2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 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 OpenCL devices.
38 * \author Anca Hamuraru <anca@streamcomputing.eu>
39 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
40 * \author Teemu Virolainen <teemu@streamcomputing.eu>
50 # include <sys/sysctl.h>
55 #include "gromacs/gpu_utils/gpu_utils.h"
56 #include "gromacs/gpu_utils/ocl_compiler.h"
57 #include "gromacs/gpu_utils/oclutils.h"
58 #include "gromacs/hardware/hw_info.h"
59 #include "gromacs/mdtypes/md_enums.h"
60 #include "gromacs/utility/cstringutil.h"
61 #include "gromacs/utility/fatalerror.h"
62 #include "gromacs/utility/smalloc.h"
64 /*! \brief Helper macro for error handling */
65 #define CALLOCLFUNC_LOGERROR(func, err_str, retval) { \
66 cl_int opencl_ret = func; \
67 if (CL_SUCCESS != opencl_ret) \
69 sprintf(err_str, "OpenCL error %d", opencl_ret); \
77 /*! \brief Return true if executing on compatible OS for AMD OpenCL.
79 * This is assumed to be true for OS X version of at least 10.10.4 and
80 * all other OS flavors.
82 * Uses the BSD sysctl() interfaces to extract the kernel version.
84 * \return true if version is 14.4 or later (= OS X version 10.10.4),
85 * or OS is not Darwin.
88 runningOnCompatibleOSForAmd()
92 char kernelVersion
[256];
93 size_t len
= sizeof(kernelVersion
);
96 mib
[1] = KERN_OSRELEASE
;
98 sysctl(mib
, sizeof(mib
)/sizeof(mib
[0]), kernelVersion
, &len
, NULL
, 0);
100 int major
= strtod(kernelVersion
, NULL
);
101 int minor
= strtod(strchr(kernelVersion
, '.')+1, NULL
);
103 // Kernel 14.4 corresponds to OS X 10.10.4
104 return (major
> 14 || (major
== 14 && minor
>= 4));
110 /*! \brief Returns true if the gpu characterized by the device properties is
111 * supported by the native gpu acceleration.
112 * \returns true if the GPU properties passed indicate a compatible
113 * GPU, otherwise false.
115 static int is_gmx_supported_gpu_id(struct gmx_device_info_t
*ocl_gpu_device
)
117 if ((getenv("GMX_OCL_DISABLE_COMPATIBILITY_CHECK")) != NULL
)
119 return egpuCompatible
;
122 /* Only AMD and NVIDIA GPUs are supported for now */
123 switch (ocl_gpu_device
->vendor_e
)
125 case OCL_VENDOR_NVIDIA
:
126 return egpuCompatible
;
128 return runningOnCompatibleOSForAmd() ? egpuCompatible
: egpuIncompatible
;
130 return egpuIncompatible
;
135 /*! \brief Returns an ocl_vendor_id_t value corresponding to the input OpenCL vendor name.
137 * \param[in] vendor_name String with OpenCL vendor name.
138 * \returns ocl_vendor_id_t value for the input vendor_name
140 static ocl_vendor_id_t
get_vendor_id(char *vendor_name
)
144 if (strstr(vendor_name
, "NVIDIA"))
146 return OCL_VENDOR_NVIDIA
;
149 if (strstr(vendor_name
, "AMD") ||
150 strstr(vendor_name
, "Advanced Micro Devices"))
152 return OCL_VENDOR_AMD
;
155 if (strstr(vendor_name
, "Intel"))
157 return OCL_VENDOR_INTEL
;
160 return OCL_VENDOR_UNKNOWN
;
164 //! This function is documented in the header file
165 int detect_gpus(gmx_gpu_info_t
*gpu_info
, char *err_str
)
168 cl_uint ocl_platform_count
;
169 cl_platform_id
*ocl_platform_ids
;
170 cl_device_type req_dev_type
= CL_DEVICE_TYPE_GPU
;
173 ocl_platform_ids
= NULL
;
175 if (getenv("GMX_OCL_FORCE_CPU") != NULL
)
177 req_dev_type
= CL_DEVICE_TYPE_CPU
;
182 CALLOCLFUNC_LOGERROR(clGetPlatformIDs(0, NULL
, &ocl_platform_count
), err_str
, retval
)
188 if (1 > ocl_platform_count
)
193 snew(ocl_platform_ids
, ocl_platform_count
);
195 CALLOCLFUNC_LOGERROR(clGetPlatformIDs(ocl_platform_count
, ocl_platform_ids
, NULL
), err_str
, retval
)
201 for (unsigned int i
= 0; i
< ocl_platform_count
; i
++)
203 cl_uint ocl_device_count
;
205 /* If requesting req_dev_type devices fails, just go to the next platform */
206 if (CL_SUCCESS
!= clGetDeviceIDs(ocl_platform_ids
[i
], req_dev_type
, 0, NULL
, &ocl_device_count
))
211 if (1 <= ocl_device_count
)
213 gpu_info
->n_dev
+= ocl_device_count
;
217 if (1 > gpu_info
->n_dev
)
222 snew(gpu_info
->gpu_dev
, gpu_info
->n_dev
);
226 cl_device_id
*ocl_device_ids
;
228 snew(ocl_device_ids
, gpu_info
->n_dev
);
231 for (unsigned int i
= 0; i
< ocl_platform_count
; i
++)
233 cl_uint ocl_device_count
;
235 /* If requesting req_dev_type devices fails, just go to the next platform */
236 if (CL_SUCCESS
!= clGetDeviceIDs(ocl_platform_ids
[i
], req_dev_type
, gpu_info
->n_dev
, ocl_device_ids
, &ocl_device_count
))
241 if (1 > ocl_device_count
)
246 for (unsigned int j
= 0; j
< ocl_device_count
; j
++)
248 gpu_info
->gpu_dev
[device_index
].ocl_gpu_id
.ocl_platform_id
= ocl_platform_ids
[i
];
249 gpu_info
->gpu_dev
[device_index
].ocl_gpu_id
.ocl_device_id
= ocl_device_ids
[j
];
251 gpu_info
->gpu_dev
[device_index
].device_name
[0] = 0;
252 clGetDeviceInfo(ocl_device_ids
[j
], CL_DEVICE_NAME
, sizeof(gpu_info
->gpu_dev
[device_index
].device_name
), gpu_info
->gpu_dev
[device_index
].device_name
, NULL
);
254 gpu_info
->gpu_dev
[device_index
].device_version
[0] = 0;
255 clGetDeviceInfo(ocl_device_ids
[j
], CL_DEVICE_VERSION
, sizeof(gpu_info
->gpu_dev
[device_index
].device_version
), gpu_info
->gpu_dev
[device_index
].device_version
, NULL
);
257 gpu_info
->gpu_dev
[device_index
].device_vendor
[0] = 0;
258 clGetDeviceInfo(ocl_device_ids
[j
], CL_DEVICE_VENDOR
, sizeof(gpu_info
->gpu_dev
[device_index
].device_vendor
), gpu_info
->gpu_dev
[device_index
].device_vendor
, NULL
);
260 gpu_info
->gpu_dev
[device_index
].compute_units
= 0;
261 clGetDeviceInfo(ocl_device_ids
[j
], CL_DEVICE_MAX_COMPUTE_UNITS
, sizeof(gpu_info
->gpu_dev
[device_index
].compute_units
), &(gpu_info
->gpu_dev
[device_index
].compute_units
), NULL
);
263 gpu_info
->gpu_dev
[device_index
].adress_bits
= 0;
264 clGetDeviceInfo(ocl_device_ids
[j
], CL_DEVICE_ADDRESS_BITS
, sizeof(gpu_info
->gpu_dev
[device_index
].adress_bits
), &(gpu_info
->gpu_dev
[device_index
].adress_bits
), NULL
);
266 gpu_info
->gpu_dev
[device_index
].vendor_e
= get_vendor_id(gpu_info
->gpu_dev
[device_index
].device_vendor
);
268 gpu_info
->gpu_dev
[device_index
].stat
= is_gmx_supported_gpu_id(gpu_info
->gpu_dev
+ device_index
);
270 if (egpuCompatible
== gpu_info
->gpu_dev
[device_index
].stat
)
272 gpu_info
->n_dev_compatible
++;
279 gpu_info
->n_dev
= device_index
;
281 /* Dummy sort of devices - AMD first, then NVIDIA, then Intel */
282 // TODO: Sort devices based on performance.
283 if (0 < gpu_info
->n_dev
)
286 for (int i
= 0; i
< gpu_info
->n_dev
; i
++)
288 if (OCL_VENDOR_AMD
== gpu_info
->gpu_dev
[i
].vendor_e
)
294 gmx_device_info_t ocl_gpu_info
;
295 ocl_gpu_info
= gpu_info
->gpu_dev
[i
];
296 gpu_info
->gpu_dev
[i
] = gpu_info
->gpu_dev
[last
];
297 gpu_info
->gpu_dev
[last
] = ocl_gpu_info
;
302 /* if more than 1 device left to be sorted */
303 if ((gpu_info
->n_dev
- 1 - last
) > 1)
305 for (int i
= 0; i
< gpu_info
->n_dev
; i
++)
307 if (OCL_VENDOR_NVIDIA
== gpu_info
->gpu_dev
[i
].vendor_e
)
313 gmx_device_info_t ocl_gpu_info
;
314 ocl_gpu_info
= gpu_info
->gpu_dev
[i
];
315 gpu_info
->gpu_dev
[i
] = gpu_info
->gpu_dev
[last
];
316 gpu_info
->gpu_dev
[last
] = ocl_gpu_info
;
323 sfree(ocl_device_ids
);
329 sfree(ocl_platform_ids
);
334 //! This function is documented in the header file
335 void free_gpu_info(const gmx_gpu_info_t gmx_unused
*gpu_info
)
337 if (gpu_info
== NULL
)
342 sfree(gpu_info
->gpu_dev
);
345 //! This function is documented in the header file
346 bool isGpuCompatible(const gmx_gpu_info_t
*gpu_info
,
351 return (index
>= gpu_info
->n_dev
?
353 gpu_info
->gpu_dev
[index
].stat
== egpuCompatible
);
356 //! This function is documented in the header file
357 const char *getGpuCompatibilityDescription(const gmx_gpu_info_t
*gpu_info
,
362 return (index
>= gpu_info
->n_dev
?
363 gpu_detect_res_str
[egpuNonexistent
] :
364 gpu_detect_res_str
[gpu_info
->gpu_dev
[index
].stat
]);
367 //! This function is documented in the header file
368 void get_gpu_device_info_string(char gmx_unused
*s
, const gmx_gpu_info_t gmx_unused
*gpu_info
, int gmx_unused index
)
373 if (index
< 0 && index
>= gpu_info
->n_dev
)
378 gmx_device_info_t
*dinfo
= &gpu_info
->gpu_dev
[index
];
381 dinfo
->stat
== egpuCompatible
||
382 dinfo
->stat
== egpuIncompatible
;
386 sprintf(s
, "#%d: %s, stat: %s",
388 gpu_detect_res_str
[dinfo
->stat
]);
392 sprintf(s
, "#%d: name: %s, vendor: %s, device version: %s, stat: %s",
393 index
, dinfo
->device_name
, dinfo
->device_vendor
,
394 dinfo
->device_version
,
395 gpu_detect_res_str
[dinfo
->stat
]);
399 //! This function is documented in the header file
400 void init_gpu(const gmx::MDLogger
& /*mdlog*/,
403 const gmx_gpu_info_t
*gpu_info
,
404 const gmx_gpu_opt_t
*gpu_opt
409 if (mygpu
< 0 || mygpu
>= gpu_opt
->n_dev_use
)
412 sprintf(sbuf
, "On rank %d trying to initialize an non-existent GPU: "
413 "there are %d selected GPU(s), but #%d was requested.",
414 rank
, gpu_opt
->n_dev_use
, mygpu
);
418 // If the device is NVIDIA, for safety reasons we disable the JIT
419 // caching as this is known to be broken at least until driver 364.19;
420 // the cache does not always get regenerated when the source code changes,
421 // e.g. if the path to the kernel sources remains the same
423 if (gpu_info
->gpu_dev
[mygpu
].vendor_e
== OCL_VENDOR_NVIDIA
)
425 // Ignore return values, failing to set the variable does not mean
426 // that something will go wrong later.
428 _putenv("CUDA_CACHE_DISABLE=1");
430 // Don't override, maybe a dev is testing.
431 setenv("CUDA_CACHE_DISABLE", "1", 0);
436 //! This function is documented in the header file
437 int get_gpu_device_id(const gmx_gpu_info_t
*,
438 const gmx_gpu_opt_t
*gpu_opt
,
442 assert(idx
>= 0 && idx
< gpu_opt
->n_dev_use
);
444 return gpu_opt
->dev_use
[idx
];
447 //! This function is documented in the header file
448 char* get_ocl_gpu_device_name(const gmx_gpu_info_t
*gpu_info
,
449 const gmx_gpu_opt_t
*gpu_opt
,
454 assert(idx
>= 0 && idx
< gpu_opt
->n_dev_use
);
456 return gpu_info
->gpu_dev
[gpu_opt
->dev_use
[idx
]].device_name
;
459 //! This function is documented in the header file
460 size_t sizeof_gpu_dev_info(void)
462 return sizeof(gmx_device_info_t
);
465 /*! \brief Prints the name of a kernel function pointer.
467 * \param[in] kernel OpenCL kernel
468 * \returns CL_SUCCESS if the operation was successful, an OpenCL error otherwise.
470 cl_int
dbg_ocl_kernel_name(const cl_kernel kernel
)
473 char kernel_name
[256];
474 cl_error
= clGetKernelInfo(kernel
, CL_KERNEL_FUNCTION_NAME
,
475 sizeof(kernel_name
), &kernel_name
, NULL
);
478 printf("No kernel found!\n");
482 printf("%s\n", kernel_name
);
487 /*! \brief Prints the name of a kernel function pointer.
489 * \param[in] kernel OpenCL kernel
490 * \returns CL_SUCCESS if the operation was successful, an OpenCL error otherwise.
492 cl_int
dbg_ocl_kernel_name_address(void* kernel
)
495 char kernel_name
[256];
496 cl_error
= clGetKernelInfo((cl_kernel
)kernel
, CL_KERNEL_FUNCTION_NAME
,
497 sizeof(kernel_name
), &kernel_name
, NULL
);
500 printf("No kernel found!\n");
504 printf("%s\n", kernel_name
);
509 void gpu_set_host_malloc_and_free(bool bUseGpuKernels
,
510 gmx_host_alloc_t
**nb_alloc
,
511 gmx_host_free_t
**nb_free
)
515 *nb_alloc
= &ocl_pmalloc
;
516 *nb_free
= &ocl_pfree
;