2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2014,2015,2016,2017, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
35 #ifndef GMX_GPU_UTILS_CUDAUTILS_CUH
36 #define GMX_GPU_UTILS_CUDAUTILS_CUH
43 #endif /* HAVE_NVML */
45 #include "gromacs/math/vec.h"
46 #include "gromacs/math/vectypes.h"
47 #include "gromacs/utility/fatalerror.h"
49 /* TODO error checking needs to be rewritten. We have 2 types of error checks needed
50 based on where they occur in the code:
51 - non performance-critical: these errors are unsafe to be ignored and must be
52 _always_ checked for, e.g. initializations
53 - performance critical: handling errors might hurt performance so care need to be taken
54 when/if we should check for them at all, e.g. in cu_upload_X. However, we should be
55 able to turn the check for these errors on!
57 Probably we'll need two sets of the macros below...
60 #define CHECK_CUDA_ERRORS
62 #ifdef CHECK_CUDA_ERRORS
64 /*! Check for CUDA error on the return status of a CUDA RT API call. */
65 #define CU_RET_ERR(status, msg) \
67 if (status != cudaSuccess) \
69 gmx_fatal(FARGS, "%s: %s\n", msg, cudaGetErrorString(status)); \
73 /*! Check for any previously occurred uncaught CUDA error. */
74 #define CU_CHECK_PREV_ERR() \
76 cudaError_t _CU_CHECK_PREV_ERR_status = cudaGetLastError(); \
77 if (_CU_CHECK_PREV_ERR_status != cudaSuccess) { \
78 gmx_warning("Just caught a previously occurred CUDA error (%s), will try to continue.", cudaGetErrorString(_CU_CHECK_PREV_ERR_status)); \
82 /*! Check for any previously occurred uncaught CUDA error
83 -- aimed at use after kernel calls. */
84 #define CU_LAUNCH_ERR(msg) \
86 cudaError_t _CU_LAUNCH_ERR_status = cudaGetLastError(); \
87 if (_CU_LAUNCH_ERR_status != cudaSuccess) { \
88 gmx_fatal(FARGS, "Error while launching kernel %s: %s\n", msg, cudaGetErrorString(_CU_LAUNCH_ERR_status)); \
92 /*! Synchronize with GPU and check for any previously occurred uncaught CUDA error
93 -- aimed at use after kernel calls. */
94 #define CU_LAUNCH_ERR_SYNC(msg) \
96 cudaError_t _CU_SYNC_LAUNCH_ERR_status = cudaThreadSynchronize(); \
97 if (_CU_SYNC_LAUNCH_ERR_status != cudaSuccess) { \
98 gmx_fatal(FARGS, "Error while launching kernel %s: %s\n", msg, cudaGetErrorString(_CU_SYNC_LAUNCH_ERR_status)); \
102 #else /* CHECK_CUDA_ERRORS */
104 #define CU_RET_ERR(status, msg) do { } while (0)
105 #define CU_CHECK_PREV_ERR() do { } while (0)
106 #define CU_LAUNCH_ERR(msg) do { } while (0)
107 #define CU_LAUNCH_ERR_SYNC(msg) do { } while (0)
108 #define HANDLE_NVML_RET_ERR(status, msg) do { } while (0)
110 #endif /* CHECK_CUDA_ERRORS */
112 /*! \brief CUDA device information.
114 * The CUDA device information is queried and set at detection and contains
115 * both information about the device/hardware returned by the runtime as well
116 * as additional data like support status.
118 * \todo extract an object to manage NVML details
120 struct gmx_device_info_t
122 int id; /* id of the CUDA device */
123 cudaDeviceProp prop; /* CUDA device properties */
124 int stat; /* result of the device check */
125 unsigned int nvml_orig_app_sm_clock; /* The original SM clock before we changed it */
126 unsigned int nvml_orig_app_mem_clock; /* The original memory clock before we changed it */
127 gmx_bool nvml_app_clocks_changed; /* If application clocks have been changed */
128 unsigned int nvml_set_app_sm_clock; /* The SM clock we set */
129 unsigned int nvml_set_app_mem_clock; /* The memory clock we set */
131 nvmlDevice_t nvml_device_id; /* NVML device id */
132 // TODO This can become a bool with a more useful name
133 nvmlEnableState_t nvml_is_restricted; /* Status of application clocks permission */
134 #endif /* HAVE_NVML */
138 /*! Launches asynchronous host to device memory copy in stream 0. */
139 int cu_copy_D2H(void * /*h_dest*/, void * /*d_src*/, size_t /*bytes*/);
141 /*! Launches asynchronous host to device memory copy in stream s. */
142 int cu_copy_D2H_async(void * /*h_dest*/, void * /*d_src*/, size_t /*bytes*/, cudaStream_t /*s = 0*/);
144 /*! Launches synchronous host to device memory copy. */
145 int cu_copy_H2D(void * /*d_dest*/, void * /*h_src*/, size_t /*bytes*/);
147 /*! Launches asynchronous host to device memory copy in stream s. */
148 int cu_copy_H2D_async(void * /*d_dest*/, void * /*h_src*/, size_t /*bytes*/, cudaStream_t /*s = 0*/);
150 /*! Frees device memory and resets the size and allocation size to -1. */
151 void cu_free_buffered(void *d_ptr, int *n = NULL, int *nalloc = NULL);
153 /*! Reallocates the device memory and copies data from the host. */
154 void cu_realloc_buffered(void **d_dest, void *h_src,
156 int *curr_size, int *curr_alloc_size,
161 // TODO: the 2 functions below are pretty much a constructor/destructor of a simple
162 // GPU table object. There is also almost self-contained fetchFromParamLookupTable()
163 // in cuda_kernel_utils.cuh. They could all live in a separate class/struct file,
164 // granted storing static texture references in there does not pose problems.
166 /*! \brief Initialize parameter lookup table.
168 * Initializes device memory, copies data from host and binds
169 * a texture to allocated device memory to be used for parameter lookup.
171 * \tparam[in] T Raw data type
172 * \param[out] d_ptr device pointer to the memory to be allocated
173 * \param[out] texObj texture object to be initialized
174 * \param[out] texRef texture reference to be initialized
175 * \param[in] h_ptr pointer to the host memory to be uploaded to the device
176 * \param[in] numElem number of elements in the h_ptr
177 * \param[in] devInfo pointer to the info struct of the device in use
179 template <typename T>
180 void initParamLookupTable(T * &d_ptr,
181 cudaTextureObject_t &texObj,
182 const struct texture<T, 1, cudaReadModeElementType> *texRef,
185 const gmx_device_info_t *devInfo);
187 /*! \brief Destroy parameter lookup table.
189 * Unbinds texture reference/object, deallocates device memory.
191 * \tparam[in] T Raw data type
192 * \param[in] d_ptr Device pointer to the memory to be deallocated
193 * \param[in] texObj Texture object to be deinitialized
194 * \param[in] texRef Texture reference to be deinitialized
195 * \param[in] devInfo Pointer to the info struct of the device in use
197 template <typename T>
198 void destroyParamLookupTable(T *d_ptr,
199 cudaTextureObject_t texObj,
200 const struct texture<T, 1, cudaReadModeElementType> *texRef,
201 const gmx_device_info_t *devInfo);
203 /*! \brief Add a triplets stored in a float3 to an rvec variable.
205 * \param[out] a Rvec to increment
206 * \param[in] b Float triplet to increment with.
208 static inline void rvec_inc(rvec a, const float3 b)
210 rvec tmp = {b.x, b.y, b.z};
214 /*! \brief Calls cudaStreamSynchronize() in the stream \p s.
216 * \param[in] s stream to synchronize with
218 static inline void gpuStreamSynchronize(cudaStream_t s)
220 cudaError_t stat = cudaStreamSynchronize(s);
221 CU_RET_ERR(stat, "cudaStreamSynchronize failed");