2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2014,2015,2016,2017,2018, 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 */
48 #include "gromacs/gpu_utils/gputraits.cuh"
49 #include "gromacs/math/vec.h"
50 #include "gromacs/math/vectypes.h"
51 #include "gromacs/utility/exceptions.h"
52 #include "gromacs/utility/fatalerror.h"
53 #include "gromacs/utility/gmxassert.h"
54 #include "gromacs/utility/stringutil.h"
61 /*! \brief Helper function to ensure no pending error silently
62 * disrupts error handling.
64 * Asserts in a debug build if an unhandled error is present. Issues a
65 * warning at run time otherwise.
67 * \todo This is similar to CU_CHECK_PREV_ERR, which should be
70 static inline void ensureNoPendingCudaError(const char *errorMessage)
72 // Ensure there is no pending error that would otherwise affect
73 // the behaviour of future error handling.
74 cudaError_t stat = cudaGetLastError();
75 if (stat == cudaSuccess)
80 // If we would find an error in a release build, we do not know
81 // what is appropriate to do about it, so assert only for debug
83 auto fullMessage = formatString("%s An unhandled error from a previous CUDA operation was detected. %s: %s",
84 errorMessage, cudaGetErrorName(stat), cudaGetErrorString(stat));
85 GMX_ASSERT(stat == cudaSuccess, fullMessage.c_str());
86 // TODO When we evolve a better logging framework, use that
87 // for release-build error reporting.
88 gmx_warning("%s", fullMessage.c_str());
94 enum class GpuApiCallBehavior;
96 /* TODO error checking needs to be rewritten. We have 2 types of error checks needed
97 based on where they occur in the code:
98 - non performance-critical: these errors are unsafe to be ignored and must be
99 _always_ checked for, e.g. initializations
100 - performance critical: handling errors might hurt performance so care need to be taken
101 when/if we should check for them at all, e.g. in cu_upload_X. However, we should be
102 able to turn the check for these errors on!
104 Probably we'll need two sets of the macros below...
107 #define CHECK_CUDA_ERRORS
109 #ifdef CHECK_CUDA_ERRORS
111 /*! Check for CUDA error on the return status of a CUDA RT API call. */
112 #define CU_RET_ERR(status, msg) \
114 if (status != cudaSuccess) \
116 gmx_fatal(FARGS, "%s: %s\n", msg, cudaGetErrorString(status)); \
120 /*! Check for any previously occurred uncaught CUDA error. */
121 #define CU_CHECK_PREV_ERR() \
123 cudaError_t _CU_CHECK_PREV_ERR_status = cudaGetLastError(); \
124 if (_CU_CHECK_PREV_ERR_status != cudaSuccess) { \
125 gmx_warning("Just caught a previously occurred CUDA error (%s), will try to continue.", cudaGetErrorString(_CU_CHECK_PREV_ERR_status)); \
129 #else /* CHECK_CUDA_ERRORS */
131 #define CU_RET_ERR(status, msg) do { } while (0)
132 #define CU_CHECK_PREV_ERR() do { } while (0)
133 #define HANDLE_NVML_RET_ERR(status, msg) do { } while (0)
135 #endif /* CHECK_CUDA_ERRORS */
137 /*! \brief CUDA device information.
139 * The CUDA device information is queried and set at detection and contains
140 * both information about the device/hardware returned by the runtime as well
141 * as additional data like support status.
143 * \todo extract an object to manage NVML details
145 struct gmx_device_info_t
147 int id; /* id of the CUDA device */
148 cudaDeviceProp prop; /* CUDA device properties */
149 int stat; /* result of the device check */
150 unsigned int nvml_orig_app_sm_clock; /* The original SM clock before we changed it */
151 unsigned int nvml_orig_app_mem_clock; /* The original memory clock before we changed it */
152 gmx_bool nvml_app_clocks_changed; /* If application clocks have been changed */
153 unsigned int nvml_set_app_sm_clock; /* The SM clock we set */
154 unsigned int nvml_set_app_mem_clock; /* The memory clock we set */
156 nvmlDevice_t nvml_device_id; /* NVML device id */
157 // TODO This can become a bool with a more useful name
158 nvmlEnableState_t nvml_is_restricted; /* Status of application clocks permission */
159 #endif /* HAVE_NVML */
162 /*! Launches synchronous or asynchronous device to host memory copy.
164 * The copy is launched in stream s or if not specified, in stream 0.
166 int cu_copy_D2H(void *h_dest, void *d_src, size_t bytes, GpuApiCallBehavior transferKind, cudaStream_t s /*= 0*/);
168 /*! Launches synchronous host to device memory copy in stream 0. */
169 int cu_copy_D2H_sync(void * /*h_dest*/, void * /*d_src*/, size_t /*bytes*/);
171 /*! Launches asynchronous host to device memory copy in stream s. */
172 int cu_copy_D2H_async(void * /*h_dest*/, void * /*d_src*/, size_t /*bytes*/, cudaStream_t /*s = 0*/);
174 /*! Launches synchronous or asynchronous host to device memory copy.
176 * The copy is launched in stream s or if not specified, in stream 0.
178 int cu_copy_H2D(void *d_dest, void *h_src, size_t bytes, GpuApiCallBehavior transferKind, cudaStream_t /*s = 0*/);
180 /*! Launches synchronous host to device memory copy. */
181 int cu_copy_H2D_sync(void * /*d_dest*/, void * /*h_src*/, size_t /*bytes*/);
183 /*! Launches asynchronous host to device memory copy in stream s. */
184 int cu_copy_H2D_async(void * /*d_dest*/, void * /*h_src*/, size_t /*bytes*/, cudaStream_t /*s = 0*/);
186 // TODO: the 2 functions below are pretty much a constructor/destructor of a simple
187 // GPU table object. There is also almost self-contained fetchFromParamLookupTable()
188 // in cuda_kernel_utils.cuh. They could all live in a separate class/struct file.
190 /*! \brief Initialize parameter lookup table.
192 * Initializes device memory, copies data from host and binds
193 * a texture to allocated device memory to be used for parameter lookup.
195 * \tparam[in] T Raw data type
196 * \param[out] d_ptr device pointer to the memory to be allocated
197 * \param[out] texObj texture object to be initialized
198 * \param[in] h_ptr pointer to the host memory to be uploaded to the device
199 * \param[in] numElem number of elements in the h_ptr
200 * \param[in] devInfo pointer to the info struct of the device in use
202 template <typename T>
203 void initParamLookupTable(T * &d_ptr,
204 cudaTextureObject_t &texObj,
207 const gmx_device_info_t *devInfo);
209 /*! \brief Destroy parameter lookup table.
211 * Unbinds texture object, deallocates device memory.
213 * \tparam[in] T Raw data type
214 * \param[in] d_ptr Device pointer to the memory to be deallocated
215 * \param[in] texObj Texture object to be deinitialized
216 * \param[in] devInfo Pointer to the info struct of the device in use
218 template <typename T>
219 void destroyParamLookupTable(T *d_ptr,
220 cudaTextureObject_t texObj,
221 const gmx_device_info_t *devInfo);
223 /*! \brief Add a triplets stored in a float3 to an rvec variable.
225 * \param[out] a Rvec to increment
226 * \param[in] b Float triplet to increment with.
228 static inline void rvec_inc(rvec a, const float3 b)
230 rvec tmp = {b.x, b.y, b.z};
234 /*! \brief Wait for all taks in stream \p s to complete.
236 * \param[in] s stream to synchronize with
238 static inline void gpuStreamSynchronize(cudaStream_t s)
240 cudaError_t stat = cudaStreamSynchronize(s);
241 CU_RET_ERR(stat, "cudaStreamSynchronize failed");
244 /*! \brief Returns true if all tasks in \p s have completed.
246 * \param[in] s stream to check
248 * \returns True if all tasks enqueued in the stream \p s (at the time of this call) have completed.
250 static inline bool haveStreamTasksCompleted(cudaStream_t s)
252 cudaError_t stat = cudaStreamQuery(s);
254 if (stat == cudaErrorNotReady)
256 // work is still in progress in the stream
260 GMX_ASSERT(stat != cudaErrorInvalidResourceHandle, "Stream idnetifier not valid");
262 // cudaSuccess and cudaErrorNotReady are the expected return values
263 CU_RET_ERR(stat, "Unexpected cudaStreamQuery failure");
265 GMX_ASSERT(stat == cudaSuccess, "Values other than cudaSuccess should have been explicitly handled");
270 /* Kernel launch helpers */
273 * A function for setting up a single CUDA kernel argument.
274 * This is the tail of the compile-time recursive function below.
275 * It has to be seen by the compiler first.
277 * \tparam totalArgsCount Number of the kernel arguments
278 * \tparam KernelPtr Kernel function handle type
279 * \param[in] argIndex Index of the current argument
281 template <size_t totalArgsCount, typename KernelPtr>
282 void prepareGpuKernelArgument(KernelPtr /*kernel*/,
283 std::array<void *, totalArgsCount> */* kernelArgsPtr */,
284 size_t gmx_used_in_debug argIndex)
286 GMX_ASSERT(argIndex == totalArgsCount, "Tail expansion");
290 * Compile-time recursive function for setting up a single CUDA kernel argument.
291 * This function copies a kernel argument pointer \p argPtr into \p kernelArgsPtr,
292 * and calls itself on the next argument, eventually calling the tail function above.
294 * \tparam CurrentArg Type of the current argument
295 * \tparam RemainingArgs Types of remaining arguments after the current one
296 * \tparam totalArgsCount Number of the kernel arguments
297 * \tparam KernelPtr Kernel function handle type
298 * \param[in] kernel Kernel function handle
299 * \param[in,out] kernelArgsPtr Pointer to the argument array to be filled in
300 * \param[in] argIndex Index of the current argument
301 * \param[in] argPtr Pointer to the current argument
302 * \param[in] otherArgsPtrs Pack of pointers to arguments remaining to process after the current one
304 template <typename CurrentArg, typename ... RemainingArgs, size_t totalArgsCount, typename KernelPtr>
305 void prepareGpuKernelArgument(KernelPtr kernel,
306 std::array<void *, totalArgsCount> *kernelArgsPtr,
308 const CurrentArg *argPtr,
309 const RemainingArgs *... otherArgsPtrs)
311 (*kernelArgsPtr)[argIndex] = (void *)argPtr;
312 prepareGpuKernelArgument(kernel, kernelArgsPtr, argIndex + 1, otherArgsPtrs ...);
316 * A wrapper function for setting up all the CUDA kernel arguments.
317 * Calls the recursive functions above.
319 * \tparam Args Types of all the kernel arguments
320 * \param[in] kernel Kernel function handle
321 * \param[in] argsPtrs Pointers to all the kernel arguments
322 * \returns A prepared parameter pack to be used with launchGpuKernel() as the last argument.
324 template <typename ... Args>
325 std::array<void *, sizeof ... (Args)> prepareGpuKernelArguments(void (*kernel)(Args...),
326 const KernelLaunchConfig & /*config */,
327 const Args *... argsPtrs)
329 std::array<void *, sizeof ... (Args)> kernelArgs;
330 prepareGpuKernelArgument(kernel, &kernelArgs, 0, argsPtrs ...);
334 /*! \brief Launches the CUDA kernel and handles the errors.
336 * \tparam Args Types of all the kernel arguments
337 * \param[in] kernel Kernel function handle
338 * \param[in] config Kernel configuration for launching
339 * \param[in] kernelName Human readable kernel description, for error handling only
340 * \param[in] kernelArgs Array of the pointers to the kernel arguments, prepared by prepareGpuKernelArguments()
341 * \throws gmx::InternalError on kernel launch failure
343 template <typename... Args>
344 void launchGpuKernel(void (*kernel)(Args...),
345 const KernelLaunchConfig &config,
346 CommandEvent */*timingEvent */,
347 const char *kernelName,
348 const std::array<void *, sizeof ... (Args)> &kernelArgs)
350 dim3 blockSize(config.blockSize[0], config.blockSize[1], config.blockSize[2]);
351 dim3 gridSize(config.gridSize[0], config.gridSize[1], config.gridSize[2]);
352 cudaLaunchKernel((void *)kernel, gridSize, blockSize, const_cast<void **>(kernelArgs.data()), config.sharedMemorySize, config.stream);
354 cudaError_t status = cudaGetLastError();
355 if (cudaSuccess != status)
357 const std::string errorMessage = "GPU kernel (" + std::string(kernelName) +
358 ") failed to launch: " + std::string(cudaGetErrorString(status));
359 GMX_THROW(gmx::InternalError(errorMessage));