cuda.hpp

// A C++ wrapper for CUDA




#ifndef _AFJDFJSDFSD_PYCUDA_HEADER_SEEN_CUDA_HPP
#define _AFJDFJSDFSD_PYCUDA_HEADER_SEEN_CUDA_HPP




#include <cuda.h>
#include <stdexcept>
#include <boost/shared_ptr.hpp>
#include <boost/foreach.hpp>
#include <utility>
#include <stack>
#include <iostream>
#include <vector>
#include <boost/python.hpp>
#include <boost/thread/thread.hpp>
#include <boost/thread/tss.hpp>
#include <boost/version.hpp>

#if (BOOST_VERSION/100) < 1035
#warning *****************************************************************
#warning **** Your version of Boost C++ is likely too old for PyCUDA. ****
#warning *****************************************************************
#endif




// #define CUDAPP_TRACE_CUDA
#define CUDAPP_POST_30_BETA




#ifdef CUDAPP_TRACE_CUDA
  #define CUDAPP_PRINT_CALL_TRACE(NAME) \
    std::cerr << NAME << std::endl;
  #define CUDAPP_PRINT_CALL_TRACE_INFO(NAME, EXTRA_INFO) \
    std::cerr << NAME << " (" << EXTRA_INFO << ')' << std::endl;
#else
  #define CUDAPP_PRINT_CALL_TRACE(NAME) /*nothing*/
  #define CUDAPP_PRINT_CALL_TRACE_INFO(NAME, EXTRA_INFO) /*nothing*/
#endif

#define CUDAPP_CALL_GUARDED_THREADED_WITH_TRACE_INFO(NAME, ARGLIST, TRACE_INFO) \
  { \
    CUDAPP_PRINT_CALL_TRACE_INFO(#NAME, TRACE_INFO); \
    CUresult cu_status_code; \
    Py_BEGIN_ALLOW_THREADS \
      cu_status_code = NAME ARGLIST; \
    Py_END_ALLOW_THREADS \
    if (cu_status_code != CUDA_SUCCESS) \
      throw cuda::error(#NAME, cu_status_code);\
  }

#define CUDAPP_CALL_GUARDED_WITH_TRACE_INFO(NAME, ARGLIST, TRACE_INFO) \
  { \
    CUDAPP_PRINT_CALL_TRACE_INFO(#NAME, TRACE_INFO); \
    CUresult cu_status_code; \
    cu_status_code = NAME ARGLIST; \
    if (cu_status_code != CUDA_SUCCESS) \
      throw cuda::error(#NAME, cu_status_code);\
  }

#define CUDAPP_CALL_GUARDED_THREADED(NAME, ARGLIST) \
  { \
    CUDAPP_PRINT_CALL_TRACE(#NAME); \
    CUresult cu_status_code; \
    Py_BEGIN_ALLOW_THREADS \
      cu_status_code = NAME ARGLIST; \
    Py_END_ALLOW_THREADS \
    if (cu_status_code != CUDA_SUCCESS) \
      throw cuda::error(#NAME, cu_status_code);\
  }

#define CUDAPP_CALL_GUARDED(NAME, ARGLIST) \
  { \
    CUDAPP_PRINT_CALL_TRACE(#NAME); \
    CUresult cu_status_code; \
    cu_status_code = NAME ARGLIST; \
    if (cu_status_code != CUDA_SUCCESS) \
      throw cuda::error(#NAME, cu_status_code);\
  }
#define CUDAPP_CALL_GUARDED_CLEANUP(NAME, ARGLIST) \
  { \
    CUDAPP_PRINT_CALL_TRACE(#NAME); \
    CUresult cu_status_code; \
    cu_status_code = NAME ARGLIST; \
    if (cu_status_code != CUDA_SUCCESS) \
      std::cerr \
        << "PyCUDA WARNING: a clean-up operation failed (dead context maybe?)" \
        << std::endl \
        << cuda::error::make_message(#NAME, cu_status_code) \
        << std::endl; \
  }
#define CUDAPP_CATCH_CLEANUP_ON_DEAD_CONTEXT(TYPE) \
  catch (cuda::cannot_activate_out_of_thread_context) \
  { } \
  catch (cuda::cannot_activate_dead_context) \
  { \
    /* PyErr_Warn( \
        PyExc_UserWarning, #TYPE " in dead context was implicitly cleaned up");*/ \
  }
  // In all likelihood, this TYPE's managing thread has exited, and
  // therefore its context has already been deleted. No need to harp
  // on the fact that we still thought there was cleanup to do.




namespace cuda
{
  namespace py = boost::python;




  class error : public std::runtime_error
  {
    private:
      const char *m_routine;
      CUresult m_code;

    public:
      static std::string make_message(const char *rout, CUresult c, const char *msg=0)
      {
        std::string result = rout;
        result += " failed: ";
        result += curesult_to_str(c);
        if (msg)
        {
          result += " - ";
          result += msg;
        }
        return result;
      }

      error(const char *rout, CUresult c, const char *msg=0)
        : std::runtime_error(make_message(rout, c, msg)),
        m_routine(rout), m_code(c)
      { }

      const char *routine() const
      {
        return m_routine;
      }

      CUresult code() const
      {
        return m_code;
      }

      static const char *curesult_to_str(CUresult e)
      {
        switch (e)
        {
          case CUDA_SUCCESS: return "success";
          case CUDA_ERROR_INVALID_VALUE: return "invalid value";
          case CUDA_ERROR_OUT_OF_MEMORY: return "out of memory";
          case CUDA_ERROR_NOT_INITIALIZED: return "not initialized";
#if CUDA_VERSION >= 2000
          case CUDA_ERROR_DEINITIALIZED: return "deinitialized";
#endif

          case CUDA_ERROR_NO_DEVICE: return "no device";
          case CUDA_ERROR_INVALID_DEVICE: return "invalid device";

          case CUDA_ERROR_INVALID_IMAGE: return "invalid image";
          case CUDA_ERROR_INVALID_CONTEXT: return "invalid context";
          case CUDA_ERROR_CONTEXT_ALREADY_CURRENT: return "context already current";
          case CUDA_ERROR_MAP_FAILED: return "map failed";
          case CUDA_ERROR_UNMAP_FAILED: return "unmap failed";
          case CUDA_ERROR_ARRAY_IS_MAPPED: return "array is mapped";
          case CUDA_ERROR_ALREADY_MAPPED: return "already mapped";
          case CUDA_ERROR_NO_BINARY_FOR_GPU: return "no binary for gpu";
          case CUDA_ERROR_ALREADY_ACQUIRED: return "already acquired";
          case CUDA_ERROR_NOT_MAPPED: return "not mapped";
#if CUDA_VERSION >= 3000
          case CUDA_ERROR_NOT_MAPPED_AS_ARRAY: return "not mapped as array";
          case CUDA_ERROR_NOT_MAPPED_AS_POINTER: return "not mapped as pointer";
#ifdef CUDAPP_POST_30_BETA
          case CUDA_ERROR_ECC_UNCORRECTABLE: return "ECC uncorrectable";
#endif
#endif
#if CUDA_VERSION >= 3010
          case CUDA_ERROR_UNSUPPORTED_LIMIT: return "unsupported limit";
#endif

          case CUDA_ERROR_INVALID_SOURCE: return "invalid source";
          case CUDA_ERROR_FILE_NOT_FOUND: return "file not found";
#if CUDA_VERSION >= 3010
          case CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND:
            return "shared object symbol not found";
          case CUDA_ERROR_SHARED_OBJECT_INIT_FAILED:
            return "shared object init failed";