2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2010,2012 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 static int cuda_max_device_count = 32; /*! Max number of devices supported by CUDA (for consistency checking).
64 In reality it 16 with CUDA <=v5.0, but let's stay on the safe side. */
66 /*! Dummy kernel used for sanity checking. */
67 __global__ void k_dummy_test(){}
70 /*! Bit-flags which refer to memtestG80 test types and are used in do_memtest to specify which tests to run. */
71 enum memtest_G80_test_types {
72 MOVING_INVERSIONS_10 = 0x1,
73 MOVING_INVERSIONS_RAND = 0x2,
74 WALKING_8BIT_M86 = 0x4,
76 WALKING_1_8BIT = 0x10,
77 WALKING_0_32BIT = 0x20,
78 WALKING_1_32BIT = 0x40,
83 LOGIC_1_ITER_SHMEM = 0x800,
84 LOGIC_4_ITER_SHMEM = 0x1000
89 * \brief Runs GPU sanity checks.
91 * Runs a series of checks to determine that the given GPU and underlying CUDA
92 * driver/runtime functions properly.
93 * Returns properties of a device with given ID or the one that has
94 * already been initialized earlier in the case if of \dev_id == -1.
96 * \param[in] dev_id the device ID of the GPU or -1 if the device has already been initialized
97 * \param[out] dev_prop pointer to the structure in which the device properties will be returned
98 * \returns 0 if the device looks OK
100 * TODO: introduce errors codes and handle errors more smoothly.
102 static int do_sanity_checks(int dev_id, cudaDeviceProp *dev_prop)
107 cu_err = cudaGetDeviceCount(&dev_count);
108 if (cu_err != cudaSuccess)
110 fprintf(stderr, "Error %d while querying device count: %s\n", cu_err,
111 cudaGetErrorString(cu_err));
115 /* no CUDA compatible device at all */
119 /* things might go horribly wrong if cudart is not compatible with the driver */
120 if (dev_count < 0 || dev_count > cuda_max_device_count)
123 if (dev_id == -1) /* device already selected let's not destroy the context */
125 cu_err = cudaGetDevice(&id);
126 if (cu_err != cudaSuccess)
128 fprintf(stderr, "Error %d while querying device id: %s\n", cu_err,
129 cudaGetErrorString(cu_err));
136 if (id > dev_count - 1) /* pfff there's no such device */
138 fprintf(stderr, "The requested device with id %d does not seem to exist (device count=%d)\n",
144 memset(dev_prop, 0, sizeof(cudaDeviceProp));
145 cu_err = cudaGetDeviceProperties(dev_prop, id);
146 if (cu_err != cudaSuccess)
148 fprintf(stderr, "Error %d while querying device properties: %s\n", cu_err,
149 cudaGetErrorString(cu_err));
153 /* both major & minor is 9999 if no CUDA capable devices are present */
154 if (dev_prop->major == 9999 && dev_prop->minor == 9999)
156 /* we don't care about emulation mode */
157 if (dev_prop->major == 0)
162 cu_err = cudaSetDevice(id);
163 if (cu_err != cudaSuccess)
165 fprintf(stderr, "Error %d while switching to device #%d: %s\n",
166 cu_err, id, cudaGetErrorString(cu_err));
171 /* try to execute a dummy kernel */
172 k_dummy_test<<<1, 512>>>();
173 if (cudaThreadSynchronize() != cudaSuccess)
178 /* destroy context if we created one */
181 #if CUDA_VERSION < 4000
182 cu_err = cudaThreadExit();
183 CU_RET_ERR(cu_err, "cudaThreadExit failed");
185 cu_err = cudaDeviceReset();
186 CU_RET_ERR(cu_err, "cudaDeviceReset failed");
195 * \brief Runs a set of memory tests specified by the given bit-flags.
196 * Tries to allocate and do the test on \p megs Mb memory or
197 * the greatest amount that can be allocated (>10Mb).
198 * In case if an error is detected it stops without finishing the remaining
199 * steps/iterations and returns greater then zero value.
200 * In case of other errors (e.g. kernel launch errors, device querying errors)
203 * \param[in] which_tests variable with bit-flags of the requested tests
204 * \param[in] megs amount of memory that will be tested in MB
205 * \param[in] iter number of iterations
206 * \returns 0 if no error was detected, otherwise >0
208 static int do_memtest(unsigned int which_tests, int megs, int iter)
212 uint err_count; //, err_iter;
214 // no parameter check as this fn won't be called externally
216 // let's try to allocate the mem
217 while (!tester.allocate(megs) && (megs - 10 > 0))
218 { megs -= 10; tester.deallocate(); }
222 fprintf(stderr, "Unable to allocate GPU memory!\n");
226 // clear the first 18 bits
227 which_tests &= 0x3FFF;
228 for (i = 0; i < iter; i++)
230 // Moving Inversions (ones and zeros)
231 if ((MOVING_INVERSIONS_10 & which_tests) == MOVING_INVERSIONS_10)
233 tester.gpuMovingInversionsOnesZeros(err_count);
235 return MOVING_INVERSIONS_10;
237 // Moving Inversions (random)
238 if ((MOVING_INVERSIONS_RAND & which_tests) == MOVING_INVERSIONS_RAND)
240 tester.gpuMovingInversionsRandom(err_count);
242 return MOVING_INVERSIONS_RAND;
244 // Memtest86 Walking 8-bit
245 if ((WALKING_8BIT_M86 & which_tests) == WALKING_8BIT_M86)
247 for (uint shift = 0; shift < 8; shift++)
249 tester.gpuWalking8BitM86(err_count, shift);
251 return WALKING_8BIT_M86;
254 // True Walking zeros (8-bit)
255 if ((WALKING_0_8BIT & which_tests) == WALKING_0_8BIT)
257 for (uint shift = 0; shift < 8; shift++)
259 tester.gpuWalking8Bit(err_count, false, shift);
261 return WALKING_0_8BIT;
264 // True Walking ones (8-bit)
265 if ((WALKING_1_8BIT & which_tests) == WALKING_1_8BIT)
267 for (uint shift = 0; shift < 8; shift++)
269 tester.gpuWalking8Bit(err_count, true, shift);
271 return WALKING_1_8BIT;
274 // Memtest86 Walking zeros (32-bit)
275 if ((WALKING_0_32BIT & which_tests) == WALKING_0_32BIT)
277 for (uint shift = 0; shift < 32; shift++)
279 tester.gpuWalking32Bit(err_count, false, shift);
281 return WALKING_0_32BIT;
284 // Memtest86 Walking ones (32-bit)
285 if ((WALKING_1_32BIT & which_tests) == WALKING_1_32BIT)
287 for (uint shift = 0; shift < 32; shift++)
289 tester.gpuWalking32Bit(err_count, true, shift);
291 return WALKING_1_32BIT;
295 if ((RANDOM_BLOCKS & which_tests) == RANDOM_BLOCKS)
297 tester.gpuRandomBlocks(err_count,rand());
299 return RANDOM_BLOCKS;
303 // Memtest86 Modulo-20
304 if ((MOD_20_32BIT & which_tests) == MOD_20_32BIT)
306 for (uint shift = 0; shift < 20; shift++)
308 tester.gpuModuloX(err_count, shift, rand(), 20, 2);
313 // Logic (one iteration)
314 if ((LOGIC_1_ITER & which_tests) == LOGIC_1_ITER)
316 tester.gpuShortLCG0(err_count,1);
320 // Logic (4 iterations)
321 if ((LOGIC_4_ITER & which_tests) == LOGIC_4_ITER)
323 tester.gpuShortLCG0(err_count,4);
328 // Logic (shared memory, one iteration)
329 if ((LOGIC_1_ITER_SHMEM & which_tests) == LOGIC_1_ITER_SHMEM)
331 tester.gpuShortLCG0Shmem(err_count,1);
333 return LOGIC_1_ITER_SHMEM;
335 // Logic (shared-memory, 4 iterations)
336 if ((LOGIC_4_ITER_SHMEM & which_tests) == LOGIC_4_ITER_SHMEM)
338 tester.gpuShortLCG0Shmem(err_count,4);
340 return LOGIC_4_ITER_SHMEM;
348 /*! \brief Runs a quick memory test and returns 0 in case if no error is detected.
349 * If an error is detected it stops before completing the test and returns a
350 * value greater then 0. In case of other errors (e.g. kernel launch errors,
351 * device querying errors) -1 is returned.
353 * \param[in] dev_id the device id of the GPU or -1 if the device has already been selected
354 * \returns 0 if no error was detected, otherwise >0
356 int do_quick_memtest(int dev_id)
358 cudaDeviceProp dev_prop;
359 int devmem, res, time=0;
361 if (debug) { time = getTimeMilliseconds(); }
363 if (do_sanity_checks(dev_id, &dev_prop) != 0)
365 // something went wrong
371 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
372 fprintf(debug, ">> Running QUICK memtests on %d MiB (out of total %d MiB), %d iterations\n",
373 QUICK_MEM, devmem, QUICK_ITER);
376 res = do_memtest(QUICK_TESTS, QUICK_MEM, QUICK_ITER);
380 fprintf(debug, "Q-RES = %d\n", res);
381 fprintf(debug, "Q-runtime: %d ms\n", getTimeMilliseconds() - time);
384 /* destroy context only if we created it */
385 if (dev_id !=-1) cudaThreadExit();
389 /*! \brief Runs a full memory test and returns 0 in case if no error is detected.
390 * If an error is detected it stops before completing the test and returns a
391 * value greater then 0. In case of other errors (e.g. kernel launch errors,
392 * device querying errors) -1 is returned.
394 * \param[in] dev_id the device id of the GPU or -1 if the device has already been selected
395 * \returns 0 if no error was detected, otherwise >0
398 int do_full_memtest(int dev_id)
400 cudaDeviceProp dev_prop;
401 int devmem, res, time=0;
403 if (debug) { time = getTimeMilliseconds(); }
405 if (do_sanity_checks(dev_id, &dev_prop) != 0)
407 // something went wrong
411 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
415 fprintf(debug, ">> Running FULL memtests on %d MiB (out of total %d MiB), %d iterations\n",
416 devmem, devmem, FULL_ITER);
419 /* do all test on the entire memory */
420 res = do_memtest(FULL_TESTS, devmem, FULL_ITER);
424 fprintf(debug, "F-RES = %d\n", res);
425 fprintf(debug, "F-runtime: %d ms\n", getTimeMilliseconds() - time);
428 /* destroy context only if we created it */
429 if (dev_id != -1) cudaThreadExit();
433 /*! \brief Runs a time constrained memory test and returns 0 in case if no error is detected.
434 * If an error is detected it stops before completing the test and returns a value greater
435 * than zero. In case of other errors (e.g. kernel launch errors, device querying errors) -1
436 * is returned. Note, that test iterations are not interrupted therefor the total runtime of
437 * the test will always be multipple of one iteration's runtime.
439 * \param[in] dev_id the device id of the GPU or -1 if the device has laredy been selected
440 * \param[in] time_constr the time limit of the testing
441 * \returns 0 if no error was detected, otherwise >0
443 int do_timed_memtest(int dev_id, int time_constr)
445 cudaDeviceProp dev_prop;
446 int devmem, res=0, time=0, startt;
448 if (debug) { time = getTimeMilliseconds(); }
450 time_constr *= 1000; /* convert to ms for convenience */
451 startt = getTimeMilliseconds();
453 if (do_sanity_checks(dev_id, &dev_prop) != 0)
455 // something went wrong
459 devmem = dev_prop.totalGlobalMem/(1024*1024); // in MiB
463 fprintf(debug, ">> Running time constrained memtests on %d MiB (out of total %d MiB), time limit of %d s \n",
464 devmem, devmem, time_constr);
467 /* do the TIMED_TESTS set, one step at a time on the entire memory
468 that can be allocated, and stop when the given time is exceeded */
469 while ( ((int)getTimeMilliseconds() - startt) < time_constr)
471 res = do_memtest(TIMED_TESTS, devmem, 1);
477 fprintf(debug, "T-RES = %d\n", res);
478 fprintf(debug, "T-runtime: %d ms\n", getTimeMilliseconds() - time);
481 /* destroy context only if we created it */
482 if (dev_id != -1) cudaThreadExit();
486 /*! \brief Initializes the GPU with the given index.
488 * The varible \mygpu is the index of the GPU to initialize in the
489 * gpu_info.cuda_dev array.
491 * \param[in] mygpu index of the GPU to initialize
492 * \param[out] result_str the message related to the error that occurred
493 * during the initialization (if there was any).
494 * \param[in] gpu_info GPU info of all detected devices in the system.
495 * \param[in] gpu_opt options for using the GPUs in gpu_info
496 * \returns true if no error occurs during initialization.
498 gmx_bool init_gpu(int mygpu, char *result_str,
499 const gmx_gpu_info_t *gpu_info,
500 const gmx_gpu_opt_t *gpu_opt)
509 if (mygpu < 0 || mygpu >= gpu_opt->ncuda_dev_use)
511 sprintf(sbuf, "Trying to initialize an inexistent GPU: "
512 "there are %d %s-selected GPU(s), but #%d was requested.",
513 gpu_opt->ncuda_dev_use, gpu_opt->bUserSet ? "user" : "auto", mygpu);
517 gpuid = gpu_info->cuda_dev[gpu_opt->cuda_dev_use[mygpu]].id;
519 stat = cudaSetDevice(gpuid);
520 strncpy(result_str, cudaGetErrorString(stat), STRLEN);
524 fprintf(stderr, "Initialized GPU ID #%d: %s\n", gpuid, gpu_info->cuda_dev[gpuid].prop.name);
527 return (stat == cudaSuccess);
530 /*! \brief Frees up the CUDA GPU used by the active context at the time of calling.
532 * The context is explicitly destroyed and therefore all data uploaded to the GPU
533 * is lost. This should only be called when none of this data is required anymore.
535 * \param[out] result_str the message related to the error that occurred
536 * during the initialization (if there was any).
537 * \returns true if no error occurs during the freeing.
539 gmx_bool free_gpu(char *result_str)
548 stat = cudaGetDevice(&gpuid);
549 CU_RET_ERR(stat, "cudaGetDevice failed");
550 fprintf(stderr, "Cleaning up context on GPU ID #%d\n", gpuid);
553 #if CUDA_VERSION < 4000
554 stat = cudaThreadExit();
556 stat = cudaDeviceReset();
558 strncpy(result_str, cudaGetErrorString(stat), STRLEN);
560 return (stat == cudaSuccess);
563 /*! \brief Returns true if the gpu characterized by the device properties is
564 * supported by the native gpu acceleration.
566 * \param[in] dev_prop the CUDA device properties of the gpus to test.
567 * \returns true if the GPU properties passed indicate a compatible
568 * GPU, otherwise false.
570 static bool is_gmx_supported_gpu(const cudaDeviceProp *dev_prop)
572 return (dev_prop->major >= 2);
575 /*! \brief Helper function that checks whether a given GPU status indicates compatible GPU.
577 * \param[in] stat GPU status.
578 * \returns true if the provided status is egpuCompatible, otherwise false.
580 static bool is_compatible_gpu(int stat)
582 return (stat == egpuCompatible);
585 /*! \brief Checks if a GPU with a given ID is supported by the native GROMACS acceleration.
587 * Returns a status value which indicates compatibility or one of the following
588 * errors: incompatibility, insistence, or insanity (=unexpected behavior).
589 * It also returns the respective device's properties in \dev_prop (if applicable).
591 * \param[in] dev_id the ID of the GPU to check.
592 * \param[out] dev_prop the CUDA device properties of the device checked.
593 * \returns the status of the requested device
595 static int is_gmx_supported_gpu_id(int dev_id, cudaDeviceProp *dev_prop)
600 stat = cudaGetDeviceCount(&ndev);
601 if (stat != cudaSuccess)
606 if (dev_id > ndev - 1)
608 return egpuNonexistent;
611 /* TODO: currently we do not make a distinction between the type of errors
612 * that can appear during sanity checks. This needs to be improved, e.g if
613 * the dummy test kernel fails to execute with a "device busy message" we
614 * should appropriately report that the device is busy instead of insane.
616 if (do_sanity_checks(dev_id, dev_prop) == 0)
618 if (is_gmx_supported_gpu(dev_prop))
620 return egpuCompatible;
624 return egpuIncompatible;
634 /*! \brief Detect all NVIDIA GPUs in the system.
636 * Will detect every NVIDIA GPU supported by the device driver in use. Also
637 * check for the compatibility of each and fill the gpu_info->cuda_dev array
638 * with the required information on each the device: ID, device properties,
641 * \param[in] gpu_info pointer to structure holding GPU information.
642 * \param[out] err_str The error message of any CUDA API error that caused
643 * the detection to fail (if there was any). The memory
644 * the pointer points to should be managed externally.
645 * \returns non-zero if the detection encountered a failure, zero otherwise.
647 int detect_cuda_gpus(gmx_gpu_info_t *gpu_info, char *err_str)
649 int i, ndev, checkres, retval;
652 cuda_dev_info_t *devs;
657 gpu_info->ncuda_dev_compatible = 0;
662 stat = cudaGetDeviceCount(&ndev);
663 if (stat != cudaSuccess)
667 /* cudaGetDeviceCount failed which means that there is something
668 * wrong with the machine: driver-runtime mismatch, all GPUs being
669 * busy in exclusive mode, or some other condition which should
670 * result in us issuing a warning a falling back to CPUs. */
672 s = cudaGetErrorString(stat);
673 strncpy(err_str, s, STRLEN*sizeof(err_str[0]));
678 for (i = 0; i < ndev; i++)
680 checkres = is_gmx_supported_gpu_id(i, &prop);
684 devs[i].stat = checkres;
686 if (checkres == egpuCompatible)
688 gpu_info->ncuda_dev_compatible++;
694 gpu_info->ncuda_dev = ndev;
695 gpu_info->cuda_dev = devs;
700 /*! \brief Select the GPUs compatible with the native GROMACS acceleration.
702 * This function selects the compatible gpus and initializes
703 * gpu_info->cuda_dev_use and gpu_info->ncuda_dev_use.
705 * Given the list of GPUs available in the system check each device in
706 * gpu_info->cuda_dev and place the indices of the compatible GPUs into
707 * cuda_dev_use with this marking the respective GPUs as "available for use."
708 * Note that \detect_cuda_gpus must have been called before.
710 * \param[in] gpu_info pointer to structure holding GPU information
711 * \param[in,out] gpu_opt pointer to structure holding GPU options
713 void pick_compatible_gpus(const gmx_gpu_info_t *gpu_info,
714 gmx_gpu_opt_t *gpu_opt)
720 /* cuda_dev/ncuda_dev have to be either NULL/0 or not (NULL/0) */
721 assert((gpu_info->ncuda_dev != 0 ? 0 : 1) ^ (gpu_info->cuda_dev == NULL ? 0 : 1));
723 snew(compat, gpu_info->ncuda_dev);
725 for (i = 0; i < gpu_info->ncuda_dev; i++)
727 if (is_compatible_gpu(gpu_info->cuda_dev[i].stat))
730 compat[ncompat - 1] = i;
734 gpu_opt->ncuda_dev_use = ncompat;
735 snew(gpu_opt->cuda_dev_use, ncompat);
736 memcpy(gpu_opt->cuda_dev_use, compat, ncompat*sizeof(*compat));
740 /*! \brief Check the existence/compatibility of a set of GPUs specified by their device IDs.
742 * Given the a list of gpu->ncuda_dev_use GPU device IDs stored in
743 * gpu_opt->cuda_dev_use check the existence and compatibility
744 * of the respective GPUs. Also provide the caller with an array containing
745 * the result of checks in \checkres.
747 * \param[out] checkres check result for each ID passed in \requested_devs
748 * \param[in] gpu_info pointer to structure holding GPU information
749 * \param[out] gpu_opt pointer to structure holding GPU options
750 * \returns TRUE if every the requested GPUs are compatible
752 gmx_bool check_selected_cuda_gpus(int *checkres,
753 const gmx_gpu_info_t *gpu_info,
754 gmx_gpu_opt_t *gpu_opt)
761 assert(gpu_opt->ncuda_dev_use >= 0);
763 if (gpu_opt->ncuda_dev_use == 0)
768 assert(gpu_opt->cuda_dev_use);
770 /* we will assume that all GPUs requested are valid IDs,
771 otherwise we'll bail anyways */
774 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
776 id = gpu_opt->cuda_dev_use[i];
778 /* devices are stored in increasing order of IDs in cuda_dev */
779 gpu_opt->cuda_dev_use[i] = id;
781 checkres[i] = (id >= gpu_info->ncuda_dev) ?
782 egpuNonexistent : gpu_info->cuda_dev[id].stat;
784 bAllOk = bAllOk && is_compatible_gpu(checkres[i]);
790 /*! \brief Frees the cuda_dev and cuda_dev_use array fields of \gpu_info.
792 * \param[in] gpu_info pointer to structure holding GPU information
794 void free_gpu_info(const gmx_gpu_info_t *gpu_info)
796 if (gpu_info == NULL)
801 sfree(gpu_info->cuda_dev);
804 /*! \brief Formats and returns a device information string for a given GPU.
806 * Given an index *directly* into the array of available GPUs (cuda_dev)
807 * returns a formatted info string for the respective GPU which includes
808 * ID, name, compute capability, and detection status.
810 * \param[out] s pointer to output string (has to be allocated externally)
811 * \param[in] gpu_info pointer to structure holding GPU information
812 * \param[in] index an index *directly* into the array of available GPUs
814 void get_gpu_device_info_string(char *s, const gmx_gpu_info_t *gpu_info, int index)
819 if (index < 0 && index >= gpu_info->ncuda_dev)
824 cuda_dev_info_t *dinfo = &gpu_info->cuda_dev[index];
827 dinfo->stat == egpuCompatible ||
828 dinfo->stat == egpuIncompatible;
832 sprintf(s, "#%d: %s, stat: %s",
834 gpu_detect_res_str[dinfo->stat]);
838 sprintf(s, "#%d: NVIDIA %s, compute cap.: %d.%d, ECC: %3s, stat: %s",
839 dinfo->id, dinfo->prop.name,
840 dinfo->prop.major, dinfo->prop.minor,
841 dinfo->prop.ECCEnabled ? "yes" : " no",
842 gpu_detect_res_str[dinfo->stat]);
846 /*! \brief Returns the device ID of the GPU with a given index into the array of used GPUs.
848 * Getter function which, given an index into the array of GPUs in use
849 * (cuda_dev_use) -- typically a tMPI/MPI rank --, returns the device ID of the
850 * respective CUDA GPU.
852 * \param[in] gpu_info pointer to structure holding GPU information
853 * \param[in] gpu_opt pointer to structure holding GPU options
854 * \param[in] idx index into the array of used GPUs
855 * \returns device ID of the requested GPU
857 int get_gpu_device_id(const gmx_gpu_info_t *gpu_info,
858 const gmx_gpu_opt_t *gpu_opt,
863 if (idx < 0 && idx >= gpu_opt->ncuda_dev_use)
868 return gpu_info->cuda_dev[gpu_opt->cuda_dev_use[idx]].id;
871 /*! \brief Returns the device ID of the GPU currently in use.
873 * The GPU used is the one that is active at the time of the call in the active context.
875 * \param[in] gpu_info pointer to structure holding GPU information
876 * \returns device ID of the GPU in use at the time of the call
878 int get_current_gpu_device_id(void)
881 CU_RET_ERR(cudaGetDevice(&gpuid), "cudaGetDevice failed");