2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2010,2011,2012,2013, 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.
42 #include "types/hw_info.h"
44 #include "gpu_utils.h"
45 #include "../cuda_tools/cudautils.cuh"
46 #include "memtestG80_core.h"
48 /** Amount of memory to be used in quick memtest. */
50 /** Bit flag with type of tests to run in quick memtest. */
51 #define QUICK_TESTS MOD_20_32BIT | LOGIC_4_ITER_SHMEM | RANDOM_BLOCKS
52 /** Number of iterations in quick memtest. */
55 /** Bitflag with all test set on for full memetest. */
56 #define FULL_TESTS 0x3FFF
57 /** Number of iterations in full memtest. */
60 /** Bit flag with type of tests to run in time constrained memtest. */
61 #define TIMED_TESTS MOD_20_32BIT | LOGIC_4_ITER_SHMEM | RANDOM_BLOCKS
64 * Max number of devices supported by CUDA (for consistency checking).
66 * In reality it is 16 with CUDA <=v5.0, but let's stay on the safe side.
68 static int cuda_max_device_count = 32;
70 /** Dummy kernel used for sanity checking. */
71 __global__ void k_dummy_test()
76 /** Bit-flags which refer to memtestG80 test types and are used in do_memtest
77 * to specify which tests to run. */
78 enum memtest_G80_test_types {
79 MOVING_INVERSIONS_10 = 0x1,
80 MOVING_INVERSIONS_RAND = 0x2,
81 WALKING_8BIT_M86 = 0x4,
83 WALKING_1_8BIT = 0x10,
84 WALKING_0_32BIT = 0x20,
85 WALKING_1_32BIT = 0x40,
90 LOGIC_1_ITER_SHMEM = 0x800,
91 LOGIC_4_ITER_SHMEM = 0x1000
96 * \brief Runs GPU sanity checks.
98 * Runs a series of checks to determine that the given GPU and underlying CUDA
99 * driver/runtime functions properly.
100 * Returns properties of a device with given ID or the one that has
101 * already been initialized earlier in the case if of \dev_id == -1.
103 * \param[in] dev_id the device ID of the GPU or -1 if the device has already been initialized
104 * \param[out] dev_prop pointer to the structure in which the device properties will be returned
105 * \returns 0 if the device looks OK
107 * TODO: introduce errors codes and handle errors more smoothly.
109 static int do_sanity_checks(int dev_id, cudaDeviceProp *dev_prop)
114 cu_err = cudaGetDeviceCount(&dev_count);
115 if (cu_err != cudaSuccess)
117 fprintf(stderr, "Error %d while querying device count: %s\n", cu_err,
118 cudaGetErrorString(cu_err));
122 /* no CUDA compatible device at all */
128 /* things might go horribly wrong if cudart is not compatible with the driver */
129 if (dev_count < 0 || dev_count > cuda_max_device_count)
134 if (dev_id == -1) /* device already selected let's not destroy the context */
136 cu_err = cudaGetDevice(&id);
137 if (cu_err != cudaSuccess)
139 fprintf(stderr, "Error %d while querying device id: %s\n", cu_err,
140 cudaGetErrorString(cu_err));
147 if (id > dev_count - 1) /* pfff there's no such device */
149 fprintf(stderr, "The requested device with id %d does not seem to exist (device count=%d)\n",
155 memset(dev_prop, 0, sizeof(cudaDeviceProp));
156 cu_err = cudaGetDeviceProperties(dev_prop, id);
157 if (cu_err != cudaSuccess)
159 fprintf(stderr, "Error %d while querying device properties: %s\n", cu_err,
160 cudaGetErrorString(cu_err));
164 /* both major & minor is 9999 if no CUDA capable devices are present */
165 if (dev_prop->major == 9999 && dev_prop->minor == 9999)
169 /* we don't care about emulation mode */
170 if (dev_prop->major == 0)
177 cu_err = cudaSetDevice(id);
178 if (cu_err != cudaSuccess)
180 fprintf(stderr, "Error %d while switching to device #%d: %s\n",
181 cu_err, id, cudaGetErrorString(cu_err));
186 /* try to execute a dummy kernel */
187 k_dummy_test<<< 1, 512>>> ();
188 if (cudaThreadSynchronize() != cudaSuccess)
193 /* destroy context if we created one */
196 #if CUDA_VERSION < 4000
197 cu_err = cudaThreadExit();
198 CU_RET_ERR(cu_err, "cudaThreadExit failed");
200 cu_err = cudaDeviceReset();
201 CU_RET_ERR(cu_err, "cudaDeviceReset failed");
210 * \brief Runs a set of memory tests specified by the given bit-flags.
211 * Tries to allocate and do the test on \p megs Mb memory or
212 * the greatest amount that can be allocated (>10Mb).
213 * In case if an error is detected it stops without finishing the remaining
214 * steps/iterations and returns greater then zero value.
215 * In case of other errors (e.g. kernel launch errors, device querying errors)
218 * \param[in] which_tests variable with bit-flags of the requested tests
219 * \param[in] megs amount of memory that will be tested in MB
220 * \param[in] iter number of iterations
221 * \returns 0 if no error was detected, otherwise >0
223 static int do_memtest(unsigned int which_tests, int megs, int iter)
227 uint err_count; //, err_iter;
229 // no parameter check as this fn won't be called externally
231 // let's try to allocate the mem
232 while (!tester.allocate(megs) && (megs - 10 > 0))
234 megs -= 10; tester.deallocate();
239 fprintf(stderr, "Unable to allocate GPU memory!\n");
243 // clear the first 18 bits
244 which_tests &= 0x3FFF;
245 for (i = 0; i < iter; i++)
247 // Moving Inversions (ones and zeros)
248 if ((MOVING_INVERSIONS_10 & which_tests) == MOVING_INVERSIONS_10)
250 tester.gpuMovingInversionsOnesZeros(err_count);
253 return MOVING_INVERSIONS_10;
256 // Moving Inversions (random)
257 if ((MOVING_INVERSIONS_RAND & which_tests) == MOVING_INVERSIONS_RAND)
259 tester.gpuMovingInversionsRandom(err_count);
262 return MOVING_INVERSIONS_RAND;
265 // Memtest86 Walking 8-bit
266 if ((WALKING_8BIT_M86 & which_tests) == WALKING_8BIT_M86)
268 for (uint shift = 0; shift < 8; shift++)
270 tester.gpuWalking8BitM86(err_count, shift);
273 return WALKING_8BIT_M86;
277 // True Walking zeros (8-bit)
278 if ((WALKING_0_8BIT & which_tests) == WALKING_0_8BIT)
280 for (uint shift = 0; shift < 8; shift++)
282 tester.gpuWalking8Bit(err_count, false, shift);
285 return WALKING_0_8BIT;
289 // True Walking ones (8-bit)
290 if ((WALKING_1_8BIT & which_tests) == WALKING_1_8BIT)
292 for (uint shift = 0; shift < 8; shift++)
294 tester.gpuWalking8Bit(err_count, true, shift);
297 return WALKING_1_8BIT;
301 // Memtest86 Walking zeros (32-bit)
302 if ((WALKING_0_32BIT & which_tests) == WALKING_0_32BIT)
304 for (uint shift = 0; shift < 32; shift++)
306 tester.gpuWalking32Bit(err_count, false, shift);
309 return WALKING_0_32BIT;
313 // Memtest86 Walking ones (32-bit)
314 if ((WALKING_1_32BIT & which_tests) == WALKING_1_32BIT)
316 for (uint shift = 0; shift < 32; shift++)
318 tester.gpuWalking32Bit(err_count, true, shift);
321 return WALKING_1_32BIT;
326 if ((RANDOM_BLOCKS & which_tests) == RANDOM_BLOCKS)
328 tester.gpuRandomBlocks(err_count, rand());
331 return RANDOM_BLOCKS;
336 // Memtest86 Modulo-20
337 if ((MOD_20_32BIT & which_tests) == MOD_20_32BIT)
339 for (uint shift = 0; shift < 20; shift++)
341 tester.gpuModuloX(err_count, shift, rand(), 20, 2);
348 // Logic (one iteration)
349 if ((LOGIC_1_ITER & which_tests) == LOGIC_1_ITER)
351 tester.gpuShortLCG0(err_count, 1);
357 // Logic (4 iterations)
358 if ((LOGIC_4_ITER & which_tests) == LOGIC_4_ITER)
360 tester.gpuShortLCG0(err_count, 4);
367 // Logic (shared memory, one iteration)
368 if ((LOGIC_1_ITER_SHMEM & which_tests) == LOGIC_1_ITER_SHMEM)
370 tester.gpuShortLCG0Shmem(err_count, 1);
373 return LOGIC_1_ITER_SHMEM;
376 // Logic (shared-memory, 4 iterations)
377 if ((LOGIC_4_ITER_SHMEM & which_tests) == LOGIC_4_ITER_SHMEM)
379 tester.gpuShortLCG0Shmem(err_count, 4);
382 return LOGIC_4_ITER_SHMEM;
391 /*! \brief Runs a quick memory test and returns 0 in case if no error is detected.
392 * If an error is detected it stops before completing the test and returns a
393 * value greater then 0. In case of other errors (e.g. kernel launch errors,
394 * device querying errors) -1 is returned.
396 * \param[in] dev_id the device id of the GPU or -1 if the device has already been selected
397 * \returns 0 if no error was detected, otherwise >0
399 int do_quick_memtest(int dev_id)
401 cudaDeviceProp dev_prop;
402 int devmem, res, time = 0;
406 time = getTimeMilliseconds();
409 if (do_sanity_checks(dev_id, &dev_prop) != 0)
411 // something went wrong
417 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
418 fprintf(debug, ">> Running QUICK memtests on %d MiB (out of total %d MiB), %d iterations\n",
419 QUICK_MEM, devmem, QUICK_ITER);
422 res = do_memtest(QUICK_TESTS, QUICK_MEM, QUICK_ITER);
426 fprintf(debug, "Q-RES = %d\n", res);
427 fprintf(debug, "Q-runtime: %d ms\n", getTimeMilliseconds() - time);
430 /* destroy context only if we created it */
438 /*! \brief Runs a full memory test and returns 0 in case if no error is detected.
439 * If an error is detected it stops before completing the test and returns a
440 * value greater then 0. In case of other errors (e.g. kernel launch errors,
441 * device querying errors) -1 is returned.
443 * \param[in] dev_id the device id of the GPU or -1 if the device has already been selected
444 * \returns 0 if no error was detected, otherwise >0
447 int do_full_memtest(int dev_id)
449 cudaDeviceProp dev_prop;
450 int devmem, res, time = 0;
454 time = getTimeMilliseconds();
457 if (do_sanity_checks(dev_id, &dev_prop) != 0)
459 // something went wrong
463 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
467 fprintf(debug, ">> Running FULL memtests on %d MiB (out of total %d MiB), %d iterations\n",
468 devmem, devmem, FULL_ITER);
471 /* do all test on the entire memory */
472 res = do_memtest(FULL_TESTS, devmem, FULL_ITER);
476 fprintf(debug, "F-RES = %d\n", res);
477 fprintf(debug, "F-runtime: %d ms\n", getTimeMilliseconds() - time);
480 /* destroy context only if we created it */
488 /*! \brief Runs a time constrained memory test and returns 0 in case if no error is detected.
489 * If an error is detected it stops before completing the test and returns a value greater
490 * than zero. In case of other errors (e.g. kernel launch errors, device querying errors) -1
491 * is returned. Note, that test iterations are not interrupted therefor the total runtime of
492 * the test will always be multipple of one iteration's runtime.
494 * \param[in] dev_id the device id of the GPU or -1 if the device has laredy been selected
495 * \param[in] time_constr the time limit of the testing
496 * \returns 0 if no error was detected, otherwise >0
498 int do_timed_memtest(int dev_id, int time_constr)
500 cudaDeviceProp dev_prop;
501 int devmem, res = 0, time = 0, startt;
505 time = getTimeMilliseconds();
508 time_constr *= 1000; /* convert to ms for convenience */
509 startt = getTimeMilliseconds();
511 if (do_sanity_checks(dev_id, &dev_prop) != 0)
513 // something went wrong
517 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
521 fprintf(debug, ">> Running time constrained memtests on %d MiB (out of total %d MiB), time limit of %d s \n",
522 devmem, devmem, time_constr);
525 /* do the TIMED_TESTS set, one step at a time on the entire memory
526 that can be allocated, and stop when the given time is exceeded */
527 while ( ((int)getTimeMilliseconds() - startt) < time_constr)
529 res = do_memtest(TIMED_TESTS, devmem, 1);
538 fprintf(debug, "T-RES = %d\n", res);
539 fprintf(debug, "T-runtime: %d ms\n", getTimeMilliseconds() - time);
542 /* destroy context only if we created it */
550 /*! \brief Initializes the GPU with the given index.
552 * The varible \mygpu is the index of the GPU to initialize in the
553 * gpu_info.cuda_dev array.
555 * \param[in] mygpu index of the GPU to initialize
556 * \param[out] result_str the message related to the error that occurred
557 * during the initialization (if there was any).
558 * \param[in] gpu_info GPU info of all detected devices in the system.
559 * \param[in] gpu_opt options for using the GPUs in gpu_info
560 * \returns true if no error occurs during initialization.
562 gmx_bool init_gpu(int mygpu, char *result_str,
563 const gmx_gpu_info_t *gpu_info,
564 const gmx_gpu_opt_t *gpu_opt)
573 if (mygpu < 0 || mygpu >= gpu_opt->ncuda_dev_use)
575 sprintf(sbuf, "Trying to initialize an inexistent GPU: "
576 "there are %d %s-selected GPU(s), but #%d was requested.",
577 gpu_opt->ncuda_dev_use, gpu_opt->bUserSet ? "user" : "auto", mygpu);
581 gpuid = gpu_info->cuda_dev[gpu_opt->cuda_dev_use[mygpu]].id;
583 stat = cudaSetDevice(gpuid);
584 strncpy(result_str, cudaGetErrorString(stat), STRLEN);
588 fprintf(stderr, "Initialized GPU ID #%d: %s\n", gpuid, gpu_info->cuda_dev[gpuid].prop.name);
591 return (stat == cudaSuccess);
594 /*! \brief Frees up the CUDA GPU used by the active context at the time of calling.
596 * The context is explicitly destroyed and therefore all data uploaded to the GPU
597 * is lost. This should only be called when none of this data is required anymore.
599 * \param[out] result_str the message related to the error that occurred
600 * during the initialization (if there was any).
601 * \returns true if no error occurs during the freeing.
603 gmx_bool free_gpu(char *result_str)
612 stat = cudaGetDevice(&gpuid);
613 CU_RET_ERR(stat, "cudaGetDevice failed");
614 fprintf(stderr, "Cleaning up context on GPU ID #%d\n", gpuid);
617 #if CUDA_VERSION < 4000
618 stat = cudaThreadExit();
620 stat = cudaDeviceReset();
622 strncpy(result_str, cudaGetErrorString(stat), STRLEN);
624 return (stat == cudaSuccess);
627 /*! \brief Returns true if the gpu characterized by the device properties is
628 * supported by the native gpu acceleration.
630 * \param[in] dev_prop the CUDA device properties of the gpus to test.
631 * \returns true if the GPU properties passed indicate a compatible
632 * GPU, otherwise false.
634 static bool is_gmx_supported_gpu(const cudaDeviceProp *dev_prop)
636 return (dev_prop->major >= 2);
639 /*! \brief Helper function that checks whether a given GPU status indicates compatible GPU.
641 * \param[in] stat GPU status.
642 * \returns true if the provided status is egpuCompatible, otherwise false.
644 static bool is_compatible_gpu(int stat)
646 return (stat == egpuCompatible);
649 /*! \brief Checks if a GPU with a given ID is supported by the native GROMACS acceleration.
651 * Returns a status value which indicates compatibility or one of the following
652 * errors: incompatibility, insistence, or insanity (=unexpected behavior).
653 * It also returns the respective device's properties in \dev_prop (if applicable).
655 * \param[in] dev_id the ID of the GPU to check.
656 * \param[out] dev_prop the CUDA device properties of the device checked.
657 * \returns the status of the requested device
659 static int is_gmx_supported_gpu_id(int dev_id, cudaDeviceProp *dev_prop)
664 stat = cudaGetDeviceCount(&ndev);
665 if (stat != cudaSuccess)
670 if (dev_id > ndev - 1)
672 return egpuNonexistent;
675 /* TODO: currently we do not make a distinction between the type of errors
676 * that can appear during sanity checks. This needs to be improved, e.g if
677 * the dummy test kernel fails to execute with a "device busy message" we
678 * should appropriately report that the device is busy instead of insane.
680 if (do_sanity_checks(dev_id, dev_prop) == 0)
682 if (is_gmx_supported_gpu(dev_prop))
684 return egpuCompatible;
688 return egpuIncompatible;
698 /*! \brief Detect all NVIDIA GPUs in the system.
700 * Will detect every NVIDIA GPU supported by the device driver in use. Also
701 * check for the compatibility of each and fill the gpu_info->cuda_dev array
702 * with the required information on each the device: ID, device properties,
705 * \param[in] gpu_info pointer to structure holding GPU information.
706 * \param[out] err_str The error message of any CUDA API error that caused
707 * the detection to fail (if there was any). The memory
708 * the pointer points to should be managed externally.
709 * \returns non-zero if the detection encountered a failure, zero otherwise.
711 int detect_cuda_gpus(gmx_gpu_info_t *gpu_info, char *err_str)
713 int i, ndev, checkres, retval;
716 cuda_dev_info_t *devs;
721 gpu_info->ncuda_dev_compatible = 0;
726 stat = cudaGetDeviceCount(&ndev);
727 if (stat != cudaSuccess)
731 /* cudaGetDeviceCount failed which means that there is something
732 * wrong with the machine: driver-runtime mismatch, all GPUs being
733 * busy in exclusive mode, or some other condition which should
734 * result in us issuing a warning a falling back to CPUs. */
736 s = cudaGetErrorString(stat);
737 strncpy(err_str, s, STRLEN*sizeof(err_str[0]));
742 for (i = 0; i < ndev; i++)
744 checkres = is_gmx_supported_gpu_id(i, &prop);
748 devs[i].stat = checkres;
750 if (checkres == egpuCompatible)
752 gpu_info->ncuda_dev_compatible++;
758 gpu_info->ncuda_dev = ndev;
759 gpu_info->cuda_dev = devs;
764 /*! \brief Select the GPUs compatible with the native GROMACS acceleration.
766 * This function selects the compatible gpus and initializes
767 * gpu_info->cuda_dev_use and gpu_info->ncuda_dev_use.
769 * Given the list of GPUs available in the system check each device in
770 * gpu_info->cuda_dev and place the indices of the compatible GPUs into
771 * cuda_dev_use with this marking the respective GPUs as "available for use."
772 * Note that \detect_cuda_gpus must have been called before.
774 * \param[in] gpu_info pointer to structure holding GPU information
775 * \param[in,out] gpu_opt pointer to structure holding GPU options
777 void pick_compatible_gpus(const gmx_gpu_info_t *gpu_info,
778 gmx_gpu_opt_t *gpu_opt)
784 /* cuda_dev/ncuda_dev have to be either NULL/0 or not (NULL/0) */
785 assert((gpu_info->ncuda_dev != 0 ? 0 : 1) ^ (gpu_info->cuda_dev == NULL ? 0 : 1));
787 snew(compat, gpu_info->ncuda_dev);
789 for (i = 0; i < gpu_info->ncuda_dev; i++)
791 if (is_compatible_gpu(gpu_info->cuda_dev[i].stat))
794 compat[ncompat - 1] = i;
798 gpu_opt->ncuda_dev_use = ncompat;
799 snew(gpu_opt->cuda_dev_use, ncompat);
800 memcpy(gpu_opt->cuda_dev_use, compat, ncompat*sizeof(*compat));
804 /*! \brief Check the existence/compatibility of a set of GPUs specified by their device IDs.
806 * Given the a list of gpu->ncuda_dev_use GPU device IDs stored in
807 * gpu_opt->cuda_dev_use check the existence and compatibility
808 * of the respective GPUs. Also provide the caller with an array containing
809 * the result of checks in \checkres.
811 * \param[out] checkres check result for each ID passed in \requested_devs
812 * \param[in] gpu_info pointer to structure holding GPU information
813 * \param[out] gpu_opt pointer to structure holding GPU options
814 * \returns TRUE if every the requested GPUs are compatible
816 gmx_bool check_selected_cuda_gpus(int *checkres,
817 const gmx_gpu_info_t *gpu_info,
818 gmx_gpu_opt_t *gpu_opt)
825 assert(gpu_opt->ncuda_dev_use >= 0);
827 if (gpu_opt->ncuda_dev_use == 0)
832 assert(gpu_opt->cuda_dev_use);
834 /* we will assume that all GPUs requested are valid IDs,
835 otherwise we'll bail anyways */
838 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
840 id = gpu_opt->cuda_dev_use[i];
842 /* devices are stored in increasing order of IDs in cuda_dev */
843 gpu_opt->cuda_dev_use[i] = id;
845 checkres[i] = (id >= gpu_info->ncuda_dev) ?
846 egpuNonexistent : gpu_info->cuda_dev[id].stat;
848 bAllOk = bAllOk && is_compatible_gpu(checkres[i]);
854 /*! \brief Frees the cuda_dev and cuda_dev_use array fields of \gpu_info.
856 * \param[in] gpu_info pointer to structure holding GPU information
858 void free_gpu_info(const gmx_gpu_info_t *gpu_info)
860 if (gpu_info == NULL)
865 sfree(gpu_info->cuda_dev);
868 /*! \brief Formats and returns a device information string for a given GPU.
870 * Given an index *directly* into the array of available GPUs (cuda_dev)
871 * returns a formatted info string for the respective GPU which includes
872 * ID, name, compute capability, and detection status.
874 * \param[out] s pointer to output string (has to be allocated externally)
875 * \param[in] gpu_info pointer to structure holding GPU information
876 * \param[in] index an index *directly* into the array of available GPUs
878 void get_gpu_device_info_string(char *s, const gmx_gpu_info_t *gpu_info, int index)
883 if (index < 0 && index >= gpu_info->ncuda_dev)
888 cuda_dev_info_t *dinfo = &gpu_info->cuda_dev[index];
891 dinfo->stat == egpuCompatible ||
892 dinfo->stat == egpuIncompatible;
896 sprintf(s, "#%d: %s, stat: %s",
898 gpu_detect_res_str[dinfo->stat]);
902 sprintf(s, "#%d: NVIDIA %s, compute cap.: %d.%d, ECC: %3s, stat: %s",
903 dinfo->id, dinfo->prop.name,
904 dinfo->prop.major, dinfo->prop.minor,
905 dinfo->prop.ECCEnabled ? "yes" : " no",
906 gpu_detect_res_str[dinfo->stat]);
910 /*! \brief Returns the device ID of the GPU with a given index into the array of used GPUs.
912 * Getter function which, given an index into the array of GPUs in use
913 * (cuda_dev_use) -- typically a tMPI/MPI rank --, returns the device ID of the
914 * respective CUDA GPU.
916 * \param[in] gpu_info pointer to structure holding GPU information
917 * \param[in] gpu_opt pointer to structure holding GPU options
918 * \param[in] idx index into the array of used GPUs
919 * \returns device ID of the requested GPU
921 int get_gpu_device_id(const gmx_gpu_info_t *gpu_info,
922 const gmx_gpu_opt_t *gpu_opt,
927 assert(idx >= 0 && idx < gpu_opt->ncuda_dev_use);
929 return gpu_info->cuda_dev[gpu_opt->cuda_dev_use[idx]].id;
932 /*! \brief Returns the device ID of the GPU currently in use.
934 * The GPU used is the one that is active at the time of the call in the active context.
936 * \param[in] gpu_info pointer to structure holding GPU information
937 * \returns device ID of the GPU in use at the time of the call
939 int get_current_gpu_device_id(void)
942 CU_RET_ERR(cudaGetDevice(&gpuid), "cudaGetDevice failed");
947 /*! \brief Returns the size of the cuda_dev_info struct.
949 * The size of cuda_dev_info can be used for allocation and communication.
951 * \returns size in bytes of cuda_dev_info
953 size_t sizeof_cuda_dev_info(void)
955 return sizeof(cuda_dev_info);