CUDAService.cc File Reference

#include <iomanip>
#include <iostream>
#include <limits>
#include <cuda.h>
#include "FWCore/MessageLogger/interface/MessageLogger.h"
#include "FWCore/ParameterSet/interface/ConfigurationDescriptions.h"
#include "FWCore/ParameterSet/interface/ParameterSet.h"
#include "FWCore/ParameterSet/interface/ParameterSetDescription.h"
#include "FWCore/Utilities/interface/ReusableObjectHolder.h"
#include "HeterogeneousCore/CUDAServices/interface/CUDAService.h"
#include "HeterogeneousCore/CUDAUtilities/interface/cudaCheck.h"
#include "HeterogeneousCore/CUDAUtilities/interface/EventCache.h"
#include "HeterogeneousCore/CUDAUtilities/interface/StreamCache.h"
#include "HeterogeneousCore/CUDAUtilities/interface/device_unique_ptr.h"
#include "HeterogeneousCore/CUDAUtilities/interface/host_unique_ptr.h"
#include "HeterogeneousCore/CUDAUtilities/interface/currentDevice.h"
#include "HeterogeneousCore/CUDAUtilities/src/getCachingDeviceAllocator.h"
#include "HeterogeneousCore/CUDAUtilities/src/getCachingHostAllocator.h"

Go to the source code of this file.

Functions
constexpr unsigned int	getCudaCoresPerSM (unsigned int major, unsigned int minor)
void	setCudaLimit (cudaLimit limit, const char *name, size_t request)

Function Documentation

getCudaCoresPerSM()

constexpr unsigned int getCudaCoresPerSM ( unsigned int  major, unsigned int  minor  )

constexpr

Definition at line 44 of file CUDAService.cc.

                                                                                 {
  switch (major * 10 + minor) {
    // Fermi architecture
    case 20:  // SM 2.0: GF100 class
      return 32;
    case 21:  // SM 2.1: GF10x class
      return 48;
 
    // Kepler architecture
    case 30:  // SM 3.0: GK10x class
    case 32:  // SM 3.2: GK10x class
    case 35:  // SM 3.5: GK11x class
    case 37:  // SM 3.7: GK21x class
      return 192;
 
    // Maxwell architecture
    case 50:  // SM 5.0: GM10x class
    case 52:  // SM 5.2: GM20x class
    case 53:  // SM 5.3: GM20x class
      return 128;
 
    // Pascal architecture
    case 60:  // SM 6.0: GP100 class
      return 64;
    case 61:  // SM 6.1: GP10x class
    case 62:  // SM 6.2: GP10x class
      return 128;
 
    // Volta architecture
    case 70:  // SM 7.0: GV100 class
    case 72:  // SM 7.2: GV11b class
      return 64;
 
    // Turing architecture
    case 75:  // SM 7.5: TU10x class
      return 64;
 
    // unknown architecture, return a default value
    default:
      return 64;
  }
}

Referenced by CUDAService::CUDAService().

setCudaLimit()

void setCudaLimit	(	cudaLimit	limit,
		const char *	name,
		size_t	request
	)

Definition at line 22 of file CUDAService.cc.

                                                                     {
  // read the current device
  int device;
  cudaCheck(cudaGetDevice(&device));
  // try to set the requested limit
  auto result = cudaDeviceSetLimit(limit, request);
  if (cudaErrorUnsupportedLimit == result) {
    edm::LogWarning("CUDAService") << "CUDA device " << device << ": unsupported limit \"" << name << "\"";
    return;
  }
  // read back the limit value
  size_t value;
  cudaCheck(cudaDeviceGetLimit(&value, limit));
  if (cudaSuccess != result) {
    edm::LogWarning("CUDAService") << "CUDA device " << device << ": failed to set limit \"" << name << "\" to "
                                   << request << ", current value is " << value;
  } else if (value != request) {
    edm::LogWarning("CUDAService") << "CUDA device " << device << ": limit \"" << name << "\" set to " << value
                                   << " instead of requested " << request;
  }
}

References cudaCheck, remoteMonitoring_LED_IterMethod_cfg::limit, Skims_PA_cff::name, mps_fire::result, and relativeConstraints::value.

Referenced by CUDAService::CUDAService().