2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016, by the GROMACS development team.
5 * Copyright (c) 2017,2018,2019,2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
37 * \brief Defines the OpenCL implementations of the device management.
39 * \author Anca Hamuraru <anca@streamcomputing.eu>
40 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
41 * \author Teemu Virolainen <teemu@streamcomputing.eu>
42 * \author Mark Abraham <mark.j.abraham@gmail.com>
43 * \author Szilárd Páll <pall.szilard@gmail.com>
44 * \author Artem Zhmurov <zhmurov@gmail.com>
46 * \ingroup module_hardware
52 #include "gromacs/gpu_utils/oclraii.h"
53 #include "gromacs/gpu_utils/oclutils.h"
54 #include "gromacs/hardware/device_management.h"
55 #include "gromacs/utility/fatalerror.h"
56 #include "gromacs/utility/smalloc.h"
57 #include "gromacs/utility/stringutil.h"
59 #include "device_information.h"
64 /*! \brief Returns an DeviceVendor value corresponding to the input OpenCL vendor name.
66 * \returns DeviceVendor value for the input vendor name
68 static DeviceVendor getDeviceVendor(const char* vendorName)
72 if (strstr(vendorName, "NVIDIA"))
74 return DeviceVendor::Nvidia;
76 else if (strstr(vendorName, "AMD") || strstr(vendorName, "Advanced Micro Devices"))
78 return DeviceVendor::Amd;
80 else if (strstr(vendorName, "Intel"))
82 return DeviceVendor::Intel;
85 return DeviceVendor::Unknown;
88 /*! \brief Return true if executing on compatible OS for AMD OpenCL.
90 * This is assumed to be true for OS X version of at least 10.10.4 and
91 * all other OS flavors.
93 * \return true if version is 14.4 or later (= OS X version 10.10.4),
94 * or OS is not Darwin.
96 static bool runningOnCompatibleOSForAmd()
100 char kernelVersion[256];
101 size_t len = sizeof(kernelVersion);
105 int major = strtod(kernelVersion, NULL);
106 int minor = strtod(strchr(kernelVersion, '.') + 1, NULL);
108 // Kernel 14.4 corresponds to OS X 10.10.4
109 return (major > 14 || (major == 14 && minor >= 4));
115 /*! \brief Return true if executing on compatible GPU for NVIDIA OpenCL.
117 * There are known issues with OpenCL when running on NVIDIA Volta or newer (CC 7+).
118 * As a workaround, we recommend using CUDA on such hardware.
120 * This function relies on cl_nv_device_attribute_query. In case it's not functioning properly,
121 * we trust the user and mark the device as compatible.
123 * \return true if running on Pascal (CC 6.x) or older, or if we can not determine device generation.
125 static bool runningOnCompatibleHWForNvidia(const DeviceInformation& deviceInfo)
127 // The macro is defined in Intel's and AMD's headers, but it's not strictly required to be there.
128 #ifndef CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV
131 static const unsigned int ccMajorBad = 7; // Volta and Turing
132 unsigned int ccMajor;
133 cl_device_id devId = deviceInfo.oclDeviceId;
134 const cl_int err = clGetDeviceInfo(devId, CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV,
135 sizeof(ccMajor), &ccMajor, nullptr);
136 if (err != CL_SUCCESS)
138 return true; // Err on a side of trusting the user to know what they are doing.
140 return ccMajor < ccMajorBad;
145 * \brief Checks that device \c deviceInfo is compatible with GROMACS.
147 * Vendor and OpenCL version support checks are executed an the result
150 * \param[in] deviceInfo The device info pointer.
151 * \returns The status enumeration value for the checked device:
153 static DeviceStatus isDeviceFunctional(const DeviceInformation& deviceInfo)
155 if (getenv("GMX_GPU_DISABLE_COMPATIBILITY_CHECK") != nullptr)
157 // Assume the device is compatible because checking has been disabled.
158 return DeviceStatus::Compatible;
160 if (getenv("GMX_OCL_DISABLE_COMPATIBILITY_CHECK") != nullptr)
163 "Environment variable GMX_OCL_DISABLE_COMPATIBILITY_CHECK is deprecated and will "
164 "be removed in release 2022. Please use GMX_GPU_DISABLE_COMPATIBILITY_CHECK "
166 return DeviceStatus::Compatible;
169 // OpenCL device version check, ensure >= REQUIRED_OPENCL_MIN_VERSION
170 constexpr unsigned int minVersionMajor = REQUIRED_OPENCL_MIN_VERSION_MAJOR;
171 constexpr unsigned int minVersionMinor = REQUIRED_OPENCL_MIN_VERSION_MINOR;
173 // Based on the OpenCL spec we're checking the version supported by
174 // the device which has the following format:
175 // OpenCL<space><major_version.minor_version><space><vendor-specific information>
176 unsigned int deviceVersionMinor, deviceVersionMajor;
177 const int valuesScanned = std::sscanf(deviceInfo.device_version, "OpenCL %u.%u",
178 &deviceVersionMajor, &deviceVersionMinor);
179 const bool versionLargeEnough =
180 ((valuesScanned == 2)
181 && ((deviceVersionMajor > minVersionMajor)
182 || (deviceVersionMajor == minVersionMajor && deviceVersionMinor >= minVersionMinor)));
183 if (!versionLargeEnough)
185 return DeviceStatus::Incompatible;
188 /* Only AMD, Intel, and NVIDIA GPUs are supported for now */
189 switch (deviceInfo.deviceVendor)
191 case DeviceVendor::Nvidia:
192 return runningOnCompatibleHWForNvidia(deviceInfo) ? DeviceStatus::Compatible
193 : DeviceStatus::IncompatibleNvidiaVolta;
194 case DeviceVendor::Amd:
195 return runningOnCompatibleOSForAmd() ? DeviceStatus::Compatible : DeviceStatus::Incompatible;
196 case DeviceVendor::Intel:
197 return GMX_OPENCL_NB_CLUSTER_SIZE == 4 ? DeviceStatus::Compatible
198 : DeviceStatus::IncompatibleClusterSize;
199 default: return DeviceStatus::Incompatible;
203 /*! \brief Make an error string following an OpenCL API call.
205 * It is meant to be called with \p status != CL_SUCCESS, but it will
206 * work correctly even if it is called with no OpenCL failure.
208 * \todo Make use of this function more.
210 * \param[in] message Supplies context, e.g. the name of the API call that returned the error.
211 * \param[in] status OpenCL API status code
212 * \returns A string describing the OpenCL error.
214 inline std::string makeOpenClInternalErrorString(const char* message, cl_int status)
216 if (message != nullptr)
218 return gmx::formatString("%s did %ssucceed %d: %s", message,
219 ((status != CL_SUCCESS) ? "not " : ""), status,
220 ocl_get_error_string(status).c_str());
224 return gmx::formatString("%sOpenCL error encountered %d: %s",
225 ((status != CL_SUCCESS) ? "" : "No "), status,
226 ocl_get_error_string(status).c_str());
231 * \brief Checks that device \c deviceInfo is sane (ie can run a kernel).
233 * Compiles and runs a dummy kernel to determine whether the given
234 * OpenCL device functions properly.
237 * \param[in] deviceInfo The device info pointer.
238 * \param[out] errorMessage An error message related to a failing OpenCL API call.
239 * \throws std::bad_alloc When out of memory.
240 * \returns Whether the device passed sanity checks
242 static bool isDeviceFunctional(const DeviceInformation& deviceInfo, std::string* errorMessage)
244 cl_context_properties properties[] = {
245 CL_CONTEXT_PLATFORM, reinterpret_cast<cl_context_properties>(deviceInfo.oclPlatformId), 0
247 // uncrustify spacing
250 auto deviceId = deviceInfo.oclDeviceId;
251 ClContext context(clCreateContext(properties, 1, &deviceId, nullptr, nullptr, &status));
252 if (status != CL_SUCCESS)
254 errorMessage->assign(makeOpenClInternalErrorString("clCreateContext", status));
257 ClCommandQueue commandQueue(clCreateCommandQueue(context, deviceId, 0, &status));
258 if (status != CL_SUCCESS)
260 errorMessage->assign(makeOpenClInternalErrorString("clCreateCommandQueue", status));
264 // Some compilers such as Apple's require kernel functions to have at least one argument
265 const char* lines[] = { "__kernel void dummyKernel(__global void* input){}" };
266 ClProgram program(clCreateProgramWithSource(context, 1, lines, nullptr, &status));
267 if (status != CL_SUCCESS)
269 errorMessage->assign(makeOpenClInternalErrorString("clCreateProgramWithSource", status));
273 if ((status = clBuildProgram(program, 0, nullptr, nullptr, nullptr, nullptr)) != CL_SUCCESS)
275 errorMessage->assign(makeOpenClInternalErrorString("clBuildProgram", status));
279 ClKernel kernel(clCreateKernel(program, "dummyKernel", &status));
280 if (status != CL_SUCCESS)
282 errorMessage->assign(makeOpenClInternalErrorString("clCreateKernel", status));
286 clSetKernelArg(kernel, 0, sizeof(void*), nullptr);
288 const size_t localWorkSize = 1, globalWorkSize = 1;
289 if ((status = clEnqueueNDRangeKernel(commandQueue, kernel, 1, nullptr, &globalWorkSize,
290 &localWorkSize, 0, nullptr, nullptr))
293 errorMessage->assign(makeOpenClInternalErrorString("clEnqueueNDRangeKernel", status));
299 /*! \brief Check whether the \c ocl_gpu_device is suitable for use by mdrun
301 * Runs sanity checks: checking that the runtime can compile a dummy kernel
302 * and this can be executed;
303 * Runs compatibility checks verifying the device OpenCL version requirement
304 * and vendor/OS support.
306 * \param[in] deviceId The runtime-reported numeric ID of the device.
307 * \param[in] deviceInfo The device info pointer.
308 * \returns A DeviceStatus to indicate if the GPU device is supported and if it was able to run
309 * basic functionality checks.
311 static DeviceStatus checkGpu(size_t deviceId, const DeviceInformation& deviceInfo)
314 DeviceStatus supportStatus = isDeviceFunctional(deviceInfo);
315 if (supportStatus != DeviceStatus::Compatible)
317 return supportStatus;
320 std::string errorMessage;
321 if (!isDeviceFunctional(deviceInfo, &errorMessage))
323 gmx_warning("While sanity checking device #%zu, %s", deviceId, errorMessage.c_str());
324 return DeviceStatus::NonFunctional;
327 return DeviceStatus::Compatible;
332 bool isDeviceDetectionFunctional(std::string* errorMessage)
334 cl_uint numPlatforms;
335 cl_int status = clGetPlatformIDs(0, nullptr, &numPlatforms);
336 GMX_ASSERT(status != CL_INVALID_VALUE, "Incorrect call of clGetPlatformIDs detected");
338 if (status == CL_PLATFORM_NOT_FOUND_KHR)
340 // No valid ICDs found
341 if (errorMessage != nullptr)
343 errorMessage->assign("No valid OpenCL driver found");
349 status == CL_SUCCESS,
350 gmx::formatString("An unexpected value was returned from clGetPlatformIDs %d: %s",
351 status, ocl_get_error_string(status).c_str())
353 bool foundPlatform = (numPlatforms > 0);
354 if (!foundPlatform && errorMessage != nullptr)
356 errorMessage->assign("No OpenCL platforms found even though the driver was valid");
358 return foundPlatform;
361 std::vector<std::unique_ptr<DeviceInformation>> findDevices()
363 cl_uint ocl_platform_count;
364 cl_platform_id* ocl_platform_ids;
365 cl_device_type req_dev_type = CL_DEVICE_TYPE_GPU;
367 ocl_platform_ids = nullptr;
369 if (getenv("GMX_OCL_FORCE_CPU") != nullptr)
371 req_dev_type = CL_DEVICE_TYPE_CPU;
375 std::vector<std::unique_ptr<DeviceInformation>> deviceInfoList(0);
379 cl_int status = clGetPlatformIDs(0, nullptr, &ocl_platform_count);
380 if (CL_SUCCESS != status)
382 GMX_THROW(gmx::InternalError(
383 gmx::formatString("An unexpected value %d was returned from clGetPlatformIDs: ", status)
384 + ocl_get_error_string(status)));
387 if (1 > ocl_platform_count)
389 // TODO this should have a descriptive error message that we only support one OpenCL platform
393 snew(ocl_platform_ids, ocl_platform_count);
395 status = clGetPlatformIDs(ocl_platform_count, ocl_platform_ids, nullptr);
396 if (CL_SUCCESS != status)
398 GMX_THROW(gmx::InternalError(
399 gmx::formatString("An unexpected value %d was returned from clGetPlatformIDs: ", status)
400 + ocl_get_error_string(status)));
403 for (unsigned int i = 0; i < ocl_platform_count; i++)
405 cl_uint ocl_device_count;
407 /* If requesting req_dev_type devices fails, just go to the next platform */
408 if (CL_SUCCESS != clGetDeviceIDs(ocl_platform_ids[i], req_dev_type, 0, nullptr, &ocl_device_count))
413 if (1 <= ocl_device_count)
415 numDevices += ocl_device_count;
424 deviceInfoList.resize(numDevices);
428 cl_device_id* ocl_device_ids;
430 snew(ocl_device_ids, numDevices);
433 for (unsigned int i = 0; i < ocl_platform_count; i++)
435 cl_uint ocl_device_count;
437 /* If requesting req_dev_type devices fails, just go to the next platform */
439 != clGetDeviceIDs(ocl_platform_ids[i], req_dev_type, numDevices, ocl_device_ids,
445 if (1 > ocl_device_count)
450 for (unsigned int j = 0; j < ocl_device_count; j++)
452 deviceInfoList[device_index] = std::make_unique<DeviceInformation>();
454 deviceInfoList[device_index]->id = device_index;
456 deviceInfoList[device_index]->oclPlatformId = ocl_platform_ids[i];
457 deviceInfoList[device_index]->oclDeviceId = ocl_device_ids[j];
459 deviceInfoList[device_index]->device_name[0] = 0;
460 clGetDeviceInfo(ocl_device_ids[j], CL_DEVICE_NAME,
461 sizeof(deviceInfoList[device_index]->device_name),
462 deviceInfoList[device_index]->device_name, nullptr);
464 deviceInfoList[device_index]->device_version[0] = 0;
465 clGetDeviceInfo(ocl_device_ids[j], CL_DEVICE_VERSION,
466 sizeof(deviceInfoList[device_index]->device_version),
467 deviceInfoList[device_index]->device_version, nullptr);
469 deviceInfoList[device_index]->vendorName[0] = 0;
470 clGetDeviceInfo(ocl_device_ids[j], CL_DEVICE_VENDOR,
471 sizeof(deviceInfoList[device_index]->vendorName),
472 deviceInfoList[device_index]->vendorName, nullptr);
474 deviceInfoList[device_index]->compute_units = 0;
475 clGetDeviceInfo(ocl_device_ids[j], CL_DEVICE_MAX_COMPUTE_UNITS,
476 sizeof(deviceInfoList[device_index]->compute_units),
477 &(deviceInfoList[device_index]->compute_units), nullptr);
479 deviceInfoList[device_index]->adress_bits = 0;
480 clGetDeviceInfo(ocl_device_ids[j], CL_DEVICE_ADDRESS_BITS,
481 sizeof(deviceInfoList[device_index]->adress_bits),
482 &(deviceInfoList[device_index]->adress_bits), nullptr);
484 deviceInfoList[device_index]->deviceVendor =
485 gmx::getDeviceVendor(deviceInfoList[device_index]->vendorName);
487 clGetDeviceInfo(ocl_device_ids[j], CL_DEVICE_MAX_WORK_ITEM_SIZES, 3 * sizeof(size_t),
488 &deviceInfoList[device_index]->maxWorkItemSizes, nullptr);
490 clGetDeviceInfo(ocl_device_ids[j], CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t),
491 &deviceInfoList[device_index]->maxWorkGroupSize, nullptr);
493 deviceInfoList[device_index]->status =
494 gmx::checkGpu(device_index, *deviceInfoList[device_index]);
500 numDevices = device_index;
502 /* Dummy sort of devices - AMD first, then NVIDIA, then Intel */
503 // TODO: Sort devices based on performance.
507 for (int i = 0; i < numDevices; i++)
509 if (deviceInfoList[i]->deviceVendor == DeviceVendor::Amd)
515 std::swap(deviceInfoList[i], deviceInfoList[last]);
520 /* if more than 1 device left to be sorted */
521 if ((numDevices - 1 - last) > 1)
523 for (int i = 0; i < numDevices; i++)
525 if (deviceInfoList[i]->deviceVendor == DeviceVendor::Nvidia)
531 std::swap(deviceInfoList[i], deviceInfoList[last]);
538 sfree(ocl_device_ids);
544 sfree(ocl_platform_ids);
545 return deviceInfoList;
548 void setActiveDevice(const DeviceInformation& deviceInfo)
550 // If the device is NVIDIA, for safety reasons we disable the JIT
551 // caching as this is known to be broken at least until driver 364.19;
552 // the cache does not always get regenerated when the source code changes,
553 // e.g. if the path to the kernel sources remains the same
555 if (deviceInfo.deviceVendor == DeviceVendor::Nvidia)
557 // Ignore return values, failing to set the variable does not mean
558 // that something will go wrong later.
560 _putenv("CUDA_CACHE_DISABLE=1");
562 // Don't override, maybe a dev is testing.
563 setenv("CUDA_CACHE_DISABLE", "1", 0);
568 void releaseDevice(DeviceInformation* /* deviceInfo */) {}
570 std::string getDeviceInformationString(const DeviceInformation& deviceInfo)
572 bool gpuExists = (deviceInfo.status != DeviceStatus::Nonexistent
573 && deviceInfo.status != DeviceStatus::NonFunctional);
577 return gmx::formatString("#%d: %s, status: %s", deviceInfo.id, "N/A",
578 c_deviceStateString[deviceInfo.status]);
582 return gmx::formatString("#%d: name: %s, vendor: %s, device version: %s, status: %s",
583 deviceInfo.id, deviceInfo.device_name, deviceInfo.vendorName,
584 deviceInfo.device_version, c_deviceStateString[deviceInfo.status]);