Simplified handling of GPU device IDs
[gromacs.git] / src / gromacs / gpu_utils / gpu_utils_ocl.cpp
blob9ffb558588ee3252c14927e3ede0fb9ceebecafd
1 /*
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.
35 /*! \internal \file
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>
43 #include "gmxpre.h"
45 #include <assert.h>
46 #include <stdio.h>
47 #include <stdlib.h>
48 #include <string.h>
49 #ifdef __APPLE__
50 # include <sys/sysctl.h>
51 #endif
53 #include <memory.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) \
68 { \
69 sprintf(err_str, "OpenCL error %d", opencl_ret); \
70 retval = -1; \
71 } \
72 else{ \
73 retval = 0; } \
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.
87 static bool
88 runningOnCompatibleOSForAmd()
90 #ifdef __APPLE__
91 int mib[2];
92 char kernelVersion[256];
93 size_t len = sizeof(kernelVersion);
95 mib[0] = CTL_KERN;
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));
105 #else
106 return true;
107 #endif
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;
127 case OCL_VENDOR_AMD:
128 return runningOnCompatibleOSForAmd() ? egpuCompatible : egpuIncompatible;
129 default:
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)
142 if (vendor_name)
144 if (strstr(vendor_name, "NVIDIA"))
146 return OCL_VENDOR_NVIDIA;
148 else
149 if (strstr(vendor_name, "AMD") ||
150 strstr(vendor_name, "Advanced Micro Devices"))
152 return OCL_VENDOR_AMD;
154 else
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)
167 int retval;
168 cl_uint ocl_platform_count;
169 cl_platform_id *ocl_platform_ids;
170 cl_device_type req_dev_type = CL_DEVICE_TYPE_GPU;
172 retval = 0;
173 ocl_platform_ids = NULL;
175 if (getenv("GMX_OCL_FORCE_CPU") != NULL)
177 req_dev_type = CL_DEVICE_TYPE_CPU;
180 while (1)
182 CALLOCLFUNC_LOGERROR(clGetPlatformIDs(0, NULL, &ocl_platform_count), err_str, retval)
183 if (0 != retval)
185 break;
188 if (1 > ocl_platform_count)
190 break;
193 snew(ocl_platform_ids, ocl_platform_count);
195 CALLOCLFUNC_LOGERROR(clGetPlatformIDs(ocl_platform_count, ocl_platform_ids, NULL), err_str, retval)
196 if (0 != retval)
198 break;
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))
208 continue;
211 if (1 <= ocl_device_count)
213 gpu_info->n_dev += ocl_device_count;
217 if (1 > gpu_info->n_dev)
219 break;
222 snew(gpu_info->gpu_dev, gpu_info->n_dev);
225 int device_index;
226 cl_device_id *ocl_device_ids;
228 snew(ocl_device_ids, gpu_info->n_dev);
229 device_index = 0;
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))
238 continue;
241 if (1 > ocl_device_count)
243 break;
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++;
275 device_index++;
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)
285 int last = -1;
286 for (int i = 0; i < gpu_info->n_dev; i++)
288 if (OCL_VENDOR_AMD == gpu_info->gpu_dev[i].vendor_e)
290 last++;
292 if (last < i)
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)
309 last++;
311 if (last < i)
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);
326 break;
329 sfree(ocl_platform_ids);
331 return retval;
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)
339 return;
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,
347 int index)
349 return (index >= gpu_info.n_dev ?
350 false :
351 gpu_info.gpu_dev[index].stat == egpuCompatible);
354 //! This function is documented in the header file
355 const char *getGpuCompatibilityDescription(const gmx_gpu_info_t &gpu_info,
356 int index)
358 return (index >= gpu_info.n_dev ?
359 gpu_detect_res_str[egpuNonexistent] :
360 gpu_detect_res_str[gpu_info.gpu_dev[index].stat]);
363 //! This function is documented in the header file
364 void get_gpu_device_info_string(char *s, const gmx_gpu_info_t &gpu_info, int index)
366 assert(s);
368 if (index < 0 && index >= gpu_info.n_dev)
370 return;
373 gmx_device_info_t *dinfo = &gpu_info.gpu_dev[index];
375 bool bGpuExists =
376 dinfo->stat == egpuCompatible ||
377 dinfo->stat == egpuIncompatible;
379 if (!bGpuExists)
381 sprintf(s, "#%d: %s, stat: %s",
382 index, "N/A",
383 gpu_detect_res_str[dinfo->stat]);
385 else
387 sprintf(s, "#%d: name: %s, vendor: %s, device version: %s, stat: %s",
388 index, dinfo->device_name, dinfo->device_vendor,
389 dinfo->device_version,
390 gpu_detect_res_str[dinfo->stat]);
394 //! This function is documented in the header file
395 void init_gpu(const gmx::MDLogger & /*mdlog*/,
396 int /* rank */,
397 gmx_device_info_t *deviceInfo)
399 assert(deviceInfo);
401 // If the device is NVIDIA, for safety reasons we disable the JIT
402 // caching as this is known to be broken at least until driver 364.19;
403 // the cache does not always get regenerated when the source code changes,
404 // e.g. if the path to the kernel sources remains the same
406 if (deviceInfo->vendor_e == OCL_VENDOR_NVIDIA)
408 // Ignore return values, failing to set the variable does not mean
409 // that something will go wrong later.
410 #ifdef _MSC_VER
411 _putenv("CUDA_CACHE_DISABLE=1");
412 #else
413 // Don't override, maybe a dev is testing.
414 setenv("CUDA_CACHE_DISABLE", "1", 0);
415 #endif
419 //! This function is documented in the header file
420 gmx_device_info_t *getDeviceInfo(const gmx_gpu_info_t &gpu_info,
421 int deviceId)
423 if (deviceId < 0 || deviceId >= gpu_info.n_dev)
425 gmx_incons("Invalid GPU deviceId requested");
427 return &gpu_info.gpu_dev[deviceId];
430 //! This function is documented in the header file
431 size_t sizeof_gpu_dev_info(void)
433 return sizeof(gmx_device_info_t);
436 /*! \brief Prints the name of a kernel function pointer.
438 * \param[in] kernel OpenCL kernel
439 * \returns CL_SUCCESS if the operation was successful, an OpenCL error otherwise.
441 cl_int dbg_ocl_kernel_name(const cl_kernel kernel)
443 cl_int cl_error;
444 char kernel_name[256];
445 cl_error = clGetKernelInfo(kernel, CL_KERNEL_FUNCTION_NAME,
446 sizeof(kernel_name), &kernel_name, NULL);
447 if (cl_error)
449 printf("No kernel found!\n");
451 else
453 printf("%s\n", kernel_name);
455 return cl_error;
458 /*! \brief Prints the name of a kernel function pointer.
460 * \param[in] kernel OpenCL kernel
461 * \returns CL_SUCCESS if the operation was successful, an OpenCL error otherwise.
463 cl_int dbg_ocl_kernel_name_address(void* kernel)
465 cl_int cl_error;
466 char kernel_name[256];
467 cl_error = clGetKernelInfo((cl_kernel)kernel, CL_KERNEL_FUNCTION_NAME,
468 sizeof(kernel_name), &kernel_name, NULL);
469 if (cl_error)
471 printf("No kernel found!\n");
473 else
475 printf("%s\n", kernel_name);
477 return cl_error;
480 void gpu_set_host_malloc_and_free(bool bUseGpuKernels,
481 gmx_host_alloc_t **nb_alloc,
482 gmx_host_free_t **nb_free)
484 if (bUseGpuKernels)
486 *nb_alloc = &ocl_pmalloc;
487 *nb_free = &ocl_pfree;
489 else
491 *nb_alloc = NULL;
492 *nb_free = NULL;