wrap_cl.hpp

        mapped = clEnqueueMapImage(
              cq.data(), img.data(),
              PYOPENCL_CAST_BOOL(is_blocking), flags,
              origin, region, &row_pitch, &slice_pitch,
              PYOPENCL_WAITLIST_ARGS, &evt,
              &status_code);
        Py_END_ALLOW_THREADS
        if (status_code != CL_SUCCESS)
          throw pyopencl::error("clEnqueueMapImage", status_code);
      } );

    event evt_handle(evt, false);

    std::unique_ptr<memory_map> map;
    try
    {
       map = std::unique_ptr<memory_map>(new memory_map(cq, img, mapped));
    }
    catch (...)
    {
      PYOPENCL_CALL_GUARDED_CLEANUP(clEnqueueUnmapMemObject, (
            cq.data(), img.data(), mapped, 0, 0, 0));
      throw;
    }

    py::object result = py::reinterpret_steal<py::object>(PyArray_NewFromDescr(
        &PyArray_Type, tp_descr,
        shape.size(),
        shape.empty() ? NULL : &shape.front(),
        strides.empty() ? NULL : &strides.front(),
        mapped, ary_flags, /*obj*/NULL));

    py::object map_py(handle_from_new_ptr(map.release()));
    PyArray_BASE(result.ptr()) = map_py.ptr();
    Py_INCREF(map_py.ptr());

    return py::make_tuple(
        result,
        handle_from_new_ptr(new event(evt_handle)),
        row_pitch, slice_pitch);
  }

  // }}}


  // {{{ sampler

  class sampler : noncopyable
  {
    private:
      cl_sampler m_sampler;

    public:
      sampler(context const &ctx, bool normalized_coordinates,
          cl_addressing_mode am, cl_filter_mode fm)
      {
        cl_int status_code;
        PYOPENCL_PRINT_CALL_TRACE("clCreateSampler");
        m_sampler = clCreateSampler(
            ctx.data(),
            normalized_coordinates,
            am, fm, &status_code);

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

      sampler(cl_sampler samp, bool retain)
        : m_sampler(samp)
      {
        if (retain)
          PYOPENCL_CALL_GUARDED(clRetainSampler, (samp));
      }

      ~sampler()
      {
        PYOPENCL_CALL_GUARDED_CLEANUP(clReleaseSampler, (m_sampler));
      }

      cl_sampler data() const
      {
        return m_sampler;
      }

      PYOPENCL_EQUALITY_TESTS(sampler);

      py::object get_info(cl_sampler_info param_name) const
      {
        switch (param_name)
        {
          case CL_SAMPLER_REFERENCE_COUNT:
            PYOPENCL_GET_INTEGRAL_INFO(Sampler, m_sampler, param_name,
                cl_uint);
          case CL_SAMPLER_CONTEXT:
            PYOPENCL_GET_OPAQUE_INFO(Sampler, m_sampler, param_name,
                cl_context, context);
          case CL_SAMPLER_ADDRESSING_MODE:
            PYOPENCL_GET_INTEGRAL_INFO(Sampler, m_sampler, param_name,
                cl_addressing_mode);
          case CL_SAMPLER_FILTER_MODE:
            PYOPENCL_GET_INTEGRAL_INFO(Sampler, m_sampler, param_name,
                cl_filter_mode);
          case CL_SAMPLER_NORMALIZED_COORDS:
            PYOPENCL_GET_INTEGRAL_INFO(Sampler, m_sampler, param_name,
                cl_bool);

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

  // }}}


  // {{{ program

  class program : noncopyable
  {
    public:
      enum program_kind_type { KND_UNKNOWN, KND_SOURCE, KND_BINARY };

    private:
      cl_program m_program;
      program_kind_type m_program_kind;

    public:
      program(cl_program prog, bool retain, program_kind_type progkind=KND_UNKNOWN)
        : m_program(prog), m_program_kind(progkind)
      {
        if (retain)
          PYOPENCL_CALL_GUARDED(clRetainProgram, (prog));
      }

      ~program()
      {
        PYOPENCL_CALL_GUARDED_CLEANUP(clReleaseProgram, (m_program));
      }

      cl_program data() const
      {
        return m_program;
      }

      program_kind_type kind() const
      {
        return m_program_kind;
      }

      PYOPENCL_EQUALITY_TESTS(program);

      py::object get_info(cl_program_info param_name) const
      {
        switch (param_name)
        {
          case CL_PROGRAM_REFERENCE_COUNT:
            PYOPENCL_GET_INTEGRAL_INFO(Program, m_program, param_name,
                cl_uint);
          case CL_PROGRAM_CONTEXT:
            PYOPENCL_GET_OPAQUE_INFO(Program, m_program, param_name,
                cl_context, context);
          case CL_PROGRAM_NUM_DEVICES:
            PYOPENCL_GET_INTEGRAL_INFO(Program, m_program, param_name,
                cl_uint);
          case CL_PROGRAM_DEVICES:
            {
              std::vector<cl_device_id> result;
              PYOPENCL_GET_VEC_INFO(Program, m_program, param_name, result);

              py::list py_result;
              for (cl_device_id did: result)
                py_result.append(handle_from_new_ptr(
                      new pyopencl::device(did)));
              return py_result;
            }
          case CL_PROGRAM_SOURCE:
            PYOPENCL_GET_STR_INFO(Program, m_program, param_name);
          case CL_PROGRAM_BINARY_SIZES:
            {
              std::vector<size_t> result;
              PYOPENCL_GET_VEC_INFO(Program, m_program, param_name, result);
              PYOPENCL_RETURN_VECTOR(size_t, result);
            }
          case CL_PROGRAM_BINARIES:
            // {{{
            {
              std::vector<size_t> sizes;
              PYOPENCL_GET_VEC_INFO(Program, m_program, CL_PROGRAM_BINARY_SIZES, sizes);

              size_t total_size = std::accumulate(sizes.begin(), sizes.end(), 0);

              std::unique_ptr<unsigned char []> result(
                  new unsigned char[total_size]);
              std::vector<unsigned char *> result_ptrs;

              unsigned char *ptr = result.get();
              for (unsigned i = 0; i < sizes.size(); ++i)
              {
                result_ptrs.push_back(ptr);
                ptr += sizes[i];
              }

              PYOPENCL_CALL_GUARDED(clGetProgramInfo,
                  (m_program, param_name, sizes.size()*sizeof(unsigned char *),
                   result_ptrs.empty( ) ? NULL : &result_ptrs.front(), 0)); \

              py::list py_result;
              ptr = result.get();
              for (unsigned i = 0; i < sizes.size(); ++i)
              {
                py::object binary_pyobj(
                    py::reinterpret_steal<py::object>(
#if PY_VERSION_HEX >= 0x03000000
                    PyBytes_FromStringAndSize(
                      reinterpret_cast<char *>(ptr), sizes[i])
#else
                    PyString_FromStringAndSize(
                      reinterpret_cast<char *>(ptr), sizes[i])
#endif
                    ));
                py_result.append(binary_pyobj);
                ptr += sizes[i];
              }
              return py_result;
            }
            // }}}
#if PYOPENCL_CL_VERSION >= 0x1020
          case CL_PROGRAM_NUM_KERNELS:
            PYOPENCL_GET_INTEGRAL_INFO(Program, m_program, param_name,
                size_t);
          case CL_PROGRAM_KERNEL_NAMES:
            PYOPENCL_GET_STR_INFO(Program, m_program, param_name);
#endif

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

      py::object get_build_info(
          device const &dev,
          cl_program_build_info param_name) const
      {
        switch (param_name)
        {
#define PYOPENCL_FIRST_ARG m_program, dev.data() // hackety hack
          case CL_PROGRAM_BUILD_STATUS:
            PYOPENCL_GET_INTEGRAL_INFO(ProgramBuild,
                PYOPENCL_FIRST_ARG, param_name,
                cl_build_status);
          case CL_PROGRAM_BUILD_OPTIONS:
          case CL_PROGRAM_BUILD_LOG:
            PYOPENCL_GET_STR_INFO(ProgramBuild,
                PYOPENCL_FIRST_ARG, param_name);
#if PYOPENCL_CL_VERSION >= 0x1020
          case CL_PROGRAM_BINARY_TYPE:
            PYOPENCL_GET_INTEGRAL_INFO(ProgramBuild,
                PYOPENCL_FIRST_ARG, param_name,
                cl_program_binary_type);
#endif
#undef PYOPENCL_FIRST_ARG

          default:
            throw error("Program.get_build_info", CL_INVALID_VALUE);
        }
      }

      void build(std::string options, py::object py_devices)
      {
        PYOPENCL_PARSE_PY_DEVICES;

        PYOPENCL_CALL_GUARDED_THREADED(clBuildProgram,
            (m_program, num_devices, devices,
             options.c_str(), 0 ,0));
      }

#if PYOPENCL_CL_VERSION >= 0x1020
      void compile(std::string options, py::object py_devices,
          py::object py_headers)
      {
        PYOPENCL_PARSE_PY_DEVICES;

        // {{{ pick apart py_headers
        // py_headers is a list of tuples *(name, program)*

        std::vector<std::string> header_names;
        std::vector<cl_program> programs;
        PYTHON_FOREACH(name_hdr_tup_py, py_headers)
        {
          py::tuple name_hdr_tup = py::reinterpret_borrow<py::tuple>(name_hdr_tup_py);
          if (py::len(name_hdr_tup) != 2)
            throw error("Program.compile", CL_INVALID_VALUE,
                "epxected (name, header) tuple in headers list");
          std::string name = (name_hdr_tup[0]).cast<std::string>();
          program &prg = (name_hdr_tup[1]).cast<program &>();

          header_names.push_back(name);
          programs.push_back(prg.data());
        }

        std::vector<const char *> header_name_ptrs;
        for (std::string const &name: header_names)
          header_name_ptrs.push_back(name.c_str());

        // }}}

        PYOPENCL_CALL_GUARDED_THREADED(clCompileProgram,
            (m_program, num_devices, devices,
             options.c_str(), header_names.size(),
             programs.empty() ? NULL : &programs.front(),
             header_name_ptrs.empty() ? NULL : &header_name_ptrs.front(),
             0, 0));
      }
#endif
  };




  inline
  program *create_program_with_source(
      context &ctx,
      std::string const &src)
  {
    const char *string = src.c_str();
    size_t length = src.size();

    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clCreateProgramWithSource");
    cl_program result = clCreateProgramWithSource(
        ctx.data(), 1, &string, &length, &status_code);
    if (status_code != CL_SUCCESS)
      throw pyopencl::error("clCreateProgramWithSource", status_code);

    try
    {
      return new program(result, false, program::KND_SOURCE);
    }
    catch (...)
    {
      clReleaseProgram(result);
      throw;
    }
  }





  inline
  program *create_program_with_binary(
      context &ctx,
      py::sequence py_devices,
      py::sequence py_binaries)
  {
    std::vector<cl_device_id> devices;
    std::vector<const unsigned char *> binaries;
    std::vector<size_t> sizes;
    std::vector<cl_int> binary_statuses;

    size_t num_devices = len(py_devices);
    if (len(py_binaries) != num_devices)
      throw error("create_program_with_binary", CL_INVALID_VALUE,
          "device and binary counts don't match");

    for (size_t i = 0; i < num_devices; ++i)
    {
      devices.push_back(
          (py_devices[i]).cast<device const &>().data());
      const void *buf;
      PYOPENCL_BUFFER_SIZE_T len;

#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
      py_buffer_wrapper buf_wrapper;

      buf_wrapper.get(py::object(py_binaries[i]).ptr(), PyBUF_ANY_CONTIGUOUS);

      buf = buf_wrapper.m_buf.buf;
      len = buf_wrapper.m_buf.len;
#else
      if (PyObject_AsReadBuffer(
            py::object(py_binaries[i]).ptr(), &buf, &len))
        throw py::error_already_set();
#endif

      binaries.push_back(reinterpret_cast<const unsigned char *>(buf));
      sizes.push_back(len);
    }

    binary_statuses.resize(num_devices);

    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clCreateProgramWithBinary");
    cl_program result = clCreateProgramWithBinary(
        ctx.data(), num_devices,
        devices.empty( ) ? NULL : &devices.front(),
        sizes.empty( ) ? NULL : &sizes.front(),
        binaries.empty( ) ? NULL : &binaries.front(),
        binary_statuses.empty( ) ? NULL : &binary_statuses.front(),
        &status_code);
    if (status_code != CL_SUCCESS)
      throw pyopencl::error("clCreateProgramWithBinary", status_code);

    /*
    for (int i = 0; i < num_devices; ++i)
      printf("%d:%d\n", i, binary_statuses[i]);
      */

    try
    {
      return new program(result, false, program::KND_BINARY);
    }
    catch (...)
    {
      clReleaseProgram(result);
      throw;
    }
  }



#if (PYOPENCL_CL_VERSION >= 0x1020) && \
      ((PYOPENCL_CL_VERSION >= 0x1030) && defined(__APPLE__))
  inline
  program *create_program_with_built_in_kernels(
      context &ctx,
      py::object py_devices,
      std::string const &kernel_names)
  {
    PYOPENCL_PARSE_PY_DEVICES;

    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clCreateProgramWithBuiltInKernels");
    cl_program result = clCreateProgramWithBuiltInKernels(
        ctx.data(), num_devices, devices,
        kernel_names.c_str(), &status_code);
    if (status_code != CL_SUCCESS)
      throw pyopencl::error("clCreateProgramWithBuiltInKernels", status_code);

    try
    {
      return new program(result, false);
    }
    catch (...)
    {
      clReleaseProgram(result);
      throw;
    }
  }
#endif



#if PYOPENCL_CL_VERSION >= 0x1020
  inline
  program *link_program(
      context &ctx,
      py::object py_programs,
      std::string const &options,
      py::object py_devices
      )
  {
    PYOPENCL_PARSE_PY_DEVICES;

    std::vector<cl_program> programs;
    PYTHON_FOREACH(py_prg, py_programs)
    {
      program &prg = (py_prg).cast<program &>();
      programs.push_back(prg.data());
    }

    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clLinkProgram");
    cl_program result = clLinkProgram(
        ctx.data(), num_devices, devices,
        options.c_str(),
        programs.size(),
        programs.empty() ? NULL : &programs.front(),
        0, 0,
        &status_code);

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

    try
    {
      return new program(result, false);
    }
    catch (...)
    {
      clReleaseProgram(result);
      throw;
    }
  }

#endif


#if PYOPENCL_CL_VERSION >= 0x1020
  inline
  void unload_platform_compiler(platform &plat)
  {
    PYOPENCL_CALL_GUARDED(clUnloadPlatformCompiler, (plat.data()));
  }
#endif

  // }}}


  // {{{ kernel
  class local_memory
  {
    private:
      size_t m_size;

    public:
      local_memory(size_t size)
        : m_size(size)
      { }

      size_t size() const
      { return m_size; }
  };




  class kernel : noncopyable
  {
    private:
      cl_kernel m_kernel;

    public:
      kernel(cl_kernel knl, bool retain)
        : m_kernel(knl)
      {
        if (retain)
          PYOPENCL_CALL_GUARDED(clRetainKernel, (knl));
      }

      kernel(program const &prg, std::string const &kernel_name)
      {
        cl_int status_code;

        PYOPENCL_PRINT_CALL_TRACE("clCreateKernel");
        m_kernel = clCreateKernel(prg.data(), kernel_name.c_str(),
            &status_code);
        if (status_code != CL_SUCCESS)
          throw pyopencl::error("clCreateKernel", status_code);
      }

      ~kernel()
      {
        PYOPENCL_CALL_GUARDED_CLEANUP(clReleaseKernel, (m_kernel));
      }

      cl_kernel data() const
      {
        return m_kernel;
      }

      PYOPENCL_EQUALITY_TESTS(kernel);

      void set_arg_null(cl_uint arg_index)
      {
        cl_mem m = 0;
        PYOPENCL_CALL_GUARDED(clSetKernelArg, (m_kernel, arg_index,
              sizeof(cl_mem), &m));
      }

      void set_arg_mem(cl_uint arg_index, memory_object_holder &moh)
      {
        cl_mem m = moh.data();
        PYOPENCL_CALL_GUARDED(clSetKernelArg,
            (m_kernel, arg_index, sizeof(cl_mem), &m));
      }

      void set_arg_local(cl_uint arg_index, local_memory const &loc)
      {
        PYOPENCL_CALL_GUARDED(clSetKernelArg,
            (m_kernel, arg_index, loc.size(), 0));
      }

      void set_arg_sampler(cl_uint arg_index, sampler const &smp)
      {
        cl_sampler s = smp.data();
        PYOPENCL_CALL_GUARDED(clSetKernelArg,
            (m_kernel, arg_index, sizeof(cl_sampler), &s));
      }

      void set_arg_buf(cl_uint arg_index, py::object py_buffer)
      {
        const void *buf;
        PYOPENCL_BUFFER_SIZE_T len;

#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
        py_buffer_wrapper buf_wrapper;

        try
        {
          buf_wrapper.get(py_buffer.ptr(), PyBUF_ANY_CONTIGUOUS);
        }
        catch (py::error_already_set &)
        {
          PyErr_Clear();
          throw error("Kernel.set_arg", CL_INVALID_VALUE,
              "invalid kernel argument");
        }

        buf = buf_wrapper.m_buf.buf;
        len = buf_wrapper.m_buf.len;
#else
        if (PyObject_AsReadBuffer(py_buffer.ptr(), &buf, &len))
        {
          PyErr_Clear();
          throw error("Kernel.set_arg", CL_INVALID_VALUE,
              "invalid kernel argument");
        }
#endif

        PYOPENCL_CALL_GUARDED(clSetKernelArg,
            (m_kernel, arg_index, len, buf));
      }

      void set_arg(cl_uint arg_index, py::object arg)
      {
        if (arg.ptr() == Py_None)
        {
          set_arg_null(arg_index);
          return;
        }

        try
        {
          set_arg_mem(arg_index, arg.cast<memory_object_holder &>());
          return;
        }
        catch (py::cast_error &) { }

        try
        {
          set_arg_local(arg_index, arg.cast<local_memory>());
          return;
        }
        catch (py::cast_error &) { }

        try
        {
          set_arg_sampler(arg_index, arg.cast<const sampler &>());
          return;
        }
        catch (py::cast_error &) { }

        set_arg_buf(arg_index, arg);
      }

      py::object get_info(cl_kernel_info param_name) const
      {
        switch (param_name)
        {
          case CL_KERNEL_FUNCTION_NAME:
            PYOPENCL_GET_STR_INFO(Kernel, m_kernel, param_name);
          case CL_KERNEL_NUM_ARGS:
          case CL_KERNEL_REFERENCE_COUNT:
            PYOPENCL_GET_INTEGRAL_INFO(Kernel, m_kernel, param_name,
                cl_uint);
          case CL_KERNEL_CONTEXT:
            PYOPENCL_GET_OPAQUE_INFO(Kernel, m_kernel, param_name,
                cl_context, context);
          case CL_KERNEL_PROGRAM:
            PYOPENCL_GET_OPAQUE_INFO(Kernel, m_kernel, param_name,
                cl_program, program);
#if PYOPENCL_CL_VERSION >= 0x1020
          case CL_KERNEL_ATTRIBUTES:
            PYOPENCL_GET_STR_INFO(Kernel, m_kernel, param_name);
#endif
          default:
            throw error("Kernel.get_info", CL_INVALID_VALUE);
        }
      }

      py::object get_work_group_info(
          cl_kernel_work_group_info param_name,
          device const &dev
          ) const
      {
        switch (param_name)
        {
#define PYOPENCL_FIRST_ARG m_kernel, dev.data() // hackety hack
          case CL_KERNEL_WORK_GROUP_SIZE:
            PYOPENCL_GET_INTEGRAL_INFO(KernelWorkGroup,
                PYOPENCL_FIRST_ARG, param_name,
                size_t);
          case CL_KERNEL_COMPILE_WORK_GROUP_SIZE:
            {
              std::vector<size_t> result;
              PYOPENCL_GET_VEC_INFO(KernelWorkGroup,
                  PYOPENCL_FIRST_ARG, param_name, result);

              PYOPENCL_RETURN_VECTOR(size_t, result);
            }
          case CL_KERNEL_LOCAL_MEM_SIZE:
#if PYOPENCL_CL_VERSION >= 0x1010
          case CL_KERNEL_PRIVATE_MEM_SIZE:
#endif
            PYOPENCL_GET_INTEGRAL_INFO(KernelWorkGroup,
                PYOPENCL_FIRST_ARG, param_name,
                cl_ulong);

#if PYOPENCL_CL_VERSION >= 0x1010
          case CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE:
            PYOPENCL_GET_INTEGRAL_INFO(KernelWorkGroup,
                PYOPENCL_FIRST_ARG, param_name,
                size_t);
#endif
          default:
            throw error("Kernel.get_work_group_info", CL_INVALID_VALUE);
#undef PYOPENCL_FIRST_ARG
        }
      }

#if PYOPENCL_CL_VERSION >= 0x1020
      py::object get_arg_info(
          cl_uint arg_index,
          cl_kernel_arg_info param_name
          ) const
      {
        switch (param_name)
        {
#define PYOPENCL_FIRST_ARG m_kernel, arg_index // hackety hack
          case CL_KERNEL_ARG_ADDRESS_QUALIFIER:
            PYOPENCL_GET_INTEGRAL_INFO(KernelArg,
                PYOPENCL_FIRST_ARG, param_name,
                cl_kernel_arg_address_qualifier);

          case CL_KERNEL_ARG_ACCESS_QUALIFIER:
            PYOPENCL_GET_INTEGRAL_INFO(KernelArg,
                PYOPENCL_FIRST_ARG, param_name,
                cl_kernel_arg_access_qualifier);

          case CL_KERNEL_ARG_TYPE_NAME:
          case CL_KERNEL_ARG_NAME:
            PYOPENCL_GET_STR_INFO(KernelArg, PYOPENCL_FIRST_ARG, param_name);
#undef PYOPENCL_FIRST_ARG
          default:
            throw error("Kernel.get_arg_info", CL_INVALID_VALUE);
        }
      }
#endif
  };


  inline
  py::list create_kernels_in_program(program &pgm)
  {
    cl_uint num_kernels;
    PYOPENCL_CALL_GUARDED(clCreateKernelsInProgram, (
          pgm.data(), 0, 0, &num_kernels));

    std::vector<cl_kernel> kernels(num_kernels);
    PYOPENCL_CALL_GUARDED(clCreateKernelsInProgram, (
          pgm.data(), num_kernels,
          kernels.empty( ) ? NULL : &kernels.front(), &num_kernels));

    py::list result;
    for (cl_kernel knl: kernels)
      result.append(handle_from_new_ptr(new kernel(knl, true)));

    return result;
  }



  inline
  event *enqueue_nd_range_kernel(
      command_queue &cq,
      kernel &knl,
      py::object py_global_work_size,
      py::object py_local_work_size,
      py::object py_global_work_offset,
      py::object py_wait_for,
      bool g_times_l)
  {
    PYOPENCL_PARSE_WAIT_FOR;

    cl_uint work_dim = len(py_global_work_size);

    std::vector<size_t> global_work_size;
    COPY_PY_LIST(size_t, global_work_size);

    size_t *local_work_size_ptr = 0;
    std::vector<size_t> local_work_size;
    if (py_local_work_size.ptr() != Py_None)
    {
      if (g_times_l)
        work_dim = std::max(work_dim, unsigned(len(py_local_work_size)));
      else
        if (work_dim != unsigned(len(py_local_work_size)))
          throw error("enqueue_nd_range_kernel", CL_INVALID_VALUE,
              "global/local work sizes have differing dimensions");

      COPY_PY_LIST(size_t, local_work_size);

      while (local_work_size.size() < work_dim)
        local_work_size.push_back(1);
      while (global_work_size.size() < work_dim)
        global_work_size.push_back(1);

      local_work_size_ptr = local_work_size.empty( ) ? NULL : &local_work_size.front();
    }

    if (g_times_l && local_work_size_ptr)
    {
      for (cl_uint work_axis = 0; work_axis < work_dim; ++work_axis)
        global_work_size[work_axis] *= local_work_size[work_axis];
    }

    size_t *global_work_offset_ptr = 0;
    std::vector<size_t> global_work_offset;
    if (py_global_work_offset.ptr() != Py_None)
    {
      if (work_dim != unsigned(len(py_global_work_offset)))
        throw error("enqueue_nd_range_kernel", CL_INVALID_VALUE,
            "global work size and offset have differing dimensions");

      COPY_PY_LIST(size_t, global_work_offset);

      if (g_times_l && local_work_size_ptr)
      {
        for (cl_uint work_axis = 0; work_axis < work_dim; ++work_axis)
          global_work_offset[work_axis] *= local_work_size[work_axis];
      }

      global_work_offset_ptr = global_work_offset.empty( ) ? NULL :  &global_work_offset.front();
    }

    PYOPENCL_RETRY_RETURN_IF_MEM_ERROR( {
          cl_event evt;
          PYOPENCL_CALL_GUARDED(clEnqueueNDRangeKernel, (
                cq.data(),
                knl.data(),
                work_dim,
                global_work_offset_ptr,
                global_work_size.empty( ) ? NULL : &global_work_size.front(),
                local_work_size_ptr,
                PYOPENCL_WAITLIST_ARGS, &evt
                ));
          PYOPENCL_RETURN_NEW_EVENT(evt);
        } );
  }






  inline
  event *enqueue_task(
      command_queue &cq,
      kernel &knl,
      py::object py_wait_for)
  {
    PYOPENCL_PARSE_WAIT_FOR;

    PYOPENCL_RETRY_RETURN_IF_MEM_ERROR( {
      cl_event evt;
      PYOPENCL_CALL_GUARDED(clEnqueueTask, (
            cq.data(),
            knl.data(),
            PYOPENCL_WAITLIST_ARGS, &evt
            ));
      PYOPENCL_RETURN_NEW_EVENT(evt);
    } );
  }

  // }}}


  // {{{ gl interop
  inline
  bool have_gl()
  {
#ifdef HAVE_GL
    return true;
#else
    return false;
#endif
  }




#ifdef HAVE_GL

#ifdef __APPLE__
  inline
  cl_context_properties get_apple_cgl_share_group()
  {
    CGLContextObj kCGLContext = CGLGetCurrentContext();
    CGLShareGroupObj kCGLShareGroup = CGLGetShareGroup(kCGLContext);

    return (cl_context_properties) kCGLShareGroup;
  }
#endif /* __APPLE__ */




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




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




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

      py::object get_gl_texture_info(cl_gl_texture_info param_name)
      {
        switch (param_name)
        {
          case CL_GL_TEXTURE_TARGET:
            PYOPENCL_GET_INTEGRAL_INFO(GLTexture, data(), param_name, GLenum);
          case CL_GL_MIPMAP_LEVEL:
            PYOPENCL_GET_INTEGRAL_INFO(GLTexture, data(), param_name, GLint);

          default:
            throw error("MemoryObject.get_gl_texture_info", CL_INVALID_VALUE);
        }
      }
  };




#define PYOPENCL_WRAP_BUFFER_CREATOR(TYPE, NAME, CL_NAME, ARGS, CL_ARGS) \
  inline \
  TYPE *NAME ARGS \
  { \
    cl_int status_code; \
    PYOPENCL_PRINT_CALL_TRACE(#CL_NAME); \
    cl_mem mem = CL_NAME CL_ARGS; \
    \
    if (status_code != CL_SUCCESS) \
      throw pyopencl::error(#CL_NAME, status_code); \
    \
    try \
    { \
      return new TYPE(mem, false); \
    } \
    catch (...) \
    { \
      PYOPENCL_CALL_GUARDED(clReleaseMemObject, (mem)); \
      throw; \
    } \
  }




  PYOPENCL_WRAP_BUFFER_CREATOR(gl_buffer,
      create_from_gl_buffer, clCreateFromGLBuffer,
      (context &ctx, cl_mem_flags flags, GLuint bufobj),
      (ctx.data(), flags, bufobj, &status_code));
  PYOPENCL_WRAP_BUFFER_CREATOR(gl_texture,
      create_from_gl_texture_2d, clCreateFromGLTexture2D,
      (context &ctx, cl_mem_flags flags,
         GLenum texture_target, GLint miplevel, GLuint texture),
      (ctx.data(), flags, texture_target, miplevel, texture, &status_code));
  PYOPENCL_WRAP_BUFFER_CREATOR(gl_texture,
      create_from_gl_texture_3d, clCreateFromGLTexture3D,
      (context &ctx, cl_mem_flags flags,
         GLenum texture_target, GLint miplevel, GLuint texture),
      (ctx.data(), flags, texture_target, miplevel, texture, &status_code));
  PYOPENCL_WRAP_BUFFER_CREATOR(gl_renderbuffer,
      create_from_gl_renderbuffer, clCreateFromGLRenderbuffer,
      (context &ctx, cl_mem_flags flags, GLuint renderbuffer),
      (ctx.data(), flags, renderbuffer, &status_code));

  inline