_iterator;
typedef typename T::size_type size_type;
const_iterator begin() const
{
return m_map.cbegin(n);
}
const_iterator end() const
{
return m_map.cend(n);
}
bucket_print_wrapper(const T & m, size_type bucket) : m_map(m), n(bucket) { }
private:
const T & m_map;
const size_type n;
};
} // namespace pretty_print
// Global accessor functions for the convenience wrappers
template<typename T>
inline pretty_print::array_wrapper_n<T> pretty_print_array(const T * const a, size_t n)
{
return pretty_print::array_wrapper_n<T>(a, n);
}
template <typename T> pretty_print::bucket_print_wrapper<T>
bucket_print(const T & m, typename T::size_type n)
{
return pretty_print::bucket_print_wrapper<T>(m, n);
}
// Main magic entry point: An overload snuck into namespace std.
// Can we do better?
namespace std
{
// Prints a container to the stream using default delimiters
template<typename T, typename TChar, typename TCharTraits>
inline typename enable_if< ::pretty_print::is_container<T>::value,
basic_ostream<TChar, TCharTraits> &>::type
operator<<(basic_ostream<TChar, TCharTraits> & stream, const T & container)
{
return stream << ::pretty_print::print_container_helper<T, TChar, TCharTraits>(container);
}
}
#endif // H_PRETTY_PRINT
include/pybind11_utils.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <algorithm>
#include <iostream>
#include <pybind11/embed.h> // everything needed for embedding
#include <pybind11/functional.h>
#include <pybind11/numpy.h>
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
#include <tensorview/tensorview.h>
namespace py = pybind11;
template <typename T, typename TPyObject>
std::vector<T> array2Vector(TPyObject arr){
py::array arr_np = arr;
size_t size = arr.attr("size").template cast<size_t>();
py::array_t<T> arr_cc = arr_np;
std::vector<T> data(arr_cc.data(), arr_cc.data() + size);
return data;
}
template <typename T>
std::vector<T> arrayT2Vector(py::array_t<T> arr)
{
std::vector<T> data(arr.data(), arr.data() + arr.size());
return data;
}
template <typename T, typename TPyObject>
tv::TensorView<T> array2TensorView(TPyObject arr){
py::array arr_np = arr;
py::array_t<T> arr_cc = arr_np;
tv::Shape shape;
for (int i = 0; i < arr_cc.ndim(); ++i){
shape.push_back(arr_cc.shape(i));
}
return tv::TensorView<T>(arr_cc.mutable_data(), shape);
}
template <typename T>
tv::TensorView<T> arrayT2TensorView(py::array_t<T> arr){
tv::Shape shape;
for (int i = 0; i < arr.ndim(); ++i){
shape.push_back(arr.shape(i));
}
return tv::TensorView<T>(arr.mutable_data(), shape);
}
\ No newline at end of file
include/spconv/box_iou.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef BOX_IOU_H
#define BOX_IOU_H
#include <pybind11/pybind11.h>
// must include pybind11/eigen.h if using eigen matrix as arguments.
#include <algorithm>
#include <boost/geometry.hpp>
#include <pybind11/numpy.h>
namespace spconv {
// #include "voxelnet/core/cc/pybind11_helper.h"
namespace py = pybind11;
using namespace pybind11::literals;
template <typename DType, typename ShapeContainer>
inline py::array_t<DType> constant(ShapeContainer shape, DType value) {
// create ROWMAJOR array.
py::array_t<DType> array(shape);
std::fill(array.mutable_data(), array.mutable_data() + array.size(), value);
return array;
}
template <typename DType>
inline py::array_t<DType> zeros(std::vector<long int> shape) {
return constant<DType, std::vector<long int>>(shape, 0);
}
template <typename DType>
py::array_t<DType>
rbbox_iou(py::array_t<DType> box_corners, py::array_t<DType> qbox_corners,
py::array_t<DType> standup_iou, DType standup_thresh) {
namespace bg = boost::geometry;
typedef bg::model::point<DType, 2, bg::cs::cartesian> point_t;
typedef bg::model::polygon<point_t> polygon_t;
polygon_t poly, qpoly;
std::vector<polygon_t> poly_inter, poly_union;
DType inter_area, union_area;
auto box_corners_r = box_corners.template unchecked<3>();
auto qbox_corners_r = qbox_corners.template unchecked<3>();
auto standup_iou_r = standup_iou.template unchecked<2>();
auto N = box_corners_r.shape(0);
auto K = qbox_corners_r.shape(0);
py::array_t<DType> overlaps = zeros<DType>({N, K});
auto overlaps_rw = overlaps.template mutable_unchecked<2>();
if (N == 0 || K == 0) {
return overlaps;
}
for (int k = 0; k < K; ++k) {
for (int n = 0; n < N; ++n) {
if (standup_iou_r(n, k) <= standup_thresh)
continue;
bg::append(poly, point_t(box_corners_r(n, 0, 0), box_corners_r(n, 0, 1)));
bg::append(poly, point_t(box_corners_r(n, 1, 0), box_corners_r(n, 1, 1)));
bg::append(poly, point_t(box_corners_r(n, 2, 0), box_corners_r(n, 2, 1)));
bg::append(poly, point_t(box_corners_r(n, 3, 0), box_corners_r(n, 3, 1)));
bg::append(poly, point_t(box_corners_r(n, 0, 0), box_corners_r(n, 0, 1)));
bg::append(qpoly,
point_t(qbox_corners_r(k, 0, 0), qbox_corners_r(k, 0, 1)));
bg::append(qpoly,
point_t(qbox_corners_r(k, 1, 0), qbox_corners_r(k, 1, 1)));
bg::append(qpoly,
point_t(qbox_corners_r(k, 2, 0), qbox_corners_r(k, 2, 1)));
bg::append(qpoly,
point_t(qbox_corners_r(k, 3, 0), qbox_corners_r(k, 3, 1)));
bg::append(qpoly,
point_t(qbox_corners_r(k, 0, 0), qbox_corners_r(k, 0, 1)));
bg::intersection(poly, qpoly, poly_inter);
if (!poly_inter.empty()) {
inter_area = bg::area(poly_inter.front());
bg::union_(poly, qpoly, poly_union);
if (!poly_union.empty()) {
union_area = bg::area(poly_union.front());
overlaps_rw(n, k) = inter_area / union_area;
}
poly_union.clear();
}
poly.clear();
qpoly.clear();
poly_inter.clear();
}
}
return overlaps;
}
} // namespace spconv
#endif
\ No newline at end of file
include/spconv/geometry.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef SPCONV_GEOMETRY_H_
#define SPCONV_GEOMETRY_H_
#include <iostream>
#include <limits>
#include <tensorview/tensorview.h>
namespace spconv {
template <typename Index, unsigned NDim>
TV_HOST_DEVICE Index getValidOutPos(const Index *input_pos,
const Index *kernelSize,
const Index *stride, const Index *padding,
const Index *dilation,
const Index *outSpatialShape, Index *out) {
Index lowers[NDim];
Index uppers[NDim];
Index counter[NDim];
Index counterSize[NDim];
Index pointCounter = 0;
Index val;
Index numPoints = 1;
Index m, offset;
bool valid = false;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
lowers[i] = (input_pos[i] - (kernelSize[i] - 1) * dilation[i] - 1 +
stride[i] + padding[i]) /
stride[i];
uppers[i] = (input_pos[i] + padding[i]) / stride[i];
}
#pragma unroll
for (unsigned i = 0; i < NDim; ++i) {
counterSize[i] = ((uppers[i] - lowers[i]) / dilation[i] + 1);
numPoints *= counterSize[i];
}
#pragma unroll
for (int i = 0; i < NDim; ++i) {
counter[i] = 0;
}
for (int i = 0; i < numPoints; ++i) {
valid = true;
m = 1;
offset = 0;
#pragma unroll
for (int j = NDim - 1; j >= 0; --j) {
val = uppers[j] - counter[j] * dilation[j];
out[pointCounter * (NDim + 1) + j] = val;
if (val < 0 || (val > outSpatialShape[j] - 1)) {
valid = false;
// break;
}
offset += m * (input_pos[j] - val * stride[j] + padding[j]) / dilation[j];
m *= kernelSize[j];
}
out[pointCounter * (NDim + 1) + NDim] = offset;
if (valid)
++pointCounter;
counter[NDim - 1] += 1;
#pragma unroll
for (int c = NDim - 1; c >= 0; --c) {
if (counter[c] == counterSize[c] && c > 0) {
counter[c - 1] += 1;
counter[c] = 0;
}
}
}
return pointCounter;
}
template <typename Index, unsigned NDim>
TV_HOST_DEVICE Index getValidOutPosTranspose(
const Index *input_pos, const Index *kernelSize, const Index *stride,
const Index *padding, const Index *dilation, const Index *outSpatialShape,
Index *out) {
Index lowers[NDim];
Index uppers[NDim];
Index counter[NDim];
Index counterSize[NDim];
Index pointCounter = 0;
Index val;
Index numPoints = 1;
Index m, offset;
bool valid = false;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
lowers[i] = input_pos[i] * stride[i] - padding[i];
uppers[i] = lowers[i] + (kernelSize[i] - 1) * dilation[i];
}
#pragma unroll
for (unsigned i = 0; i < NDim; ++i) {
counterSize[i] = ((uppers[i] - lowers[i]) / dilation[i] + 1);
numPoints *= counterSize[i];
}
#pragma unroll
for (int i = 0; i < NDim; ++i) {
counter[i] = 0;
}
for (int i = 0; i < numPoints; ++i) {
valid = true;
m = 1;
offset = 0;
#pragma unroll
for (int j = NDim - 1; j >= 0; --j) {
val = uppers[j] - counter[j] * dilation[j];
out[pointCounter * (NDim + 1) + j] = val;
if (val < 0 || (val > outSpatialShape[j] - 1)) {
valid = false;
// break;
}
offset += m * (val - lowers[j]) / dilation[j];
m *= kernelSize[j];
}
out[pointCounter * (NDim + 1) + NDim] = offset;
if (valid)
++pointCounter;
counter[NDim - 1] += 1;
#pragma unroll
for (int c = NDim - 1; c >= 0; --c) {
if (counter[c] == counterSize[c] && c > 0) {
counter[c - 1] += 1;
counter[c] = 0;
}
}
}
return pointCounter;
}
template <typename Index, typename IndexGrid, unsigned NDim>
Index getIndicePairsConv(tv::TensorView<const Index> indicesIn,
tv::TensorView<Index> indicesOut,
tv::TensorView<IndexGrid> gridsOut,
tv::TensorView<Index> indicePairs,
tv::TensorView<Index> indiceNum,
const Index *kernelSize, const Index *stride,
const Index *padding, const Index *dilation,
const Index *outSpatialShape) {
// indicesOut: num_active * kernelVolume * (NDim + 1)
Index numAct = 0;
auto numActIn = indicesIn.dim(0);
Index batchIdx = 0;
Index spatialVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
spatialVolume *= outSpatialShape[i];
}
Index kernelVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
kernelVolume *= kernelSize[i];
}
Index numValidPoints = 0;
Index validPoints[kernelVolume * (NDim + 1)];
Index *pointPtr = nullptr;
for (int j = 0; j < numActIn; ++j) {
batchIdx = indicesIn(j, 0);
numValidPoints = getValidOutPos<Index, NDim>(
indicesIn.data() + j * (NDim + 1) + 1, kernelSize, stride, padding,
dilation, outSpatialShape, validPoints);
for (Index i = 0; i < numValidPoints; ++i) {
pointPtr = validPoints + i * (NDim + 1);
auto offset = pointPtr[NDim];
auto index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape) +
spatialVolume * batchIdx;
if (gridsOut[index] == -1) {
for (unsigned k = 1; k < NDim + 1; ++k) {
indicesOut(numAct, k) = pointPtr[k - 1];
}
indicesOut(numAct, 0) = batchIdx;
gridsOut[index] = numAct++;
}
// indicePairs: [K, 2, L]
indicePairs(offset, 0, indiceNum[offset]) = j;
indicePairs(offset, 1, indiceNum[offset]++) = gridsOut[index];
}
}
return numAct;
}
template <typename Index, typename IndexGrid, unsigned NDim>
Index getIndicePairsDeConv(tv::TensorView<const Index> indicesIn,
tv::TensorView<Index> indicesOut,
tv::TensorView<IndexGrid> gridsOut,
tv::TensorView<Index> indicePairs,
tv::TensorView<Index> indiceNum,
const Index *kernelSize, const Index *stride,
const Index *padding, const Index *dilation,
const Index *outSpatialShape) {
Index numAct = 0;
auto numActIn = indicesIn.dim(0);
Index batchIdx = 0;
Index spatialVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
spatialVolume *= outSpatialShape[i];
}
Index kernelVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
kernelVolume *= kernelSize[i];
}
Index numValidPoints = 0;
Index validPoints[kernelVolume * (NDim + 1)];
Index *pointPtr = nullptr;
for (int j = 0; j < numActIn; ++j) {
batchIdx = indicesIn(j, 0);
numValidPoints = getValidOutPosTranspose<Index, NDim>(
indicesIn.data() + j * (NDim + 1) + 1, kernelSize, stride, padding,
dilation, outSpatialShape, validPoints);
for (Index i = 0; i < numValidPoints; ++i) {
pointPtr = validPoints + i * (NDim + 1);
auto offset = pointPtr[NDim];
auto index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape) +
spatialVolume * batchIdx;
if (gridsOut[index] == -1) {
for (unsigned k = 1; k < NDim + 1; ++k) {
indicesOut(numAct, k) = pointPtr[k - 1];
}
indicesOut(numAct, 0) = batchIdx;
gridsOut[index] = numAct++;
}
// indicePairs: [K, 2, L]
indicePairs(offset, 0, indiceNum[offset]) = j;
indicePairs(offset, 1, indiceNum[offset]++) = gridsOut[index];
}
}
return numAct;
}
template <typename Index, typename IndexGrid, unsigned NDim>
Index getIndicePairsSubM(tv::TensorView<const Index> indicesIn,
tv::TensorView<IndexGrid> gridsOut,
tv::TensorView<Index> indicePairs,
tv::TensorView<Index> indiceNum,
const Index *const kernelSize,
const Index *const stride, const Index *const padding,
const Index *dilation, const Index *const outSpatialShape) {
Index numAct = 0;
auto numActIn = indicesIn.dim(0);
Index batchIdx = 0;
Index spatialVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
spatialVolume *= outSpatialShape[i];
}
Index kernelVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
kernelVolume *= kernelSize[i];
}
Index numValidPoints = 0;
Index validPoints[kernelVolume * (NDim + 1)];
Index *pointPtr = nullptr;
Index index = 0;
for (int j = 0; j < numActIn; ++j) {
index = tv::rowArrayIdx<Index, NDim>(indicesIn.data() + j * (NDim + 1) + 1,
outSpatialShape) +
spatialVolume * indicesIn(j, 0);
gridsOut[index] = j;
}
for (int j = 0; j < numActIn; ++j) {
numValidPoints = getValidOutPos<Index, NDim>(
indicesIn.data() + j * (NDim + 1) + 1, kernelSize, stride, padding,
dilation, outSpatialShape, validPoints);
for (Index i = 0; i < numValidPoints; ++i) {
pointPtr = validPoints + i * (NDim + 1);
auto offset = pointPtr[NDim];
index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape) +
spatialVolume * indicesIn(j, 0);
if (gridsOut[index] > -1) {
indicePairs(offset, 0, indiceNum[offset]) = j;
indicePairs(offset, 1, indiceNum[offset]++) = gridsOut[index];
}
}
}
return numActIn;
}
} // namespace spconv
#endif
\ No newline at end of file
include/spconv/indice.cu.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef INDICE_CU_H_
#define INDICE_CU_H_
#include <tensorview/tensorview.h>
#include <tensorview/helper_kernel.cu.h>
namespace spconv {
template <typename Index, typename IndexGrid, unsigned NDim,
int KernelMaxVolume = 256>
__global__ void prepareIndicePairsKernel(
tv::TensorView<const Index> indicesIn, tv::TensorView<Index> indicesOut,
tv::TensorView<IndexGrid> gridsOut, tv::TensorView<Index> indicePairs,
tv::TensorView<Index> indiceNum, tv::TensorView<Index> indicePairUnique,
const tv::SimpleVector<Index, NDim> kernelSize,
const tv::SimpleVector<Index, NDim> stride,
const tv::SimpleVector<Index, NDim> padding,
const tv::SimpleVector<Index, NDim> dilation,
const tv::SimpleVector<Index, NDim> outSpatialShape) {
auto numActIn = indicesIn.dim(0);
Index spatialVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
spatialVolume *= outSpatialShape[i];
}
Index kernelVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
kernelVolume *= kernelSize[i];
}
Index numValidPoints = 0;
Index validPoints[KernelMaxVolume * (NDim + 1)];
Index *pointPtr = nullptr;
auto indicePairsDim2 = indicePairs.dim(2);
Index index;
for (int ix : tv::KernelLoopX<int>(numActIn)) {
numValidPoints = getValidOutPos<Index, NDim>(
indicesIn.data() + ix * (NDim + 1) + 1, kernelSize.data(),
stride.data(), padding.data(), dilation.data(), outSpatialShape.data(),
validPoints);
for (Index i = 0; i < numValidPoints; ++i) {
pointPtr = validPoints + i * (NDim + 1);
auto offset = pointPtr[NDim];
auto oldNum = atomicAdd(indiceNum.data() + offset, Index(1));
indicePairs(offset, 0, oldNum) = ix;
index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape.data()) +
spatialVolume * indicesIn(ix, 0);
indicePairs(offset, 1, oldNum) = index;
indicePairUnique[offset * indicePairsDim2 + oldNum] = index;
}
}
}
template <typename Index, typename IndexGrid, unsigned NDim,
int KernelMaxVolume = 256>
__global__ void prepareDeConvIndicePairsKernel(
tv::TensorView<const Index> indicesIn, tv::TensorView<Index> indicesOut,
tv::TensorView<IndexGrid> gridsOut, tv::TensorView<Index> indicePairs,
tv::TensorView<Index> indiceNum, tv::TensorView<Index> indicePairUnique,
const tv::SimpleVector<Index, NDim> kernelSize,
const tv::SimpleVector<Index, NDim> stride,
const tv::SimpleVector<Index, NDim> padding,
const tv::SimpleVector<Index, NDim> dilation,
const tv::SimpleVector<Index, NDim> outSpatialShape) {
auto numActIn = indicesIn.dim(0);
Index spatialVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
spatialVolume *= outSpatialShape[i];
}
Index kernelVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
kernelVolume *= kernelSize[i];
}
Index numValidPoints = 0;
Index validPoints[KernelMaxVolume * (NDim + 1)];
Index *pointPtr = nullptr;
auto indicePairsDim2 = indicePairs.dim(2);
Index index;
for (int ix : tv::KernelLoopX<int>(numActIn)) {
numValidPoints = getValidOutPosTranspose<Index, NDim>(
indicesIn.data() + ix * (NDim + 1) + 1, kernelSize.data(),
stride.data(), padding.data(), dilation.data(), outSpatialShape.data(),
validPoints);
for (Index i = 0; i < numValidPoints; ++i) {
pointPtr = validPoints + i * (NDim + 1);
auto offset = pointPtr[NDim];
auto oldNum = atomicAdd(indiceNum.data() + offset, Index(1));
indicePairs(offset, 0, oldNum) = ix;
index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape.data()) +
spatialVolume * indicesIn(ix, 0);
indicePairs(offset, 1, oldNum) = index;
indicePairUnique[offset * indicePairsDim2 + oldNum] = index;
}
}
}
template <typename Index, typename IndexGrid, unsigned NDim>
__global__ void assignGridAndIndiceOutKernel(
tv::TensorView<Index> indicesOut, tv::TensorView<IndexGrid> gridsOut,
int numAct, tv::TensorView<Index> indicePairs,
tv::TensorView<Index> indicePairUnique,
const tv::SimpleVector<Index, NDim> outSpatialShape, int batchSize) {
Index index;
auto indicesOutPtr = indicesOut.data();
for (int ix : tv::KernelLoopX<int>(numAct)) {
index = indicePairUnique[ix];
gridsOut[index] = ix;
index = tv::rowArrayIdxInv<Index, NDim>(
index, indicesOutPtr + ix * (NDim + 1) + 1, outSpatialShape.data());
indicesOut[ix * (NDim + 1)] = index % batchSize;
}
}
template <typename Index, typename IndexGrid, unsigned NDim>
__global__ void
assignIndicePairsKernel(tv::TensorView<Index> indicesOut,
tv::TensorView<IndexGrid> gridsOut, int numActIn,
tv::TensorView<Index> indicePairs,
tv::TensorView<Index> indicePairUnique,
const tv::SimpleVector<Index, NDim> outSpatialShape) {
Index index;
int kernelVolume = indicePairs.dim(0);
for (int ix : tv::KernelLoopX<int>(numActIn)) {
for (int i = 0; i < kernelVolume; ++i) {
index = indicePairs(i, 1, ix);
if (index > -1) {
indicePairs(i, 1, ix) = gridsOut[index];
}
}
}
}
template <typename Index, typename IndexGrid, unsigned NDim,
int KernelMaxVolume = 256>
__global__ void
prepareSubMGridKernel(tv::TensorView<const Index> indicesIn,
tv::TensorView<IndexGrid> gridsOut,
const tv::SimpleVector<Index, NDim> outSpatialShape) {
auto numActIn = indicesIn.dim(0);
Index spatialVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
spatialVolume *= outSpatialShape[i];
}
Index index = 0;
for (int ix : tv::KernelLoopX<int>(numActIn)) {
index = tv::rowArrayIdx<Index, NDim>(indicesIn.data() + ix * (NDim + 1) + 1,
outSpatialShape.data()) +
spatialVolume * indicesIn(ix, 0);
gridsOut[index] = ix;
}
}
template <typename Index, typename IndexGrid, unsigned NDim,
int KernelMaxVolume = 256>
__global__ void getSubMIndicePairsKernel(
tv::TensorView<const Index> indicesIn, tv::TensorView<IndexGrid> gridsOut,
tv::TensorView<Index> indicePairs, tv::TensorView<Index> indiceNum,
const tv::SimpleVector<Index, NDim> kernelSize,
const tv::SimpleVector<Index, NDim> stride,
const tv::SimpleVector<Index, NDim> padding,
const tv::SimpleVector<Index, NDim> dilation,
const tv::SimpleVector<Index, NDim> outSpatialShape) {
auto numActIn = indicesIn.dim(0);
Index spatialVolume = 1;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
spatialVolume *= outSpatialShape[i];
}
Index numValidPoints = 0;
Index validPoints[KernelMaxVolume * (NDim + 1)];
Index *pointPtr = nullptr;
Index index = 0;
for (int ix : tv::KernelLoopX<int>(numActIn)) {
numValidPoints = getValidOutPos<Index, NDim>(
indicesIn.data() + ix * (NDim + 1) + 1, kernelSize.data(),
stride.data(), padding.data(), dilation.data(), outSpatialShape.data(),
validPoints);
for (int i = 0; i < numValidPoints; ++i) {
pointPtr = validPoints + i * (NDim + 1);
auto offset = pointPtr[NDim];
index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape.data()) +
spatialVolume * indicesIn(ix, 0);
if (gridsOut[index] > -1) {
auto oldNum = atomicAdd(indiceNum.data() + offset, Index(1));
indicePairs(offset, 1, oldNum) = gridsOut[index];
indicePairs(offset, 0, oldNum) = ix;
}
}
}
}
template <typename Index, typename IndexGrid, unsigned NDim>
__global__ void resetGridKernel(const Index *indicePairUnique,
tv::TensorView<IndexGrid> gridsOut,
int numAct) {
for (int ix : tv::KernelLoopX<int>(numAct)) {
gridsOut[indicePairUnique[ix]] = -1;
}
}
template <typename Index, typename IndexGrid, unsigned NDim>
__global__ void
resetGridSubMKernel(const Index *indices, tv::TensorView<IndexGrid> gridsOut,
const tv::SimpleVector<Index, NDim> outSpatialShape,
int numAct) {
int outSpatialShapeReg[NDim];
for (int i = 0; i < NDim; ++i) {
outSpatialShapeReg[i] = outSpatialShape[i];
}
Index spatialVolume = 1;
auto indsPtr = indices;
#pragma unroll
for (int i = 0; i < NDim; ++i) {
spatialVolume *= outSpatialShape[i];
}
Index index;
for (int ix : tv::KernelLoopX<int>(numAct)) {
indsPtr = indices + ix * (NDim + 1);
index = tv::rowArrayIdx<Index, NDim>(indsPtr + 1, outSpatialShapeReg);
gridsOut[index + spatialVolume * indsPtr[0]] = -1;
}
}
} // namespace spconv
#endif
\ No newline at end of file
include/spconv/indice.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef SPARSE_CONV_INDICE_FUNCTOR_H_
#define SPARSE_CONV_INDICE_FUNCTOR_H_
#include <tensorview/tensorview.h>
namespace spconv
{
namespace functor
{
template <typename Device, typename Index, typename IndexGrid, unsigned NDim>
struct CreateConvIndicePairFunctorP1
{
Index operator()(
const Device& d, tv::TensorView<const Index> indicesIn,
tv::TensorView<Index> indicesOut, tv::TensorView<IndexGrid> gridsOut,
tv::TensorView<Index> indicePairs, tv::TensorView<Index> indiceNum,
tv::TensorView<Index> indicePairUnique,
const tv::SimpleVector<Index, NDim> kernelSize,
const tv::SimpleVector<Index, NDim> stride,
const tv::SimpleVector<Index, NDim> padding,
const tv::SimpleVector<Index, NDim> dilation,
const tv::SimpleVector<Index, NDim> outSpatialShape, bool transpose);
};
template <typename Device, typename Index, typename IndexGrid, unsigned NDim>
struct CreateConvIndicePairFunctorP2
{
Index operator()(
const Device& d, tv::TensorView<const Index> indicesIn,
tv::TensorView<Index> indicesOut, tv::TensorView<IndexGrid> gridsOut,
tv::TensorView<Index> indicePairs, tv::TensorView<Index> indiceNum,
tv::TensorView<Index> indicePairUnique,
const tv::SimpleVector<Index, NDim> outSpatialShape, bool transpose,
bool resetGrid=false);
};
template <typename Device, typename Index, typename IndexGrid, unsigned NDim>
struct CreateConvIndicePairFunctor
{
Index operator()(
const Device& d, tv::TensorView<const Index> indicesIn,
tv::TensorView<Index> indicesOut, tv::TensorView<IndexGrid> gridsOut,
tv::TensorView<Index> indicePairs, tv::TensorView<Index> indiceNum,
const tv::SimpleVector<Index, NDim> kernelSize,
const tv::SimpleVector<Index, NDim> stride,
const tv::SimpleVector<Index, NDim> padding,
const tv::SimpleVector<Index, NDim> dilation,
const tv::SimpleVector<Index, NDim> outSpatialShape, bool transpose, bool resetGrid=false);
};
template <typename Device, typename Index, typename IndexGrid, unsigned NDim>
struct CreateSubMIndicePairFunctor
{
Index operator()(
const Device& d, tv::TensorView<const Index> indicesIn, tv::TensorView<IndexGrid> gridsOut,
tv::TensorView<Index> indicePairs, tv::TensorView<Index> indiceNum,
const tv::SimpleVector<Index, NDim> kernelSize,
const tv::SimpleVector<Index, NDim> stride,
const tv::SimpleVector<Index, NDim> padding,
const tv::SimpleVector<Index, NDim> dilation,
const tv::SimpleVector<Index, NDim> outSpatialShape, bool transpose, bool resetGrid=false);
};
} // namespace functor
} // namespace spconv
#endif
\ No newline at end of file
include/spconv/maxpool.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef SPARSE_MAXPOOL_FUNCTOR_H_
#define SPARSE_MAXPOOL_FUNCTOR_H_
#include <tensorview/tensorview.h>
namespace spconv
{
namespace functor
{
template <typename Device, typename T, typename Index>
struct SparseMaxPoolForwardFunctor
{
void operator()(const Device& d, tv::TensorView<T> outFeatures,
tv::TensorView<const T> inFeatures,
tv::TensorView<const Index> indices, int size);
};
template <typename Device, typename T, typename Index>
struct SparseMaxPoolBackwardFunctor
{
void operator()(const Device& d, tv::TensorView<const T> outFeatures,
tv::TensorView<const T> inFeatures,
tv::TensorView<const T> dout,
tv::TensorView<T> din,
tv::TensorView<const Index> indices, int size);
};
} // namespace functor
} // namespace spconv
#endif
\ No newline at end of file
include/spconv/mp_helper.h 0 → 100644
View file @ 3f1316d5
#ifndef MP_HELPER_H_
#define MP_HELPER_H_
#include <type_traits>
#include <utility>
namespace spconv {
template <class... T> struct mp_list {};
template <class T, T... I>
using mp_list_c = mp_list<std::integral_constant<T, I>...>;
namespace detail {
template <class... T, class F>
constexpr F mp_for_each_impl(mp_list<T...>, F &&f) {
return std::initializer_list<int>{(f(T()), 0)...}, std::forward<F>(f);
}
template <class F> constexpr F mp_for_each_impl(mp_list<>, F &&f) {
return std::forward<F>(f);
}
} // namespace detail
namespace detail {
template <class A, template <class...> class B> struct mp_rename_impl {
// An error "no type named 'type'" here means that the first argument to
// mp_rename is not a list
};
template <template <class...> class A, class... T, template <class...> class B>
struct mp_rename_impl<A<T...>, B> {
using type = B<T...>;
};
} // namespace detail
template <class A, template <class...> class B>
using mp_rename = typename detail::mp_rename_impl<A, B>::type;
template <class L, class F> constexpr F mp_for_each(F &&f) {
return detail::mp_for_each_impl(mp_rename<L, mp_list>(), std::forward<F>(f));
}
} // namespace spconv
#endif
\ No newline at end of file
include/spconv/nms.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef NMS_CPU_H
#define NMS_CPU_H
#include <pybind11/pybind11.h>
// must include pybind11/stl.h if using containers in STL in arguments.
#include <algorithm>
#include <boost/geometry.hpp>
#include <pybind11/numpy.h>
#include <pybind11/stl.h>
#include <vector>
#include "box_iou.h"
#include "nms_gpu.h"
namespace spconv {
namespace py = pybind11;
using namespace pybind11::literals;
template <typename DType>
std::vector<int> non_max_suppression_cpu(py::array_t<DType> boxes,
py::array_t<int> order, DType thresh,
DType eps = 0) {
auto ndets = boxes.shape(0);
auto boxes_r = boxes.template unchecked<2>();
auto order_r = order.template unchecked<1>();
auto suppressed = zeros<int>({ndets});
auto suppressed_rw = suppressed.template mutable_unchecked<1>();
auto area = zeros<DType>({ndets});
auto area_rw = area.template mutable_unchecked<1>();
// get areas
for (int i = 0; i < ndets; ++i) {
area_rw(i) = (boxes_r(i, 2) - boxes_r(i, 0) + eps) *
(boxes_r(i, 3) - boxes_r(i, 1) + eps);
}
std::vector<int> keep;
int i, j;
DType xx1, xx2, w, h, inter, ovr;
for (int _i = 0; _i < ndets; ++_i) {
i = order_r(_i);
if (suppressed_rw(i) == 1)
continue;
keep.push_back(i);
for (int _j = _i + 1; _j < ndets; ++_j) {
j = order_r(_j);
if (suppressed_rw(j) == 1)
continue;
xx2 = std::min(boxes_r(i, 2), boxes_r(j, 2));
xx1 = std::max(boxes_r(i, 0), boxes_r(j, 0));
w = xx2 - xx1 + eps;
if (w > 0) {
xx2 = std::min(boxes_r(i, 3), boxes_r(j, 3));
xx1 = std::max(boxes_r(i, 1), boxes_r(j, 1));
h = xx2 - xx1 + eps;
if (h > 0) {
inter = w * h;
ovr = inter / (area_rw(i) + area_rw(j) - inter);
if (ovr >= thresh)
suppressed_rw(j) = 1;
}
}
}
}
return keep;
}
template <typename DType>
std::vector<int> rotate_non_max_suppression_cpu(py::array_t<DType> box_corners,
py::array_t<int> order,
py::array_t<DType> standup_iou,
DType thresh) {
auto ndets = box_corners.shape(0);
auto box_corners_r = box_corners.template unchecked<3>();
auto order_r = order.template unchecked<1>();
auto suppressed = zeros<int>({ndets});
auto suppressed_rw = suppressed.template mutable_unchecked<1>();
auto standup_iou_r = standup_iou.template unchecked<2>();
std::vector<int> keep;
int i, j;
namespace bg = boost::geometry;
typedef bg::model::point<DType, 2, bg::cs::cartesian> point_t;
typedef bg::model::polygon<point_t> polygon_t;
polygon_t poly, qpoly;
std::vector<polygon_t> poly_inter, poly_union;
DType inter_area, union_area, overlap;
for (int _i = 0; _i < ndets; ++_i) {
i = order_r(_i);
if (suppressed_rw(i) == 1)
continue;
keep.push_back(i);
for (int _j = _i + 1; _j < ndets; ++_j) {
j = order_r(_j);
if (suppressed_rw(j) == 1)
continue;
if (standup_iou_r(i, j) <= 0.0)
continue;
// std::cout << "pre_poly" << std::endl;
try {
bg::append(poly,
point_t(box_corners_r(i, 0, 0), box_corners_r(i, 0, 1)));
bg::append(poly,
point_t(box_corners_r(i, 1, 0), box_corners_r(i, 1, 1)));
bg::append(poly,
point_t(box_corners_r(i, 2, 0), box_corners_r(i, 2, 1)));
bg::append(poly,
point_t(box_corners_r(i, 3, 0), box_corners_r(i, 3, 1)));
bg::append(poly,
point_t(box_corners_r(i, 0, 0), box_corners_r(i, 0, 1)));
bg::append(qpoly,
point_t(box_corners_r(j, 0, 0), box_corners_r(j, 0, 1)));
bg::append(qpoly,
point_t(box_corners_r(j, 1, 0), box_corners_r(j, 1, 1)));
bg::append(qpoly,
point_t(box_corners_r(j, 2, 0), box_corners_r(j, 2, 1)));
bg::append(qpoly,
point_t(box_corners_r(j, 3, 0), box_corners_r(j, 3, 1)));
bg::append(qpoly,
point_t(box_corners_r(j, 0, 0), box_corners_r(j, 0, 1)));
bg::intersection(poly, qpoly, poly_inter);
} catch (const std::exception &e) {
std::cout << "box i corners:" << std::endl;
for (int k = 0; k < 4; ++k) {
std::cout << box_corners_r(i, k, 0) << " " << box_corners_r(i, k, 1)
<< std::endl;
}
std::cout << "box j corners:" << std::endl;
for (int k = 0; k < 4; ++k) {
std::cout << box_corners_r(j, k, 0) << " " << box_corners_r(j, k, 1)
<< std::endl;
}
// throw e;
continue;
}
// std::cout << "post_poly" << std::endl;
// std::cout << "post_intsec" << std::endl;
if (!poly_inter.empty()) {
inter_area = bg::area(poly_inter.front());
// std::cout << "pre_union" << " " << inter_area << std::endl;
bg::union_(poly, qpoly, poly_union);
/*
if (poly_union.empty()){
std::cout << "intsec area:" << " " << inter_area << std::endl;
std::cout << "box i corners:" << std::endl;
for(int k = 0; k < 4; ++k){
std::cout << box_corners_r(i, k, 0) << " " << box_corners_r(i,
k, 1) << std::endl;
}
std::cout << "box j corners:" << std::endl;
for(int k = 0; k < 4; ++k){
std::cout << box_corners_r(j, k, 0) << " " << box_corners_r(j,
k, 1) << std::endl;
}
}*/
// std::cout << "post_union" << poly_union.empty() << std::endl;
if (!poly_union.empty()) { // ignore invalid box
union_area = bg::area(poly_union.front());
// std::cout << "post union area" << std::endl;
// std::cout << union_area << "debug" << std::endl;
overlap = inter_area / union_area;
if (overlap >= thresh)
suppressed_rw(j) = 1;
poly_union.clear();
}
}
poly.clear();
qpoly.clear();
poly_inter.clear();
}
}
return keep;
}
constexpr int const threadsPerBlock = sizeof(unsigned long long) * 8;
template <typename DType>
int non_max_suppression(py::array_t<DType> boxes, py::array_t<int> keep_out,
DType nms_overlap_thresh, int device_id) {
py::buffer_info info = boxes.request();
auto boxes_ptr = static_cast<DType *>(info.ptr);
py::buffer_info info_k = keep_out.request();
auto keep_out_ptr = static_cast<int *>(info_k.ptr);
return _nms_gpu<DType, threadsPerBlock>(keep_out_ptr, boxes_ptr,
boxes.shape(0), boxes.shape(1),
nms_overlap_thresh, device_id);
}
} // namespace spconv
#endif
include/spconv/nms_gpu.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
template <typename DType, int BLOCK_THREADS>
int _nms_gpu(int *keep_out, const DType *boxes_host, int boxes_num,
int boxes_dim, DType nms_overlap_thresh, int device_id);
include/spconv/point2voxel.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <pybind11/pybind11.h>
// must include pybind11/eigen.h if using eigen matrix as arguments.
// must include pybind11/stl.h if using containers in STL in arguments.
#include <algorithm>
#include <pybind11/numpy.h>
#include <pybind11/stl.h>
// #include <vector>
#include <iostream>
#include <math.h>
namespace spconv {
namespace py = pybind11;
using namespace pybind11::literals;
template <typename DType, int NDim>
int points_to_voxel_3d_np(py::array_t<DType> points, py::array_t<DType> voxels,
py::array_t<int> coors,
py::array_t<int> num_points_per_voxel,
py::array_t<int> coor_to_voxelidx,
std::vector<DType> voxel_size,
std::vector<DType> coors_range, int max_points,
int max_voxels) {
auto points_rw = points.template mutable_unchecked<2>();
auto voxels_rw = voxels.template mutable_unchecked<3>();
auto coors_rw = coors.mutable_unchecked<2>();
auto num_points_per_voxel_rw = num_points_per_voxel.mutable_unchecked<1>();
auto coor_to_voxelidx_rw = coor_to_voxelidx.mutable_unchecked<NDim>();
auto N = points_rw.shape(0);
auto num_features = points_rw.shape(1);
// auto ndim = points_rw.shape(1) - 1;
constexpr int ndim_minus_1 = NDim - 1;
int voxel_num = 0;
bool failed = false;
int coor[NDim];
int c;
int grid_size[NDim];
for (int i = 0; i < NDim; ++i) {
grid_size[i] =
round((coors_range[NDim + i] - coors_range[i]) / voxel_size[i]);
}
int voxelidx, num;
for (int i = 0; i < N; ++i) {
failed = false;
for (int j = 0; j < NDim; ++j) {
c = floor((points_rw(i, j) - coors_range[j]) / voxel_size[j]);
if ((c < 0 || c >= grid_size[j])) {
failed = true;
break;
}
coor[ndim_minus_1 - j] = c;
}
if (failed)
continue;
voxelidx = coor_to_voxelidx_rw(coor[0], coor[1], coor[2]);
if (voxelidx == -1) {
voxelidx = voxel_num;
if (voxel_num >= max_voxels)
break;
voxel_num += 1;
coor_to_voxelidx_rw(coor[0], coor[1], coor[2]) = voxelidx;
for (int k = 0; k < NDim; ++k) {
coors_rw(voxelidx, k) = coor[k];
}
}
num = num_points_per_voxel_rw(voxelidx);
if (num < max_points) {
for (int k = 0; k < num_features; ++k) {
voxels_rw(voxelidx, num, k) = points_rw(i, k);
}
num_points_per_voxel_rw(voxelidx) += 1;
}
}
for (int i = 0; i < voxel_num; ++i) {
coor_to_voxelidx_rw(coors_rw(i, 0), coors_rw(i, 1), coors_rw(i, 2)) = -1;
}
return voxel_num;
}
} // namespace spconv
\ No newline at end of file
include/spconv/pool_ops.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef SPARSE_POOL_OP_H_
#define SPARSE_POOL_OP_H_
#include <cuda_runtime_api.h>
#include <spconv/maxpool.h>
#include <torch/script.h>
#include <torch_utils.h>
#include <utility/timer.h>
namespace spconv {
template <typename T>
torch::Tensor indiceMaxPool(torch::Tensor features, torch::Tensor indicePairs,
torch::Tensor indiceNum, int64_t numAct) {
auto device = features.device().type();
auto kernelVolume = indicePairs.size(0);
auto numInPlanes = features.size(1);
auto indicePairNumCpu = indiceNum.to({torch::kCPU});
auto options =
torch::TensorOptions().dtype(features.dtype()).device(features.device());
torch::Tensor output = torch::zeros({numAct, numInPlanes}, options);
double totalTime = 0;
for (int i = 0; i < kernelVolume; ++i) {
auto nHot = indicePairNumCpu.data<int>()[i];
if (nHot <= 0) {
continue;
}
// auto timer = spconv::CudaContextTimer<>();
if (device == torch::kCPU) {
functor::SparseMaxPoolForwardFunctor<tv::CPU, T, int> forwardFtor;
forwardFtor(tv::CPU(), tv::torch2tv<T>(output),
tv::torch2tv<const T>(features),
tv::torch2tv<const int>(indicePairs).subview(i), nHot);
} else {
functor::SparseMaxPoolForwardFunctor<tv::GPU, T, int> forwardFtor;
forwardFtor(tv::TorchGPU(), tv::torch2tv<T>(output),
tv::torch2tv<const T>(features),
tv::torch2tv<const int>(indicePairs).subview(i), nHot);
TV_CHECK_CUDA_ERR();
}
// totalTime += timer.report() / 1000.0;
}
// std::cout << "maxpool forward time " << totalTime << std::endl;
return output;
}
template <typename T>
torch::Tensor indiceMaxPoolBackward(torch::Tensor features,
torch::Tensor outFeatures,
torch::Tensor outGrad, torch::Tensor indicePairs,
torch::Tensor indiceNum) {
auto device = features.device().type();
auto numInPlanes = features.size(1);
auto indicePairNumCpu = indiceNum.to({torch::kCPU});
auto options =
torch::TensorOptions().dtype(features.dtype()).device(features.device());
torch::Tensor inputGrad = torch::zeros(features.sizes(), options);
auto kernelVolume = indicePairs.size(0);
for (int i = 0; i < kernelVolume; ++i) {
auto nHot = indicePairNumCpu.data<int>()[i];
if (nHot <= 0) {
continue;
}
if (device == torch::kCPU) {
functor::SparseMaxPoolBackwardFunctor<tv::CPU, T, int> backwardFtor;
backwardFtor(tv::CPU(), tv::torch2tv<const T>(outFeatures),
tv::torch2tv<const T>(features),
tv::torch2tv<const T>(outGrad), tv::torch2tv<T>(inputGrad),
tv::torch2tv<const int>(indicePairs).subview(i), nHot);
} else {
functor::SparseMaxPoolBackwardFunctor<tv::GPU, T, int> backwardFtor;
backwardFtor(tv::TorchGPU(), tv::torch2tv<const T>(outFeatures),
tv::torch2tv<const T>(features),
tv::torch2tv<const T>(outGrad), tv::torch2tv<T>(inputGrad),
tv::torch2tv<const int>(indicePairs).subview(i), nHot);
TV_CHECK_CUDA_ERR();
}
}
return inputGrad;
}
} // namespace spconv
#endif
\ No newline at end of file
include/spconv/reordering.cu.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef REORDERING_CU_H_
#define REORDERING_CU_H_
#include <tensorview/helper_kernel.cu.h>
// see http://www.nvidia.com/content/GTC-2010/pdfs/2238_GTC2010.pdf.
namespace spconv {
template <typename T, typename Index, int NumTLP, int NumILP>
__global__ void gatherGenericKernel(T *buffer, const T *features,
const Index *indices, int size,
int numPlanes) {
int ILPStrideX[NumILP];
Index inds[NumILP];
#pragma unroll
for (int ilp = 0; ilp < NumILP; ilp++)
ILPStrideX[ilp] = ilp * gridDim.x * blockDim.x;
for (int ix : tv::KernelLoopX<int, NumILP>(size)) {
#pragma unroll
for (int ilp = 0; ilp < NumILP; ilp++) {
if (ix + ILPStrideX[ilp] < size)
inds[ilp] = indices[ix + ILPStrideX[ilp]] * numPlanes;
}
for (int iy : tv::KernelLoopY<int>(numPlanes)) {
#pragma unroll
for (int ilp = 0; ilp < NumILP; ++ilp) {
if (ix + ILPStrideX[ilp] < size)
buffer[(ix + ILPStrideX[ilp]) * numPlanes + iy] =
features[inds[ilp] + iy];
}
}
}
}
template <typename T, typename Index, int NumTLP, int NumILP, typename VecType>
__global__ void gatherVecKernel(T *buffer, const T *features,
const Index *indices, int size, int numPlanes) {
int ILPStrideX[NumILP];
Index inds[NumILP];
#pragma unroll
for (int ilp = 0; ilp < NumILP; ilp++)
ILPStrideX[ilp] = ilp * gridDim.x * blockDim.x;
for (int ix : tv::KernelLoopX<int, NumILP>(size)) {
#pragma unroll
for (int ilp = 0; ilp < NumILP; ilp++) {
if (ix + ILPStrideX[ilp] < size)
inds[ilp] = indices[ix + ILPStrideX[ilp]] * numPlanes;
}
for (int iy : tv::KernelLoopY<int>(numPlanes)) {
#pragma unroll
for (int ilp = 0; ilp < NumILP; ++ilp) {
if (ix + ILPStrideX[ilp] < size)
reinterpret_cast<VecType *>(
buffer)[(ix + ILPStrideX[ilp]) * numPlanes + iy] =
reinterpret_cast<const VecType *>(features)[inds[ilp] + iy];
}
}
}
}
template <typename T, typename Index, int NumTLP, int NumILP,
typename VecType = int4>
__global__ void gatherVecBlockKernel(T *buffer, const T *features,
const Index *indices, int size,
int numPlanes) {
int ILPStrideY[NumILP];
#pragma unroll
for (int ilp = 0; ilp < NumILP; ilp++)
ILPStrideY[ilp] = ilp * gridDim.y * blockDim.y;
features += blockIdx.x * NumTLP;
buffer += blockIdx.x * NumTLP;
for (int iy : tv::KernelLoopY<int, NumILP>(size)) {
#pragma unroll
for (int ilp = 0; ilp < NumILP; ++ilp) {
reinterpret_cast<VecType *>(
buffer)[(iy + ILPStrideY[ilp]) * numPlanes + threadIdx.x] =
reinterpret_cast<const VecType *>(
features)[indices[iy + ILPStrideY[ilp]] * numPlanes +
threadIdx.x];
}
}
}
template <typename T, typename Index, int NumTLP, int NumILP>
__global__ void scatterAddGenericKernel(T *outFeatures, const T *buffer,
const Index *indices, int size,
int numPlanes) {
int ILPStrideX[NumILP];
Index inds[NumILP];
#pragma unroll
for (int ilp = 0; ilp < NumILP; ilp++)
ILPStrideX[ilp] = ilp * gridDim.x * blockDim.x;
for (int ix : tv::KernelLoopX<int, NumILP>(size)) {
#pragma unroll
for (int ilp = 0; ilp < NumILP; ilp++) {
if (ix + ILPStrideX[ilp] < size)
inds[ilp] = indices[ix + ILPStrideX[ilp]] * numPlanes;
}
for (int iy : tv::KernelLoopY<int>(numPlanes)) {
#pragma unroll
for (int ilp = 0; ilp < NumILP; ++ilp) {
if (ix + ILPStrideX[ilp] < size) {
outFeatures[inds[ilp] + iy] +=
buffer[(ix + ILPStrideX[ilp]) * numPlanes + iy];
}
}
}
}
}
template <typename T, typename Index, int NumTLP, int NumILP,
typename VecType = int4>
__global__ void scatterAddVecBlockKernel(T *outFeatures, const T *buffer,
const Index *indices, int size,
int numPlanes) {
int ILPStrideY[NumILP];
constexpr int vecloadFactor = sizeof(VecType) / sizeof(T);
#pragma unroll
for (int ilp = 0; ilp < NumILP; ilp++)
ILPStrideY[ilp] = ilp * gridDim.y * blockDim.y;
outFeatures += blockIdx.x * NumTLP;
buffer += blockIdx.x * NumTLP;
T buf[vecloadFactor];
T buf2[vecloadFactor];
Index idx;
for (int iy : tv::KernelLoopY<int, NumILP>(size)) {
#pragma unroll
for (int ilp = 0; ilp < NumILP; ++ilp) {
idx = indices[iy + ILPStrideY[ilp]] * numPlanes + threadIdx.x;
reinterpret_cast<VecType *>(buf)[0] =
reinterpret_cast<VecType *>(outFeatures)[idx];
reinterpret_cast<VecType *>(buf2)[0] = reinterpret_cast<const VecType *>(
buffer)[(iy + ILPStrideY[ilp]) * numPlanes + threadIdx.x];
#pragma unroll
for (int i = 0; i < vecloadFactor; i++) {
buf[i] += buf2[i];
}
reinterpret_cast<VecType *>(outFeatures)[idx] =
reinterpret_cast<VecType *>(buf)[0];
}
}
}
} // namespace spconv
#endif
\ No newline at end of file
include/spconv/reordering.h 0 → 100644
View file @ 3f1316d5
// Copyright 2019 Yan Yan
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef SPARSE_REORDERING_FUNCTOR_H_
#define SPARSE_REORDERING_FUNCTOR_H_
#include <tensorview/tensorview.h>
namespace spconv
{
namespace functor
{
template <typename Device, typename T, typename Index>
struct SparseGatherFunctor
{
void operator()(const Device& d, tv::TensorView<T> buffer, tv::TensorView<const T> features,
tv::TensorView<const Index> indices, int size);
};
template <typename Device, typename T, typename Index>
struct SparseScatterAddFunctor
{
void operator()(const Device& d, tv::TensorView<T> out_features,
tv::TensorView<const T> buffer, tv::TensorView<const Index> indices,
int size, bool stable=false);
};
} // namespace functor
} // namespace spconv
#endif
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment