wrap_cl.hpp

#endif
        else
          throw error("Context", CL_INVALID_VALUE, "invalid context property");
      }
      props.push_back(0);
    }

    return props;
  }


  inline
  context *create_context_inner(py::object py_devices, py::object py_properties,
      py::object py_dev_type)
  {
    std::vector<cl_context_properties> props
      = parse_context_properties(py_properties);

    cl_context_properties *props_ptr
      = props.empty( ) ? NULL : &props.front();

    cl_int status_code;

    cl_context ctx;

    // from device list
    if (py_devices.ptr() != Py_None)
    {
      if (py_dev_type.ptr() != Py_None)
        throw error("Context", CL_INVALID_VALUE,
            "one of 'devices' or 'dev_type' must be None");

      std::vector<cl_device_id> devices;
      PYTHON_FOREACH(py_dev, py_devices)
        devices.push_back(py_dev.cast<const device &>().data());

      PYOPENCL_PRINT_CALL_TRACE("clCreateContext");
      ctx = clCreateContext(
          props_ptr,
          devices.size(),
          devices.empty( ) ? NULL : &devices.front(),
          0, 0, &status_code);
    }
    // from dev_type
    else
    {
      cl_device_type dev_type = CL_DEVICE_TYPE_DEFAULT;
      if (py_dev_type.ptr() != Py_None)
        dev_type = py_dev_type.cast<cl_device_type>();

      PYOPENCL_PRINT_CALL_TRACE("clCreateContextFromType");
      ctx = clCreateContextFromType(props_ptr, dev_type, 0, 0, &status_code);
    }

    if (status_code != CL_SUCCESS)
      throw pyopencl::error("Context", status_code);

    try
    {
      return new context(ctx, false);
    }
    catch (...)
    {
      PYOPENCL_CALL_GUARDED(clReleaseContext, (ctx));
      throw;
    }
  }


  inline
  context *create_context(py::object py_devices, py::object py_properties,
      py::object py_dev_type)
  {
    PYOPENCL_RETRY_RETURN_IF_MEM_ERROR(
      return create_context_inner(py_devices, py_properties, py_dev_type);
    )
  }

  // }}}


  // {{{ command_queue

  class command_queue
  {
    private:
      cl_command_queue m_queue;

    public:
      command_queue(cl_command_queue q, bool retain)
        : m_queue(q)
      {
        if (retain)
          PYOPENCL_CALL_GUARDED(clRetainCommandQueue, (q));
      }

      command_queue(command_queue const &src)
        : m_queue(src.m_queue)
      {
        PYOPENCL_CALL_GUARDED(clRetainCommandQueue, (m_queue));
      }

      command_queue(
          const context &ctx,
          const device *py_dev=0,
          cl_command_queue_properties props=0)
      {
        cl_device_id dev;
        if (py_dev)
          dev = py_dev->data();
        else
        {
          std::vector<cl_device_id> devs;
          PYOPENCL_GET_VEC_INFO(Context, ctx.data(), CL_CONTEXT_DEVICES, devs);
          if (devs.size() == 0)
            throw pyopencl::error("CommandQueue", CL_INVALID_VALUE,
                "context doesn't have any devices? -- don't know which one to default to");
          dev = devs[0];
        }

        cl_int status_code;
        PYOPENCL_PRINT_CALL_TRACE("clCreateCommandQueue");
        m_queue = clCreateCommandQueue(
            ctx.data(), dev, props, &status_code);

        if (status_code != CL_SUCCESS)
          throw pyopencl::error("CommandQueue", status_code);
      }

      ~command_queue()
      {
        PYOPENCL_CALL_GUARDED_CLEANUP(clReleaseCommandQueue,
            (m_queue));
      }

      const cl_command_queue data() const
      { return m_queue; }

      PYOPENCL_EQUALITY_TESTS(command_queue);

      py::object get_info(cl_command_queue_info param_name) const
      {
        switch (param_name)
        {
          case CL_QUEUE_CONTEXT:
            PYOPENCL_GET_OPAQUE_INFO(CommandQueue, m_queue, param_name,
                cl_context, context);
          case CL_QUEUE_DEVICE:
            PYOPENCL_GET_OPAQUE_INFO(CommandQueue, m_queue, param_name,
                cl_device_id, device);
          case CL_QUEUE_REFERENCE_COUNT:
            PYOPENCL_GET_INTEGRAL_INFO(CommandQueue, m_queue, param_name,
                cl_uint);
          case CL_QUEUE_PROPERTIES:
            PYOPENCL_GET_INTEGRAL_INFO(CommandQueue, m_queue, param_name,
                cl_command_queue_properties);

          default:
            throw error("CommandQueue.get_info", CL_INVALID_VALUE);
        }
      }

      std::unique_ptr<context> get_context() const
      {
        cl_context param_value;
        PYOPENCL_CALL_GUARDED(clGetCommandQueueInfo,
            (m_queue, CL_QUEUE_CONTEXT, sizeof(param_value), &param_value, 0));
        return std::unique_ptr<context>(
            new context(param_value, /*retain*/ true));
      }

#if PYOPENCL_CL_VERSION < 0x1010
      cl_command_queue_properties set_property(
          cl_command_queue_properties prop,
          bool enable)
      {
        cl_command_queue_properties old_prop;
        PYOPENCL_CALL_GUARDED(clSetCommandQueueProperty,
            (m_queue, prop, PYOPENCL_CAST_BOOL(enable), &old_prop));
        return old_prop;
      }
#endif

      void flush()
      { PYOPENCL_CALL_GUARDED(clFlush, (m_queue)); }
      void finish()
      { PYOPENCL_CALL_GUARDED_THREADED(clFinish, (m_queue)); }
  };

  // }}}


  // {{{ event/synchronization

  class event : noncopyable
  {
    private:
      cl_event m_event;

    public:
      event(cl_event event, bool retain)
        : m_event(event)
      {
        if (retain)
          PYOPENCL_CALL_GUARDED(clRetainEvent, (event));
      }

      event(event const &src)
        : m_event(src.m_event)
      { PYOPENCL_CALL_GUARDED(clRetainEvent, (m_event)); }

      virtual ~event()
      {
        PYOPENCL_CALL_GUARDED_CLEANUP(clReleaseEvent,
            (m_event));
      }

      const cl_event data() const
      { return m_event; }

      PYOPENCL_EQUALITY_TESTS(event);

      py::object get_info(cl_event_info param_name) const
      {
        switch (param_name)
        {
          case CL_EVENT_COMMAND_QUEUE:
            PYOPENCL_GET_OPAQUE_INFO(Event, m_event, param_name,
                cl_command_queue, command_queue);
          case CL_EVENT_COMMAND_TYPE:
            PYOPENCL_GET_INTEGRAL_INFO(Event, m_event, param_name,
                cl_command_type);
          case CL_EVENT_COMMAND_EXECUTION_STATUS:
            PYOPENCL_GET_INTEGRAL_INFO(Event, m_event, param_name,
                cl_int);
          case CL_EVENT_REFERENCE_COUNT:
            PYOPENCL_GET_INTEGRAL_INFO(Event, m_event, param_name,
                cl_uint);
#if PYOPENCL_CL_VERSION >= 0x1010
          case CL_EVENT_CONTEXT:
            PYOPENCL_GET_OPAQUE_INFO(Event, m_event, param_name,
                cl_context, context);
#endif

          default:
            throw error("Event.get_info", CL_INVALID_VALUE);
        }
      }

      py::object get_profiling_info(cl_profiling_info param_name) const
      {
        switch (param_name)
        {
          case CL_PROFILING_COMMAND_QUEUED:
          case CL_PROFILING_COMMAND_SUBMIT:
          case CL_PROFILING_COMMAND_START:
          case CL_PROFILING_COMMAND_END:
            PYOPENCL_GET_INTEGRAL_INFO(EventProfiling, m_event, param_name,
                cl_ulong);
          default:
            throw error("Event.get_profiling_info", CL_INVALID_VALUE);
        }
      }

      virtual void wait()
      {
        PYOPENCL_CALL_GUARDED_THREADED(clWaitForEvents, (1, &m_event));
      }
  };

#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
  class nanny_event : public event
  {
    // In addition to everything an event does, the nanny event holds a reference
    // to a Python object and waits for its own completion upon destruction.

    protected:
      std::unique_ptr<py_buffer_wrapper> m_ward;

    public:

      nanny_event(cl_event evt, bool retain, std::unique_ptr<py_buffer_wrapper> &ward)
        : event(evt, retain), m_ward(std::move(ward))
      { }

      ~nanny_event()
      { wait(); }

      py::object get_ward() const
      {
        if (m_ward.get())
        {
          return py::reinterpret_borrow<py::object>(m_ward->m_buf.obj);
        }
        else
          return py::none();
      }

      virtual void wait()
      {
        event::wait();
        m_ward.reset();
      }
  };
#else
  class nanny_event : public event
  {
    // In addition to everything an event does, the nanny event holds a reference
    // to a Python object and waits for its own completion upon destruction.

    protected:
      py::object        m_ward;

    public:

      nanny_event(cl_event evt, bool retain, py::object ward)
        : event(evt, retain), m_ward(ward)
      { }

      nanny_event(nanny_event const &src)
        : event(src), m_ward(src.m_ward)
      { }

      ~nanny_event()
      { wait(); }

      py::object get_ward() const
      { return m_ward; }

      virtual void wait()
      {
        event::wait();
        m_ward = py::none();
      }
  };
#endif




  inline
  void wait_for_events(py::object events)
  {
    cl_uint num_events_in_wait_list = 0;
    std::vector<cl_event> event_wait_list(len(events));

    PYTHON_FOREACH(evt, events)
      event_wait_list[num_events_in_wait_list++] =
        evt.cast<event &>().data();

    PYOPENCL_CALL_GUARDED_THREADED(clWaitForEvents, (
          PYOPENCL_WAITLIST_ARGS));
  }




#if PYOPENCL_CL_VERSION >= 0x1020
  inline
  event *enqueue_marker_with_wait_list(command_queue &cq,
      py::object py_wait_for)
  {
    PYOPENCL_PARSE_WAIT_FOR;
    cl_event evt;

    PYOPENCL_CALL_GUARDED(clEnqueueMarkerWithWaitList, (
          cq.data(), PYOPENCL_WAITLIST_ARGS, &evt));

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }

  inline
  event *enqueue_barrier_with_wait_list(command_queue &cq,
      py::object py_wait_for)
  {
    PYOPENCL_PARSE_WAIT_FOR;
    cl_event evt;

    PYOPENCL_CALL_GUARDED(clEnqueueBarrierWithWaitList,
        (cq.data(), PYOPENCL_WAITLIST_ARGS, &evt));

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }
#endif


  // {{{ used internally for pre-OpenCL-1.2 contexts

  inline
  event *enqueue_marker(command_queue &cq)
  {
    cl_event evt;

    PYOPENCL_CALL_GUARDED(clEnqueueMarker, (
          cq.data(), &evt));

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }

  inline
  void enqueue_wait_for_events(command_queue &cq, py::object py_events)
  {
    cl_uint num_events = 0;
    std::vector<cl_event> event_list(len(py_events));

    PYTHON_FOREACH(py_evt, py_events)
      event_list[num_events++] = py_evt.cast<event &>().data();

    PYOPENCL_CALL_GUARDED(clEnqueueWaitForEvents, (
          cq.data(), num_events, event_list.empty( ) ? NULL : &event_list.front()));
  }

  inline
  void enqueue_barrier(command_queue &cq)
  {
    PYOPENCL_CALL_GUARDED(clEnqueueBarrier, (cq.data()));
  }

  // }}}


#if PYOPENCL_CL_VERSION >= 0x1010
  class user_event : public event
  {
    public:
      user_event(cl_event evt, bool retain)
        : event(evt, retain)
      { }

      void set_status(cl_int execution_status)
      {
        PYOPENCL_CALL_GUARDED(clSetUserEventStatus, (data(), execution_status));
      }
  };




  inline
  user_event *create_user_event(context &ctx)
  {
    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clCreateUserEvent");
    cl_event evt = clCreateUserEvent(ctx.data(), &status_code);

    if (status_code != CL_SUCCESS)
      throw pyopencl::error("UserEvent", status_code);

    try
    {
      return new user_event(evt, false);
    }
    catch (...)
    {
      clReleaseEvent(evt);
      throw;
    }
  }

#endif

  // }}}


  // {{{ memory_object

  py::object create_mem_object_wrapper(cl_mem mem);

  class memory_object_holder
  {
    public:
      virtual const cl_mem data() const = 0;

      PYOPENCL_EQUALITY_TESTS(memory_object_holder);

      size_t size() const
      {
        size_t param_value;
        PYOPENCL_CALL_GUARDED(clGetMemObjectInfo,
            (data(), CL_MEM_SIZE, sizeof(param_value), &param_value, 0));
        return param_value;
      }

      py::object get_info(cl_mem_info param_name) const;
  };




  class memory_object : noncopyable, public memory_object_holder
  {
    public:
#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
      typedef std::unique_ptr<py_buffer_wrapper> hostbuf_t;
#else
      typedef py::object hostbuf_t;
#endif

    private:
      bool m_valid;
      cl_mem m_mem;
      hostbuf_t m_hostbuf;

    public:
      memory_object(cl_mem mem, bool retain, hostbuf_t hostbuf=hostbuf_t())
        : m_valid(true), m_mem(mem)
      {
        if (retain)
          PYOPENCL_CALL_GUARDED(clRetainMemObject, (mem));

        m_hostbuf = PYOPENCL_STD_MOVE_IF_NEW_BUF_INTF(hostbuf);
      }

      memory_object(memory_object &src)
        : m_valid(true), m_mem(src.m_mem),
        m_hostbuf(PYOPENCL_STD_MOVE_IF_NEW_BUF_INTF(src.m_hostbuf))
      {
        PYOPENCL_CALL_GUARDED(clRetainMemObject, (m_mem));
      }

      memory_object(memory_object_holder const &src)
        : m_valid(true), m_mem(src.data())
      {
        PYOPENCL_CALL_GUARDED(clRetainMemObject, (m_mem));
      }

      void release()
      {
        if (!m_valid)
            throw error("MemoryObject.free", CL_INVALID_VALUE,
                "trying to double-unref mem object");
        PYOPENCL_CALL_GUARDED_CLEANUP(clReleaseMemObject, (m_mem));
        m_valid = false;
      }

      virtual ~memory_object()
      {
        if (m_valid)
          release();
      }

      py::object hostbuf()
      {
#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
        if (m_hostbuf.get())
          return py::reinterpret_borrow<py::object>(m_hostbuf->m_buf.obj);
        else
          return py::none();
#else
        return m_hostbuf;
#endif
      }

      const cl_mem data() const
      { return m_mem; }

  };

#if PYOPENCL_CL_VERSION >= 0x1020
  inline
  event *enqueue_migrate_mem_objects(
      command_queue &cq,
      py::object py_mem_objects,
      cl_mem_migration_flags flags,
      py::object py_wait_for)
  {
    PYOPENCL_PARSE_WAIT_FOR;

    std::vector<cl_mem> mem_objects;
    PYTHON_FOREACH(mo, py_mem_objects)
      mem_objects.push_back(mo.cast<const memory_object &>().data());

    cl_event evt;
    PYOPENCL_RETRY_IF_MEM_ERROR(
      PYOPENCL_CALL_GUARDED(clEnqueueMigrateMemObjects, (
            cq.data(),
            mem_objects.size(), mem_objects.empty( ) ? NULL : &mem_objects.front(),
            flags,
            PYOPENCL_WAITLIST_ARGS, &evt
            ));
      );
    PYOPENCL_RETURN_NEW_EVENT(evt);
  }
#endif

#ifdef cl_ext_migrate_memobject
  inline
  event *enqueue_migrate_mem_object_ext(
      command_queue &cq,
      py::object py_mem_objects,
      cl_mem_migration_flags_ext flags,
      py::object py_wait_for)
  {
    PYOPENCL_PARSE_WAIT_FOR;

#if PYOPENCL_CL_VERSION >= 0x1020
    // {{{ get platform
    cl_device_id dev;
    PYOPENCL_CALL_GUARDED(clGetCommandQueueInfo, (cq.data(), CL_QUEUE_DEVICE,
          sizeof(dev), &dev, NULL));
    cl_platform_id plat;
    PYOPENCL_CALL_GUARDED(clGetDeviceInfo, (dev, CL_DEVICE_PLATFORM,
          sizeof(plat), &plat, NULL));
    // }}}
#endif

    PYOPENCL_GET_EXT_FUN(plat,
        clEnqueueMigrateMemObjectEXT, enqueue_migrate_fn);

    std::vector<cl_mem> mem_objects;
    PYTHON_FOREACH(mo, py_mem_objects)
      mem_objects.push_back(mo.cast<memory_object &>().data());

    cl_event evt;
    PYOPENCL_RETRY_IF_MEM_ERROR(
      PYOPENCL_CALL_GUARDED(enqueue_migrate_fn, (
            cq.data(),
            mem_objects.size(), mem_objects.empty( ) ? NULL : &mem_objects.front(),
            flags,
            PYOPENCL_WAITLIST_ARGS, &evt
            ));
      );
    PYOPENCL_RETURN_NEW_EVENT(evt);
  }
#endif

  // }}}


  // {{{ buffer

  inline cl_mem create_buffer(
      cl_context ctx,
      cl_mem_flags flags,
      size_t size,
      void *host_ptr)
  {
    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clCreateBuffer");
    cl_mem mem = clCreateBuffer(ctx, flags, size, host_ptr, &status_code);

    if (status_code != CL_SUCCESS)
      throw pyopencl::error("create_buffer", status_code);

    return mem;
  }




  inline cl_mem create_buffer_gc(
      cl_context ctx,
      cl_mem_flags flags,
      size_t size,
      void *host_ptr)
  {
    PYOPENCL_RETRY_RETURN_IF_MEM_ERROR(
      return create_buffer(ctx, flags, size, host_ptr);
    );
  }



#if PYOPENCL_CL_VERSION >= 0x1010
  inline cl_mem create_sub_buffer(
      cl_mem buffer, cl_mem_flags flags, cl_buffer_create_type bct,
      const void *buffer_create_info)
  {
    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clCreateSubBuffer");
    cl_mem mem = clCreateSubBuffer(buffer, flags,
        bct, buffer_create_info, &status_code);

    if (status_code != CL_SUCCESS)
      throw pyopencl::error("clCreateSubBuffer", status_code);

    return mem;
  }




  inline cl_mem create_sub_buffer_gc(
      cl_mem buffer, cl_mem_flags flags, cl_buffer_create_type bct,
      const void *buffer_create_info)
  {
    PYOPENCL_RETRY_RETURN_IF_MEM_ERROR(
      return create_sub_buffer(buffer, flags, bct, buffer_create_info);
    );
  }
#endif



  class buffer : public memory_object
  {
    public:
      buffer(cl_mem mem, bool retain, hostbuf_t hostbuf=hostbuf_t())
        : memory_object(mem, retain, PYOPENCL_STD_MOVE_IF_NEW_BUF_INTF(hostbuf))
      { }

#if PYOPENCL_CL_VERSION >= 0x1010
      buffer *get_sub_region(
          size_t origin, size_t size, cl_mem_flags flags) const
      {
        cl_buffer_region region = { origin, size};

        cl_mem mem = create_sub_buffer_gc(
            data(), flags, CL_BUFFER_CREATE_TYPE_REGION, &region);

        try
        {
          return new buffer(mem, false);
        }
        catch (...)
        {
          PYOPENCL_CALL_GUARDED(clReleaseMemObject, (mem));
          throw;
        }
      }

      buffer *getitem(py::slice slc) const
      {
        PYOPENCL_BUFFER_SIZE_T start, end, stride, length;

        size_t my_length;
        PYOPENCL_CALL_GUARDED(clGetMemObjectInfo,
            (data(), CL_MEM_SIZE, sizeof(my_length), &my_length, 0));

#if PY_VERSION_HEX >= 0x03020000
        if (PySlice_GetIndicesEx(slc.ptr(),
#else
        if (PySlice_GetIndicesEx(reinterpret_cast<PySliceObject *>(slc.ptr()),
#endif
              my_length, &start, &end, &stride, &length) != 0)
          throw py::error_already_set();

        if (stride != 1)
          throw pyopencl::error("Buffer.__getitem__", CL_INVALID_VALUE,
              "Buffer slice must have stride 1");

        cl_mem_flags my_flags;
        PYOPENCL_CALL_GUARDED(clGetMemObjectInfo,
            (data(), CL_MEM_FLAGS, sizeof(my_flags), &my_flags, 0));

        my_flags &= ~CL_MEM_COPY_HOST_PTR;

        if (end <= start)
          throw pyopencl::error("Buffer.__getitem__", CL_INVALID_VALUE,
              "Buffer slice have end > start");

        return get_sub_region(start, end-start, my_flags);
      }
#endif
  };

  // {{{ buffer creation

  inline
  buffer *create_buffer_py(
      context &ctx,
      cl_mem_flags flags,
      size_t size,
      py::object py_hostbuf
      )
  {
    if (py_hostbuf.ptr() != Py_None &&
        !(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)))
      PyErr_Warn(PyExc_UserWarning, "'hostbuf' was passed, "
          "but no memory flags to make use of it.");

    void *buf = 0;

#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
    std::unique_ptr<py_buffer_wrapper> retained_buf_obj;
    if (py_hostbuf.ptr() != Py_None)
    {
      retained_buf_obj = std::unique_ptr<py_buffer_wrapper>(new py_buffer_wrapper);

      int py_buf_flags = PyBUF_ANY_CONTIGUOUS;
      if ((flags & CL_MEM_USE_HOST_PTR)
          && ((flags & CL_MEM_READ_WRITE)
            || (flags & CL_MEM_WRITE_ONLY)))
        py_buf_flags |= PyBUF_WRITABLE;

      retained_buf_obj->get(py_hostbuf.ptr(), py_buf_flags);

      buf = retained_buf_obj->m_buf.buf;

      if (size > size_t(retained_buf_obj->m_buf.len))
        throw pyopencl::error("Buffer", CL_INVALID_VALUE,
            "specified size is greater than host buffer size");
      if (size == 0)
        size = retained_buf_obj->m_buf.len;
    }
#else
    py::object retained_buf_obj;
    if (py_hostbuf.ptr() != Py_None)
    {
      PYOPENCL_BUFFER_SIZE_T len;
      if ((flags & CL_MEM_USE_HOST_PTR)
          && ((flags & CL_MEM_READ_WRITE)
            || (flags & CL_MEM_WRITE_ONLY)))
      {
        if (PyObject_AsWriteBuffer(py_hostbuf.ptr(), &buf, &len))
          throw py::error_already_set();
      }
      else
      {
        if (PyObject_AsReadBuffer(
              py_hostbuf.ptr(), const_cast<const void **>(&buf), &len))
          throw py::error_already_set();
      }

      if (flags & CL_MEM_USE_HOST_PTR)
        retained_buf_obj = py_hostbuf;

      if (size > size_t(len))
        throw pyopencl::error("Buffer", CL_INVALID_VALUE,
            "specified size is greater than host buffer size");
      if (size == 0)
        size = len;
    }
#endif

    cl_mem mem = create_buffer_gc(ctx.data(), flags, size, buf);

#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
    if (!(flags & CL_MEM_USE_HOST_PTR))
      retained_buf_obj.reset();
#endif

    try
    {
      return new buffer(mem, false, PYOPENCL_STD_MOVE_IF_NEW_BUF_INTF(retained_buf_obj));
    }
    catch (...)
    {
      PYOPENCL_CALL_GUARDED(clReleaseMemObject, (mem));
      throw;
    }
  }

  // }}}

  // {{{ buffer transfers

  // {{{ byte-for-byte transfers

  inline
  event *enqueue_read_buffer(
      command_queue &cq,
      memory_object_holder &mem,
      py::object buffer,
      size_t device_offset,
      py::object py_wait_for,
      bool is_blocking)
  {
    PYOPENCL_PARSE_WAIT_FOR;

    void *buf;
    PYOPENCL_BUFFER_SIZE_T len;

#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
    std::unique_ptr<py_buffer_wrapper> ward(new py_buffer_wrapper);

    ward->get(buffer.ptr(), PyBUF_ANY_CONTIGUOUS | PyBUF_WRITABLE);

    buf = ward->m_buf.buf;
    len = ward->m_buf.len;
#else
    py::object ward = buffer;
    if (PyObject_AsWriteBuffer(buffer.ptr(), &buf, &len))
      throw py::error_already_set();
#endif

    cl_event evt;
    PYOPENCL_RETRY_IF_MEM_ERROR(
      PYOPENCL_CALL_GUARDED_THREADED(clEnqueueReadBuffer, (
            cq.data(),
            mem.data(),
            PYOPENCL_CAST_BOOL(is_blocking),
            device_offset, len, buf,
            PYOPENCL_WAITLIST_ARGS, &evt
            ))
      );
    PYOPENCL_RETURN_NEW_NANNY_EVENT(evt, ward);
  }




  inline
  event *enqueue_write_buffer(
      command_queue &cq,
      memory_object_holder &mem,
      py::object buffer,
      size_t device_offset,
      py::object py_wait_for,
      bool is_blocking)
  {
    PYOPENCL_PARSE_WAIT_FOR;

    const void *buf;
    PYOPENCL_BUFFER_SIZE_T len;

#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
    std::unique_ptr<py_buffer_wrapper> ward(new py_buffer_wrapper);

    ward->get(buffer.ptr(), PyBUF_ANY_CONTIGUOUS);

    buf = ward->m_buf.buf;
    len = ward->m_buf.len;
#else
    py::object ward = buffer;
    if (PyObject_AsReadBuffer(buffer.ptr(), &buf, &len))
      throw py::error_already_set();
#endif

    cl_event evt;
    PYOPENCL_RETRY_IF_MEM_ERROR(
      PYOPENCL_CALL_GUARDED_THREADED(clEnqueueWriteBuffer, (
            cq.data(),
            mem.data(),
            PYOPENCL_CAST_BOOL(is_blocking),
            device_offset, len, buf,
            PYOPENCL_WAITLIST_ARGS, &evt
            ))
      );
    PYOPENCL_RETURN_NEW_NANNY_EVENT(evt, ward);
  }




  inline
  event *enqueue_copy_buffer(
      command_queue &cq,
      memory_object_holder &src,
      memory_object_holder &dst,
      ptrdiff_t byte_count,
      size_t src_offset,
      size_t dst_offset,
      py::object py_wait_for)
  {
    PYOPENCL_PARSE_WAIT_FOR;

    if (byte_count < 0)
    {
      size_t byte_count_src = 0;
      size_t byte_count_dst = 0;
      PYOPENCL_CALL_GUARDED(clGetMemObjectInfo,
          (src.data(), CL_MEM_SIZE, sizeof(byte_count), &byte_count_src, 0));
      PYOPENCL_CALL_GUARDED(clGetMemObjectInfo,
          (src.data(), CL_MEM_SIZE, sizeof(byte_count), &byte_count_dst, 0));
      byte_count = std::min(byte_count_src, byte_count_dst);
    }

    cl_event evt;
    PYOPENCL_RETRY_IF_MEM_ERROR(
      PYOPENCL_CALL_GUARDED(clEnqueueCopyBuffer, (
            cq.data(),
            src.data(), dst.data(),
            src_offset, dst_offset,
            byte_count,
            PYOPENCL_WAITLIST_ARGS,
            &evt
            ))
      );

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }

  // }}}

  // {{{ rectangular transfers
#if PYOPENCL_CL_VERSION >= 0x1010
  inline
  event *enqueue_read_buffer_rect(
      command_queue &cq,
      memory_object_holder &mem,
      py::object buffer,
      py::object py_buffer_origin,
      py::object py_host_origin,
      py::object py_region,
      py::sequence py_buffer_pitches,
      py::sequence py_host_pitches,
      py::object py_wait_for,
      bool is_blocking
      )
  {
    PYOPENCL_PARSE_WAIT_FOR;
    COPY_PY_COORD_TRIPLE(buffer_origin);
    COPY_PY_COORD_TRIPLE(host_origin);
    COPY_PY_REGION_TRIPLE(region);
    COPY_PY_PITCH_TUPLE(buffer_pitches);
    COPY_PY_PITCH_TUPLE(host_pitches);

    void *buf;