2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2014,2015,2016,2017 by the GROMACS development team.
5 * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
39 #include "cudautils.cuh"
44 #include "gromacs/gpu_utils/cuda_arch_utils.cuh"
45 #include "gromacs/gpu_utils/gpu_utils.h"
46 #include "gromacs/utility/gmxassert.h"
48 /*** Generic CUDA data operation wrappers ***/
50 // TODO: template on transferKind to avoid runtime conditionals
51 int cu_copy_D2H(void* h_dest, void* d_src, size_t bytes, GpuApiCallBehavior transferKind, cudaStream_t s = nullptr)
55 if (h_dest == nullptr || d_src == nullptr || bytes == 0)
62 case GpuApiCallBehavior::Async:
63 GMX_ASSERT(isHostMemoryPinned(h_dest), "Destination buffer was not pinned for CUDA");
64 stat = cudaMemcpyAsync(h_dest, d_src, bytes, cudaMemcpyDeviceToHost, s);
65 CU_RET_ERR(stat, "DtoH cudaMemcpyAsync failed");
68 case GpuApiCallBehavior::Sync:
69 stat = cudaMemcpy(h_dest, d_src, bytes, cudaMemcpyDeviceToHost);
70 CU_RET_ERR(stat, "DtoH cudaMemcpy failed");
79 int cu_copy_D2H_sync(void* h_dest, void* d_src, size_t bytes)
81 return cu_copy_D2H(h_dest, d_src, bytes, GpuApiCallBehavior::Sync);
85 * The copy is launched in stream s or if not specified, in stream 0.
87 int cu_copy_D2H_async(void* h_dest, void* d_src, size_t bytes, cudaStream_t s = nullptr)
89 return cu_copy_D2H(h_dest, d_src, bytes, GpuApiCallBehavior::Async, s);
92 // TODO: template on transferKind to avoid runtime conditionals
93 int cu_copy_H2D(void* d_dest, const void* h_src, size_t bytes, GpuApiCallBehavior transferKind, cudaStream_t s = nullptr)
97 if (d_dest == nullptr || h_src == nullptr || bytes == 0)
102 switch (transferKind)
104 case GpuApiCallBehavior::Async:
105 GMX_ASSERT(isHostMemoryPinned(h_src), "Source buffer was not pinned for CUDA");
106 stat = cudaMemcpyAsync(d_dest, h_src, bytes, cudaMemcpyHostToDevice, s);
107 CU_RET_ERR(stat, "HtoD cudaMemcpyAsync failed");
110 case GpuApiCallBehavior::Sync:
111 stat = cudaMemcpy(d_dest, h_src, bytes, cudaMemcpyHostToDevice);
112 CU_RET_ERR(stat, "HtoD cudaMemcpy failed");
121 int cu_copy_H2D_sync(void* d_dest, const void* h_src, size_t bytes)
123 return cu_copy_H2D(d_dest, h_src, bytes, GpuApiCallBehavior::Sync);
127 * The copy is launched in stream s or if not specified, in stream 0.
129 int cu_copy_H2D_async(void* d_dest, const void* h_src, size_t bytes, cudaStream_t s = nullptr)
131 return cu_copy_H2D(d_dest, h_src, bytes, GpuApiCallBehavior::Async, s);
134 /*! \brief Set up texture object for an array of type T.
136 * Set up texture object for an array of type T and bind it to the device memory
137 * \p d_ptr points to.
139 * \tparam[in] T Raw data type
140 * \param[out] texObj texture object to initialize
141 * \param[in] d_ptr pointer to device global memory to bind \p texObj to
142 * \param[in] sizeInBytes size of memory area to bind \p texObj to
145 static void setup1DTexture(cudaTextureObject_t& texObj, void* d_ptr, size_t sizeInBytes)
147 assert(!c_disableCudaTextures);
153 memset(&rd, 0, sizeof(rd));
154 rd.resType = cudaResourceTypeLinear;
155 rd.res.linear.devPtr = d_ptr;
156 rd.res.linear.desc = cudaCreateChannelDesc<T>();
157 rd.res.linear.sizeInBytes = sizeInBytes;
159 memset(&td, 0, sizeof(td));
160 td.readMode = cudaReadModeElementType;
161 stat = cudaCreateTextureObject(&texObj, &rd, &td, nullptr);
162 CU_RET_ERR(stat, "cudaCreateTextureObject failed");
166 void initParamLookupTable(T*& d_ptr, cudaTextureObject_t& texObj, const T* h_ptr, int numElem)
168 const size_t sizeInBytes = numElem * sizeof(*d_ptr);
169 cudaError_t stat = cudaMalloc((void**)&d_ptr, sizeInBytes);
170 CU_RET_ERR(stat, "cudaMalloc failed in initParamLookupTable");
171 cu_copy_H2D_sync(d_ptr, (void*)h_ptr, sizeInBytes);
173 if (!c_disableCudaTextures)
175 setup1DTexture<T>(texObj, d_ptr, sizeInBytes);
180 void destroyParamLookupTable(T* d_ptr, cudaTextureObject_t texObj)
182 if (!c_disableCudaTextures)
184 CU_RET_ERR(cudaDestroyTextureObject(texObj), "cudaDestroyTextureObject on texObj failed");
186 CU_RET_ERR(cudaFree(d_ptr), "cudaFree failed");
189 /*! \brief Add explicit instantiations of init/destroyParamLookupTable() here as needed.
190 * One should also verify that the result of cudaCreateChannelDesc<T>() during texture setup
191 * looks reasonable, when instantiating the templates for new types - just in case.
193 template void initParamLookupTable<float>(float*&, cudaTextureObject_t&, const float*, int);
194 template void destroyParamLookupTable<float>(float*, cudaTextureObject_t);
195 template void initParamLookupTable<int>(int*&, cudaTextureObject_t&, const int*, int);
196 template void destroyParamLookupTable<int>(int*, cudaTextureObject_t);