#include <algorithm>
#include <array>
-#include <chrono>
#include <memory>
#include <string>
-#include <thread>
#include <vector>
-#include "gromacs/compat/pointers.h"
-#include "gromacs/gpu_utils/gpu_utils.h"
#include "gromacs/hardware/cpuinfo.h"
+#include "gromacs/hardware/device_management.h"
#include "gromacs/hardware/hardwaretopology.h"
#include "gromacs/hardware/hw_info.h"
#include "gromacs/simd/support.h"
#include "gromacs/utility/exceptions.h"
#include "gromacs/utility/fatalerror.h"
#include "gromacs/utility/gmxassert.h"
+#include "gromacs/utility/inmemoryserializer.h"
#include "gromacs/utility/logger.h"
-#include "gromacs/utility/mutex.h"
#include "gromacs/utility/physicalnodecommunicator.h"
#include "architecture.h"
+#include "device_information.h"
+#include "prepare_detection.h"
#ifdef HAVE_UNISTD_H
# include <unistd.h> // sysconf()
{
}
-gmx_hw_info_t::~gmx_hw_info_t()
-{
- free_gpu_info(&gpu_info);
-}
+gmx_hw_info_t::~gmx_hw_info_t() = default;
namespace gmx
{
# define _SC_NPROCESSORS_CONF _SC_NPROC_CONF
#endif
-/*! \brief Information about the hardware of all nodes (common to all threads in this process).
+/*! \brief The result of device detection
*
- * This information is constructed only when required, but thereafter
- * its lifetime is that of the whole process, potentially across
- * multiple successive simulation parts. It's wise to ensure that only
- * one thread can create the information, but thereafter they can all
- * read it without e.g. needing a std::shared_ptr to ensure its
- * lifetime exceeds that of the thread. */
-static std::unique_ptr<gmx_hw_info_t> g_hardwareInfo;
-//! A mutex to protect the hwinfo structure
-static Mutex g_hardwareInfoMutex;
-
-//! Detect GPUs, if that makes sense to attempt.
-static void gmx_detect_gpus(const gmx::MDLogger& mdlog,
- const PhysicalNodeCommunicator& physicalNodeComm,
- compat::not_null<gmx_hw_info_t*> hardwareInfo)
+ * Note that non-functional device detection still produces
+ * a detection result, ie. of no devices. This might not be
+ * what the user wanted, so it makes sense to log later when
+ * that is possible. */
+struct DeviceDetectionResult
{
- hardwareInfo->gpu_info.bDetectGPUs = canPerformGpuDetection();
+ //! The device information detected
+ std::vector<std::unique_ptr<DeviceInformation>> deviceInfoList_;
+ //! Container of possible warnings to issue when that is possible
+ std::vector<std::string> deviceDetectionWarnings_;
+};
- if (!hardwareInfo->gpu_info.bDetectGPUs)
+/*! \brief Detect GPUs when that makes sense to attempt.
+ *
+ * \param[in] physicalNodeComm The communicator across this physical node
+ * \return The result of the detection, perhaps including diagnostic messages
+ * to issue later.
+ *
+ * \todo Coordinating the efficient detection of devices across
+ * multiple ranks per node should be separated from the lower-level
+ * hardware detection. See
+ * https://gitlab.com/gromacs/gromacs/-/issues/3650.
+ */
+static DeviceDetectionResult detectAllDeviceInformation(const PhysicalNodeCommunicator& physicalNodeComm)
+{
+ DeviceDetectionResult deviceDetectionResult;
+
+ if (!isDeviceDetectionEnabled())
{
- return;
+ return deviceDetectionResult;
}
+ std::string errorMessage;
+
bool isMasterRankOfPhysicalNode = true;
#if GMX_LIB_MPI
isMasterRankOfPhysicalNode = (physicalNodeComm.rank_ == 0);
#else
- // We choose to run the detection only once with thread-MPI and
- // use a mutex to enforce it.
+ // Without an MPI library, this process is trivially the only one
+ // on the physical node. This code runs before e.g. thread-MPI
+ // ranks are spawned, so detection is race-free by construction.
+ // Read-only access is enforced with providing those ranks with a
+ // handle to a const object, so usage is also free of races.
GMX_UNUSED_VALUE(physicalNodeComm);
- isMasterRankOfPhysicalNode = true;
+ isMasterRankOfPhysicalNode = true;
#endif
- /* The OpenCL support requires us to run detection on all ranks.
+ /* The SYCL and OpenCL support requires us to run detection on all
+ * ranks.
+ *
* With CUDA we don't need to, and prefer to detect on one rank
- * and send the information to the other ranks over MPI. */
- bool allRanksMustDetectGpus = (GMX_GPU == GMX_GPU_OPENCL);
- bool gpusCanBeDetected = false;
+ * and send the information to the other ranks over MPI. This
+ * avoids creating a start-up bottleneck with each MPI rank on a
+ * node making the same GPU API calls. */
+ constexpr bool allRanksMustDetectGpus = (GMX_GPU_OPENCL != 0 || GMX_GPU_SYCL != 0);
+ bool gpusCanBeDetected = false;
if (isMasterRankOfPhysicalNode || allRanksMustDetectGpus)
{
std::string errorMessage;
- gpusCanBeDetected = isGpuDetectionFunctional(&errorMessage);
+ gpusCanBeDetected = isDeviceDetectionFunctional(&errorMessage);
if (!gpusCanBeDetected)
{
- GMX_LOG(mdlog.info)
- .asParagraph()
- .appendTextFormatted(
- "NOTE: Detection of GPUs failed. The API reported:\n"
- " %s\n"
- " GROMACS cannot run tasks on a GPU.",
- errorMessage.c_str());
+ deviceDetectionResult.deviceDetectionWarnings_.emplace_back(
+ "Detection of GPUs failed. The API reported:\n" + errorMessage);
}
}
if (gpusCanBeDetected)
{
- findGpus(&hardwareInfo->gpu_info);
+ deviceDetectionResult.deviceInfoList_ = findDevices();
// No need to tell the user anything at this point, they get a
// hardware report later.
}
#if GMX_LIB_MPI
- if (!allRanksMustDetectGpus)
+ if (!allRanksMustDetectGpus && (physicalNodeComm.size_ > 1))
{
- /* Broadcast the GPU info to the other ranks within this node */
- MPI_Bcast(&hardwareInfo->gpu_info.n_dev, 1, MPI_INT, 0, physicalNodeComm.comm_);
-
- if (hardwareInfo->gpu_info.n_dev > 0)
+ // Master rank must serialize the device information list and
+ // send it to the other ranks on this node.
+ std::vector<char> buffer;
+ int sizeOfBuffer;
+ if (isMasterRankOfPhysicalNode)
{
- int dev_size;
-
- dev_size = hardwareInfo->gpu_info.n_dev * sizeof_gpu_dev_info();
-
+ gmx::InMemorySerializer writer;
+ serializeDeviceInformations(deviceDetectionResult.deviceInfoList_, &writer);
+ buffer = writer.finishAndGetBuffer();
+ sizeOfBuffer = buffer.size();
+ }
+ // Ensure all ranks agree on the size of list to be sent
+ MPI_Bcast(&sizeOfBuffer, 1, MPI_INT, 0, physicalNodeComm.comm_);
+ buffer.resize(sizeOfBuffer);
+ if (!buffer.empty())
+ {
+ // Send the list and deserialize it
+ MPI_Bcast(buffer.data(), buffer.size(), MPI_BYTE, 0, physicalNodeComm.comm_);
if (!isMasterRankOfPhysicalNode)
{
- hardwareInfo->gpu_info.gpu_dev = (struct gmx_device_info_t*)malloc(dev_size);
+ gmx::InMemoryDeserializer reader(buffer, false);
+ deviceDetectionResult.deviceInfoList_ = deserializeDeviceInformations(&reader);
}
- MPI_Bcast(hardwareInfo->gpu_info.gpu_dev, dev_size, MPI_BYTE, 0, physicalNodeComm.comm_);
- MPI_Bcast(&hardwareInfo->gpu_info.n_dev_compatible, 1, MPI_INT, 0, physicalNodeComm.comm_);
}
}
#endif
+ return deviceDetectionResult;
}
//! Reduce the locally collected \p hardwareInfo over MPI ranks
-static void gmx_collect_hardware_mpi(const gmx::CpuInfo& cpuInfo,
- const PhysicalNodeCommunicator& physicalNodeComm,
- compat::not_null<gmx_hw_info_t*> hardwareInfo)
+static void gmx_collect_hardware_mpi(const gmx::CpuInfo& cpuInfo,
+ const PhysicalNodeCommunicator& physicalNodeComm,
+ gmx_hw_info_t* hardwareInfo)
{
const int ncore = hardwareInfo->hardwareTopology->numberOfCores();
/* Zen1 is assumed for:
&& (cpuInfo.model() == 1 || cpuInfo.model() == 17
|| cpuInfo.model() == 8 || cpuInfo.model() == 24))
|| cpuInfo.vendor() == CpuInfo::Vendor::Hygon);
+
+ int numCompatibleDevices = getCompatibleDevices(hardwareInfo->deviceInfoList).size();
#if GMX_LIB_MPI
- int nhwthread, ngpu, i;
+ int nhwthread;
int gpu_hash;
nhwthread = hardwareInfo->nthreads_hw_avail;
- ngpu = hardwareInfo->gpu_info.n_dev_compatible;
/* Create a unique hash of the GPU type(s) in this node */
gpu_hash = 0;
/* Here it might be better to only loop over the compatible GPU, but we
* don't have that information available and it would also require
* removing the device ID from the device info string.
*/
- for (i = 0; i < hardwareInfo->gpu_info.n_dev; i++)
+ for (const auto& deviceInfo : hardwareInfo->deviceInfoList)
{
- char stmp[STRLEN];
-
/* Since the device ID is incorporated in the hash, the order of
* the GPUs affects the hash. Also two identical GPUs won't give
* a gpu_hash of zero after XORing.
*/
- get_gpu_device_info_string(stmp, hardwareInfo->gpu_info, i);
- gpu_hash ^= gmx_string_fullhash_func(stmp, gmx_string_hash_init);
+ std::string deviceInfoString = getDeviceInformationString(*deviceInfo);
+ gpu_hash ^= gmx_string_fullhash_func(deviceInfoString.c_str(), gmx_string_hash_init);
}
constexpr int numElementsCounts = 4;
countsLocal[0] = 1;
countsLocal[1] = ncore;
countsLocal[2] = nhwthread;
- countsLocal[3] = ngpu;
+ countsLocal[3] = numCompatibleDevices;
}
MPI_Allreduce(countsLocal.data(), countsReduced.data(), countsLocal.size(), MPI_INT,
*/
maxMinLocal[0] = ncore;
maxMinLocal[1] = nhwthread;
- maxMinLocal[2] = ngpu;
+ maxMinLocal[2] = numCompatibleDevices;
maxMinLocal[3] = static_cast<int>(gmx::simdSuggested(cpuInfo));
maxMinLocal[4] = gpu_hash;
maxMinLocal[5] = -maxMinLocal[0];
hardwareInfo->bIdenticalGPUs = (maxMinReduced[4] == -maxMinReduced[9]);
hardwareInfo->haveAmdZen1Cpu = (maxMinReduced[10] > 0);
#else
- /* All ranks use the same pointer, protected by a mutex in the caller */
hardwareInfo->nphysicalnode = 1;
hardwareInfo->ncore_tot = ncore;
hardwareInfo->ncore_min = ncore;
hardwareInfo->nhwthread_tot = hardwareInfo->nthreads_hw_avail;
hardwareInfo->nhwthread_min = hardwareInfo->nthreads_hw_avail;
hardwareInfo->nhwthread_max = hardwareInfo->nthreads_hw_avail;
- hardwareInfo->ngpu_compatible_tot = hardwareInfo->gpu_info.n_dev_compatible;
- hardwareInfo->ngpu_compatible_min = hardwareInfo->gpu_info.n_dev_compatible;
- hardwareInfo->ngpu_compatible_max = hardwareInfo->gpu_info.n_dev_compatible;
+ hardwareInfo->ngpu_compatible_tot = numCompatibleDevices;
+ hardwareInfo->ngpu_compatible_min = numCompatibleDevices;
+ hardwareInfo->ngpu_compatible_max = numCompatibleDevices;
hardwareInfo->simd_suggest_min = static_cast<int>(simdSuggested(cpuInfo));
hardwareInfo->simd_suggest_max = static_cast<int>(simdSuggested(cpuInfo));
hardwareInfo->bIdenticalGPUs = TRUE;
#endif
}
-/*! \brief Utility that does dummy computing for max 2 seconds to spin up cores
- *
- * This routine will check the number of cores configured and online
- * (using sysconf), and the spins doing dummy compute operations for up to
- * 2 seconds, or until all cores have come online. This can be used prior to
- * hardware detection for platforms that take unused processors offline.
- *
- * This routine will not throw exceptions. In principle it should be
- * declared noexcept, but at least icc 19.1 and 21-beta08 with the
- * libstdc++-7.5 has difficulty implementing a std::vector of
- * std::thread started with this function when declared noexcept. It
- * is a known compiler bug that should be fixed after 19.1.
- * Fortunately, this function is not performance sensitive,
- * and only runs on platforms other than x86 and POWER (ie ARM),
- * so the possible overhead introduced by omitting noexcept is not
- * important.
- */
-static void spinUpCore()
-{
-#if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) && defined(_SC_NPROCESSORS_ONLN)
- float dummy = 0.1;
- int countConfigured = sysconf(_SC_NPROCESSORS_CONF); // noexcept
- auto start = std::chrono::steady_clock::now(); // noexcept
-
- while (sysconf(_SC_NPROCESSORS_ONLN) < countConfigured
- && std::chrono::steady_clock::now() - start < std::chrono::seconds(2))
- {
- for (int i = 1; i < 10000; i++)
- {
- dummy /= i;
- }
- }
-
- if (dummy < 0)
- {
- printf("This cannot happen, but prevents loop from being optimized away.");
- }
-#endif
-}
-
-/*! \brief Prepare the system before hardware topology detection
- *
- * This routine should perform any actions we want to put the system in a state
- * where we want it to be before detecting the hardware topology. For most
- * processors there is nothing to do, but some architectures (in particular ARM)
- * have support for taking configured cores offline, which will make them disappear
- * from the online processor count.
- *
- * This routine checks if there is a mismatch between the number of cores
- * configured and online, and in that case we issue a small workload that
- * attempts to wake sleeping cores before doing the actual detection.
- *
- * This type of mismatch can also occur for x86 or PowerPC on Linux, if SMT has only
- * been disabled in the kernel (rather than bios). Since those cores will never
- * come online automatically, we currently skip this test for x86 & PowerPC to
- * avoid wasting 2 seconds. We also skip the test if there is no thread support.
- *
- * \note Cores will sleep relatively quickly again, so it's important to issue
- * the real detection code directly after this routine.
- */
-static void hardwareTopologyPrepareDetection()
-{
-#if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF) \
- && (defined(THREAD_PTHREADS) || defined(THREAD_WINDOWS))
-
- // Modify this conditional when/if x86 or PowerPC starts to sleep some cores
- if (c_architecture != Architecture::X86 && c_architecture != Architecture::PowerPC)
- {
- int countConfigured = sysconf(_SC_NPROCESSORS_CONF);
- std::vector<std::thread> workThreads(countConfigured);
-
- for (auto& t : workThreads)
- {
- t = std::thread(spinUpCore);
- }
-
- for (auto& t : workThreads)
- {
- t.join();
- }
- }
-#endif
-}
-
-/*! \brief Sanity check hardware topology and print some notes to log
- *
- * \param mdlog Logger.
- * \param hardwareTopology Reference to hardwareTopology object.
- */
-static void hardwareTopologyDoubleCheckDetection(const gmx::MDLogger gmx_unused& mdlog,
- const gmx::HardwareTopology gmx_unused& hardwareTopology)
+void hardwareTopologyDoubleCheckDetection(const gmx::MDLogger gmx_unused& mdlog,
+ const gmx::HardwareTopology gmx_unused& hardwareTopology)
{
#if defined HAVE_SYSCONF && defined(_SC_NPROCESSORS_CONF)
if (hardwareTopology.supportLevel() < gmx::HardwareTopology::SupportLevel::LogicalProcessorCount)
"performance.");
}
}
+#else
+ GMX_UNUSED_VALUE(mdlog);
+ GMX_UNUSED_VALUE(hardwareTopology);
#endif
}
-gmx_hw_info_t* gmx_detect_hardware(const gmx::MDLogger& mdlog, const PhysicalNodeCommunicator& physicalNodeComm)
+std::unique_ptr<gmx_hw_info_t> gmx_detect_hardware(const PhysicalNodeCommunicator& physicalNodeComm)
{
- // By construction, only one thread ever runs hardware detection,
- // but we may as well prevent issues arising if that would change.
- // Taking the lock early ensures that exactly one thread can
- // attempt to construct g_hardwareInfo.
- lock_guard<Mutex> lock(g_hardwareInfoMutex);
-
- // If we already have the information, just return a handle to it.
- if (g_hardwareInfo != nullptr)
- {
- return g_hardwareInfo.get();
- }
-
- // Make the new hardwareInfo in a temporary.
+ // Ensure all cores have spun up, where applicable.
hardwareTopologyPrepareDetection();
// TODO: We should also do CPU hardware detection only once on each
std::make_unique<CpuInfo>(CpuInfo::detect()),
std::make_unique<HardwareTopology>(HardwareTopology::detect()));
- // If we detected the topology on this system, double-check that it makes sense
- if (hardwareInfo->hardwareTopology->isThisSystem())
- {
- hardwareTopologyDoubleCheckDetection(mdlog, *hardwareInfo->hardwareTopology);
- }
-
// TODO: Get rid of this altogether.
hardwareInfo->nthreads_hw_avail = hardwareInfo->hardwareTopology->machine().logicalProcessorCount;
// Detect GPUs
- hardwareInfo->gpu_info.n_dev = 0;
- hardwareInfo->gpu_info.n_dev_compatible = 0;
- hardwareInfo->gpu_info.gpu_dev = nullptr;
-
- gmx_detect_gpus(mdlog, physicalNodeComm, compat::make_not_null(hardwareInfo));
- gmx_collect_hardware_mpi(*hardwareInfo->cpuInfo, physicalNodeComm, compat::make_not_null(hardwareInfo));
+ // Open a nested scope so no temporary variables can
+ // be mis-used later.
+ {
+ DeviceDetectionResult deviceDetectionResult = detectAllDeviceInformation(physicalNodeComm);
+ hardwareInfo->deviceInfoList.swap(deviceDetectionResult.deviceInfoList_);
+ std::swap(hardwareInfo->hardwareDetectionWarnings_, deviceDetectionResult.deviceDetectionWarnings_);
+ }
- // Now that the temporary is fully constructed, swap it to become
- // the real thing.
- g_hardwareInfo.swap(hardwareInfo);
+ gmx_collect_hardware_mpi(*hardwareInfo->cpuInfo, physicalNodeComm, hardwareInfo.get());
- return g_hardwareInfo.get();
+ return hardwareInfo;
}
-bool compatibleGpusFound(const gmx_gpu_info_t& gpu_info)
+void logHardwareDetectionWarnings(const gmx::MDLogger& mdlog, const gmx_hw_info_t& hardwareInformation)
{
- return gpu_info.n_dev_compatible > 0;
+ for (const std::string& warningString : hardwareInformation.hardwareDetectionWarnings_)
+ {
+ GMX_LOG(mdlog.warning).asParagraph().appendText(warningString);
+ }
}
} // namespace gmx