2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2010, The GROMACS development team,
6 * check out http://www.gromacs.org for more information.
7 * Copyright (c) 2012, by the GROMACS development team, led by
8 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
9 * others, as listed in the AUTHORS file in the top-level source
10 * directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
45 #include "types/hw_info.h"
47 #include "gpu_utils.h"
48 #include "../cuda_tools/cudautils.cuh"
49 #include "memtestG80_core.h"
52 #define QUICK_MEM 250 /*!< Amount of memory to be used in quick memtest. */
53 #define QUICK_TESTS MOD_20_32BIT | LOGIC_4_ITER_SHMEM | RANDOM_BLOCKS /*!< Bit flag with type of tests
54 to run in quick memtest. */
55 #define QUICK_ITER 3 /*!< Number of iterations in quick memtest. */
57 #define FULL_TESTS 0x3FFF /*!< Bitflag with all test set on for full memetest. */
58 #define FULL_ITER 25 /*!< Number of iterations in full memtest. */
60 #define TIMED_TESTS MOD_20_32BIT | LOGIC_4_ITER_SHMEM | RANDOM_BLOCKS /*!< Bit flag with type of tests to
61 run in time constrained memtest. */
63 /*! Number of supported GPUs */
64 #define NB_GPUS (sizeof(SupportedGPUs)/sizeof(SupportedGPUs[0]))
66 static int cuda_max_device_count = 32; /*! Max number of devices supported by CUDA (for consistency checking).
67 In reality it 16 with CUDA <=v5.0, but let's stay on the safe side. */
69 /*! Dummy kernel used for sanity checking. */
70 __global__ void k_dummy_test(){}
73 /*! Bit-flags which refer to memtestG80 test types and are used in do_memtest to specify which tests to run. */
74 enum memtest_G80_test_types {
75 MOVING_INVERSIONS_10 = 0x1,
76 MOVING_INVERSIONS_RAND = 0x2,
77 WALKING_8BIT_M86 = 0x4,
79 WALKING_1_8BIT = 0x10,
80 WALKING_0_32BIT = 0x20,
81 WALKING_1_32BIT = 0x40,
86 LOGIC_1_ITER_SHMEM = 0x800,
87 LOGIC_4_ITER_SHMEM = 0x1000
90 // TODO put this list into an external file and include it so that the list is easily accessible
91 /*! List of supported GPUs. */
92 static const char * const SupportedGPUs[] = {
128 "Quadro Plex 2200 D2",
129 "Quadro Plex 2200 S4",
132 "GeForce 9800 G", /* GX2, GTX, GTX+, GT */
136 "Quadro Plex 2100 D4"
141 * \brief Runs GPU sanity checks.
143 * Runs a series of checks to determine that the given GPU and underlying CUDA
144 * driver/runtime functions properly.
145 * Returns properties of a device with given ID or the one that has
146 * already been initialized earlier in the case if of \dev_id == -1.
148 * \param[in] dev_id the device ID of the GPU or -1 if the device has already been initialized
149 * \param[out] dev_prop pointer to the structure in which the device properties will be returned
150 * \returns 0 if the device looks OK
152 * TODO: introduce errors codes and handle errors more smoothly.
154 static int do_sanity_checks(int dev_id, cudaDeviceProp *dev_prop)
159 cu_err = cudaGetDeviceCount(&dev_count);
160 if (cu_err != cudaSuccess)
162 fprintf(stderr, "Error %d while querying device count: %s\n", cu_err,
163 cudaGetErrorString(cu_err));
167 /* no CUDA compatible device at all */
171 /* things might go horribly wrong if cudart is not compatible with the driver */
172 if (dev_count < 0 || dev_count > cuda_max_device_count)
175 if (dev_id == -1) /* device already selected let's not destroy the context */
177 cu_err = cudaGetDevice(&id);
178 if (cu_err != cudaSuccess)
180 fprintf(stderr, "Error %d while querying device id: %s\n", cu_err,
181 cudaGetErrorString(cu_err));
188 if (id > dev_count - 1) /* pfff there's no such device */
190 fprintf(stderr, "The requested device with id %d does not seem to exist (device count=%d)\n",
196 memset(dev_prop, 0, sizeof(cudaDeviceProp));
197 cu_err = cudaGetDeviceProperties(dev_prop, id);
198 if (cu_err != cudaSuccess)
200 fprintf(stderr, "Error %d while querying device properties: %s\n", cu_err,
201 cudaGetErrorString(cu_err));
205 /* both major & minor is 9999 if no CUDA capable devices are present */
206 if (dev_prop->major == 9999 && dev_prop->minor == 9999)
208 /* we don't care about emulation mode */
209 if (dev_prop->major == 0)
214 cu_err = cudaSetDevice(id);
215 if (cu_err != cudaSuccess)
217 fprintf(stderr, "Error %d while switching to device #%d: %s\n",
218 cu_err, id, cudaGetErrorString(cu_err));
223 /* try to execute a dummy kernel */
224 k_dummy_test<<<1, 512>>>();
225 if (cudaThreadSynchronize() != cudaSuccess)
230 /* destroy context if we created one */
233 #if CUDA_VERSION < 4000
234 cu_err = cudaThreadExit();
235 CU_RET_ERR(cu_err, "cudaThreadExit failed");
237 cu_err = cudaDeviceReset();
238 CU_RET_ERR(cu_err, "cudaDeviceReset failed");
247 * \brief Checks whether the GPU with the given name is supported in Gromacs-OpenMM.
249 * \param[in] gpu_name the name of the CUDA device
250 * \returns TRUE if the device is supported, otherwise FALSE
252 static bool is_gmx_openmm_supported_gpu_name(char *gpuName)
255 for (i = 0; i < NB_GPUS; i++)
258 if (gmx_strncasecmp(gpuName, SupportedGPUs[i], strlen(SupportedGPUs[i])) == 0)
264 /*! \brief Checks whether the GPU with the given device id is supported in Gromacs-OpenMM.
266 * \param[in] dev_id the device id of the GPU or -1 if the device has already been selected
267 * \param[out] gpu_name Set to contain the name of the CUDA device, if NULL passed, no device name is set.
268 * \returns TRUE if the device is supported, otherwise FALSE
271 gmx_bool is_gmx_openmm_supported_gpu(int dev_id, char *gpu_name)
273 cudaDeviceProp dev_prop;
275 if (debug) fprintf(debug, "Checking compatibility with device #%d, %s\n", dev_id, gpu_name);
277 if (do_sanity_checks(dev_id, &dev_prop) != 0)
280 if (gpu_name != NULL)
282 strcpy(gpu_name, dev_prop.name);
284 return is_gmx_openmm_supported_gpu_name(dev_prop.name);
289 * \brief Runs a set of memory tests specified by the given bit-flags.
290 * Tries to allocate and do the test on \p megs Mb memory or
291 * the greatest amount that can be allocated (>10Mb).
292 * In case if an error is detected it stops without finishing the remaining
293 * steps/iterations and returns greater then zero value.
294 * In case of other errors (e.g. kernel launch errors, device querying errors)
297 * \param[in] which_tests variable with bit-flags of the requested tests
298 * \param[in] megs amount of memory that will be tested in MB
299 * \param[in] iter number of iterations
300 * \returns 0 if no error was detected, otherwise >0
302 static int do_memtest(unsigned int which_tests, int megs, int iter)
306 uint err_count; //, err_iter;
308 // no parameter check as this fn won't be called externally
310 // let's try to allocate the mem
311 while (!tester.allocate(megs) && (megs - 10 > 0))
312 { megs -= 10; tester.deallocate(); }
316 fprintf(stderr, "Unable to allocate GPU memory!\n");
320 // clear the first 18 bits
321 which_tests &= 0x3FFF;
322 for (i = 0; i < iter; i++)
324 // Moving Inversions (ones and zeros)
325 if ((MOVING_INVERSIONS_10 & which_tests) == MOVING_INVERSIONS_10)
327 tester.gpuMovingInversionsOnesZeros(err_count);
329 return MOVING_INVERSIONS_10;
331 // Moving Inversions (random)
332 if ((MOVING_INVERSIONS_RAND & which_tests) == MOVING_INVERSIONS_RAND)
334 tester.gpuMovingInversionsRandom(err_count);
336 return MOVING_INVERSIONS_RAND;
338 // Memtest86 Walking 8-bit
339 if ((WALKING_8BIT_M86 & which_tests) == WALKING_8BIT_M86)
341 for (uint shift = 0; shift < 8; shift++)
343 tester.gpuWalking8BitM86(err_count, shift);
345 return WALKING_8BIT_M86;
348 // True Walking zeros (8-bit)
349 if ((WALKING_0_8BIT & which_tests) == WALKING_0_8BIT)
351 for (uint shift = 0; shift < 8; shift++)
353 tester.gpuWalking8Bit(err_count, false, shift);
355 return WALKING_0_8BIT;
358 // True Walking ones (8-bit)
359 if ((WALKING_1_8BIT & which_tests) == WALKING_1_8BIT)
361 for (uint shift = 0; shift < 8; shift++)
363 tester.gpuWalking8Bit(err_count, true, shift);
365 return WALKING_1_8BIT;
368 // Memtest86 Walking zeros (32-bit)
369 if ((WALKING_0_32BIT & which_tests) == WALKING_0_32BIT)
371 for (uint shift = 0; shift < 32; shift++)
373 tester.gpuWalking32Bit(err_count, false, shift);
375 return WALKING_0_32BIT;
378 // Memtest86 Walking ones (32-bit)
379 if ((WALKING_1_32BIT & which_tests) == WALKING_1_32BIT)
381 for (uint shift = 0; shift < 32; shift++)
383 tester.gpuWalking32Bit(err_count, true, shift);
385 return WALKING_1_32BIT;
389 if ((RANDOM_BLOCKS & which_tests) == RANDOM_BLOCKS)
391 tester.gpuRandomBlocks(err_count,rand());
393 return RANDOM_BLOCKS;
397 // Memtest86 Modulo-20
398 if ((MOD_20_32BIT & which_tests) == MOD_20_32BIT)
400 for (uint shift = 0; shift < 20; shift++)
402 tester.gpuModuloX(err_count, shift, rand(), 20, 2);
407 // Logic (one iteration)
408 if ((LOGIC_1_ITER & which_tests) == LOGIC_1_ITER)
410 tester.gpuShortLCG0(err_count,1);
414 // Logic (4 iterations)
415 if ((LOGIC_4_ITER & which_tests) == LOGIC_4_ITER)
417 tester.gpuShortLCG0(err_count,4);
422 // Logic (shared memory, one iteration)
423 if ((LOGIC_1_ITER_SHMEM & which_tests) == LOGIC_1_ITER_SHMEM)
425 tester.gpuShortLCG0Shmem(err_count,1);
427 return LOGIC_1_ITER_SHMEM;
429 // Logic (shared-memory, 4 iterations)
430 if ((LOGIC_4_ITER_SHMEM & which_tests) == LOGIC_4_ITER_SHMEM)
432 tester.gpuShortLCG0Shmem(err_count,4);
434 return LOGIC_4_ITER_SHMEM;
442 /*! \brief Runs a quick memory test and returns 0 in case if no error is detected.
443 * If an error is detected it stops before completing the test and returns a
444 * value greater then 0. In case of other errors (e.g. kernel launch errors,
445 * device querying errors) -1 is returned.
447 * \param[in] dev_id the device id of the GPU or -1 if the device has already been selected
448 * \returns 0 if no error was detected, otherwise >0
450 int do_quick_memtest(int dev_id)
452 cudaDeviceProp dev_prop;
453 int devmem, res, time=0;
455 if (debug) { time = getTimeMilliseconds(); }
457 if (do_sanity_checks(dev_id, &dev_prop) != 0)
459 // something went wrong
465 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
466 fprintf(debug, ">> Running QUICK memtests on %d MiB (out of total %d MiB), %d iterations\n",
467 QUICK_MEM, devmem, QUICK_ITER);
470 res = do_memtest(QUICK_TESTS, QUICK_MEM, QUICK_ITER);
474 fprintf(debug, "Q-RES = %d\n", res);
475 fprintf(debug, "Q-runtime: %d ms\n", getTimeMilliseconds() - time);
478 /* destroy context only if we created it */
479 if (dev_id !=-1) cudaThreadExit();
483 /*! \brief Runs a full memory test and returns 0 in case if no error is detected.
484 * If an error is detected it stops before completing the test and returns a
485 * value greater then 0. In case of other errors (e.g. kernel launch errors,
486 * device querying errors) -1 is returned.
488 * \param[in] dev_id the device id of the GPU or -1 if the device has already been selected
489 * \returns 0 if no error was detected, otherwise >0
492 int do_full_memtest(int dev_id)
494 cudaDeviceProp dev_prop;
495 int devmem, res, time=0;
497 if (debug) { time = getTimeMilliseconds(); }
499 if (do_sanity_checks(dev_id, &dev_prop) != 0)
501 // something went wrong
505 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
509 fprintf(debug, ">> Running FULL memtests on %d MiB (out of total %d MiB), %d iterations\n",
510 devmem, devmem, FULL_ITER);
513 /* do all test on the entire memory */
514 res = do_memtest(FULL_TESTS, devmem, FULL_ITER);
518 fprintf(debug, "F-RES = %d\n", res);
519 fprintf(debug, "F-runtime: %d ms\n", getTimeMilliseconds() - time);
522 /* destroy context only if we created it */
523 if (dev_id != -1) cudaThreadExit();
527 /*! \brief Runs a time constrained memory test and returns 0 in case if no error is detected.
528 * If an error is detected it stops before completing the test and returns a value greater
529 * than zero. In case of other errors (e.g. kernel launch errors, device querying errors) -1
530 * is returned. Note, that test iterations are not interrupted therefor the total runtime of
531 * the test will always be multipple of one iteration's runtime.
533 * \param[in] dev_id the device id of the GPU or -1 if the device has laredy been selected
534 * \param[in] time_constr the time limit of the testing
535 * \returns 0 if no error was detected, otherwise >0
537 int do_timed_memtest(int dev_id, int time_constr)
539 cudaDeviceProp dev_prop;
540 int devmem, res=0, time=0, startt;
542 if (debug) { time = getTimeMilliseconds(); }
544 time_constr *= 1000; /* convert to ms for convenience */
545 startt = getTimeMilliseconds();
547 if (do_sanity_checks(dev_id, &dev_prop) != 0)
549 // something went wrong
553 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
557 fprintf(debug, ">> Running time constrained memtests on %d MiB (out of total %d MiB), time limit of %d s \n",
558 devmem, devmem, time_constr);
561 /* do the TIMED_TESTS set, one step at a time on the entire memory
562 that can be allocated, and stop when the given time is exceeded */
563 while ( ((int)getTimeMilliseconds() - startt) < time_constr)
565 res = do_memtest(TIMED_TESTS, devmem, 1);
571 fprintf(debug, "T-RES = %d\n", res);
572 fprintf(debug, "T-runtime: %d ms\n", getTimeMilliseconds() - time);
575 /* destroy context only if we created it */
576 if (dev_id != -1) cudaThreadExit();
580 /*! \brief Initializes the GPU with the given index.
582 * The varible \mygpu is the index of the GPU to initialize in the
583 * gpu_info.cuda_dev array.
585 * \param[in] mygpu index of the GPU to initialize
586 * \param[out] result_str the message related to the error that occurred
587 * during the initialization (if there was any).
588 * \param[in] gpu_info GPU info of all detected devices in the system.
589 * \returns true if no error occurs during initialization.
591 gmx_bool init_gpu(int mygpu, char *result_str, const gmx_gpu_info_t *gpu_info)
600 if (mygpu < 0 || mygpu >= gpu_info->ncuda_dev_use)
602 sprintf(sbuf, "Trying to initialize an inexistent GPU: "
603 "there are %d %s-selected GPU(s), but #%d was requested.",
604 gpu_info->ncuda_dev_use, gpu_info->bUserSet ? "user" : "auto", mygpu);
608 gpuid = gpu_info->cuda_dev[gpu_info->cuda_dev_use[mygpu]].id;
610 stat = cudaSetDevice(gpuid);
611 strncpy(result_str, cudaGetErrorString(stat), STRLEN);
615 fprintf(stderr, "Initialized GPU ID #%d: %s\n", gpuid, gpu_info->cuda_dev[gpuid].prop.name);
618 return (stat == cudaSuccess);
621 /*! \brief Frees up the CUDA GPU used by the active context at the time of calling.
623 * The context is explicitly destroyed and therefore all data uploaded to the GPU
624 * is lost. This should only be called when none of this data is required anymore.
626 * \param[out] result_str the message related to the error that occurred
627 * during the initialization (if there was any).
628 * \returns true if no error occurs during the freeing.
630 gmx_bool free_gpu(char *result_str)
639 stat = cudaGetDevice(&gpuid);
640 CU_RET_ERR(stat, "cudaGetDevice failed");
641 fprintf(stderr, "Cleaning up context on GPU ID #%d\n", gpuid);
644 #if CUDA_VERSION < 4000
645 stat = cudaThreadExit();
647 stat = cudaDeviceReset();
649 strncpy(result_str, cudaGetErrorString(stat), STRLEN);
651 return (stat == cudaSuccess);
654 /*! \brief Returns true if the gpu characterized by the device properties is
655 * supported by the native gpu acceleration.
657 * \param[in] dev_prop the CUDA device properties of the gpus to test.
658 * \returns true if the GPU properties passed indicate a compatible
659 * GPU, otherwise false.
661 static bool is_gmx_supported_gpu(const cudaDeviceProp *dev_prop)
663 return (dev_prop->major >= 2);
666 /*! \brief Helper function that checks whether a given GPU status indicates compatible GPU.
668 * \param[in] stat GPU status.
669 * \returns true if the provided status is egpuCompatible, otherwise false.
671 static bool is_compatible_gpu(int stat)
673 return (stat == egpuCompatible);
676 /*! \brief Checks if a GPU with a given ID is supported by the native GROMACS acceleration.
678 * Returns a status value which indicates compatibility or one of the following
679 * errors: incompatibility, insistence, or insanity (=unexpected behavior).
680 * It also returns the respective device's properties in \dev_prop (if applicable).
682 * \param[in] dev_id the ID of the GPU to check.
683 * \param[out] dev_prop the CUDA device properties of the device checked.
684 * \returns the status of the requested device
686 static int is_gmx_supported_gpu_id(int dev_id, cudaDeviceProp *dev_prop)
691 stat = cudaGetDeviceCount(&ndev);
692 if (stat != cudaSuccess)
697 if (dev_id > ndev - 1)
699 return egpuNonexistent;
702 /* TODO: currently we do not make a distinction between the type of errors
703 * that can appear during sanity checks. This needs to be improved, e.g if
704 * the dummy test kernel fails to execute with a "device busy message" we
705 * should appropriately report that the device is busy instead of insane.
707 if (do_sanity_checks(dev_id, dev_prop) == 0)
709 if (is_gmx_supported_gpu(dev_prop))
711 return egpuCompatible;
715 return egpuIncompatible;
725 /*! \brief Detect all NVIDIA GPUs in the system.
727 * Will detect every NVIDIA GPU supported by the device driver in use. Also
728 * check for the compatibility of each and fill the gpu_info->cuda_dev array
729 * with the required information on each the device: ID, device properties,
732 * \param[in] gpu_info pointer to structure holding GPU information.
733 * \param[out] err_str The error message of any CUDA API error that caused
734 * the detection to fail (if there was any). The memory
735 * the pointer points to should be managed externally.
736 * \returns non-zero if the detection encountered a failure, zero otherwise.
738 int detect_cuda_gpus(gmx_gpu_info_t *gpu_info, char *err_str)
740 int i, ndev, checkres, retval;
743 cuda_dev_info_t *devs;
751 stat = cudaGetDeviceCount(&ndev);
752 if (stat != cudaSuccess)
756 /* cudaGetDeviceCount failed which means that there is something
757 * wrong with the machine: driver-runtime mismatch, all GPUs being
758 * busy in exclusive mode, or some other condition which should
759 * result in us issuing a warning a falling back to CPUs. */
761 s = cudaGetErrorString(stat);
762 strncpy(err_str, s, STRLEN*sizeof(err_str[0]));
767 for (i = 0; i < ndev; i++)
769 checkres = is_gmx_supported_gpu_id(i, &prop);
773 devs[i].stat = checkres;
778 gpu_info->ncuda_dev = ndev;
779 gpu_info->cuda_dev = devs;
784 /*! \brief Select the GPUs compatible with the native GROMACS acceleration.
786 * This function selects the compatible gpus and initializes
787 * gpu_info->cuda_dev_use and gpu_info->ncuda_dev_use.
789 * Given the list of GPUs available in the system the it checks each gpu in
790 * gpu_info->cuda_dev and puts the the indices (into gpu_info->cuda_dev) of
791 * the compatible ones into cuda_dev_use with this marking the respective
792 * GPUs as "available for use."
793 * Note that \detect_cuda_gpus must have been called before.
795 * \param[in] gpu_info pointer to structure holding GPU information
797 void pick_compatible_gpus(gmx_gpu_info_t *gpu_info)
803 /* cuda_dev/ncuda_dev have to be either NULL/0 or not (NULL/0) */
804 assert((gpu_info->ncuda_dev != 0 ? 0 : 1) ^ (gpu_info->cuda_dev == NULL ? 0 : 1));
806 snew(compat, gpu_info->ncuda_dev);
808 for (i = 0; i < gpu_info->ncuda_dev; i++)
810 if (is_compatible_gpu(gpu_info->cuda_dev[i].stat))
813 compat[ncompat - 1] = i;
817 gpu_info->ncuda_dev_use = ncompat;
818 snew(gpu_info->cuda_dev_use, ncompat);
819 memcpy(gpu_info->cuda_dev_use, compat, ncompat*sizeof(*compat));
823 /*! \brief Check the existence/compatibility of a set of GPUs specified by their device IDs.
825 * Given the a list of GPU devide IDs in \requested_devs, check for the
826 * existence and compatibility of the respective GPUs and fill in \gpu_info
827 * with the collected information. Also provide the caller with an array with
828 * the result of checks in \checkres.
830 * \param[out] checkres check result for each ID passed in \requested_devs
831 * \param[in] gpu_info pointer to structure holding GPU information
832 * \param[in] requested_devs array of requested device IDs
833 * \param[in] count number of IDs in \requested_devs
834 * \returns TRUE if every requested GPU is compatible
836 gmx_bool check_select_cuda_gpus(int *checkres, gmx_gpu_info_t *gpu_info,
837 const int *requested_devs, int count)
844 assert(requested_devs);
852 /* we will assume that all GPUs requested are valid IDs,
853 otherwise we'll bail anyways */
854 gpu_info->ncuda_dev_use = count;
855 snew(gpu_info->cuda_dev_use, count);
858 for (i = 0; i < count; i++)
860 id = requested_devs[i];
862 /* devices are stored in increasing order of IDs in cuda_dev */
863 gpu_info->cuda_dev_use[i] = id;
865 checkres[i] = (id >= gpu_info->ncuda_dev) ?
866 egpuNonexistent : gpu_info->cuda_dev[id].stat;
868 bAllOk = bAllOk && is_compatible_gpu(checkres[i]);
874 /*! \brief Frees the cuda_dev and cuda_dev_use array fields of \gpu_info.
876 * \param[in] gpu_info pointer to structure holding GPU information
878 void free_gpu_info(const gmx_gpu_info_t *gpu_info)
880 if (gpu_info == NULL)
885 sfree(gpu_info->cuda_dev_use);
886 sfree(gpu_info->cuda_dev);
889 /*! \brief Formats and returns a device information string for a given GPU.
891 * Given an index *directly* into the array of available GPUs (cuda_dev)
892 * returns a formatted info string for the respective GPU which includes
893 * ID, name, compute capability, and detection status.
895 * \param[out] s pointer to output string (has to be allocated externally)
896 * \param[in] gpu_info pointer to structure holding GPU information
897 * \param[in] index an index *directly* into the array of available GPUs
899 void get_gpu_device_info_string(char *s, const gmx_gpu_info_t *gpu_info, int index)
904 if (index < 0 && index >= gpu_info->ncuda_dev)
909 cuda_dev_info_t *dinfo = &gpu_info->cuda_dev[index];
912 dinfo->stat == egpuCompatible ||
913 dinfo->stat == egpuIncompatible;
917 sprintf(s, "#%d: %s, stat: %s",
919 gpu_detect_res_str[dinfo->stat]);
923 sprintf(s, "#%d: NVIDIA %s, compute cap.: %d.%d, ECC: %3s, stat: %s",
924 dinfo->id, dinfo->prop.name,
925 dinfo->prop.major, dinfo->prop.minor,
926 dinfo->prop.ECCEnabled ? "yes" : " no",
927 gpu_detect_res_str[dinfo->stat]);
931 /*! \brief Returns the device ID of the GPU with a given index into the array of used GPUs.
933 * Getter function which, given an index into the array of GPUs in use
934 * (cuda_dev_use) -- typically a tMPI/MPI rank --, returns the device ID of the
935 * respective CUDA GPU.
937 * \param[in] gpu_info pointer to structure holding GPU information
938 * \param[in] idx index into the array of used GPUs
939 * \returns device ID of the requested GPU
941 int get_gpu_device_id(const gmx_gpu_info_t *gpu_info, int idx)
944 if (idx < 0 && idx >= gpu_info->ncuda_dev_use)
949 return gpu_info->cuda_dev[gpu_info->cuda_dev_use[idx]].id;
952 /*! \brief Returns the device ID of the GPU currently in use.
954 * The GPU used is the one that is active at the time of the call in the active context.
956 * \param[in] gpu_info pointer to structure holding GPU information
957 * \returns device ID of the GPU in use at the time of the call
959 int get_current_gpu_device_id(void)
962 CU_RET_ERR(cudaGetDevice(&gpuid), "cudaGetDevice failed");