OilfieldToolsNet/opm-simulators
4
0
Fork 0
mirror of https://github.com/OPM/opm-simulators.git synced 2025-02-25 18:55:30 -06:00
Code Issues Packages Projects Releases Wiki Activity

refactor CuBuffer

Browse Source
This commit is contained in:
Tobias Meyer Andersen
2024-08-22 14:12:30 +02:00
parent 67bc9e8f34
commit 69897753e8
7 changed files with 72 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
CMakeLists.txt
View File

@@ -558,7 +558,7 @@ if(CUDA_FOUND)
cuseqilu0
cuowneroverlapcopy
solver_adapter
cubuffer
GpuBuffer
cuview
PROPERTIES LABELS ${gpu_label})
endif()

6
CMakeLists_files.cmake
View File

@@ -209,7 +209,7 @@ if (HAVE_CUDA)
ADD_CUDA_OR_HIP_FILE(MAIN_SOURCE_FILES opm/simulators/linalg detail/CuBlasHandle.cpp)
ADD_CUDA_OR_HIP_FILE(MAIN_SOURCE_FILES opm/simulators/linalg detail/cusparse_matrix_operations.cu)
ADD_CUDA_OR_HIP_FILE(MAIN_SOURCE_FILES opm/simulators/linalg detail/CuSparseHandle.cpp)
ADD_CUDA_OR_HIP_FILE(MAIN_SOURCE_FILES opm/simulators/linalg CuBuffer.cpp)
ADD_CUDA_OR_HIP_FILE(MAIN_SOURCE_FILES opm/simulators/linalg GpuBuffer.cpp)
ADD_CUDA_OR_HIP_FILE(MAIN_SOURCE_FILES opm/simulators/linalg detail/preconditionerKernels/DILUKernels.cu)
ADD_CUDA_OR_HIP_FILE(MAIN_SOURCE_FILES opm/simulators/linalg detail/preconditionerKernels/ILU0Kernels.cu)
ADD_CUDA_OR_HIP_FILE(MAIN_SOURCE_FILES opm/simulators/linalg detail/preconditionerKernels/JacKernels.cu)
@@ -233,7 +233,7 @@ if (HAVE_CUDA)
ADD_CUDA_OR_HIP_FILE(PUBLIC_HEADER_FILES opm/simulators/linalg detail/cuda_check_last_error.hpp)
ADD_CUDA_OR_HIP_FILE(PUBLIC_HEADER_FILES opm/simulators/linalg detail/CuBlasHandle.hpp)
ADD_CUDA_OR_HIP_FILE(PUBLIC_HEADER_FILES opm/simulators/linalg detail/CuSparseHandle.hpp)
ADD_CUDA_OR_HIP_FILE(PUBLIC_HEADER_FILES opm/simulators/linalg CuBuffer.hpp)
ADD_CUDA_OR_HIP_FILE(PUBLIC_HEADER_FILES opm/simulators/linalg GpuBuffer.hpp)
ADD_CUDA_OR_HIP_FILE(PUBLIC_HEADER_FILES opm/simulators/linalg detail/preconditionerKernels/DILUKernels.hpp)
ADD_CUDA_OR_HIP_FILE(PUBLIC_HEADER_FILES opm/simulators/linalg detail/preconditionerKernels/ILU0Kernels.hpp)
ADD_CUDA_OR_HIP_FILE(PUBLIC_HEADER_FILES opm/simulators/linalg detail/preconditionerKernels/JacKernels.hpp)
@@ -389,7 +389,7 @@ if (HAVE_CUDA)
ADD_CUDA_OR_HIP_FILE(TEST_SOURCE_FILES tests test_converttofloatadapter.cpp)
ADD_CUDA_OR_HIP_FILE(TEST_SOURCE_FILES tests test_cublas_handle.cpp)
ADD_CUDA_OR_HIP_FILE(TEST_SOURCE_FILES tests test_cublas_safe_call.cpp)
ADD_CUDA_OR_HIP_FILE(TEST_SOURCE_FILES tests test_cubuffer.cu)
ADD_CUDA_OR_HIP_FILE(TEST_SOURCE_FILES tests test_GpuBuffer.cu)
ADD_CUDA_OR_HIP_FILE(TEST_SOURCE_FILES tests test_cuview.cu)
ADD_CUDA_OR_HIP_FILE(TEST_SOURCE_FILES tests test_cusparse_safe_call.cpp)
ADD_CUDA_OR_HIP_FILE(TEST_SOURCE_FILES tests test_cuda_safe_call.cpp)

2
opm/simulators/linalg/cuistl/CuView.hpp
View File

@@ -48,7 +48,7 @@ namespace Opm::gpuistl
* Implementations are placed in this headerfile for functions that may be called
* inside a kernel to avoid expensive RDC (relocatable device code)
*
* The view will typically provide a view into a CuBuffer and be able to
* The view will typically provide a view into a GpuBuffer and be able to
* manipulate the data within it
*
* @param T Type of the data we store, typically int/float/double w/o const specifier

62
opm/simulators/linalg/cuistl/CuBuffer.cpp → opm/simulators/linalg/cuistl/GpuBuffer.cpp
View File

@@ -20,7 +20,7 @@
#include <cuda_runtime.h>
#include <algorithm>
#include <fmt/core.h>
#include <opm/simulators/linalg/cuistl/CuBuffer.hpp>
#include <opm/simulators/linalg/cuistl/GpuBuffer.hpp>
#include <opm/simulators/linalg/cuistl/CuView.hpp>
#include <opm/simulators/linalg/cuistl/detail/cuda_safe_call.hpp>
@@ -28,24 +28,24 @@ namespace Opm::gpuistl
{
template <class T>
CuBuffer<T>::CuBuffer(const std::vector<T>& data)
: CuBuffer(data.data(), data.size())
GpuBuffer<T>::GpuBuffer(const std::vector<T>& data)
: GpuBuffer(data.data(), data.size())
{
}
template <class T>
CuBuffer<T>::CuBuffer(const size_t numberOfElements)
GpuBuffer<T>::GpuBuffer(const size_t numberOfElements)
: m_numberOfElements(numberOfElements)
{
if (numberOfElements < 1) {
OPM_THROW(std::invalid_argument, "Setting a CuBuffer size to a non-positive number is not allowed");
OPM_THROW(std::invalid_argument, "Setting a GpuBuffer size to a non-positive number is not allowed");
}
OPM_CUDA_SAFE_CALL(cudaMalloc(&m_dataOnDevice, sizeof(T) * m_numberOfElements));
}
template <class T>
CuBuffer<T>::CuBuffer(const T* dataOnHost, const size_t numberOfElements)
: CuBuffer(numberOfElements)
GpuBuffer<T>::GpuBuffer(const T* dataOnHost, const size_t numberOfElements)
: GpuBuffer(numberOfElements)
{
OPM_CUDA_SAFE_CALL(cudaMemcpy(
@@ -53,8 +53,8 @@ CuBuffer<T>::CuBuffer(const T* dataOnHost, const size_t numberOfElements)
}
template <class T>
CuBuffer<T>::CuBuffer(const CuBuffer<T>& other)
: CuBuffer(other.m_numberOfElements)
GpuBuffer<T>::GpuBuffer(const GpuBuffer<T>& other)
: GpuBuffer(other.m_numberOfElements)
{
assertHasElements();
assertSameSize(other);
@@ -65,24 +65,24 @@ CuBuffer<T>::CuBuffer(const CuBuffer<T>& other)
}
template <class T>
CuBuffer<T>::~CuBuffer()
GpuBuffer<T>::~GpuBuffer()
{
OPM_CUDA_WARN_IF_ERROR(cudaFree(m_dataOnDevice));
}
template <typename T>
typename CuBuffer<T>::size_type
CuBuffer<T>::size() const
typename GpuBuffer<T>::size_type
GpuBuffer<T>::size() const
{
return m_numberOfElements;
}
template <typename T>
void
CuBuffer<T>::resize(size_t newSize)
GpuBuffer<T>::resize(size_t newSize)
{
if (newSize < 1) {
OPM_THROW(std::invalid_argument, "Setting a CuBuffer size to a non-positive number is not allowed");
OPM_THROW(std::invalid_argument, "Setting a GpuBuffer size to a non-positive number is not allowed");
}
// Allocate memory for the new buffer
T* tmpBuffer = nullptr;
@@ -107,7 +107,7 @@ CuBuffer<T>::resize(size_t newSize)
template <typename T>
std::vector<T>
CuBuffer<T>::asStdVector() const
GpuBuffer<T>::asStdVector() const
{
std::vector<T> temporary(m_numberOfElements);
copyToHost(temporary);
@@ -116,14 +116,14 @@ CuBuffer<T>::asStdVector() const
template <typename T>
void
CuBuffer<T>::assertSameSize(const CuBuffer<T>& x) const
GpuBuffer<T>::assertSameSize(const GpuBuffer<T>& x) const
{
assertSameSize(x.m_numberOfElements);
}
template <typename T>
void
CuBuffer<T>::assertSameSize(size_t size) const
GpuBuffer<T>::assertSameSize(size_t size) const
{
if (size != m_numberOfElements) {
OPM_THROW(std::invalid_argument,
@@ -133,7 +133,7 @@ CuBuffer<T>::assertSameSize(size_t size) const
template <typename T>
void
CuBuffer<T>::assertHasElements() const
GpuBuffer<T>::assertHasElements() const
{
if (m_numberOfElements <= 0) {
OPM_THROW(std::invalid_argument, "We have 0 elements");
@@ -142,21 +142,21 @@ CuBuffer<T>::assertHasElements() const
template <typename T>
T*
CuBuffer<T>::data()
GpuBuffer<T>::data()
{
return m_dataOnDevice;
}
template <typename T>
const T*
CuBuffer<T>::data() const
GpuBuffer<T>::data() const
{
return m_dataOnDevice;
}
template <class T>
void
CuBuffer<T>::copyFromHost(const T* dataPointer, size_t numberOfElements)
GpuBuffer<T>::copyFromHost(const T* dataPointer, size_t numberOfElements)
{
if (numberOfElements > size()) {
OPM_THROW(std::runtime_error,
@@ -169,7 +169,7 @@ CuBuffer<T>::copyFromHost(const T* dataPointer, size_t numberOfElements)
template <class T>
void
CuBuffer<T>::copyToHost(T* dataPointer, size_t numberOfElements) const
GpuBuffer<T>::copyToHost(T* dataPointer, size_t numberOfElements) const
{
assertSameSize(numberOfElements);
OPM_CUDA_SAFE_CALL(cudaMemcpy(dataPointer, data(), numberOfElements * sizeof(T), cudaMemcpyDeviceToHost));
@@ -177,28 +177,28 @@ CuBuffer<T>::copyToHost(T* dataPointer, size_t numberOfElements) const
template <class T>
void
CuBuffer<T>::copyFromHost(const std::vector<T>& data)
GpuBuffer<T>::copyFromHost(const std::vector<T>& data)
{
copyFromHost(data.data(), data.size());
}
template <class T>
void
CuBuffer<T>::copyToHost(std::vector<T>& data) const
GpuBuffer<T>::copyToHost(std::vector<T>& data) const
{
copyToHost(data.data(), data.size());
}
template class CuBuffer<double>;
template class CuBuffer<float>;
template class CuBuffer<int>;
template class GpuBuffer<double>;
template class GpuBuffer<float>;
template class GpuBuffer<int>;
template <class T>
CuView<const T> make_view(const CuBuffer<T>& buf) {
CuView<const T> make_view(const GpuBuffer<T>& buf) {
return CuView<const T>(buf.data(), buf.size());
}
template CuView<const double> make_view<double>(const CuBuffer<double>&);
template CuView<const float> make_view<float>(const CuBuffer<float>&);
template CuView<const int> make_view<int>(const CuBuffer<int>&);
template CuView<const double> make_view<double>(const GpuBuffer<double>&);
template CuView<const float> make_view<float>(const GpuBuffer<float>&);
template CuView<const int> make_view<int>(const GpuBuffer<int>&);
} // namespace Opm::gpuistl

50
opm/simulators/linalg/cuistl/CuBuffer.hpp → opm/simulators/linalg/cuistl/GpuBuffer.hpp
View File

@@ -16,8 +16,8 @@
You should have received a copy of the GNU General Public License
along with OPM. If not, see <http://www.gnu.org/licenses/>.
*/
#ifndef OPM_CUBUFFER_HEADER_HPP
#define OPM_CUBUFFER_HEADER_HPP
#ifndef OPM_GPUBUFFER_HEADER_HPP
#define OPM_GPUBUFFER_HEADER_HPP
#include <dune/common/fvector.hh>
#include <dune/istl/bvector.hh>
#include <exception>
@@ -33,18 +33,18 @@ namespace Opm::gpuistl
{
/**
* @brief The CuBuffer class is a simple container class for the GPU.
* @brief The GpuBuffer class is a simple container class for the GPU.
*
*
* Example usage:
*
* @code{.cpp}
* #include <opm/simulators/linalg/cuistl/CuBuffer.hpp>
* #include <opm/simulators/linalg/cuistl/GpuBuffer.hpp>
*
* void someFunction() {
* auto someDataOnCPU = std::vector<double>({1.0, 2.0, 42.0, 59.9451743, 10.7132692});
*
* auto dataOnGPU = CuBuffer<double>(someDataOnCPU);
* auto dataOnGPU = GpuBuffer<double>(someDataOnCPU);
*
* auto stdVectorOnCPU = dataOnGPU.asStdVector();
* }
@@ -52,7 +52,7 @@ namespace Opm::gpuistl
* @tparam T the type to store. Can be either float, double or int.
*/
template <typename T>
class CuBuffer
class GpuBuffer
{
public:
using field_type = T;
@@ -60,17 +60,17 @@ public:
using value_type = T;
/**
* @brief CuBuffer allocates new GPU memory of the same size as other and copies the content of the other buffer to
* @brief GpuBuffer allocates new GPU memory of the same size as other and copies the content of the other buffer to
* this newly allocated memory.
*
* @note This does synchronous transfer.
*
* @param other the buffer to copy from
*/
CuBuffer(const CuBuffer<T>& other);
GpuBuffer(const GpuBuffer<T>& other);
/**
* @brief CuBuffer allocates new GPU memory of the same size as data and copies the content of the data vector to
* @brief GpuBuffer allocates new GPU memory of the same size as data and copies the content of the data vector to
* this newly allocated memory.
*
* @note This does CPU to GPU transfer.
@@ -78,23 +78,23 @@ public:
*
* @param data the vector to copy from
*/
explicit CuBuffer(const std::vector<T>& data);
explicit GpuBuffer(const std::vector<T>& data);
/**
* @brief Default constructor that will initialize cublas and allocate 0 bytes of memory
*/
explicit CuBuffer();
explicit GpuBuffer();
/**
* @brief CuBuffer allocates new GPU memory of size numberOfElements * sizeof(T)
* @brief GpuBuffer allocates new GPU memory of size numberOfElements * sizeof(T)
*
* @param numberOfElements number of T elements to allocate
*/
explicit CuBuffer(const size_t numberOfElements);
explicit GpuBuffer(const size_t numberOfElements);
/**
* @brief CuBuffer allocates new GPU memory of size numberOfElements * sizeof(T) and copies numberOfElements from
* @brief GpuBuffer allocates new GPU memory of size numberOfElements * sizeof(T) and copies numberOfElements from
* data
*
* @note This assumes the data is on the CPU.
@@ -102,12 +102,12 @@ public:
* @param numberOfElements number of T elements to allocate
* @param dataOnHost data on host/CPU
*/
CuBuffer(const T* dataOnHost, const size_t numberOfElements);
GpuBuffer(const T* dataOnHost, const size_t numberOfElements);
/**
* @brief ~CuBuffer calls cudaFree
* @brief ~GpuBuffer calls cudaFree
*/
virtual ~CuBuffer();
virtual ~GpuBuffer();
/**
* @return the raw pointer to the GPU data
@@ -120,7 +120,7 @@ public:
const T* data() const;
/**
* @return fetch the first element in a CuBuffer
* @return fetch the first element in a GpuBuffer
*/
__host__ __device__ T& front()
{
@@ -131,7 +131,7 @@ public:
}
/**
* @return fetch the last element in a CuBuffer
* @return fetch the last element in a GpuBuffer
*/
__host__ __device__ T& back()
{
@@ -142,7 +142,7 @@ public:
}
/**
* @return fetch the first element in a CuBuffer
* @return fetch the first element in a GpuBuffer
*/
__host__ __device__ T front() const
{
@@ -153,7 +153,7 @@ public:
}
/**
* @return fetch the last element in a CuBuffer
* @return fetch the last element in a GpuBuffer
*/
__host__ __device__ T back() const
{
@@ -176,7 +176,7 @@ public:
// TODO: [perf] vector.size() can be replaced by bvector.N() * BlockDimension
if (m_numberOfElements != bvector.size()) {
OPM_THROW(std::runtime_error,
fmt::format("Given incompatible vector size. CuBuffer has size {}, \n"
fmt::format("Given incompatible vector size. GpuBuffer has size {}, \n"
"however, BlockVector has N() = {}, and size = {}.",
m_numberOfElements,
bvector.N(),
@@ -199,7 +199,7 @@ public:
// TODO: [perf] vector.size() can be replaced by bvector.N() * BlockDimension
if (m_numberOfElements != bvector.size()) {
OPM_THROW(std::runtime_error,
fmt::format("Given incompatible vector size. CuBuffer has size {},\n however, the BlockVector "
fmt::format("Given incompatible vector size. GpuBuffer has size {},\n however, the BlockVector "
"has has N() = {}, and size() = {}.",
m_numberOfElements,
bvector.N(),
@@ -267,14 +267,14 @@ private:
T* m_dataOnDevice = nullptr;
size_t m_numberOfElements;
void assertSameSize(const CuBuffer<T>& other) const;
void assertSameSize(const GpuBuffer<T>& other) const;
void assertSameSize(size_t size) const;
void assertHasElements() const;
};
template <class T>
CuView<const T> make_view(const CuBuffer<T>&);
CuView<const T> make_view(const GpuBuffer<T>&);
} // namespace Opm::gpuistl
#endif

8
tests/cuistl/test_cubuffer.cu → tests/cuistl/test_GpuBuffer.cu
View File

@@ -18,12 +18,12 @@
*/
#include <config.h>
#define BOOST_TEST_MODULE TestCuBuffer
#define BOOST_TEST_MODULE TestGpuBuffer
#include <boost/test/unit_test.hpp>
#include <cuda_runtime.h>
#include <opm/simulators/linalg/cuistl/CuBuffer.hpp>
#include <opm/simulators/linalg/cuistl/GpuBuffer.hpp>
#include <opm/simulators/linalg/cuistl/CuView.hpp>
#include <opm/simulators/linalg/cuistl/detail/cuda_safe_call.hpp>
@@ -35,13 +35,13 @@ BOOST_AUTO_TEST_CASE(TestMakeView)
{
// test that we can create buffers and make views of the buffers using the pointer constructor
auto buf = std::vector<int>({1, 2, 3, 4, 5, 6});
const auto gpubuf = ::Opm::gpuistl::CuBuffer<int>(buf);
const auto gpubuf = ::Opm::gpuistl::GpuBuffer<int>(buf);
auto gpuview = ::Opm::gpuistl::CuView<int>(buf.data(), buf.size());
bool gpuBufCreatedView = std::is_same<::Opm::gpuistl::CuView<int>, decltype(gpuview)>::value;
BOOST_CHECK(gpuBufCreatedView);
// test that we can make views of buffers by using the cubuffer constructor
// test that we can make views of buffers by using the GpuBuffer constructor
auto gpuview2 = ::Opm::gpuistl::make_view(gpubuf);
bool gpuBufCreatedView2 = std::is_same<::Opm::gpuistl::CuView<const int>, decltype(gpuview2)>::value;

14
tests/cuistl/test_cuview.cu
View File

@@ -25,7 +25,7 @@
#include <dune/common/fvector.hh>
#include <dune/istl/bvector.hh>
#include <opm/simulators/linalg/cuistl/CuView.hpp>
#include <opm/simulators/linalg/cuistl/CuBuffer.hpp>
#include <opm/simulators/linalg/cuistl/GpuBuffer.hpp>
#include <opm/simulators/linalg/cuistl/detail/cuda_safe_call.hpp>
#include <random>
#include <array>
@@ -33,7 +33,7 @@
#include <type_traits>
using CuViewDouble = ::Opm::gpuistl::CuView<double>;
using CuBufferDouble = ::Opm::gpuistl::CuBuffer<double>;
using GpuBufferDouble = ::Opm::gpuistl::GpuBuffer<double>;
__global__ void useCuViewOnGPU(CuViewDouble a, CuViewDouble b){
b[0] = a.front();
@@ -48,9 +48,9 @@ BOOST_AUTO_TEST_CASE(TestCreationAndIndexing)
{
// A simple test to check that we can move data to and from the GPU
auto cpubuffer = std::vector<double>({1.0, 2.0, 42.0, 59.9451743, 10.7132692});
auto cubuffer = CuBufferDouble(cpubuffer);
auto cuview = CuViewDouble(cubuffer.data(), cubuffer.size());
const auto const_cuview = CuViewDouble(cubuffer.data(), cubuffer.size());
auto GpuBuffer = GpuBufferDouble(cpubuffer);
auto cuview = CuViewDouble(GpuBuffer.data(), GpuBuffer.size());
const auto const_cuview = CuViewDouble(GpuBuffer.data(), GpuBuffer.size());
auto stdVecOfCuView = cuview.asStdVector();
auto const_stdVecOfCuView = cuview.asStdVector();
@@ -94,9 +94,9 @@ BOOST_AUTO_TEST_CASE(TestCuViewOnCPUWithSTLIteratorAlgorithm)
BOOST_AUTO_TEST_CASE(TestCuViewOnGPU)
{
auto buf = std::vector<double>({1.0, 2.0, 42.0, 59.9451743, 10.7132692});
auto cubufA = CuBufferDouble(buf);
auto cubufA = GpuBufferDouble(buf);
auto cuviewA = CuViewDouble(cubufA.data(), cubufA.size());
auto cubufB = CuBufferDouble(4);
auto cubufB = GpuBufferDouble(4);
auto cuviewB = CuViewDouble(cubufB.data(), cubufB.size());
useCuViewOnGPU<<<1,1>>>(cuviewA, cuviewB);