Autonomous Space Robotics Lab Speeded Up SURF CudaSynchronizedMemory.hpp Go to the documentation of this file. /* Copyright (c) 2010, Paul Furgale and Chi Hay Tong All rights reserved. Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. * The names of its contributors may not be used to endorse or promote products derived from this software without specific prior written permission. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #ifndef ASRL_CUDA_SYNCHRONIZED_MEMORY_HPP #define ASRL_CUDA_SYNCHRONIZED_MEMORY_HPP #include <builtin_types.h> #include <cuda_runtime_api.h> #include "assert_macros.hpp" #include <limits> #include <iostream> #include <typeinfo> // typeid() #include <cstring> // memset() namespace asrl { template<typename T> class CudaSynchronizedMemory { CudaSynchronizedMemory(CudaSynchronizedMemory const & rhs); CudaSynchronizedMemory & operator=(CudaSynchronizedMemory const & rhs); public: typedef T value_type; CudaSynchronizedMemory(); CudaSynchronizedMemory(unsigned N, bool pageLocked = false); ~CudaSynchronizedMemory(); void init(unsigned N, bool pageLocked = false); void reset(); T * begin(); T * end(); T * h_get(); T const * h_get() const; T * d_get(); T const * d_get() const; size_t size() const; T & operator[](size_t i); T const & operator[](size_t i) const; void pushToDevice(size_t nElements = std::numeric_limits<size_t>::max()); void pullFromDevice(size_t nElements = std::numeric_limits<size_t>::max()); void pullFromDeviceAsync(cudaStream_t stream, size_t nElements = std::numeric_limits<size_t>::max()); void memsetHost(int val); void memsetDevice(int val); void memset(int val); bool isPageLocked(); private: void freeHostMemory(); void freeDeviceMemory(); T * m_host; T * m_device; size_t m_size; bool m_pageLocked; }; template<typename T> CudaSynchronizedMemory<T>::CudaSynchronizedMemory() : m_host(0), m_device(0), m_size(0), m_pageLocked(false) {} template<typename T> CudaSynchronizedMemory<T>::CudaSynchronizedMemory(unsigned N, bool pageLocked) : m_host(0), m_device(0), m_size(0), m_pageLocked(false) { init(N, pageLocked); } template<typename T> CudaSynchronizedMemory<T>::~CudaSynchronizedMemory() { try { reset(); } catch(std::exception const & e) { std::cerr << "Error in " << __FUNCTION__ << ": " << e.what() << std::endl; } } template<typename T> void CudaSynchronizedMemory<T>::reset() { m_size = 0; freeHostMemory(); freeDeviceMemory(); m_pageLocked = false; } template<typename T> void CudaSynchronizedMemory<T>::freeHostMemory() { if(m_host != NULL) { if(m_pageLocked) { cudaError_t err = cudaFreeHost(m_host); ASRL_ASSERT_EQ(err, cudaSuccess, "Unable to free page-locked host memory for " << m_size << " elements of type " << typeid(T).name() << ": " << cudaGetErrorString(err)); } else { delete [] m_host; } m_host = NULL; } } template<typename T> void CudaSynchronizedMemory<T>::freeDeviceMemory() { if(m_device != NULL) { cudaError_t err = cudaFree(m_device); ASRL_ASSERT_EQ(err, cudaSuccess, "Unable to free device memory for " << m_size << " elements of type " << typeid(T).name() << ": " << cudaGetErrorString(err)); m_device = NULL; } } template<typename T> void CudaSynchronizedMemory<T>::memset(int val) { ASRL_ASSERT_GT_DBG(m_size, 0, "The array is empty"); memsetHost(val); memsetDevice(val); } template<typename T> void CudaSynchronizedMemory<T>::memsetHost(int val) { ASRL_ASSERT_GT_DBG(m_size, 0, "The array is empty"); ::memset(m_host, val, m_size * sizeof(T)); } template<typename T> void CudaSynchronizedMemory<T>::memsetDevice(int val) { ASRL_ASSERT_GT_DBG(m_size, 0, "The array is empty"); cudaError_t err = cudaMemset(m_device,val,m_size*sizeof(T)); ASRL_ASSERT_EQ(err, cudaSuccess, "Unable to set device memory"); } template<typename T> void CudaSynchronizedMemory<T>::init(unsigned N, bool pageLocked) { try { if(pageLocked) { cudaError_t e = cudaMallocHost((void**)&m_host,N*sizeof(T)); ASRL_ASSERT_EQ(e, cudaSuccess, "Unable to allocate page-locked host memory for " << N << " elements of type " << typeid(T).name() << ": " << cudaGetErrorString(e)); } else { try { m_host = new T[N]; } catch(std::exception const & e) { ASRL_THROW( "Unable to allocate host memory for " << N << " elements of type " << typeid(T).name() << ": " << e.what()); } } cudaError_t err = cudaMalloc((void**)&m_device,N*sizeof(T)); ASRL_ASSERT_EQ(err, cudaSuccess, "Unable to allocate device memory for " << N << " elements of type " << typeid(T).name() << ": " << cudaGetErrorString(err)); m_size = N; m_pageLocked = pageLocked; } catch(std::exception const &) { reset(); throw; } } template<typename T> bool CudaSynchronizedMemory<T>::isPageLocked() { return m_pageLocked; } template<typename T> T * CudaSynchronizedMemory<T>::begin() { ASRL_ASSERT_GT_DBG(m_size, 0, "The array is empty"); return m_host; } template<typename T> T * CudaSynchronizedMemory<T>::end() { ASRL_ASSERT_GT_DBG(m_size, 0, "The array is empty"); return m_host + m_size; } template<typename T> T * CudaSynchronizedMemory<T>::h_get() { return m_host; } template<typename T> T const * CudaSynchronizedMemory<T>::h_get() const { return m_host; } template<typename T> T * CudaSynchronizedMemory<T>::d_get() { return m_device; } template<typename T> T const * CudaSynchronizedMemory<T>::d_get() const { return m_device; } template<typename T> size_t CudaSynchronizedMemory<T>::size() const { return m_size; } template<typename T> T & CudaSynchronizedMemory<T>::operator[](size_t i) { ASRL_ASSERT_GT_DBG(m_size, 0, "The array is empty"); ASRL_ASSERT_LT_DBG(i, m_size, "Index out of range: " << i << " >= " << m_size);//@} return m_host[i]; } template<typename T> T const & CudaSynchronizedMemory<T>::operator[](size_t i) const { ASRL_ASSERT_GT_DBG(m_size, 0, "The array is empty"); ASRL_ASSERT_LT_DBG(i, m_size, "Index out of range: " << i << " >= " << m_size); return m_host[i]; } template<typename T> void CudaSynchronizedMemory<T>::pushToDevice(size_t nElements) { ASRL_ASSERT_GT_DBG(m_size, 0, "The array is empty"); if(nElements > m_size) nElements = m_size; cudaError_t err = (cudaMemcpy((void*) m_device, (void *)m_host, nElements*sizeof(T), cudaMemcpyHostToDevice)); ASRL_ASSERT_EQ(err, cudaSuccess, "Unable to copy " << typeid(T).name() << " array of size " << m_size << " to device (" << err << "): " << cudaGetErrorString(err)); } template<typename T> void CudaSynchronizedMemory<T>::pullFromDevice(size_t nElements) { ASRL_ASSERT_GT_DBG(m_size,0, "The array is empty"); if(nElements > m_size) nElements = m_size; cudaError_t err = (cudaMemcpy((void*) m_host, (void *)m_device, nElements*sizeof(T), cudaMemcpyDeviceToHost)); ASRL_ASSERT_EQ(err, cudaSuccess, "Unable to copy " << typeid(T).name() << " array of size " << m_size << " from device (" << err << "): " << cudaGetErrorString(err)); } template<typename T> void CudaSynchronizedMemory<T>::pullFromDeviceAsync(cudaStream_t stream, size_t nElements) { ASRL_ASSERT_GT(m_size,0, "The array is empty"); ASRL_ASSERT(m_pageLocked, "Asynchronous transfer is only valid for page-locked host memory"); if(nElements > m_size) nElements = m_size; cudaError_t err = (cudaMemcpyAsync((void*) m_host, (void *)m_device, nElements*sizeof(T), cudaMemcpyDeviceToHost, stream)); ASRL_ASSERT_EQ(err,cudaSuccess, "Unable to copy " << typeid(T).name() << " array of size " << m_size << " from device. Stream " << stream << ": (" << err << "): " << cudaGetErrorString(err)); } } // namespace asrl #endif // ASRL_CUDA_SYNCHRONIZED_MEMORY_HPP