Is this page helpful?

User-defined Functions#

group User-defined Functions

Typedefs

using lto_fragments = std::span<lto_fragment const>#

Device modules linked into a JIT kernel.

Enums

enum class lto_binary_type : uint8_t#

The type of LTO Binary.

Values:

enumerator LTO_IR#
enumerator FATBIN#
enumerator PTX#

Functions

compiled_fragment compile_cuda_to_lto_ir(
char const *source_code,
std::span<char const*const> include_names,
std::span<char const*const> includes,
std::span<char const*const> include_paths,
bool use_pch = true
)#

Compiles CUDA device source into an owning LTO IR fragment.

Compiled fragments are cached by source, in-memory headers, ordered include paths and PCH use.

Filesystem headers and their dependencies must remain unchanged while caches may be reused; use versioned include paths when changing filesystem dependencies.

When PCH is enabled, headers used to build it must remain unchanged while it may be reused. Place changing headers after #pragma nv_hdrstop or disable PCH.

Parameters:

  • source_code – Null-terminated CUDA source

  • include_names – Null-terminated names of additional in-memory headers

  • includes – Matching null-terminated header contents

  • include_paths – Filesystem header directories, searched before bundled header directories

  • use_pch – Enable automatic precompiled headers on supported NVRTC versions

Throws:

  • std::invalid_argument – for mismatched header spans

  • std::runtime_error – if compilation fails

Returns:

Shared ownership of the compiled bytes

struct lto_fragment#
#include <cudf/udf.hpp>

A compiled device module to link into a JIT kernel.

A module may define multiple functions and reference functions in other modules. Keep its bytes alive until evaluation completes or a reusable transform program has been constructed.

Public Members

lto_binary_type type = lto_binary_type::LTO_IR#

Binary format of the module.

std::span<uint8_t const> data = {}#

The module’s bytes.

struct compiled_fragment#
#include <cudf/udf.hpp>

Shares ownership of compiled device code; copies retain the same storage.

Byte views do not own storage. Keep an owner alive while passing a view to an LTO API.

Public Functions

lto_fragment view() const noexcept#

Returns a non-owning view of the compiled fragment.

Returns:

Non-owning view of the compiled fragment