wrap_cl.hpp

      cl_map_flags flags,
      size_t offset,
      py::object py_shape, py::object dtype,
      py::object py_order, py::object py_strides,
      py::object py_wait_for,
      bool is_blocking
      )
  {
    PYOPENCL_PARSE_WAIT_FOR;
    PYOPENCL_PARSE_NUMPY_ARRAY_SPEC;

    npy_uintp size_in_bytes = tp_descr->elsize;
    for (npy_intp sdim: shape)
      size_in_bytes *= sdim;

    py::object result;

    cl_event evt;
    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clEnqueueMapBuffer");
    void *mapped;

    PYOPENCL_RETRY_IF_MEM_ERROR(
        {
          {
            py::gil_scoped_release release;
            mapped = clEnqueueMapBuffer(
                  cq->data(), buf.data(),
                  PYOPENCL_CAST_BOOL(is_blocking), flags,
                  offset, size_in_bytes,
                  PYOPENCL_WAITLIST_ARGS, &evt,
                  &status_code);
          }
          if (status_code != CL_SUCCESS)
            throw pyopencl::error("clEnqueueMapBuffer", status_code);
        } );

    event evt_handle(evt, false);

    std::unique_ptr<memory_map> map;
    try
    {
      result = py::object(py::reinterpret_steal<py::object>(PyArray_NewFromDescr(
          &PyArray_Type, tp_descr,
          shape.size(),
          shape.empty() ? nullptr : &shape.front(),
          strides.empty() ? nullptr : &strides.front(),
          mapped, ary_flags, /*obj*/nullptr)));

      if (size_in_bytes != (npy_uintp) PyArray_NBYTES(result.ptr()))
        throw pyopencl::error("enqueue_map_buffer", CL_INVALID_VALUE,
            "miscalculated numpy array size (not contiguous?)");

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

    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)));
  }




  inline
  py::object enqueue_map_image(
      std::shared_ptr<command_queue> cq,
      memory_object_holder &img,
      cl_map_flags flags,
      py::object py_origin,
      py::object py_region,
      py::object py_shape, py::object dtype,
      py::object py_order, py::object py_strides,
      py::object py_wait_for,
      bool is_blocking
      )
  {
    PYOPENCL_PARSE_WAIT_FOR;
    PYOPENCL_PARSE_NUMPY_ARRAY_SPEC;
    COPY_PY_COORD_TRIPLE(origin);
    COPY_PY_REGION_TRIPLE(region);

    cl_event evt;
    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clEnqueueMapImage");
    size_t row_pitch, slice_pitch;
    void *mapped;
    PYOPENCL_RETRY_IF_MEM_ERROR(
      {
        {
          py::gil_scoped_release release;
          mapped = clEnqueueMapImage(
                cq->data(), img.data(),
                PYOPENCL_CAST_BOOL(is_blocking), flags,
                origin, region, &row_pitch, &slice_pitch,
                PYOPENCL_WAITLIST_ARGS, &evt,
                &status_code);
        }
        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() ? nullptr : &shape.front(),
        strides.empty() ? nullptr : &strides.front(),
        mapped, ary_flags, /*obj*/nullptr));

    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);
  }

  // }}}


  // {{{ svm

#if PYOPENCL_CL_VERSION >= 0x2000

  class svm_arg_wrapper
  {
    private:
      void *m_ptr;
      PYOPENCL_BUFFER_SIZE_T m_size;
#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
        std::unique_ptr<py_buffer_wrapper> ward;
#endif

    public:
      svm_arg_wrapper(py::object holder)
      {
#ifdef PYOPENCL_USE_NEW_BUFFER_INTERFACE
        ward = std::unique_ptr<py_buffer_wrapper>(new py_buffer_wrapper);
        ward->get(holder.ptr(), PyBUF_ANY_CONTIGUOUS | PyBUF_WRITABLE);
        m_ptr = ward->m_buf.buf;
        m_size = ward->m_buf.len;
#else
        py::object ward = buffer;
        if (PyObject_AsWriteBuffer(buffer.ptr(), &m_ptr, &m_size))
          throw py::error_already_set();
#endif
      }

      void *ptr() const
      {
        return m_ptr;
      }
      size_t size() const
      {
        return m_size;
      }
  };


  class svm_allocation : noncopyable
  {
    private:
      std::shared_ptr<context> m_context;
      void *m_allocation;

    public:
      svm_allocation(std::shared_ptr<context> const &ctx, size_t size, cl_uint alignment, cl_svm_mem_flags flags)
        : m_context(ctx)
      {
        PYOPENCL_PRINT_CALL_TRACE("clSVMalloc");
        m_allocation = clSVMAlloc(
            ctx->data(),
            flags, size, alignment);

        if (!m_allocation)
          throw pyopencl::error("clSVMAlloc", CL_OUT_OF_RESOURCES);
      }

      ~svm_allocation()
      {
        if (m_allocation)
          release();
      }

      void release()
      {
        if (!m_allocation)
          throw error("SVMAllocation.release", CL_INVALID_VALUE,
              "trying to double-unref svm allocation");

        clSVMFree(m_context->data(), m_allocation);
        m_allocation = nullptr;
      }

      void enqueue_release(command_queue &queue, py::object py_wait_for)
      {
        PYOPENCL_PARSE_WAIT_FOR;

        if (!m_allocation)
          throw error("SVMAllocation.release", CL_INVALID_VALUE,
              "trying to double-unref svm allocation");

        cl_event evt;

        PYOPENCL_CALL_GUARDED_CLEANUP(clEnqueueSVMFree, (
              queue.data(), 1, &m_allocation,
              nullptr, nullptr,
              PYOPENCL_WAITLIST_ARGS, &evt));

        m_allocation = nullptr;
      }

      void *ptr() const
      {
        return m_allocation;
      }

      intptr_t ptr_as_int() const
      {
        return (intptr_t) m_allocation;
      }

      bool operator==(svm_allocation const &other) const
      {
        return m_allocation == other.m_allocation;
      }

      bool operator!=(svm_allocation const &other) const
      {
        return m_allocation != other.m_allocation;
      }
  };


  inline
  event *enqueue_svm_memcpy(
      command_queue &cq,
      cl_bool is_blocking,
      svm_arg_wrapper &dst, svm_arg_wrapper &src,
      py::object py_wait_for
      )
  {
    PYOPENCL_PARSE_WAIT_FOR;

    if (src.size() != dst.size())
      throw error("_enqueue_svm_memcpy", CL_INVALID_VALUE,
          "sizes of source and destination buffer do not match");

    cl_event evt;
    PYOPENCL_CALL_GUARDED(
        clEnqueueSVMMemcpy,
        (
          cq.data(),
          is_blocking,
          dst.ptr(), src.ptr(),
          dst.size(),
          PYOPENCL_WAITLIST_ARGS,
          &evt
        ));

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }


  inline
  event *enqueue_svm_memfill(
      command_queue &cq,
      svm_arg_wrapper &dst, py::object py_pattern,
      py::object byte_count,
      py::object py_wait_for
      )
  {
    PYOPENCL_PARSE_WAIT_FOR;

    void *pattern_ptr;
    PYOPENCL_BUFFER_SIZE_T pattern_len;

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

    pattern_ward->get(py_pattern.ptr(), PyBUF_ANY_CONTIGUOUS);

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

    size_t fill_size = dst.size();
    if (!byte_count.is_none())
      fill_size = py::cast<size_t>(byte_count);

    cl_event evt;
    PYOPENCL_CALL_GUARDED(
        clEnqueueSVMMemFill,
        (
          cq.data(),
          dst.ptr(), pattern_ptr,
          pattern_len,
          fill_size,
          PYOPENCL_WAITLIST_ARGS,
          &evt
        ));

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }


  inline
  event *enqueue_svm_map(
      command_queue &cq,
      cl_bool is_blocking,
      cl_map_flags flags,
      svm_arg_wrapper &svm,
      py::object py_wait_for
      )
  {
    PYOPENCL_PARSE_WAIT_FOR;

    cl_event evt;
    PYOPENCL_CALL_GUARDED(
        clEnqueueSVMMap,
        (
          cq.data(),
          is_blocking,
          flags,
          svm.ptr(), svm.size(),
          PYOPENCL_WAITLIST_ARGS,
          &evt
        ));

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }


  inline
  event *enqueue_svm_unmap(
      command_queue &cq,
      svm_arg_wrapper &svm,
      py::object py_wait_for
      )
  {
    PYOPENCL_PARSE_WAIT_FOR;

    cl_event evt;
    PYOPENCL_CALL_GUARDED(
        clEnqueueSVMUnmap,
        (
          cq.data(),
          svm.ptr(),
          PYOPENCL_WAITLIST_ARGS,
          &evt
        ));

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }
#endif


#if PYOPENCL_CL_VERSION >= 0x2010
  inline
  event *enqueue_svm_migratemem(
      command_queue &cq,
      py::sequence svms,
      cl_mem_migration_flags flags,
      py::object py_wait_for
      )
  {
    PYOPENCL_PARSE_WAIT_FOR;

    std::vector<const void *> svm_pointers;
    std::vector<size_t> sizes;

    for (py::handle py_svm: svms)
    {
      svm_arg_wrapper &svm(py::cast<svm_arg_wrapper &>(py_svm));

      svm_pointers.push_back(svm.ptr());
      sizes.push_back(svm.size());
    }

    cl_event evt;
    PYOPENCL_CALL_GUARDED(
        clEnqueueSVMMigrateMem,
        (
         cq.data(),
         svm_pointers.size(),
         svm_pointers.empty() ? nullptr : &svm_pointers.front(),
         sizes.empty() ? nullptr : &sizes.front(),
         flags,
         PYOPENCL_WAITLIST_ARGS,
         &evt
        ));

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }
#endif

  // }}}


  // {{{ 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( ) ? nullptr : &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
#if PYOPENCL_CL_VERSION >= 0x2000
          case CL_PROGRAM_BUILD_GLOBAL_VARIABLE_TOTAL_SIZE:
            PYOPENCL_GET_INTEGRAL_INFO(ProgramBuild,
                PYOPENCL_FIRST_ARG, param_name,
                size_t);
#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;
        for (py::handle 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() ? nullptr : &programs.front(),
             header_name_ptrs.empty() ? nullptr : &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( ) ? nullptr : &devices.front(),
        sizes.empty( ) ? nullptr : &sizes.front(),
        binaries.empty( ) ? nullptr : &binaries.front(),
        binary_statuses.empty( ) ? nullptr : &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;
    for (py::handle 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() ? nullptr : &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;