2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2010,2011,2012,2013,2014,2015,2016,2017, 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.
36 * \brief Define functions for detection and initialization for CUDA devices.
38 * \author Szilard Pall <pall.szilard@gmail.com>
43 #include "gpu_utils.h"
51 #include <cuda_profiler_api.h>
53 #include "gromacs/gpu_utils/cudautils.cuh"
54 #include "gromacs/gpu_utils/pmalloc_cuda.h"
55 #include "gromacs/hardware/gpu_hw_info.h"
56 #include "gromacs/utility/basedefinitions.h"
57 #include "gromacs/utility/cstringutil.h"
58 #include "gromacs/utility/exceptions.h"
59 #include "gromacs/utility/fatalerror.h"
60 #include "gromacs/utility/gmxassert.h"
61 #include "gromacs/utility/logger.h"
62 #include "gromacs/utility/programcontext.h"
63 #include "gromacs/utility/smalloc.h"
64 #include "gromacs/utility/snprintf.h"
65 #include "gromacs/utility/stringutil.h"
69 #define HAVE_NVML_APPLICATION_CLOCKS (NVML_API_VERSION >= 6)
71 #define HAVE_NVML_APPLICATION_CLOCKS 0
72 #endif /* HAVE_NVML */
74 #if defined(CHECK_CUDA_ERRORS) && HAVE_NVML_APPLICATION_CLOCKS
75 /*! Check for NVML error on the return status of a NVML API call. */
76 # define HANDLE_NVML_RET_ERR(status, msg) \
78 if (status != NVML_SUCCESS) \
80 gmx_warning("%s: %s\n", msg, nvmlErrorString(status)); \
83 #else /* defined(CHECK_CUDA_ERRORS) && HAVE_NVML_APPLICATION_CLOCKS */
84 # define HANDLE_NVML_RET_ERR(status, msg) do { } while (0)
85 #endif /* defined(CHECK_CUDA_ERRORS) && HAVE_NVML_APPLICATION_CLOCKS */
87 #if HAVE_NVML_APPLICATION_CLOCKS
88 static const gmx_bool bCompiledWithApplicationClockSupport = true;
90 static const gmx_bool gmx_unused bCompiledWithApplicationClockSupport = false;
94 * Max number of devices supported by CUDA (for consistency checking).
96 * In reality it is 16 with CUDA <=v5.0, but let's stay on the safe side.
98 static int cuda_max_device_count = 32;
100 static bool cudaProfilerRun = ((getenv("NVPROF_ID") != NULL));
102 /** Dummy kernel used for sanity checking. */
103 static __global__ void k_dummy_test(void)
107 static void checkCompiledTargetCompatibility(const gmx_device_info_t *devInfo)
111 cudaFuncAttributes attributes;
112 cudaError_t stat = cudaFuncGetAttributes(&attributes, k_dummy_test);
114 if (cudaErrorInvalidDeviceFunction == stat)
117 "The %s binary does not include support for the CUDA architecture "
118 "of the selected GPU (device ID #%d, compute capability %d.%d). "
119 "By default, GROMACS supports all common architectures, so your GPU "
120 "might be rare, or some architectures were disabled in the build. ",
121 "Consult the install guide for how to use the GMX_CUDA_TARGET_SM and ",
122 "GMX_CUDA_TARGET_COMPUTE CMake variables to add this architecture.",
123 gmx::getProgramContext().displayName(), devInfo->id,
124 devInfo->prop.major, devInfo->prop.minor);
127 CU_RET_ERR(stat, "cudaFuncGetAttributes failed");
129 if (devInfo->prop.major >= 3 && attributes.ptxVersion < 30)
132 "The GPU device code was compiled at runtime from 2.0 source which is "
133 "not compatible with the selected GPU (device ID #%d, compute capability %d.%d). "
134 "Pass the appropriate target in GMX_CUDA_TARGET_SM or a >=30 value to GMX_CUDA_TARGET_COMPUTE.",
136 devInfo->prop.major, devInfo->prop.minor);
140 bool isHostMemoryPinned(void *h_ptr)
142 cudaPointerAttributes memoryAttributes;
143 cudaError_t stat = cudaPointerGetAttributes(&memoryAttributes, h_ptr);
152 case cudaErrorInvalidValue:
153 // If the buffer was not pinned, then it will not be recognized by CUDA at all
155 // Reset the last error status
160 CU_RET_ERR(stat, "Unexpected CUDA error");
166 * \brief Runs GPU sanity checks.
168 * Runs a series of checks to determine that the given GPU and underlying CUDA
169 * driver/runtime functions properly.
170 * Returns properties of a device with given ID or the one that has
171 * already been initialized earlier in the case if of \dev_id == -1.
173 * \param[in] dev_id the device ID of the GPU or -1 if the device has already been initialized
174 * \param[out] dev_prop pointer to the structure in which the device properties will be returned
175 * \returns 0 if the device looks OK
177 * TODO: introduce errors codes and handle errors more smoothly.
179 static int do_sanity_checks(int dev_id, cudaDeviceProp *dev_prop)
184 cu_err = cudaGetDeviceCount(&dev_count);
185 if (cu_err != cudaSuccess)
187 fprintf(stderr, "Error %d while querying device count: %s\n", cu_err,
188 cudaGetErrorString(cu_err));
192 /* no CUDA compatible device at all */
198 /* things might go horribly wrong if cudart is not compatible with the driver */
199 if (dev_count < 0 || dev_count > cuda_max_device_count)
204 if (dev_id == -1) /* device already selected let's not destroy the context */
206 cu_err = cudaGetDevice(&id);
207 if (cu_err != cudaSuccess)
209 fprintf(stderr, "Error %d while querying device id: %s\n", cu_err,
210 cudaGetErrorString(cu_err));
217 if (id > dev_count - 1) /* pfff there's no such device */
219 fprintf(stderr, "The requested device with id %d does not seem to exist (device count=%d)\n",
225 memset(dev_prop, 0, sizeof(cudaDeviceProp));
226 cu_err = cudaGetDeviceProperties(dev_prop, id);
227 if (cu_err != cudaSuccess)
229 fprintf(stderr, "Error %d while querying device properties: %s\n", cu_err,
230 cudaGetErrorString(cu_err));
234 /* both major & minor is 9999 if no CUDA capable devices are present */
235 if (dev_prop->major == 9999 && dev_prop->minor == 9999)
239 /* we don't care about emulation mode */
240 if (dev_prop->major == 0)
247 cu_err = cudaSetDevice(id);
248 if (cu_err != cudaSuccess)
250 fprintf(stderr, "Error %d while switching to device #%d: %s\n",
251 cu_err, id, cudaGetErrorString(cu_err));
256 /* try to execute a dummy kernel */
257 k_dummy_test<<< 1, 512>>> ();
258 if (cudaThreadSynchronize() != cudaSuccess)
263 /* destroy context if we created one */
266 cu_err = cudaDeviceReset();
267 CU_RET_ERR(cu_err, "cudaDeviceReset failed");
273 #if HAVE_NVML_APPLICATION_CLOCKS
274 /*! \brief Determines and adds the NVML device ID to the passed \cuda_dev.
276 * Determines and adds the NVML device ID to the passed \cuda_dev. This is done by
277 * matching PCI-E information from \cuda_dev with the available NVML devices.
279 * \param[in,out] cuda_dev CUDA device information to enrich with NVML device info
280 * \returns true if \cuda_dev could be enriched with matching NVML device information.
282 static bool addNVMLDeviceId(gmx_device_info_t* cuda_dev)
284 nvmlDevice_t nvml_device_id;
285 unsigned int nvml_device_count = 0;
286 nvmlReturn_t nvml_stat = nvmlDeviceGetCount ( &nvml_device_count );
287 bool nvmlWasInitialized = false;
288 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlDeviceGetCount failed" );
289 for (unsigned int nvml_device_idx = 0; nvml_stat == NVML_SUCCESS && nvml_device_idx < nvml_device_count; ++nvml_device_idx)
291 nvml_stat = nvmlDeviceGetHandleByIndex ( nvml_device_idx, &nvml_device_id );
292 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlDeviceGetHandleByIndex failed" );
293 if (nvml_stat != NVML_SUCCESS)
298 nvmlPciInfo_t nvml_pci_info;
299 nvml_stat = nvmlDeviceGetPciInfo ( nvml_device_id, &nvml_pci_info );
300 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlDeviceGetPciInfo failed" );
301 if (nvml_stat != NVML_SUCCESS)
305 if (static_cast<unsigned int>(cuda_dev->prop.pciBusID) == nvml_pci_info.bus &&
306 static_cast<unsigned int>(cuda_dev->prop.pciDeviceID) == nvml_pci_info.device &&
307 static_cast<unsigned int>(cuda_dev->prop.pciDomainID) == nvml_pci_info.domain)
309 nvmlWasInitialized = true;
310 cuda_dev->nvml_device_id = nvml_device_id;
314 return nvmlWasInitialized;
317 /*! \brief Reads and returns the application clocks for device.
319 * \param[in] device The GPU device
320 * \param[out] app_sm_clock The current application SM clock
321 * \param[out] app_mem_clock The current application memory clock
322 * \returns if applacation clocks are supported
324 static bool getApplicationClocks(const gmx_device_info_t *cuda_dev,
325 unsigned int *app_sm_clock,
326 unsigned int *app_mem_clock)
328 nvmlReturn_t nvml_stat;
330 nvml_stat = nvmlDeviceGetApplicationsClock(cuda_dev->nvml_device_id, NVML_CLOCK_SM, app_sm_clock);
331 if (NVML_ERROR_NOT_SUPPORTED == nvml_stat)
335 HANDLE_NVML_RET_ERR(nvml_stat, "nvmlDeviceGetApplicationsClock failed for NVIDIA_CLOCK_SM");
336 nvml_stat = nvmlDeviceGetApplicationsClock(cuda_dev->nvml_device_id, NVML_CLOCK_MEM, app_mem_clock);
337 HANDLE_NVML_RET_ERR(nvml_stat, "nvmlDeviceGetApplicationsClock failed for NVIDIA_CLOCK_MEM");
341 #endif /* HAVE_NVML_APPLICATION_CLOCKS */
343 /*! \brief Tries to set application clocks for the GPU with the given index.
345 * Application clocks are set to the max supported value to increase
346 * performance if application clock permissions allow this. For future
347 * GPU architectures a more sophisticated scheme might be required.
349 * \todo Refactor this into a detection phase and a work phase. Also
350 * refactor to remove compile-time dependence on logging header.
352 * \param mdlog log file to write to
353 * \param[in] cuda_dev GPU device info for the GPU in use
354 * \returns true if no error occurs during application clocks handling.
356 static gmx_bool init_gpu_application_clocks(
357 const gmx::MDLogger &mdlog,
358 gmx_device_info_t *cuda_dev)
360 const cudaDeviceProp *prop = &cuda_dev->prop;
361 int cuda_compute_capability = prop->major * 10 + prop->minor;
362 gmx_bool bGpuCanUseApplicationClocks =
363 ((0 == gmx_wcmatch("*Tesla*", prop->name) && cuda_compute_capability >= 35 ) ||
364 (0 == gmx_wcmatch("*Quadro*", prop->name) && cuda_compute_capability >= 52 ));
365 if (!bGpuCanUseApplicationClocks)
370 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
371 "NOTE: GROMACS was configured without NVML support hence it can not exploit\n"
372 " application clocks of the detected %s GPU to improve performance.\n"
373 " Recompile with the NVML library (compatible with the driver used) or set application clocks manually.",
377 if (!bCompiledWithApplicationClockSupport)
379 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
380 "NOTE: GROMACS was compiled with an old NVML library which does not support\n"
381 " managing application clocks of the detected %s GPU to improve performance.\n"
382 " If your GPU supports application clocks, upgrade NVML (and driver) and recompile or set the clocks manually.",
387 /* We've compiled with NVML application clocks support, and have a GPU that can use it */
388 nvmlReturn_t nvml_stat = NVML_SUCCESS;
390 //TODO: GMX_GPU_APPLICATION_CLOCKS is currently only used to enable/disable setting of application clocks
391 // this variable can be later used to give a user more fine grained control.
392 env = getenv("GMX_GPU_APPLICATION_CLOCKS");
393 if (env != NULL && ( strcmp( env, "0") == 0 ||
394 gmx_strcasecmp( env, "OFF") == 0 ||
395 gmx_strcasecmp( env, "DISABLE") == 0 ))
399 nvml_stat = nvmlInit();
400 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlInit failed." );
401 if (nvml_stat != NVML_SUCCESS)
406 if (!addNVMLDeviceId(cuda_dev))
410 //get current application clocks setting
411 if (!getApplicationClocks(cuda_dev,
412 &cuda_dev->nvml_orig_app_sm_clock,
413 &cuda_dev->nvml_orig_app_mem_clock))
417 //get max application clocks
418 unsigned int max_sm_clock = 0;
419 unsigned int max_mem_clock = 0;
420 nvml_stat = nvmlDeviceGetMaxClockInfo(cuda_dev->nvml_device_id, NVML_CLOCK_SM, &max_sm_clock);
421 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlDeviceGetMaxClockInfo failed" );
422 nvml_stat = nvmlDeviceGetMaxClockInfo(cuda_dev->nvml_device_id, NVML_CLOCK_MEM, &max_mem_clock);
423 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlDeviceGetMaxClockInfo failed" );
425 cuda_dev->nvml_is_restricted = NVML_FEATURE_ENABLED;
426 cuda_dev->nvml_app_clocks_changed = false;
428 if (cuda_dev->nvml_orig_app_sm_clock >= max_sm_clock)
430 //TODO: This should probably be integrated into the GPU Properties table.
431 GMX_LOG(mdlog.info).appendTextFormatted(
432 "Application clocks (GPU clocks) for %s are (%d,%d)",
433 cuda_dev->prop.name, cuda_dev->nvml_orig_app_mem_clock, cuda_dev->nvml_orig_app_sm_clock);
437 if (cuda_compute_capability >= 60)
439 // Only warn about not being able to change clocks if they are not already at the max values
440 if (max_mem_clock > cuda_dev->nvml_orig_app_mem_clock || max_sm_clock > cuda_dev->nvml_orig_app_sm_clock)
442 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
443 "Cannot change application clocks for %s to optimal values due to insufficient permissions. Current values are (%d,%d), max values are (%d,%d).\nPlease contact your admin to change application clocks.\n",
444 cuda_dev->prop.name, cuda_dev->nvml_orig_app_mem_clock, cuda_dev->nvml_orig_app_sm_clock, max_mem_clock, max_sm_clock);
449 nvml_stat = nvmlDeviceGetAPIRestriction(cuda_dev->nvml_device_id, NVML_RESTRICTED_API_SET_APPLICATION_CLOCKS, &(cuda_dev->nvml_is_restricted));
450 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlDeviceGetAPIRestriction failed" );
452 if (nvml_stat != NVML_SUCCESS)
454 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
455 "Cannot change GPU application clocks to optimal values due to NVML error (%d): %s.",
456 nvml_stat, nvmlErrorString(nvml_stat));
460 if (cuda_dev->nvml_is_restricted != NVML_FEATURE_DISABLED)
462 // Only warn about not being able to change clocks if they are not already at the max values
463 if (max_mem_clock > cuda_dev->nvml_orig_app_mem_clock || max_sm_clock > cuda_dev->nvml_orig_app_sm_clock)
465 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
466 "Cannot change application clocks for %s to optimal values due to insufficient permissions. Current values are (%d,%d), max values are (%d,%d).\nUse sudo nvidia-smi -acp UNRESTRICTED or contact your admin to change application clocks.",
467 cuda_dev->prop.name, cuda_dev->nvml_orig_app_mem_clock, cuda_dev->nvml_orig_app_sm_clock, max_mem_clock, max_sm_clock);
472 /* Note: Distinguishing between different types of GPUs here might be necessary in the future,
473 e.g. if max application clocks should not be used for certain GPUs. */
474 GMX_LOG(mdlog.warning).appendTextFormatted(
475 "Changing GPU application clocks for %s to (%d,%d)",
476 cuda_dev->prop.name, max_mem_clock, max_sm_clock);
477 nvml_stat = nvmlDeviceSetApplicationsClocks(cuda_dev->nvml_device_id, max_mem_clock, max_sm_clock);
478 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlDeviceGetApplicationsClock failed" );
479 cuda_dev->nvml_app_clocks_changed = true;
480 cuda_dev->nvml_set_app_sm_clock = max_sm_clock;
481 cuda_dev->nvml_set_app_mem_clock = max_mem_clock;
484 #endif /* HAVE_NVML */
487 /*! \brief Resets application clocks if changed and cleans up NVML for the passed \gpu_dev.
489 * \param[in] gpu_dev CUDA device information
491 static gmx_bool reset_gpu_application_clocks(const gmx_device_info_t gmx_unused * cuda_dev)
493 #if !HAVE_NVML_APPLICATION_CLOCKS
494 GMX_UNUSED_VALUE(cuda_dev);
496 #else /* HAVE_NVML_APPLICATION_CLOCKS */
497 nvmlReturn_t nvml_stat = NVML_SUCCESS;
499 cuda_dev->nvml_is_restricted == NVML_FEATURE_DISABLED &&
500 cuda_dev->nvml_app_clocks_changed)
502 /* Check if the clocks are still what we set them to.
503 * If so, set them back to the state we originally found them in.
504 * If not, don't touch them, because something else set them later.
506 unsigned int app_sm_clock, app_mem_clock;
507 getApplicationClocks(cuda_dev, &app_sm_clock, &app_mem_clock);
508 if (app_sm_clock == cuda_dev->nvml_set_app_sm_clock &&
509 app_mem_clock == cuda_dev->nvml_set_app_mem_clock)
511 nvml_stat = nvmlDeviceSetApplicationsClocks(cuda_dev->nvml_device_id, cuda_dev->nvml_orig_app_mem_clock, cuda_dev->nvml_orig_app_sm_clock);
512 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlDeviceSetApplicationsClock failed" );
515 nvml_stat = nvmlShutdown();
516 HANDLE_NVML_RET_ERR( nvml_stat, "nvmlShutdown failed" );
517 return (nvml_stat == NVML_SUCCESS);
518 #endif /* HAVE_NVML_APPLICATION_CLOCKS */
521 void init_gpu(const gmx::MDLogger &mdlog,
522 gmx_device_info_t *deviceInfo)
528 stat = cudaSetDevice(deviceInfo->id);
529 if (stat != cudaSuccess)
531 auto message = gmx::formatString("Failed to initialize GPU #%d", deviceInfo->id);
532 CU_RET_ERR(stat, message.c_str());
537 fprintf(stderr, "Initialized GPU ID #%d: %s\n", deviceInfo->id, deviceInfo->prop.name);
540 checkCompiledTargetCompatibility(deviceInfo);
542 //Ignoring return value as NVML errors should be treated not critical.
543 init_gpu_application_clocks(mdlog, deviceInfo);
546 void free_gpu(const gmx_device_info_t *deviceInfo)
548 // One should only attempt to clear the device context when
549 // it has been used, but currently the only way to know that a GPU
550 // device was used is that deviceInfo will be non-null.
551 if (deviceInfo == nullptr)
561 stat = cudaGetDevice(&gpuid);
562 CU_RET_ERR(stat, "cudaGetDevice failed");
563 fprintf(stderr, "Cleaning up context on GPU ID #%d\n", gpuid);
566 if (!reset_gpu_application_clocks(deviceInfo))
568 gmx_warning("Failed to reset GPU application clocks on GPU #%d", deviceInfo->id);
571 stat = cudaDeviceReset();
572 if (stat != cudaSuccess)
574 gmx_warning("Failed to free GPU #%d: %s", deviceInfo->id, cudaGetErrorString(stat));
578 gmx_device_info_t *getDeviceInfo(const gmx_gpu_info_t &gpu_info,
581 if (deviceId < 0 || deviceId >= gpu_info.n_dev)
583 gmx_incons("Invalid GPU deviceId requested");
585 return &gpu_info.gpu_dev[deviceId];
588 /*! \brief Returns true if the gpu characterized by the device properties is
589 * supported by the native gpu acceleration.
591 * \param[in] dev_prop the CUDA device properties of the gpus to test.
592 * \returns true if the GPU properties passed indicate a compatible
593 * GPU, otherwise false.
595 static bool is_gmx_supported_gpu(const cudaDeviceProp *dev_prop)
597 return (dev_prop->major >= 2);
600 /*! \brief Checks if a GPU with a given ID is supported by the native GROMACS acceleration.
602 * Returns a status value which indicates compatibility or one of the following
603 * errors: incompatibility, insistence, or insanity (=unexpected behavior).
604 * It also returns the respective device's properties in \dev_prop (if applicable).
606 * \param[in] dev_id the ID of the GPU to check.
607 * \param[out] dev_prop the CUDA device properties of the device checked.
608 * \returns the status of the requested device
610 static int is_gmx_supported_gpu_id(int dev_id, cudaDeviceProp *dev_prop)
615 stat = cudaGetDeviceCount(&ndev);
616 if (stat != cudaSuccess)
621 if (dev_id > ndev - 1)
623 return egpuNonexistent;
626 /* TODO: currently we do not make a distinction between the type of errors
627 * that can appear during sanity checks. This needs to be improved, e.g if
628 * the dummy test kernel fails to execute with a "device busy message" we
629 * should appropriately report that the device is busy instead of insane.
631 if (do_sanity_checks(dev_id, dev_prop) == 0)
633 if (is_gmx_supported_gpu(dev_prop))
635 return egpuCompatible;
639 return egpuIncompatible;
648 bool canDetectGpus(std::string *errorMessage)
651 int driverVersion = -1;
652 stat = cudaDriverGetVersion(&driverVersion);
653 GMX_ASSERT(stat != cudaErrorInvalidValue, "An impossible null pointer was passed to cudaDriverGetVersion");
654 GMX_RELEASE_ASSERT(stat == cudaSuccess,
655 gmx::formatString("An unexpected value was returned from cudaDriverGetVersion %s: %s",
656 cudaGetErrorName(stat), cudaGetErrorString(stat)).c_str());
657 bool foundDriver = (driverVersion > 0);
660 // Can't detect GPUs if there is no driver
661 if (errorMessage != nullptr)
663 errorMessage->assign("No valid CUDA driver found");
669 stat = cudaGetDeviceCount(&numDevices);
670 if (stat != cudaSuccess)
672 if (errorMessage != nullptr)
674 /* cudaGetDeviceCount failed which means that there is
675 * something wrong with the machine: driver-runtime
676 * mismatch, all GPUs being busy in exclusive mode,
677 * invalid CUDA_VISIBLE_DEVICES, or some other condition
678 * which should result in GROMACS issuing a warning a
679 * falling back to CPUs. */
680 errorMessage->assign(cudaGetErrorString(stat));
683 // Consume the error now that we have prepared to handle
684 // it. This stops it reappearing next time we check for
685 // errors. Note that if CUDA_VISIBLE_DEVICES does not contain
686 // valid devices, then cudaGetLastError returns the
687 // (undocumented) cudaErrorNoDevice, but this should not be a
688 // problem as there should be no future CUDA API calls.
689 // NVIDIA bug report #2038718 has been filed.
695 // We don't actually use numDevices here, that's not the job of
700 void findGpus(gmx_gpu_info_t *gpu_info)
702 int i, ndev, checkres;
705 gmx_device_info_t *devs;
709 gpu_info->n_dev_compatible = 0;
714 stat = cudaGetDeviceCount(&ndev);
715 if (stat != cudaSuccess)
717 GMX_THROW(gmx::InternalError("Invalid call of findGpus() when CUDA API returned an error, perhaps "
718 "canDetectGpus() was not called appropriately beforehand."));
722 for (i = 0; i < ndev; i++)
724 checkres = is_gmx_supported_gpu_id(i, &prop);
728 devs[i].stat = checkres;
730 if (checkres == egpuCompatible)
732 gpu_info->n_dev_compatible++;
735 GMX_RELEASE_ASSERT(cudaSuccess == cudaPeekAtLastError(), "Should be cudaSuccess");
737 gpu_info->n_dev = ndev;
738 gpu_info->gpu_dev = devs;
741 std::vector<int> getCompatibleGpus(const gmx_gpu_info_t &gpu_info)
743 // Possible minor over-allocation here, but not important for anything
744 std::vector<int> compatibleGpus;
745 compatibleGpus.reserve(gpu_info.n_dev);
746 for (int i = 0; i < gpu_info.n_dev; i++)
748 assert(gpu_info.gpu_dev);
749 if (gpu_info.gpu_dev[i].stat == egpuCompatible)
751 compatibleGpus.push_back(i);
754 return compatibleGpus;
757 const char *getGpuCompatibilityDescription(const gmx_gpu_info_t &gpu_info,
760 return (index >= gpu_info.n_dev ?
761 gpu_detect_res_str[egpuNonexistent] :
762 gpu_detect_res_str[gpu_info.gpu_dev[index].stat]);
765 void free_gpu_info(const gmx_gpu_info_t *gpu_info)
767 if (gpu_info == NULL)
772 sfree(gpu_info->gpu_dev);
775 void get_gpu_device_info_string(char *s, const gmx_gpu_info_t &gpu_info, int index)
779 if (index < 0 && index >= gpu_info.n_dev)
784 gmx_device_info_t *dinfo = &gpu_info.gpu_dev[index];
787 dinfo->stat == egpuCompatible ||
788 dinfo->stat == egpuIncompatible;
792 sprintf(s, "#%d: %s, stat: %s",
794 gpu_detect_res_str[dinfo->stat]);
798 sprintf(s, "#%d: NVIDIA %s, compute cap.: %d.%d, ECC: %3s, stat: %s",
799 dinfo->id, dinfo->prop.name,
800 dinfo->prop.major, dinfo->prop.minor,
801 dinfo->prop.ECCEnabled ? "yes" : " no",
802 gpu_detect_res_str[dinfo->stat]);
806 int get_current_cuda_gpu_device_id(void)
809 CU_RET_ERR(cudaGetDevice(&gpuid), "cudaGetDevice failed");
814 size_t sizeof_gpu_dev_info(void)
816 return sizeof(gmx_device_info_t);
819 void gpu_set_host_malloc_and_free(bool bUseGpuKernels,
820 gmx_host_alloc_t **nb_alloc,
821 gmx_host_free_t **nb_free)
825 *nb_alloc = &pmalloc;
835 void startGpuProfiler(void)
837 /* The NVPROF_ID environment variable is set by nvprof and indicates that
838 mdrun is executed in the CUDA profiler.
839 If nvprof was run is with "--profile-from-start off", the profiler will
840 be started here. This way we can avoid tracing the CUDA events from the
841 first part of the run. Starting the profiler again does nothing.
846 stat = cudaProfilerStart();
847 CU_RET_ERR(stat, "cudaProfilerStart failed");
851 void stopGpuProfiler(void)
853 /* Stopping the nvidia here allows us to eliminate the subsequent
854 API calls from the trace, e.g. uninitialization and cleanup. */
858 stat = cudaProfilerStop();
859 CU_RET_ERR(stat, "cudaProfilerStop failed");
863 void resetGpuProfiler(void)
865 /* With CUDA <=7.5 the profiler can't be properly reset; we can only start
866 * the profiling here (can't stop it) which will achieve the desired effect if
867 * the run was started with the profiling disabled.
869 * TODO: add a stop (or replace it with reset) when this will work correctly in CUDA.