wrap_cl.hpp

      command_queue &cq,
      cl_bool is_blocking,
      cl_map_flags flags,
      svm_pointer &svm,
      py::object py_wait_for,
      py::object user_size_py
      )
  {
    PYOPENCL_PARSE_WAIT_FOR;

    // {{{ process size

    PYOPENCL_GET_SVM_SIZE(svm);

    size_t size;
    bool have_size = false;
    if (svm_has_size)
    {
      size = svm_size;
      have_size = true;
    }
    if (!user_size_py.is_none())
    {
      size_t user_size = py::cast<size_t>(user_size_py);
      if (have_size && user_size > size)
        throw error("enqueue_svm_memfill", CL_INVALID_VALUE,
            "user-provided size too large for specified SVM buffer");
    }

    if (!have_size)
    {
      throw error("enqueue_svm_mem_map", CL_INVALID_VALUE,
          "size not passed and could not be determined");
    }

    // }}}

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

    PYOPENCL_RETURN_NEW_EVENT(evt);
  }


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

    cl_event evt;
    PYOPENCL_CALL_GUARDED(
        clEnqueueSVMUnmap,
        (
          cq.data(),
          svm.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_pointer &svm(py::cast<svm_pointer &>(py_svm));

      svm_pointers.push_back(svm.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:
#if PYOPENCL_CL_VERSION >= 0x2000
      sampler(context const &ctx, py::sequence py_props)
      {
        int hex_plat_version = ctx.get_hex_platform_version();

        if (hex_plat_version  < 0x2000)
        {
          std::cerr <<
            "sampler properties given as an iterable, "
            "which uses an OpenCL 2+-only interface, "
            "but the context's platform does not "
            "declare OpenCL 2 support. Proceeding "
            "as requested, but the next thing you see "
            "may be a crash." << std:: endl;
        }

        PYOPENCL_STACK_CONTAINER(cl_sampler_properties, props, py::len(py_props) + 1);
        {
          size_t i = 0;
          for (auto prop: py_props)
            props[i++] = py::cast<cl_sampler_properties>(prop);
          props[i++] = 0;
        }

        cl_int status_code;
        PYOPENCL_PRINT_CALL_TRACE("clCreateSamplerWithProperties");

        m_sampler = clCreateSamplerWithProperties(
            ctx.data(),
            PYOPENCL_STACK_CONTAINER_GET_PTR(props),
            &status_code);

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

      sampler(context const &ctx, bool normalized_coordinates,
          cl_addressing_mode am, cl_filter_mode fm)
      {
        PYOPENCL_PRINT_CALL_TRACE("clCreateSampler");

        int hex_plat_version = ctx.get_hex_platform_version();
#if PYOPENCL_CL_VERSION >= 0x2000
        if (hex_plat_version  >= 0x2000)
        {
            cl_sampler_properties props_list[] = {
              CL_SAMPLER_NORMALIZED_COORDS, normalized_coordinates,
              CL_SAMPLER_ADDRESSING_MODE, am,
              CL_SAMPLER_FILTER_MODE, fm,
              0,
            };

            cl_int status_code;

            PYOPENCL_PRINT_CALL_TRACE("clCreateSamplerWithProperties");
            m_sampler = clCreateSamplerWithProperties(
                ctx.data(), props_list, &status_code);

            if (status_code != CL_SUCCESS)
              throw pyopencl::error("Sampler", status_code);
        }
        else
#endif
        {
          cl_int status_code;

#if defined(__GNUG__) && !defined(__clang__)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
#endif
          m_sampler = clCreateSampler(
              ctx.data(),
              normalized_coordinates,
              am, fm, &status_code);
#if defined(__GNUG__) && !defined(__clang__)
#pragma GCC diagnostic pop
#endif

          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_TYPED_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_TYPED_INFO(Sampler, m_sampler, param_name,
                cl_addressing_mode);
          case CL_SAMPLER_FILTER_MODE:
            PYOPENCL_GET_TYPED_INFO(Sampler, m_sampler, param_name,
                cl_filter_mode);
          case CL_SAMPLER_NORMALIZED_COORDS:
            PYOPENCL_GET_TYPED_INFO(Sampler, m_sampler, param_name,
                cl_bool);
#if PYOPENCL_CL_VERSION >= 0x3000
          case CL_SAMPLER_PROPERTIES:
            {
              std::vector<cl_sampler_properties> result;
              PYOPENCL_GET_VEC_INFO(Sampler, m_sampler, param_name, result);
              PYOPENCL_RETURN_VECTOR(cl_sampler_properties, result);
            }
#endif

#ifdef CL_SAMPLER_MIP_FILTER_MODE_KHR
          case CL_SAMPLER_MIP_FILTER_MODE_KHR:
            PYOPENCL_GET_TYPED_INFO(Sampler, m_sampler, param_name,
                cl_filter_mode);
          case CL_SAMPLER_LOD_MIN_KHR:
          case CL_SAMPLER_LOD_MAX_KHR:
            PYOPENCL_GET_TYPED_INFO(Sampler, m_sampler, param_name, float);
#endif

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

  // }}}


  // {{{ program

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

    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_TYPED_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_TYPED_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_TYPED_INFO(Program, m_program, param_name,
                size_t);
          case CL_PROGRAM_KERNEL_NAMES:
            PYOPENCL_GET_STR_INFO(Program, m_program, param_name);
#endif
#if PYOPENCL_CL_VERSION >= 0x2010
          case CL_PROGRAM_IL:
            PYOPENCL_GET_STR_INFO(Program, m_program, param_name);
#endif
#if PYOPENCL_CL_VERSION >= 0x2020
          case CL_PROGRAM_SCOPE_GLOBAL_CTORS_PRESENT:
          case CL_PROGRAM_SCOPE_GLOBAL_DTORS_PRESENT:
            PYOPENCL_GET_TYPED_INFO(Program, m_program, param_name, cl_bool);
#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_TYPED_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_TYPED_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_TYPED_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 = py::cast<std::string>(name_hdr_tup[0]);
          program &prg = py::cast<program &>(name_hdr_tup[1]);

          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

#if PYOPENCL_CL_VERSION >= 0x2020
      void set_specialization_constant(cl_uint spec_id, py::object py_buffer)
      {
        py_buffer_wrapper bufwrap;
        bufwrap.get(py_buffer.ptr(), PyBUF_ANY_CONTIGUOUS);
        PYOPENCL_CALL_GUARDED(clSetProgramSpecializationConstant,
            (m_program, spec_id, bufwrap.m_buf.len, bufwrap.m_buf.buf));
      }
#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;

    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::cast<device const &>(py_devices[i]).data());
      const void *buf;
      PYOPENCL_BUFFER_SIZE_T len;

      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;

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

    PYOPENCL_STACK_CONTAINER(cl_int, binary_statuses, 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(),
        PYOPENCL_STACK_CONTAINER_GET_PTR(binary_statuses),
        &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 >= 0x2010)
  inline
  program *create_program_with_il(
      context &ctx,
      std::string const &src)
  {
    cl_int status_code;
    PYOPENCL_PRINT_CALL_TRACE("clCreateProgramWithIL");
    cl_program result = clCreateProgramWithIL(
        ctx.data(), src.c_str(), src.size(), &status_code);
    if (status_code != CL_SUCCESS)
      throw pyopencl::error("clCreateProgramWithIL", status_code);

    try
    {
      return new program(result, false, program::KND_IL);
    }
    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::cast<program &>(py_prg);
      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("clLinkProgram", result, 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;
      bool m_set_arg_prefer_svm;

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

      kernel(program const &prg, std::string const &kernel_name)
        : m_set_arg_prefer_svm(false)
      {
        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);

#if PYOPENCL_CL_VERSION >= 0x2010
      kernel *clone()
      {
        cl_int status_code;

        PYOPENCL_PRINT_CALL_TRACE("clCloneKernel");
        cl_kernel result = clCloneKernel(m_kernel, &status_code);
        if (status_code != CL_SUCCESS)
          throw pyopencl::error("clCloneKernel", status_code);

        try
        {
          return new kernel(result, /* retain */ false);
        }
        catch (...)
        {
          PYOPENCL_CALL_GUARDED_CLEANUP(clReleaseKernel, (result));
          throw;
        }
      }
#endif

      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_command_queue(cl_uint arg_index, command_queue const &queue)
      {
        cl_command_queue q = queue.data();
        PYOPENCL_CALL_GUARDED(clSetKernelArg,
            (m_kernel, arg_index, sizeof(cl_command_queue), &q));
      }

      void set_arg_buf_pack(cl_uint arg_index, py::handle py_typechar, py::handle obj)
      {
        std::string typechar_str(py::cast<std::string>(py_typechar));
        if (typechar_str.size() != 1)
          throw error("Kernel.set_arg_buf_pack", CL_INVALID_VALUE,
              "type char argument must have exactly one character");

        char typechar = typechar_str[0];

#define PYOPENCL_KERNEL_PACK_AND_SET_ARG(TYPECH_VAL, TYPE) \
        case TYPECH_VAL: \
          { \
            TYPE val = py::cast<TYPE>(obj); \
            PYOPENCL_CALL_GUARDED(clSetKernelArg, (m_kernel, arg_index, sizeof(val), &val)); \
            break; \
          }
        switch (typechar)
        {
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('c', char)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('b', signed char)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('B', unsigned char)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('h', short)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('H', unsigned short)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('i', int)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('I', unsigned int)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('l', long)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('L', unsigned long)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('f', float)
          PYOPENCL_KERNEL_PACK_AND_SET_ARG('d', double)
          default:
            throw error("Kernel.set_arg_buf_pack", CL_INVALID_VALUE,
                "invalid type char");
        }
#undef PYOPENCL_KERNEL_PACK_AND_SET_ARG
      }

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

        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;

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

#if PYOPENCL_CL_VERSION >= 0x2000
      void set_arg_svm(cl_uint arg_index, svm_pointer const &wrp)
      {
        PYOPENCL_CALL_GUARDED(clSetKernelArgSVMPointer,
            (m_kernel, arg_index, wrp.svm_ptr()));
      }
#endif

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

        // It turns out that a taken 'catch' has a relatively high cost, so
        // in deciding which of "mem object" and "svm" to try first, we use
        // whatever we were given last time around.
        if (m_set_arg_prefer_svm)
        {
#if PYOPENCL_CL_VERSION >= 0x2000
          try
          {
            set_arg_svm(arg_index, py::cast<svm_pointer const &>(arg));
            return;
          }
          catch (py::cast_error &) { }
#endif

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

#if PYOPENCL_CL_VERSION >= 0x2000
          try
          {
            set_arg_svm(arg_index, py::cast<svm_pointer const &>(arg));
            m_set_arg_prefer_svm = true;
            return;
          }
          catch (py::cast_error &) { }
#endif
        }

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

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

        try
        {
          set_arg_command_queue(arg_index, py::cast<const command_queue &>(arg));
          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_TYPED_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
      {