cuda.core.ProgramOptions

class cuda.core.ProgramOptions(

name: str | None = 'default_program',

arch: str | None = None,

relocatable_device_code: bool | None = None,

extensible_whole_program: bool | None = None,

debug: bool | None = None,

lineinfo: bool | None = None,

device_code_optimize: bool | None = None,

ptxas_options: str | list[str] | tuple[str] | None = None,

max_register_count: int | None = None,

ftz: bool | None = None,

prec_sqrt: bool | None = None,

prec_div: bool | None = None,

fma: bool | None = None,

use_fast_math: bool | None = None,

extra_device_vectorization: bool | None = None,

link_time_optimization: bool | None = None,

gen_opt_lto: bool | None = None,

define_macro: str | tuple[str, str] | list[str | tuple[str, str]] | tuple[str | tuple[str, str], ...] | None = None,

undefine_macro: str | list[str] | tuple[str] | None = None,

include_path: str | list[str] | tuple[str] | None = None,

use_bundled_headers: bool | None = None,

pre_include: str | list[str] | tuple[str] | None = None,

no_source_include: bool | None = None,

std: str | None = None,

builtin_move_forward: bool | None = None,

builtin_initializer_list: bool | None = None,

disable_warnings: bool | None = None,

restrict: bool | None = None,

device_as_default_execution_space: bool | None = None,

device_int128: bool | None = None,

optimization_info: str | None = None,

no_display_error_number: bool | None = None,

diag_error: int | list[int] | tuple[int] | None = None,

diag_suppress: int | list[int] | tuple[int] | None = None,

diag_warn: int | list[int] | tuple[int] | None = None,

brief_diagnostics: bool | None = None,

time: str | None = None,

split_compile: int | None = None,

fdevice_syntax_only: bool | None = None,

minimal: bool | None = None,

no_cache: bool | None = None,

fdevice_time_trace: str | None = None,

device_float128: bool | None = None,

frandom_seed: str | None = None,

ofast_compile: str | None = None,

pch: bool | None = None,

create_pch: str | None = None,

use_pch: str | None = None,

pch_dir: str | None = None,

pch_verbose: bool | None = None,

pch_messages: bool | None = None,

instantiate_templates_in_pch: bool | None = None,

extra_sources: list[tuple[str, str | bytes | bytearray]] | tuple[tuple[str, str | bytes | bytearray], ...] | None = None,

use_libdevice: bool | None = None,

numba_debug: bool | None = None,

)

Customizable options for configuring Program.

name

Name of the program. If the compilation succeeds, the name is passed down to the generated ObjectCode. When set to None, "default_program" is used.

Type:

str, optional

arch

Pass the SM architecture value, such as sm_<CC> (for generating CUBIN) or compute_<CC> (for generating PTX). If not provided, the current device’s architecture will be used.

Type:

str, optional

relocatable_device_code

Enable (disable) the generation of relocatable device code. Default: False

Type:

bool, optional

extensible_whole_program

Do extensible whole program compilation of device code. Default: False

Type:

bool, optional

debug

Generate debug information. If –dopt is not specified, then turns off all optimizations. Default: False

Type:

bool, optional

lineinfo

Generate line-number information. Default: False

Type:

bool, optional

device_code_optimize

Enable device code optimization. When specified along with ‘-G’, enables limited debug information generation for optimized device code. Default: None

Type:

bool, optional

ptxas_options

Specify one or more options directly to ptxas, the PTX optimizing assembler. Options should be strings. For example [“-v”, “-O2”]. Default: None

Type:

str | list[str], optional

max_register_count

Specify the maximum amount of registers that GPU functions can use. Default: None

Type:

int, optional

ftz

When performing single-precision floating-point operations, flush denormal values to zero or preserve denormal values. Default: False

Type:

bool, optional

prec_sqrt

For single-precision floating-point square root, use IEEE round-to-nearest mode or use a faster approximation. Default: True

Type:

bool, optional

prec_div

For single-precision floating-point division and reciprocals, use IEEE round-to-nearest mode or use a faster approximation. Default: True

Type:

bool, optional

fma

Enables (disables) the contraction of floating-point multiplies and adds/subtracts into floating-point multiply-add operations. Default: True

Type:

bool, optional

use_fast_math

Make use of fast math operations. Default: False

Type:

bool, optional

extra_device_vectorization

Enables more aggressive device code vectorization in the NVVM optimizer. Default: False

Type:

bool, optional

link_time_optimization

Generate intermediate code for later link-time optimization. Default: False

Type:

bool, optional

gen_opt_lto

Run the optimizer passes before generating the LTO IR. Default: False

Type:

bool, optional

define_macro

Predefine a macro. Can be either a string, in which case that macro will be set to 1, a 2 element tuple of strings, in which case the first element is defined as the second, or a list of strings or tuples. Default: None

Type:

str | tuple[str, str] | list[str | tuple[str, str]], optional

undefine_macro

Cancel any previous definition of a macro, or list of macros. Default: None

Type:

str | list[str], optional

include_path

Add the directory or directories to the list of directories to be searched for headers. Default: None

Type:

str | list[str], optional

use_bundled_headers

Use the CUDA and CCCL headers bundled with NVRTC, installed into a per-user cache directory, instead of requiring a full CUDA Toolkit installation. Implemented via NVRTC’s --use-bundled-headers=<dir> compiler option, which installs the headers into the cache directory (skipping installation if already present and up to date) and adds that directory to the include search path. NVRTC only. Default: False

Type:

bool, optional

pre_include

Preinclude one or more headers during preprocessing. Can be either a string or a list of strings. Default: None

Type:

str | list[str], optional

no_source_include

Disable the default behavior of adding the directory of each input source to the include path. Default: False

Type:

bool, optional

std

Set language dialect to C++03, C++11, C++14, C++17 or C++20. Default: c++17

Type:

str, optional

builtin_move_forward

Provide builtin definitions of std::move and std::forward. Default: True

Type:

bool, optional

builtin_initializer_list

Provide builtin definitions of std::initializer_list class and member functions. Default: True

Type:

bool, optional

disable_warnings

Inhibit all warning messages. Default: False

Type:

bool, optional

restrict

Programmer assertion that all kernel pointer parameters are restrict pointers. Default: False

Type:

bool, optional

device_as_default_execution_space

Treat entities with no execution space annotation as __device__ entities. Default: False

Type:

bool, optional

device_int128

Allow the __int128 type in device code. Default: False

Type:

bool, optional

optimization_info

Provide optimization reports for the specified kind of optimization. Default: None

Type:

str, optional

no_display_error_number

Disable the display of a diagnostic number for warning messages. Default: False

Type:

bool, optional

diag_error

Emit error for a specified diagnostic message number or comma-separated list of numbers. Default: None

Type:

int | list[int], optional

diag_suppress

Suppress a specified diagnostic message number or comma-separated list of numbers. Default: None

Type:

int | list[int], optional

diag_warn

Emit warning for a specified diagnostic message number or comma-separated list of numbers. Default: None

Type:

int | list[int], optional

brief_diagnostics

Disable or enable showing source line and column info in a diagnostic. Default: False

Type:

bool, optional

time

Generate a CSV table with the time taken by each compilation phase. Default: None

Type:

str, optional

split_compile

Perform compiler optimizations in parallel. Default: 1

Type:

int, optional

fdevice_syntax_only

Ends device compilation after front-end syntax checking. Default: False

Type:

bool, optional

minimal

Omit certain language features to reduce compile time for small programs. Default: False

Type:

bool, optional

no_cache

Disable compiler caching. Default: False

Type:

bool, optional

fdevice_time_trace

Generate time trace JSON for profiling compilation (NVRTC only). Default: None

Type:

str, optional

device_float128

Allow __float128 type in device code (NVRTC only). Default: False

Type:

bool, optional

frandom_seed

Set random seed for randomized optimizations (NVRTC only). Default: None

Type:

str, optional

ofast_compile

Fast compilation mode: “0”, “min”, “mid”, or “max” (NVRTC only). Default: None

Type:

str, optional

pch

Use default precompiled header (NVRTC only, CUDA 12.8+). Default: False

Type:

bool, optional

create_pch

Create precompiled header file (NVRTC only, CUDA 12.8+). Default: None

Type:

str, optional

use_pch

Use specific precompiled header file (NVRTC only, CUDA 12.8+). Default: None

Type:

str, optional

pch_dir

PCH directory location (NVRTC only, CUDA 12.8+). Default: None

Type:

str, optional

pch_verbose

Verbose PCH output (NVRTC only, CUDA 12.8+). Default: False

Type:

bool, optional

pch_messages

Control PCH diagnostic messages (NVRTC only, CUDA 12.8+). Default: False

Type:

bool, optional

instantiate_templates_in_pch

Control template instantiation in PCH (NVRTC only, CUDA 12.8+). Default: False

Type:

bool, optional

extra_sources

Additional NVVM IR modules to compile together with the main program, specified as ((name1, source1), (name2, source2), ...). Each name is a string identifier used in diagnostic messages. Each source can be a string (textual LLVM IR) or bytes/bytearray (LLVM bitcode). Only supported for the NVVM backend. Default: None

Type:

list of 2-tuples or tuple of 2-tuples, optional

use_libdevice

Load NVIDIA’s libdevice math builtins library. Only supported for the NVVM backend. Default: False

Type:

bool, optional

numba_debug

Emit the debug information layout expected by Numba. Recognized only by newer toolkits; compilers that do not support it reject the option with an error. Applies only to the NVVM and NVRTC compilation backends – code_type="ptx" is processed by the linker, which cannot honor it, so enabling this option there emits a UserWarning and the option is ignored. Default: None

Type:

bool, optional