nvrtc

The User guide for the NVRTC library.

1. Introduction

NVRTC is a runtime compilation library for CUDA C++. It accepts CUDA C++ source code in character string form and creates handles that can be used to obtain the PTX. The PTX string generated by NVRTC can be loaded by cuModuleLoadData and cuModuleLoadDataEx, and linked with other modules by using the nvJitLink library or using cuLinkAddData of the CUDA Driver API. This facility can often provide optimizations and performance not possible in a purely offline static compilation.

In the absence of NVRTC (or any runtime compilation support in CUDA), users needed to spawn a separate process to execute nvcc at runtime if they wished to implement runtime compilation in their applications or libraries, and, unfortunately, this approach has the following drawbacks:

  • The compilation overhead tends to be higher than necessary.

  • End users are required to install nvcc and related tools which make it complicated to distribute applications that use runtime compilation.

NVRTC addresses these issues by providing a library interface that eliminates overhead associated with spawning separate processes, disk I/O,and so on, while keeping application deployment simple.

2. Getting Started

2.1. System Requirements

NVRTC is supported on the following platforms: Linux x86_64, Linux ppc64le, Linux aarch64, Windows x86_64.

Note: NVRTC does not depend on any other libraries or headers from the CUDA toolkit, and can be run on a system without a GPU.

2.2. Installation

NVRTC is part of the CUDA Toolkit release and the components are organized as follows in the CUDA toolkit installation directory:

  • On Windows:

    • include\nvrtc.h

    • bin\nvrtc64_Major Release VersionMinor Release Version_0.dll

    • bin\nvrtc-builtins64_Major Release VersionMinor Release Version.dll

    • lib\x64\nvrtc.lib

    • lib\x64\nvrtc_static.lib

    • lib\x64\nvrtc-builtins_static.lib

    • doc\pdf\NVRTC_User_Guide.pdf

  • On Linux:

    • include/nvrtc.h

    • lib64/libnvrtc.so

    • lib64/libnvrtc.so.Major Release Version.Minor Release Version

    • lib64/libnvrtc.so.Major Release Version.Minor Release Version.<build version>

    • lib64/libnvrtc-builtins.so

    • lib64/libnvrtc-builtins.so.Major Release Version.Minor Release Version

    • lib64/libnvrtc-builtins.so.Major Release Version.Minor Release Version.<build version>

    • lib64/libnvrtc_static.a

    • lib64/libnvrtc-builtins_static.a

    • doc/pdf/NVRTC_User_Guide.pdf

3. User Interface

This chapter presents the API of NVRTC. Basic usage of the API is explained in Basic Usage.

3.1. Error Handling

NVRTC defines the following enumeration type and function for API call error handling.

Enumerations

nvrtcResult

The enumerated type nvrtcResult defines API call result codes.

Functions

const char * nvrtcGetErrorString(nvrtcResult result)

nvrtcGetErrorString is a helper function that returns a string describing the given nvrtcResult code, e.g., NVRTC_SUCCESS to "NVRTC_SUCCESS" .

3.1.1. Enumerations

enum nvrtcResult

The enumerated type nvrtcResult defines API call result codes.

NVRTC API functions return nvrtcResult to indicate the call result.

Values:

enumerator NVRTC_SUCCESS
enumerator NVRTC_ERROR_OUT_OF_MEMORY
enumerator NVRTC_ERROR_PROGRAM_CREATION_FAILURE
enumerator NVRTC_ERROR_INVALID_INPUT
enumerator NVRTC_ERROR_INVALID_PROGRAM
enumerator NVRTC_ERROR_INVALID_OPTION
enumerator NVRTC_ERROR_COMPILATION
enumerator NVRTC_ERROR_BUILTIN_OPERATION_FAILURE
enumerator NVRTC_ERROR_NO_NAME_EXPRESSIONS_AFTER_COMPILATION
enumerator NVRTC_ERROR_NO_LOWERED_NAMES_BEFORE_COMPILATION
enumerator NVRTC_ERROR_NAME_EXPRESSION_NOT_VALID
enumerator NVRTC_ERROR_INTERNAL_ERROR
enumerator NVRTC_ERROR_TIME_FILE_WRITE_FAILED

3.1.2. Functions

const char *nvrtcGetErrorString(nvrtcResult result)

nvrtcGetErrorString is a helper function that returns a string describing the given nvrtcResult code, e.g., NVRTC_SUCCESS to "NVRTC_SUCCESS".

For unrecognized enumeration values, it returns "NVRTC_ERROR unknown".

Parameters

result – [in] CUDA Runtime Compilation API result code.

Returns

Message string for the given nvrtcResult code.

3.2. General Information Query

NVRTC defines the following function for general information query.

Functions

nvrtcResult nvrtcGetNumSupportedArchs(int *numArchs)

nvrtcGetNumSupportedArchs sets the output parameter numArchs with the number of architectures supported by NVRTC.

nvrtcResult nvrtcGetSupportedArchs(int *supportedArchs)

nvrtcGetSupportedArchs populates the array passed via the output parameter supportedArchs with the architectures supported by NVRTC.

nvrtcResult nvrtcVersion(int *major, int *minor)

nvrtcVersion sets the output parameters major and minor with the CUDA Runtime Compilation version number.

3.2.1. Functions

nvrtcResult nvrtcGetNumSupportedArchs(int *numArchs)

nvrtcGetNumSupportedArchs sets the output parameter numArchs with the number of architectures supported by NVRTC.

This can then be used to pass an array to nvrtcGetSupportedArchs to get the supported architectures.

see nvrtcGetSupportedArchs

Parameters

numArchs – [out] number of supported architectures.

Returns

nvrtcResult nvrtcGetSupportedArchs(int *supportedArchs)

nvrtcGetSupportedArchs populates the array passed via the output parameter supportedArchs with the architectures supported by NVRTC.

The array is sorted in the ascending order. The size of the array to be passed can be determined using nvrtcGetNumSupportedArchs.

see nvrtcGetNumSupportedArchs

Parameters

supportedArchs – [out] sorted array of supported architectures.

Returns

nvrtcResult nvrtcVersion(int *major, int *minor)

nvrtcVersion sets the output parameters major and minor with the CUDA Runtime Compilation version number.

Parameters

  • major – [out] CUDA Runtime Compilation major version number.

  • minor – [out] CUDA Runtime Compilation minor version number.

Returns

3.3. Compilation

NVRTC defines the following type and functions for actual compilation.

Functions

nvrtcResult nvrtcAddNameExpression(nvrtcProgram prog, const char *const name_expression)

nvrtcAddNameExpression notes the given name expression denoting the address of a global function or device /__constant__ variable.

nvrtcResult nvrtcCompileProgram(nvrtcProgram prog, int numOptions, const char *const *options)

nvrtcCompileProgram compiles the given program.

nvrtcResult nvrtcCreateProgram(nvrtcProgram *prog, const char *src, const char *name, int numHeaders, const char *const *headers, const char *const *includeNames)

nvrtcCreateProgram creates an instance of nvrtcProgram with the given input parameters, and sets the output parameter prog with it.

nvrtcResult nvrtcDestroyProgram(nvrtcProgram *prog)

nvrtcDestroyProgram destroys the given program.

nvrtcResult nvrtcGetCUBIN(nvrtcProgram prog, char *cubin)

nvrtcGetCUBIN stores the cubin generated by the previous compilation of prog in the memory pointed by cubin .

nvrtcResult nvrtcGetCUBINSize(nvrtcProgram prog, size_t *cubinSizeRet)

nvrtcGetCUBINSize sets the value of cubinSizeRet with the size of the cubin generated by the previous compilation of prog .

nvrtcResult nvrtcGetLTOIR(nvrtcProgram prog, char *LTOIR)

nvrtcGetLTOIR stores the LTO IR generated by the previous compilation of prog in the memory pointed by LTOIR .

nvrtcResult nvrtcGetLTOIRSize(nvrtcProgram prog, size_t *LTOIRSizeRet)

nvrtcGetLTOIRSize sets the value of LTOIRSizeRet with the size of the LTO IR generated by the previous compilation of prog .

nvrtcResult nvrtcGetLoweredName(nvrtcProgram prog, const char *const name_expression, const char **lowered_name)

nvrtcGetLoweredName extracts the lowered (mangled) name for a global function or device /__constant__ variable, and updates *lowered_name to point to it.

nvrtcResult nvrtcGetNVVM(nvrtcProgram prog, char *nvvm)

DEPRECATION NOTICE: This function will be removed in a future release.

nvrtcResult nvrtcGetNVVMSize(nvrtcProgram prog, size_t *nvvmSizeRet)

DEPRECATION NOTICE: This function will be removed in a future release.

nvrtcResult nvrtcGetOptiXIR(nvrtcProgram prog, char *optixir)

nvrtcGetOptiXIR stores the OptiX IR generated by the previous compilation of prog in the memory pointed by optixir .

nvrtcResult nvrtcGetOptiXIRSize(nvrtcProgram prog, size_t *optixirSizeRet)

nvrtcGetOptiXIRSize sets the value of optixirSizeRet with the size of the OptiX IR generated by the previous compilation of prog .

nvrtcResult nvrtcGetPTX(nvrtcProgram prog, char *ptx)

nvrtcGetPTX stores the PTX generated by the previous compilation of prog in the memory pointed by ptx .

nvrtcResult nvrtcGetPTXSize(nvrtcProgram prog, size_t *ptxSizeRet)

nvrtcGetPTXSize sets the value of ptxSizeRet with the size of the PTX generated by the previous compilation of prog (including the trailing NULL ).

nvrtcResult nvrtcGetProgramLog(nvrtcProgram prog, char *log)

nvrtcGetProgramLog stores the log generated by the previous compilation of prog in the memory pointed by log .

nvrtcResult nvrtcGetProgramLogSize(nvrtcProgram prog, size_t *logSizeRet)

nvrtcGetProgramLogSize sets logSizeRet with the size of the log generated by the previous compilation of prog (including the trailing NULL ).

Typedefs

nvrtcProgram

nvrtcProgram is the unit of compilation, and an opaque handle for a program.

3.3.1. Functions

nvrtcResult nvrtcAddNameExpression(nvrtcProgram prog, const char *const name_expression)

nvrtcAddNameExpression notes the given name expression denoting the address of a global function or device/__constant__ variable.

The identical name expression string must be provided on a subsequent call to nvrtcGetLoweredName to extract the lowered name.

See also

nvrtcGetLoweredName

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • name_expression – [in] constant expression denoting the address of a global function or device/__constant__ variable.

Returns

nvrtcResult nvrtcCompileProgram(nvrtcProgram prog, int numOptions, const char *const *options)

nvrtcCompileProgram compiles the given program.

It supports compile options listed in Supported Compile Options.

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • numOptions – [in] Number of compiler options passed.

  • options – [in] Compiler options in the form of C string array.options can be NULL when numOptions is 0.

Returns

nvrtcResult nvrtcCreateProgram(nvrtcProgram *prog, const char *src, const char *name, int numHeaders, const char *const *headers, const char *const *includeNames)

nvrtcCreateProgram creates an instance of nvrtcProgram with the given input parameters, and sets the output parameter prog with it.

See also

nvrtcDestroyProgram

Parameters

  • prog – [out] CUDA Runtime Compilation program.

  • src – [in] CUDA program source.

  • name – [in] CUDA program name.name can be NULL; "default_program" is used when name is NULL or “”.

  • numHeaders – [in] Number of headers used.numHeaders must be greater than or equal to 0.

  • headers – [in] Sources of the headers.headers can be NULL when numHeaders is 0.

  • includeNames – [in] Name of each header by which they can be included in the CUDA program source.includeNames can be NULL when numHeaders is 0. These headers must be included with the exact names specified here.

Returns

nvrtcResult nvrtcDestroyProgram(nvrtcProgram *prog)

nvrtcDestroyProgram destroys the given program.

See also

nvrtcCreateProgram

Parameters

prog – [in] CUDA Runtime Compilation program.

Returns

nvrtcResult nvrtcGetCUBIN(nvrtcProgram prog, char *cubin)

nvrtcGetCUBIN stores the cubin generated by the previous compilation of prog in the memory pointed by cubin.

No cubin is available if the value specified to -arch is a virtual architecture instead of an actual architecture.

See also

nvrtcGetCUBINSize

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • cubin – [out] Compiled and assembled result.

Returns

nvrtcResult nvrtcGetCUBINSize(nvrtcProgram prog, size_t *cubinSizeRet)

nvrtcGetCUBINSize sets the value of cubinSizeRet with the size of the cubin generated by the previous compilation of prog.

The value of cubinSizeRet is set to 0 if the value specified to -arch is a virtual architecture instead of an actual architecture.

See also

nvrtcGetCUBIN

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • cubinSizeRet – [out] Size of the generated cubin.

Returns

nvrtcResult nvrtcGetLTOIR(nvrtcProgram prog, char *LTOIR)

nvrtcGetLTOIR stores the LTO IR generated by the previous compilation of prog in the memory pointed by LTOIR.

No LTO IR is available if the program was compiled without -dlto.

See also

nvrtcGetLTOIRSize

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • LTOIR – [out] Compiled result.

Returns

nvrtcResult nvrtcGetLTOIRSize(nvrtcProgram prog, size_t *LTOIRSizeRet)

nvrtcGetLTOIRSize sets the value of LTOIRSizeRet with the size of the LTO IR generated by the previous compilation of prog.

The value of LTOIRSizeRet is set to 0 if the program was not compiled with -dlto.

See also

nvrtcGetLTOIR

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • LTOIRSizeRet – [out] Size of the generated LTO IR.

Returns

nvrtcResult nvrtcGetLoweredName(nvrtcProgram prog, const char *const name_expression, const char **lowered_name)

nvrtcGetLoweredName extracts the lowered (mangled) name for a global function or device/__constant__ variable, and updates *lowered_name to point to it.

The memory containing the name is released when the NVRTC program is destroyed by nvrtcDestroyProgram. The identical name expression must have been previously provided to nvrtcAddNameExpression.

See also

nvrtcAddNameExpression

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • name_expression – [in] constant expression denoting the address of a global function or device/__constant__ variable.

  • lowered_name – [out] initialized by the function to point to a C string containing the lowered (mangled) name corresponding to the provided name expression.

Returns

nvrtcResult nvrtcGetNVVM(nvrtcProgram prog, char *nvvm)

DEPRECATION NOTICE: This function will be removed in a future release.

Please use nvrtcGetLTOIR (and nvrtcGetLTOIRSize) instead.

nvrtcResult nvrtcGetNVVMSize(nvrtcProgram prog, size_t *nvvmSizeRet)

DEPRECATION NOTICE: This function will be removed in a future release.

Please use nvrtcGetLTOIRSize (and nvrtcGetLTOIR) instead.

nvrtcResult nvrtcGetOptiXIR(nvrtcProgram prog, char *optixir)

nvrtcGetOptiXIR stores the OptiX IR generated by the previous compilation of prog in the memory pointed by optixir.

No OptiX IR is available if the program was compiled with options incompatible with OptiX IR generation.

See also

nvrtcGetOptiXIRSize

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • Optix – [out] IR Compiled result.

Returns

nvrtcResult nvrtcGetOptiXIRSize(nvrtcProgram prog, size_t *optixirSizeRet)

nvrtcGetOptiXIRSize sets the value of optixirSizeRet with the size of the OptiX IR generated by the previous compilation of prog.

The value of nvrtcGetOptiXIRSize is set to 0 if the program was compiled with options incompatible with OptiX IR generation.

See also

nvrtcGetOptiXIR

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • optixirSizeRet – [out] Size of the generated LTO IR.

Returns

nvrtcResult nvrtcGetPTX(nvrtcProgram prog, char *ptx)

nvrtcGetPTX stores the PTX generated by the previous compilation of prog in the memory pointed by ptx.

See also

nvrtcGetPTXSize

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • ptx – [out] Compiled result.

Returns

nvrtcResult nvrtcGetPTXSize(nvrtcProgram prog, size_t *ptxSizeRet)

nvrtcGetPTXSize sets the value of ptxSizeRet with the size of the PTX generated by the previous compilation of prog (including the trailing NULL).

See also

nvrtcGetPTX

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • ptxSizeRet – [out] Size of the generated PTX (including the trailing NULL).

Returns

nvrtcResult nvrtcGetProgramLog(nvrtcProgram prog, char *log)

nvrtcGetProgramLog stores the log generated by the previous compilation of prog in the memory pointed by log.

See also

nvrtcGetProgramLogSize

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • log – [out] Compilation log.

Returns

nvrtcResult nvrtcGetProgramLogSize(nvrtcProgram prog, size_t *logSizeRet)

nvrtcGetProgramLogSize sets logSizeRet with the size of the log generated by the previous compilation of prog (including the trailing NULL).

Note that compilation log may be generated with warnings and informative messages, even when the compilation of prog succeeds.

See also

nvrtcGetProgramLog

Parameters

  • prog – [in] CUDA Runtime Compilation program.

  • logSizeRet – [out] Size of the compilation log (including the trailing NULL).

Returns

3.3.2. Typedefs

typedef struct _nvrtcProgram *nvrtcProgram

nvrtcProgram is the unit of compilation, and an opaque handle for a program.

To compile a CUDA program string, an instance of nvrtcProgram must be created first with nvrtcCreateProgram, then compiled with nvrtcCompileProgram.

3.4. Supported Compile Options

NVRTC supports the compile options below.

Option names with two preceding dashs (--) are long option names and option names with one preceding dash (-) are short option names. Short option names can be used instead of long option names. When a compile option takes an argument, an assignment operator (=) is used to separate the compile option argument from the compile option name, e.g., "--gpu-architecture=compute_60". Alternatively, the compile option name and the argument can be specified in separate strings without an assignment operator, .e.g, "--gpu-architecture" "compute_60". Single-character short option names, such as -D, -U, and -I, do not require an assignment operator, and the compile option name and the argument can be present in the same string with or without spaces between them. For instance, "-D=<def>", "-D<def>", and "-D <def>" are all supported.

The valid compiler options are:

  • Compilation targets

    • --gpu-architecture=<arch> (-arch

      )

      Specify the name of the class of GPU architectures for which the input must be compiled.

      • Valid <arch>s:

        • compute_50

        • compute_52

        • compute_53

        • compute_60

        • compute_61

        • compute_62

        • compute_70

        • compute_72

        • compute_75

        • compute_80

        • compute_87

        • compute_89

        • compute_90

        • compute_90a

        • sm_50

        • sm_52

        • sm_53

        • sm_60

        • sm_61

        • sm_62

        • sm_70

        • sm_72

        • sm_75

        • sm_80

        • sm_87

        • sm_89

        • sm_90

        • sm_90a

      • Default: compute_52

  • Separate compilation / whole-program compilation

    • --device-c (-dc

      )

      Generate relocatable code that can be linked with other relocatable device code. It is equivalent to

      &#8212;relocatable-device-code=true.

    • --device-w (-dw

      )

      Generate non-relocatable code. It is equivalent to

      --relocatable-device-code=false.

    • --relocatable-device-code={true|false} (-rdc

      )

      Enable (disable) the generation of relocatable device code.

      • Default: false

    • --extensible-whole-program (-ewp

      )

      Do extensible whole program compilation of device code.

      • Default: false

  • Debugging support

    • --device-debug (-G

      )

      Generate debug information. If

      &#8212;dopt is not specified, then turns off all optimizations.

    • --generate-line-info (-lineinfo

      )

      Generate line-number information.

  • Code generation

    • --dopt on (-dopt)

    • --dopt=on Enable device code optimization. When specified along with ‘-G’, enables limited debug information generation for optimized device code (currently, only line number information). When ‘-G’ is not specified, ‘-dopt=on’ is implicit.

    • --ptxas-options <options> (-Xptxas)

    • --ptxas-options=<options> Specify options directly to ptxas, the PTX optimizing assembler.

    • --maxrregcount=<N> (-maxrregcount

      )

      Specify the maximum amount of registers that GPU functions can use. Until a function-specific limit, a higher value will generally increase the performance of individual GPU threads that execute this function. However, because thread registers are allocated from a global register pool on each GPU, a higher value of this option will also reduce the maximum thread block size, thereby reducing the amount of thread parallelism. Hence, a good maxrregcount value is the result of a trade-off. If this option is not specified, then no maximum is assumed. Value less than the minimum registers required by ABI will be bumped up by the compiler to ABI minimum limit.

    • --ftz={true|false} (-ftz

      )

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

      --use_fast_math implies --ftz=true.

      • Default: false

    • --prec-sqrt={true|false} (-prec-sqrt

      )

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

      --use_fast_math implies --prec-sqrt=false.

      • Default: true

    • --prec-div={true|false} (-prec-div

      )

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

      --use_fast_math implies --prec-div=false.

      • Default: true

    • --fmad={true|false} (-fmad

      )

      Enables (disables) the contraction of floating-point multiplies and adds/subtracts into floating-point multiply-add operations (FMAD, FFMA, or DFMA).

      --use_fast_math implies --fmad=true.

      • Default: true

    • --use_fast_math (-use_fast_math

      )

      Make use of fast math operations.

      --use_fast_math implies --ftz=true --prec-div=false --prec-sqrt=false --fmad=true.

    • --extra-device-vectorization (-extra-device-vectorization

      )

      Enables more aggressive device code vectorization in the NVVM optimizer.

    • --modify-stack-limit={true|false} (-modify-stack-limit

      )

      On Linux, during compilation, use

      setrlimit() to increase stack size to maximum allowed. The limit is reset to the previous value at the end of compilation. Note: setrlimit() changes the value for the entire process.

      • Default: true

    • --dlink-time-opt (-dlto

      )

      Generate intermediate code for later link-time optimization. It implies

      -rdc=true. Note: when this option is used the nvrtcGetLTOIR API should be used, as PTX or Cubin will not be generated.

    • --gen-opt-lto (-gen-opt-lto

      )

      Run the optimizer passes before generating the LTO IR.

    • --optix-ir (-optix-ir

      )

      Generate OptiX IR. The Optix IR is only intended for consumption by OptiX through appropriate APIs. This feature is not supported with link-time-optimization (

      -dlto

      )

      . Note: when this option is used the nvrtcGetOptiX API should be used, as PTX or Cubin will not be generated.

    • --jump-table-density=[0-101] (-jtd

      )

      Specify the case density percentage in switch statements, and use it as a minimal threshold to determine whether jump table(brx.idx instruction) will be used to implement a switch statement. Default value is 101. The percentage ranges from 0 to 101 inclusively.

    • --device-stack-protector={true|false} (-device-stack-protector

      )

      Enable (disable) the generation of stack canaries in device code.

      • Default: false

  • Preprocessing

    • --define-macro=<def> (-D)<def> can be either <name> or <name=definitions>.

      • <name> Predefine <name> as a macro with definition 1.

      • <name>=<definition> The contents of <definition> are tokenized and preprocessed as if they appeared during translation phase three in a #define directive. In particular, the definition will be truncated by embedded new line characters.

    • --undefine-macro=<def> (-U

      )

      Cancel any previous definition of

      <def>.

    • --include-path=<dir> (-I

      )

      Add the directory

      <dir> to the list of directories to be searched for headers. These paths are searched after the list of headers given to nvrtcCreateProgram.

    • --pre-include=<header> (-include

      )

      Preinclude

      <header> during preprocessing.

    • --no-source-include (-no-source-include) The preprocessor by default adds the directory of each input sources to the include path. This option disables this feature and only considers the path specified explicitly.

  • Language Dialect

    • --std={c++03|c++11|c++14|c++17|c++20} (-std={c++11|c++14|c++17|c++20}

      )

      Set language dialect to C++03, C++11, C++14, C++17 or C++20

      • Default: c++17

    • --builtin-move-forward={true|false} (-builtin-move-forward

      )

      Provide builtin definitions of

      std::move and std::forward, when C++11 or later language dialect is selected.

      • Default: true

    • --builtin-initializer-list={true|false} (-builtin-initializer-list

      )

      Provide builtin definitions of

      std::initializer_list class and member functions when C++11 or later language dialect is selected.

      • Default: true

  • Misc.

    • --disable-warnings (-w

      )

      Inhibit all warning messages.

    • --restrict (-restrict

      )

      Programmer assertion that all kernel pointer parameters are restrict pointers.

    • --device-as-default-execution-space (-default-device

      )

      Treat entities with no execution space annotation as

      __device__ entities.

    • --device-int128 (-device-int128

      )

      Allow the

      __int128 type in device code. Also causes the macro __CUDACC_RTC_INT128__ to be defined.

    • --optimization-info=<kind> (-opt-info

      )

      Provide optimization reports for the specified kind of optimization. The following kind tags are supported:

      • inline : emit a remark when a function is inlined.

    • --display-error-number (-err-no

      )

      Display diagnostic number for warning messages. (Default)

    • --no-display-error-number (-no-err-no

      )

      Disables the display of a diagnostic number for warning messages.

    • --diag-error=<error-number>,… (-diag-error

      )

      Emit error for specified diagnostic message number(s). Message numbers can be separated by comma.

    • --diag-suppress=<error-number>,… (-diag-suppress

      )

      Suppress specified diagnostic message number(s). Message numbers can be separated by comma.

    • --diag-warn=<error-number>,… (-diag-warn

      )

      Emit warning for specified diagnostic message number(s). Message numbers can be separated by comma.

    • --brief-diagnostics={true|false} (-brief-diag

      )

      This option disables or enables showing source line and column info in a diagnostic. The

      &#8212;brief-diagnostics=true will not show the source line and column info.

      • Default: false

    • --time=<file-name> (-time

      )

      Generate a comma separated value table with the time taken by each compilation phase, and append it at the end of the file given as the option argument. If the file does not exist, the column headings are generated in the first row of the table. If the file name is ‘-’, the timing data is written to the compilation log.

    • --split-compile= <number of threads> (-split-compile=

      <number of threads>)

      Perform compiler optimizations in parallel. Split compilation attempts to reduce compile time by enabling the compiler to run certain optimization passes concurrently. This option accepts a numerical value that specifies the maximum number of threads the compiler can use. One can also allow the compiler to use the maximum threads available on the system by setting

      &#8212;split-compile=0. Setting &#8212;split-compile=1 will cause this option to be ignored.

    • --fdevice-syntax-only (-fdevice-syntax-only

      )

      Ends device compilation after front-end syntax checking. This option does not generate valid device code.

    • --minimal (-minimal

      )

      Omit certain language features to reduce compile time for small programs. In particular, the following are omitted:

      • Texture and surface functions and associated types, e.g., cudaTextureObject_t.

      • CUDA Runtime Functions that are provided by the cudadevrt device code library, typically named with prefix “cuda”, e.g., cudaMalloc.

      • Kernel launch from device code.

      • Types and macros associated with CUDA Runtime and Driver APIs, provided by cuda/tools/cudart/driver_types.h, typically named with prefix “cuda”, e.g., cudaError_t.

    • --device-stack-protector (-device-stack-protector

      )

      Enable stack canaries in device code. Stack canaries make it more difficult to exploit certain types of memory safety bugs involving stack-local variables. The compiler uses heuristics to assess the risk of such a bug in each function. Only those functions which are deemed high-risk make use of a stack canary.

3.5. Host Helper

NVRTC defines the following functions for easier interaction with host code.

Functions

nvrtcResult nvrtcGetTypeName(const std::type_info &tinfo, std::string *result)

nvrtcGetTypeName stores the source level name of a type in the given std::string location.

nvrtcResult nvrtcGetTypeName(std::string *result)

nvrtcGetTypeName stores the source level name of the template type argument T in the given std::string location.

3.5.1. Functions

inline nvrtcResult nvrtcGetTypeName(const std::type_info &tinfo, std::string *result)

nvrtcGetTypeName stores the source level name of a type in the given std::string location.

This function is only provided when the macro NVRTC_GET_TYPE_NAME is defined with a non-zero value. It uses abi::__cxa_demangle or UnDecorateSymbolName function calls to extract the type name, when using gcc/clang or cl.exe compilers, respectively. If the name extraction fails, it will return NVRTC_INTERNAL_ERROR, otherwise *result is initialized with the extracted name.

Windows-specific notes:

  • nvrtcGetTypeName() is not multi-thread safe because it calls UnDecorateSymbolName(), which is not multi-thread safe.

  • The returned string may contain Microsoft-specific keywords such as __ptr64 and __cdecl.

Parameters

  • tinfo – [in] reference to object of type std::type_info for a given type.

  • result – [in] pointer to std::string in which to store the type name.

Returns

template<typename T>
nvrtcResult nvrtcGetTypeName(std::string *result)

nvrtcGetTypeName stores the source level name of the template type argument T in the given std::string location.

This function is only provided when the macro NVRTC_GET_TYPE_NAME is defined with a non-zero value. It uses abi::__cxa_demangle or UnDecorateSymbolName function calls to extract the type name, when using gcc/clang or cl.exe compilers, respectively. If the name extraction fails, it will return NVRTC_INTERNAL_ERROR, otherwise *result is initialized with the extracted name.

Windows-specific notes:

  • nvrtcGetTypeName() is not multi-thread safe because it calls UnDecorateSymbolName(), which is not multi-thread safe.

  • The returned string may contain Microsoft-specific keywords such as __ptr64 and __cdecl.

Parameters

result – [in] pointer to std::string in which to store the type name.

Returns

4. Language

Unlike the offline nvcc compiler, NVRTC is meant for compiling only device CUDA C++ code. It does not accept host code or host compiler extensions in the input code, unless otherwise noted.

4.1. Execution Space

NVRTC uses __host__ as the default execution space, and it generates an error if it encounters any host code in the input. That is, if the input contains entities with explicit __host__ annotations or no execution space annotation, NVRTC will emit an error. __host__ __device__ functions are treated as device functions.

NVRTC provides a compile option, --device-as-default-execution-space (refer to Supported Compile Options), that enables an alternative compilation mode, in which entities with no execution space annotations are treated as __device__ entities.

4.2. Separate Compilation

NVRTC itself does not provide any linker. Users can, however, use the nvJitLink library or cuLinkAddData in the CUDA Driver API to link the generated relocatable PTX code with other relocatable code. To generate relocatable PTX code, the compile option --relocatable-device-code=true or --device-c is required.

4.3. Dynamic Parallelism

NVRTC supports dynamic parallelism under the following conditions:

  • Compilation target must be compute 35 or higher.

  • Either separate compilation (--relocatable-device-code=true or --device-c) or extensible whole program compilation (--extensible-whole-program) must be enabled.

  • Generated PTX must be linked against the CUDA device runtime (cudadevrt) library (refer to Separate Compilation).

Example: Dynamic Parallelism provides a simple example.

4.4. Integer Size

Different operating systems define integer type sizes differently. Linux x86_64 implements LP64, and Windows x86_64 implements LLP64.

Table 1. Integer sizes in bits for LLP64 and LP64

short

int

long

long long

pointers and size_t

LLP64

16

32

32

64

64

LP64

16

32

64

64

64

NVRTC implements LP64 on Linux and LLP64 on Windows.

NVRTC supports 128-bit integer types through the __int128 type. This can be enabled with the --device-int128 flag. 128-bit integer support is not available on Windows.

4.5. Include Syntax

When nvrtcCompileProgram() is called, the current working directory is added to the header search path used for locating files included with the quoted syntax (for example, #include "foo.h"), before the code is compiled.

4.6. Predefined Macros

  • __CUDACC_RTC__: useful for distinguishing between runtime and offline nvcc compilation in user code.

  • __CUDACC__: defined with same semantics as with offline nvcc compilation.

  • __CUDACC_RDC__: defined with same semantics as with offline nvcc compilation.

  • __CUDACC_EWP__: defined with same semantics as with offline nvcc compilation.

  • __CUDACC_DEBUG__: defined with same semantics as with offline nvcc compilation.

  • __CUDA_ARCH__: defined with same semantics as with offline nvcc compilation.

  • __CUDA_ARCH_LIST__: defined with same semantics as with offline nvcc compilation.

  • __CUDACC_VER_MAJOR__: defined with the major version number as returned by nvrtcVersion.

  • __CUDACC_VER_MINOR__: defined with the minor version number as returned by nvrtcVersion.

  • __CUDACC_VER_BUILD__: defined with the build version number.

  • __NVCC_DIAG_PRAGMA_SUPPORT__: defined with same semantics as with offline nvcc compilation.

  • __CUDACC_RTC_INT128__: defined when -device-int128 flag is specified during compilation, and indicates that __int128 type is supported.

  • NULL: null pointer constant.

  • va_start

  • va_end

  • va_arg

  • va_copy : defined when language dialect C++11 or later is selected.

  • __cplusplus

  • _WIN64 : defined on Windows platforms.

  • __LP64__ : defined on non-Windows platforms where long int and pointer types are 64-bits.

  • __cdecl : defined to empty on all platforms.

  • __ptr64 : defined to empty on Windows platforms.

  • __CUDACC_RTC_MINIMAL__: defined when -minimal flag is specified during compilation (since CUDA 12.4).

  • Macros defined in nv/target header are implicitly provided, e.g., NV_IF_TARGET.

4.7. Predefined Types

  • clock_t

  • size_t

  • ptrdiff_t

  • va_list: Note that the definition of this type may be different than the one selected by nvcc when compiling CUDA code.

  • Predefined types such as dim3, char4, etc., that are available in the CUDA Runtime headers when compiling offline with nvcc are also available, unless otherwise noted.

  • std::initializer_list<T>: implicitly provided in C++11 and later dialects, unless -builtin-initializer-list=false is specified.

  • std::move<T>, std::forward<T>: implicitly provided in C++11 and later dialects, unless -builtin-move-forward=false is specified.

4.8. Builtin Functions

Builtin functions provided by the CUDA Runtime headers when compiling offline with nvcc are available, unless otherwise noted.

4.9. Default C++ Dialect

The default C++ dialect is C++17. Other dialects can be selected using the -std flag.

5. Basic Usage

This section of the document uses a simple example, Single-Precision α⋅X Plus Y (SAXPY), shown in Figure 1 to explain what is involved in runtime compilation with NVRTC. For brevity and readability, error checks on the API return values are not shown. The complete code listing is available in Example: SAXPY.

Figure 1. CUDA source string for SAXPY

const char *saxpy = "                                          \n\
extern \"C\" __global__                                        \n\
void saxpy(float a, float *x, float *y, float *out, size_t n)  \n\
{                                                              \n\
   size_t tid = blockIdx.x * blockDim.x + threadIdx.x;         \n\
   if (tid < n) {                                              \n\
      out[tid] = a * x[tid] + y[tid];                          \n\
   }                                                           \n\
}                                                              \n";

First, an instance of nvrtcProgram needs to be created. Figure 2 shows creation of nvrtcProgram for SAXPY. As SAXPY does not require any header, 0 is passed as numHeaders, and NULL as headers and includeNames.

Figure 2. nvrtcProgram creation for SAXPY

nvrtcProgram prog;
nvrtcCreateProgram(&prog, // prog
        saxpy,         // buffer
        "saxpy.cu",    // name
        0,             // numHeaders
        NULL,          // headers
        NULL);         // includeNames

If SAXPY had any #include directives, the contents of the files that are #include’d can be passed as elements of headers, and their names as elements of includeNames. For example, #include <foo.h> and #include <bar.h> would require 2 as numHeaders, { "<contents of foo.h>", "<contents of bar.h>" } as headers, and { "foo.h", "bar.h" } as includeNames (<contents of foo.h> and <contents of bar.h> must be replaced by the actual contents of foo.h and bar.h). Alternatively, the compile option -I can be used if the header is guaranteed to exist in the file system at runtime.

Once the instance of nvrtcProgram for compilation is created, it can be compiled by nvrtcCompileProgram as shown in Figure 3. Two compile options are used in this example, --gpu-architecture=compute_80 and --fmad=false, to generate code for the compute_80 architecture and to disable the contraction of floating-point multiplies and adds/subtracts into floating-point multiply-add operations. Other combinations of compile options can be used as needed and Supported Compile Options lists valid compile options.

Figure 3. Compilation of SAXPY for compute_80 with FMAD enabled

const char *opts[] = {"--gpu-architecture=compute_80",
         "--fmad=false"};
nvrtcCompileProgram(prog,     // prog
         2,        // numOptions
         opts);    // options

After the compilation completes, users can obtain the program compilation log and the generated PTX as Figure 4 shows. NVRTC does not generate valid PTX when the compilation fails, and it may generate program compilation log even when the compilation succeeds if needed.

An nvrtcProgram can be compiled by nvrtcCompileProgram multiple times with different compile options, and users can only retrieve the PTX and the log generated by the last compilation.

Figure 4. Obtaining generated PTX and program compilation log

// Obtain compilation log from the program.

size_t logSize;

nvrtcGetProgramLogSize(prog, &logSize);
char *log = new char[logSize];
nvrtcGetProgramLog(prog, log);
// Obtain PTX from the program.
size_t ptxSize;
nvrtcGetPTXSize(prog, &ptxSize);
char *ptx = new char[ptxSize];
nvrtcGetPTX(prog, ptx);

When the instance of nvrtcProgram is no longer needed, it can be destroyed by nvrtcDestroyProgram as shown in Figure 5.

Figure 5. Destruction of nvrtcProgram

nvrtcDestroyProgram(&prog);

The generated PTX can be further manipulated by the CUDA Driver API for execution or linking. Figure 6 shows an example code sequence for execution of the generated PTX.

Figure 6. Execution of SAXPY using the PTX generated by NVRTC

CUdevice cuDevice;
CUcontext context;
CUmodule module;
CUfunction kernel;
cuInit(0);
cuDeviceGet(&cuDevice, 0);
cuCtxCreate(&context, 0, cuDevice);
cuModuleLoadDataEx(&module, ptx, 0, 0, 0);
cuModuleGetFunction(&kernel, module, "saxpy");
size_t n = size_t n = NUM_THREADS * NUM_BLOCKS;
size_t bufferSize = n * sizeof(float);
float a = ...;
float *hX = ..., *hY = ..., *hOut = ...;
CUdeviceptr dX, dY, dOut;
cuMemAlloc(&dX, bufferSize);
cuMemAlloc(&dY, bufferSize);
cuMemAlloc(&dOut, bufferSize);
cuMemcpyHtoD(dX, hX, bufferSize);
cuMemcpyHtoD(dY, hY, bufferSize);
void *args[] = { &a, &dX, &dY, &dOut, &n };
cuLaunchKernel(kernel,
            NUM_THREADS, 1, 1,   // grid dim
            NUM_BLOCKS, 1, 1,    // block dim
            0, NULL,             // shared mem and stream
            args,                // arguments
            0);
cuCtxSynchronize();
cuMemcpyDtoH(hOut, dOut, bufferSize);

6. Accessing Lowered Names

NVRTC will mangle __global__ function names and names of __device__ and __constant__ variables as specified by the IA64 ABI. If the generated PTX is being loaded using the CUDA Driver API, the kernel function or __device__/__constant__ variable must be looked up by name, but this is hard to do when the name has been mangled. To address this problem, NVRTC provides API functions that map source level __global__ function or __device__/__constant__ variable names to the mangled names present in the generated PTX.

The two API functions nvrtcAddNameExpression and nvrtcGetLoweredName work together to provide this functionality. First, a ‘name expression’ string denoting the address for the __global__ function or __device__/__constant__ variable is provided to nvrtcAddNameExpression. Then, the program is compiled with nvrtcCompileProgram. During compilation, NVRTC will parse the name expression string as a C++ constant expression at the end of the user program. The constant expression must provide the address of the __global__ function or __device__/__constant__ variable. Finally, the function nvrtcGetLoweredName is called with the original name expression and it returns a pointer to the lowered name. The lowered name can be used to refer to the kernel or variable in the CUDA Driver API.

NVRTC guarantees that any __global__ function or __device__/__constant__ variable referenced in a call to nvrtcAddNameExpression will be present in the generated PTX (if the definition is available in the input source code).

6.1. Example

Example: Using Lowered Name`_ lists a complete runnable example. Some relevant snippets:

  1. The GPU source code (‘gpu_program’) contains definitions of various __global__ functions/function templates and __device__/__constant__ variables:

    const char *gpu_program = "                                     \n\
__device__ int V1; // set from host code                        \n\
static __global__ void f1(int *result) { *result = V1 + 10; }   \n\
namespace N1 {                                                  \n\
   namespace N2 {                                               \n\
      __constant__ int V2; // set from host code                \n\
      __global__ void f2(int *result) { *result = V2 + 20; }    \n\
   }                                                            \n\
}                                                               \n\
template<typename T>                                            \n\
__global__ void f3(int *result) { *result = sizeof(T); }        \n\

  2. The host source code invokes nvrtcAddNameExpression with various name expressions referring to the address of __global__ functions and __device__/__constant__ variables:

    kernel_name_vec.push_back("&f1");
..
kernel_name_vec.push_back("N1::N2::f2");
..
kernel_name_vec.push_back("f3<int>");
..
kernel_name_vec.push_back("f3<double>");

// add name expressions to NVRTC. Note this must be done before
// the program is compiled.
for (size_t i = 0; i < name_vec.size(); ++i)
NVRTC_SAFE_CALL(nvrtcAddNameExpression(prog, kernel_name_vec[i].c_str()));
..
// add expressions for  __device__ / __constant__ variables to NVRTC
variable_name_vec.push_back("&V1");
..
variable_name_vec.push_back("&N1::N2::V2");
..
for (size_t i = 0; i < variable_name_vec.size(); ++i)
NVRTC_SAFE_CALL(nvrtcAddNameExpression(prog,
variable_name_vec[i].c_str()));

  3. The GPU program is then compiled with nvrtcCompileProgram. The generated PTX is loaded on the GPU. The mangled names of the __device__/__constant__ variables and __global__ functions are looked up:

    // note: this call must be made after NVRTC program has been
// compiled and before it has been destroyed.
NVRTC_SAFE_CALL(nvrtcGetLoweredName(
prog,
variable_name_vec[i].c_str(), // name expression
&name                         // lowered name
));
..
NVRTC_SAFE_CALL(nvrtcGetLoweredName(
prog,
kernel_name_vec[i].c_str(), // name expression
&name // lowered name
));

  4. The mangled name of the __device__/__constant__ variable is then used to lookup the variable in the module and update its value using the CUDA Driver API:

    CUdeviceptr variable_addr;
CUDA_SAFE_CALL(cuModuleGetGlobal(&variable_addr, NULL, module, name));
CUDA_SAFE_CALL(cuMemcpyHtoD(variable_addr,
&initial_value, sizeof(initial_value)));

  5. The mangled name of the kernel is then used to launch it using the CUDA Driver API:

    CUfunction kernel;
CUDA_SAFE_CALL(cuModuleGetFunction(&kernel, module, name));
...
CUDA_SAFE_CALL(
cuLaunchKernel(kernel,
1, 1, 1, // grid dim
1, 1, 1, // block dim
0, NULL, // shared mem and stream
args, 0));

6.2. Notes

  • Sequence of calls: All name expressions must be added using nvrtcAddNameExpression before the NVRTC program is compiled with nvrtcCompileProgram. This is required because the name expressions are parsed at the end of the user program, and may trigger template instantiations. The lowered names must be looked up by calling nvrtcGetLoweredName only after the NVRTC program has been compiled, and before it has been destroyed. The pointer returned by nvrtcGetLoweredName points to memory owned by NVRTC, and this memory is freed when the NVRTC program has been destroyed (nvrtcDestroyProgram). Thus the correct sequence of calls is: nvrtcAddNameExpression, nvrtcCompileProgram, nvrtcGetLoweredName, nvrtcDestroyProgram.

  • Identical Name Expressions: The name expression string passed to nvrtcAddNameExpression and nvrtcGetLoweredName must have identical characters. For example, “foo” and “foo ” are not identical strings, even though semantically they refer to the same entity (foo), because the second string has a extra whitespace character.

  • Constant Expressions: The characters in the name expression string are parsed as a C++ constant expression at the end of the user program. Any errors during parsing will cause compilation failure and compiler diagnostics will be generated in the compilation log. The constant expression must refer to the address of a __global__ function or __device__/__constant__ variable.

  • Address of overloaded function: If the NVRTC source code has multiple overloaded __global__ functions, then the name expression must use a cast operation to disambiguate. However, casts are not allowed in constant expression for C++ dialects before C++11. If using such name expressions, please compile the code in C++11 or later dialect using the -std command line flag. Example: Consider that the GPU code string contains:

    __global__ void foo(int) { }
__global__ void foo(char) { }

    The name expression (void(*)(int))foo correctly disambiguates foo(int), but the program must be compiled in C++11 or later dialect (such as -std=c++11) because casts are not allowed in pre-C++11 constant expressions.

7. Interfacing With Template Host Code

In some scenarios, it is useful to instantiate __global__ function templates in device code based on template arguments in host code. The NVRTC helper function nvrtcGetTypeName can be used to extract the source level name of a type in host code, and this string can be used to instantiate a __global__ function template and get the mangled name of the instantiation using the nvrtcAddNameExpression and nvrtcGetLoweredName functions.

nvrtcGetTypeName is defined inline in the NVRTC header file, and is available when the macro NVRTC_GET_TYPE_NAME is defined with a non-zero value. It uses the abi::__cxa_demangle and UnDecorateSymbolName host code functions when using gcc/clang and cl.exe compilers, respectively. Users may need to specify additional header paths and libraries to find the host functions used (abi::__cxa_demangle / UnDecorateSymbolName). Refer to the build instructions for the example below for reference (nvrtcGetTypeName Build Instructions).

7.1. Template Host Code Example

Example: Using nvrtcGetTypeName lists a complete runnable example. Some relevant snippets:

  1. The GPU source code (gpu_program) contains definitions of a __global__ function template:

    const char *gpu_program = " \n\
namespace N1 { struct S1_t { int i; double d; }; } \n\
template<typename T> \n\
__global__ void f3(int *result) { *result = sizeof(T); } \n\
\n";

  2. The host code function getKernelNameForType creates the name expression for a __global__ function template instantiation based on the host template type T. The name of the type T is extracted using nvrtcGetTypeName:

    template <typename T>
std::string getKernelNameForType(void)
{
// Look up the source level name string for the type "T" using
// nvrtcGetTypeName() and use it to create the kernel name
std::string type_name;
NVRTC_SAFE_CALL(nvrtcGetTypeName<T>(&type_name));
return std::string("f3<") + type_name + ">";
}

  3. The name expressions are presented to NVRTC using the nvrtcAddNameExpression function:

    name_vec.push_back(getKernelNameForType<int>());
..
name_vec.push_back(getKernelNameForType<double>());
..
name_vec.push_back(getKernelNameForType<N1::S1_t>());
..
for (size_t i = 0; i < name_vec.size(); ++i)
NVRTC_SAFE_CALL(nvrtcAddNameExpression(prog, name_vec[i].c_str()));

  4. The GPU program is then compiled with nvrtcCompileProgram. The generated PTX is loaded on the GPU. The mangled names of the __global__ function template instantiations are looked up:

    // note: this call must be made after NVRTC program has been
// compiled and before it has been destroyed.
NVRTC_SAFE_CALL(nvrtcGetLoweredName(
prog,
name_vec[i].c_str(), // name expression
&name // lowered name
));

  5. The mangled name is then used to launch the kernel using the CUDA Driver API:

    CUfunction kernel;
CUDA_SAFE_CALL(cuModuleGetFunction(&kernel, module, name));
...
CUDA_SAFE_CALL(
cuLaunchKernel(kernel,
1, 1, 1, // grid dim
1, 1, 1, // block dim
0, NULL, // shared mem and stream
args, 0));

8. Versioning Scheme

8.1. NVRTC Shared Library Versioning

In the following, MAJOR and MINOR denote the major and minor versions of the CUDA Toolkit. For example, for CUDA 11.2, MAJOR is “11” and MINOR is “2”.

  • Linux:

    • In CUDA toolkits prior to CUDA 11.3, the soname was set to “MAJOR.MINOR”.

    • In CUDA 11.3 and later 11.x toolkits, the soname field is set to “11.2”.

    • In CUDA toolkits with major version > 11 (e.g. CUDA 12.x), the soname field is set to “MAJOR”.

  • Windows:

    • In CUDA toolkits prior to cuda 11.3, the DLL name was of the form “nvrtc64_XY_0.dll”, where X = MAJOR, Y = MINOR.

    • In CUDA 11.3 and later 11.x toolkits, the DLL name is “nvrtc64_112_0.dll”.

    • In CUDA toolkits with major version > 11 (e.g. CUDA 12.x), the DLL name is of the form “nvrtc64_X0_0.dll” where X = MAJOR.

Consider a CUDA toolkit with major version > 11. The NVRTC shared library in this CUDA toolkit will have the same soname (Linux) or DLL name (Windows) as an NVRTC shared library in a previous minor version of the same CUDA toolkit. Similarly, the NVRTC shared library in CUDA 11.3 and later 11.x releases will have the same soname (Linux) or DLL name (Windows) as the NVRTC shared library in CUDA 11.2.

As a consequence of the versioning scheme described above, an NVRTC client that links against a particular NVRTC shared library will continue to work with a future NVRTC shared library with a matching soname (Linux) or DLL name (Windows). This allows the NVRTC client to take advantage of bug fixes and enhancements available in the more recent NVRTC shared library 1. However, the more recent NVRTC shared library may generate PTX with a version that is not accepted by the CUDA Driver API functions of an older CUDA driver, as explained in the Best Practices Guide.

Some approaches to resolving this issue:

  • Install a more recent CUDA driver that is compatible with the CUDA toolkit containing the NVRTC library being used.

  • Compile directly to SASS instead of PTX with NVRTC (refer to Best Practices Guide).

Alternately, an NVRTC client can either link against the static NVRTC library or redistribute a specific version of the NVRTC shared library and use dlopen (Linux) or LoadLibrary (Windows) functions to use that library at run time. Either approach allows the NVRTC client to maintain control over the version of NVRTC being used during deployment, to ensure predictable functionality and performance.

8.2. NVRTC-builtins Library

The NVRTC-builtins library contains helper code that is part of the NVRTC package. It is only used by the NVRTC library internally. Each NVRTC library is only compatible with the NVRTC-builtins library from the same CUDA toolkit.

9. Miscellaneous Notes

9.1. Thread Safety

Multiple threads can invoke NVRTC API functions concurrently, as long as there is no race condition. In this context, a race condition is defined to occur if multiple threads concurrently invoke NVRTC API functions with the same nvrtcProgram argument, where at least one thread is invoking either nvrtcCompileProgram or nvrtcAddNameExpression 2.

Since CUDA 12.3, NVRTC allows concurrent invocations of nvrtcCompileProgram to potentially concurrently also invoke the embedded NVVM optimizer/codegen phase. Setting the environment variable NVRTC_DISABLE_CONCURRENT_NVVM disables this behavior, i.e., invocations of the embedded NVVM optimizer/codegen phase will be serialized.

9.2. Stack Size

On Linux, NVRTC will increase the stack size to the maximum allowed using the setrlimit() function during compilation. This reduces the chance that the compiler will run out of stack when processing complex input sources. The stack size is reset to the previous value when compilation is completed.

Because setrlimit() changes the stack size for the entire process, it will also affect other application threads that may be executing concurrently. The command line flag -modify-stack-limit=false will prevent NVRTC from modifying the stack limit.

9.3. NVRTC Static Library

The NVRTC static library references functions defined in the NVRTC-builtins static library and the PTX compiler static library. Please see Build Instructions for an example.

10. Example: SAXPY

10.1. Code (saxpy.cpp) 

#include <nvrtc.h>
#include <cuda.h>
#include <iostream>

#define NUM_THREADS 128
#define NUM_BLOCKS 32
#define NVRTC_SAFE_CALL(x)                                        \
  do {                                                            \
    nvrtcResult result = x;                                       \
    if (result != NVRTC_SUCCESS) {                                \
      std::cerr << "\nerror: " #x " failed with error "           \
                << nvrtcGetErrorString(result) << '\n';           \
      exit(1);                                                    \
    }                                                             \
} while(0)
#define CUDA_SAFE_CALL(x)                                         \
  do {                                                            \
    CUresult result = x;                                          \
    if (result != CUDA_SUCCESS) {                                 \
      const char *msg;                                            \
      cuGetErrorName(result, &msg);                               \
      std::cerr << "\nerror: " #x " failed with error "           \
                << msg << '\n';                                   \
      exit(1);                                                    \
    }                                                             \
} while(0)

const char *saxpy = "                                           \n\
extern \"C\" __global__                                         \n\
void saxpy(float a, float *x, float *y, float *out, size_t n)   \n\
{                                                               \n\
  size_t tid = blockIdx.x * blockDim.x + threadIdx.x;           \n\
  if (tid < n) {                                                \n\
    out[tid] = a * x[tid] + y[tid];                             \n\
  }                                                             \n\
}                                                               \n";

int main()
{
   // Create an instance of nvrtcProgram with the SAXPY code string.
   nvrtcProgram prog;
   NVRTC_SAFE_CALL(
      nvrtcCreateProgram(&prog,         // prog
                        saxpy,         // buffer
                        "saxpy.cu",    // name
                        0,             // numHeaders
                        NULL,          // headers
                        NULL));        // includeNames
   // Compile the program with fmad disabled.
   // Note: Can specify GPU target architecture explicitly with '-arch' flag.
   const char *opts[] = {"--fmad=false"};
   nvrtcResult compileResult = nvrtcCompileProgram(prog,  // prog
                                                   1,     // numOptions
                                                   opts); // options
   // Obtain compilation log from the program.
   size_t logSize;
   NVRTC_SAFE_CALL(nvrtcGetProgramLogSize(prog, &logSize));
   char *log = new char[logSize];
   NVRTC_SAFE_CALL(nvrtcGetProgramLog(prog, log));
   std::cout << log << '\n';
   delete[] log;
   if (compileResult != NVRTC_SUCCESS) {
      exit(1);
   }
   // Obtain PTX from the program.
   size_t ptxSize;
   NVRTC_SAFE_CALL(nvrtcGetPTXSize(prog, &ptxSize));
   char *ptx = new char[ptxSize];
   NVRTC_SAFE_CALL(nvrtcGetPTX(prog, ptx));
   // Destroy the program.
   NVRTC_SAFE_CALL(nvrtcDestroyProgram(&prog));
   // Load the generated PTX and get a handle to the SAXPY kernel.
   CUdevice cuDevice;
   CUcontext context;
   CUmodule module;
   CUfunction kernel;
   CUDA_SAFE_CALL(cuInit(0));
   CUDA_SAFE_CALL(cuDeviceGet(&cuDevice, 0));
   CUDA_SAFE_CALL(cuCtxCreate(&context, 0, cuDevice));
   CUDA_SAFE_CALL(cuModuleLoadDataEx(&module, ptx, 0, 0, 0));
   CUDA_SAFE_CALL(cuModuleGetFunction(&kernel, module, "saxpy"));
   // Generate input for execution, and create output buffers.
   size_t n = NUM_THREADS * NUM_BLOCKS;
   size_t bufferSize = n * sizeof(float);
   float a = 5.1f;
   float *hX = new float[n], *hY = new float[n], *hOut = new float[n];
   for (size_t i = 0; i < n; ++i) {
      hX[i] = static_cast<float>(i);
      hY[i] = static_cast<float>(i * 2);
   }
   CUdeviceptr dX, dY, dOut;
   CUDA_SAFE_CALL(cuMemAlloc(&dX, bufferSize));
   CUDA_SAFE_CALL(cuMemAlloc(&dY, bufferSize));
   CUDA_SAFE_CALL(cuMemAlloc(&dOut, bufferSize));
   CUDA_SAFE_CALL(cuMemcpyHtoD(dX, hX, bufferSize));
   CUDA_SAFE_CALL(cuMemcpyHtoD(dY, hY, bufferSize));
   // Execute SAXPY.
   void *args[] = { &a, &dX, &dY, &dOut, &n };
   CUDA_SAFE_CALL(
      cuLaunchKernel(kernel,
                     NUM_BLOCKS, 1, 1,    // grid dim
                     NUM_THREADS, 1, 1,   // block dim
                     0, NULL,             // shared mem and stream
                     args, 0));           // arguments
   CUDA_SAFE_CALL(cuCtxSynchronize());
   // Retrieve and print output.
   CUDA_SAFE_CALL(cuMemcpyDtoH(hOut, dOut, bufferSize));
   for (size_t i = 0; i < n; ++i) {
      std::cout << a << " * " << hX[i] << " + " << hY[i]
               << " = " << hOut[i] << '\n';
   }
   // Release resources.
   CUDA_SAFE_CALL(cuMemFree(dX));
   CUDA_SAFE_CALL(cuMemFree(dY));
   CUDA_SAFE_CALL(cuMemFree(dOut));
   CUDA_SAFE_CALL(cuModuleUnload(module));
   CUDA_SAFE_CALL(cuCtxDestroy(context));
   delete[] hX;
   delete[] hY;
   delete[] hOut;
   delete[] ptx;
   return 0;
}

10.2. Host Type Name Build Instructions

Assuming the environment variable CUDA_PATH points to the CUDA Toolkit installation directory, build this example as:

  • With NVRTC shared library:

    • Windows:

      cl.exe saxpy.cpp /Fesaxpy ^
   /I "%CUDA_PATH%"\include ^
   "%CUDA_PATH%"\lib\x64\nvrtc.lib "%CUDA_PATH%"\lib\x64\cuda.lib

    • Linux:

      g++ saxpy.cpp -o saxpy \
   -I $CUDA_PATH/include \
   -L $CUDA_PATH/lib64 \
   -lnvrtc -lcuda \
   -Wl,-rpath,$CUDA_PATH/lib64

  • With NVRTC static library:

    • Windows:

      cl.exe saxpy.cpp /Fesaxpy  ^
    /I "%CUDA_PATH%"\include ^
    "%CUDA_PATH%"\lib\x64\nvrtc_static.lib ^
    "%CUDA_PATH%"\lib\x64\nvrtc-builtins_static.lib ^
    "%CUDA_PATH%"\lib\x64\nvptxcompiler_static.lib ^
    "%CUDA_PATH%"\lib\x64\cuda.lib user32.lib Ws2_32.lib

    • Linux:

      g++ saxpy.cpp -o saxpy \
   -I $CUDA_PATH/include \
   -L $CUDA_PATH/lib64 \
   -lnvrtc_static -lnvrtc-builtins_static -lnvptxcompiler_static -lcuda \
   -lpthread

11. Example: Using Lowered Name

11.1. Code (lowered-name.cpp) 

#include <nvrtc.h>
#include <cuda.h>
#include <iostream>
#include <vector>
#include <string>

#define NVRTC_SAFE_CALL(x)                                        \
do {                                                              \
   nvrtcResult result = x;                                        \
   if (result != NVRTC_SUCCESS) {                                 \
      std::cerr << "\nerror: " #x " failed with error "           \
               << nvrtcGetErrorString(result) << '\n';            \
      exit(1);                                                    \
   }                                                              \
} while(0)
#define CUDA_SAFE_CALL(x)                                         \
do {                                                              \
   CUresult result = x;                                           \
   if (result != CUDA_SUCCESS) {                                  \
      const char *msg;                                            \
      cuGetErrorName(result, &msg);                               \
      std::cerr << "\nerror: " #x " failed with error "           \
               << msg << '\n';                                    \
      exit(1);                                                    \
   }                                                              \
} while(0)

const char *gpu_program = "
  __device__ int V1; // set from host code                        \n\
  static __global__ void f1(int *result) { *result = V1 + 10; }   \n\
  namespace N1 {                                                  \n\
    namespace N2 {                                                \n\
       __constant__ int V2; // set from host code                 \n\
       __global__ void f2(int *result) { *result = V2 + 20; }     \n\
      }                                                           \n\
   }                                                              \n\
   template<typename T>                                           \n\
   __global__ void f3(int *result) { *result = sizeof(T); }       \n\
                                                                  \n";

int main()
{
   // Create an instance of nvrtcProgram
   nvrtcProgram prog;
   NVRTC_SAFE_CALL(nvrtcCreateProgram(&prog,         // prog
                                       gpu_program,   // buffer
                                       "prog.cu",     // name
                                       0,             // numHeaders
                                       NULL,          // headers
                                       NULL));        // includeNames

   // add all name expressions for kernels
   std::vector<std::string> kernel_name_vec;
   std::vector<std::string> variable_name_vec;
   std::vector<int> variable_initial_value;

   std::vector<int> expected_result;

   // note the name expressions are parsed as constant expressions
   kernel_name_vec.push_back("&f1");
   expected_result.push_back(10 + 100);

   kernel_name_vec.push_back("N1::N2::f2");
   expected_result.push_back(20 + 200);

   kernel_name_vec.push_back("f3<int>");
   expected_result.push_back(sizeof(int));

   kernel_name_vec.push_back("f3<double>");
   expected_result.push_back(sizeof(double));

   // add kernel name expressions to NVRTC. Note this must be done before
   // the program is compiled.
   for (size_t i = 0; i < kernel_name_vec.size(); ++i)
      NVRTC_SAFE_CALL(nvrtcAddNameExpression(prog, kernel_name_vec[i].c_str()));

   // add expressions for  __device__ / __constant__ variables to NVRTC
   variable_name_vec.push_back("&V1");
   variable_initial_value.push_back(100);

   variable_name_vec.push_back("&N1::N2::V2");
   variable_initial_value.push_back(200);

   for (size_t i = 0; i < variable_name_vec.size(); ++i)
      NVRTC_SAFE_CALL(nvrtcAddNameExpression(prog, variable_name_vec[i].c_str()));

   nvrtcResult compileResult = nvrtcCompileProgram(prog,  // prog
                                                   0,     // numOptions
                                                   NULL); // options
   // Obtain compilation log from the program.
   size_t logSize;
   NVRTC_SAFE_CALL(nvrtcGetProgramLogSize(prog, &logSize));
   char *log = new char[logSize];
   NVRTC_SAFE_CALL(nvrtcGetProgramLog(prog, log));
   std::cout << log << '\n';
   delete[] log;
   if (compileResult != NVRTC_SUCCESS) {
      exit(1);
   }
   // Obtain PTX from the program.
   size_t ptxSize;
   NVRTC_SAFE_CALL(nvrtcGetPTXSize(prog, &ptxSize));
   char *ptx = new char[ptxSize];
   NVRTC_SAFE_CALL(nvrtcGetPTX(prog, ptx));
   // Load the generated PTX
   CUdevice cuDevice;
   CUcontext context;
   CUmodule module;

   CUDA_SAFE_CALL(cuInit(0));
   CUDA_SAFE_CALL(cuDeviceGet(&cuDevice, 0));
   CUDA_SAFE_CALL(cuCtxCreate(&context, 0, cuDevice));
   CUDA_SAFE_CALL(cuModuleLoadDataEx(&module, ptx, 0, 0, 0));

   CUdeviceptr dResult;
   int hResult = 0;
   CUDA_SAFE_CALL(cuMemAlloc(&dResult, sizeof(hResult)));
   CUDA_SAFE_CALL(cuMemcpyHtoD(dResult, &hResult, sizeof(hResult)));

   // for each of the __device__/__constant__ variable address
   // expressions provided to NVRTC, extract the lowered name for the
   // corresponding variable, and set its value
   for (size_t i = 0; i < variable_name_vec.size(); ++i) {
      const char *name;

      // note: this call must be made after NVRTC program has been
      // compiled and before it has been destroyed.
      NVRTC_SAFE_CALL(nvrtcGetLoweredName(
                           prog,
            variable_name_vec[i].c_str(), // name expression
            &name                         // lowered name
                                          ));
      int initial_value = variable_initial_value[i];

      // get pointer to variable using lowered name, and set its
      // initial value
      CUdeviceptr variable_addr;
      CUDA_SAFE_CALL(cuModuleGetGlobal(&variable_addr, NULL, module, name));
      CUDA_SAFE_CALL(cuMemcpyHtoD(variable_addr, &initial_value, sizeof(initial_value)));
   }


   // for each of the kernel name expressions previously provided to NVRTC,
   // extract the lowered name for corresponding __global__ function,
   // and launch it.

   for (size_t i = 0; i < kernel_name_vec.size(); ++i) {
      const char *name;

      // note: this call must be made after NVRTC program has been
      // compiled and before it has been destroyed.
      NVRTC_SAFE_CALL(nvrtcGetLoweredName(
                           prog,
            kernel_name_vec[i].c_str(), // name expression
            &name                // lowered name
                                          ));

      // get pointer to kernel from loaded PTX
      CUfunction kernel;
      CUDA_SAFE_CALL(cuModuleGetFunction(&kernel, module, name));

      // launch the kernel
      std::cout << "\nlaunching " << name << " ("
            << kernel_name_vec[i] << ")" << std::endl;

      void *args[] = { &dResult };
      CUDA_SAFE_CALL(
         cuLaunchKernel(kernel,
            1, 1, 1,             // grid dim
            1, 1, 1,             // block dim
            0, NULL,             // shared mem and stream
            args, 0));           // arguments
      CUDA_SAFE_CALL(cuCtxSynchronize());

      // Retrieve the result
      CUDA_SAFE_CALL(cuMemcpyDtoH(&hResult, dResult, sizeof(hResult)));

      // check against expected value
      if (expected_result[i] != hResult) {
         std::cout << "\n Error: expected result = " << expected_result[i]
                  << " , actual result = " << hResult << std::endl;
         exit(1);
      }
   }  // for

   // Release resources.
   CUDA_SAFE_CALL(cuMemFree(dResult));
   CUDA_SAFE_CALL(cuModuleUnload(module));
   CUDA_SAFE_CALL(cuCtxDestroy(context));
   delete[] ptx;

   // Destroy the program.
   NVRTC_SAFE_CALL(nvrtcDestroyProgram(&prog));

   return 0;
}

11.2. Lowered Name Build Instructions

Assuming the environment variable CUDA_PATH points to CUDA Toolkit installation directory, build this example as:

  • With NVRTC shared library:

    • Windows:

      cl.exe lowered-name.cpp /Felowered-name ^
/I "%CUDA_PATH%"\include ^
"%CUDA_PATH%"\lib\x64\nvrtc.lib "%CUDA_PATH%"\lib\x64\cuda.lib

    • Linux:

      g++ lowered-name.cpp -o lowered-name \
-I $CUDA_PATH/include \
-L $CUDA_PATH/lib64 \
-lnvrtc -lcuda \
-Wl,-rpath,$CUDA_PATH/lib64

  • With NVRTC static library:

    • Windows:

      cl.exe lowered-name.cpp /Felowered-name  ^
/I "%CUDA_PATH%"\include ^
"%CUDA_PATH%"\lib\x64\nvrtc_static.lib ^
"%CUDA_PATH%"\lib\x64\nvrtc-builtins_static.lib ^
"%CUDA_PATH%"\lib\x64\nvptxcompiler_static.lib ^
"%CUDA_PATH%"\lib\x64\cuda.lib user32.lib Ws2_32.lib

    • Linux:

      g++ lowered-name.cpp -o lowered-name \
-I $CUDA_PATH/include \
-L $CUDA_PATH/lib64 \
-lnvrtc_static -lnvrtc-builtins_static -lnvptxcompiler_static \
-lcuda -lpthread

12. Example: Using nvrtcGetTypeName

12.1. Code (host-type-name.cpp) 

#include <nvrtc.h>
#include <cuda.h>
#include <iostream>
#include <vector>
#include <string>

#define NVRTC_SAFE_CALL(x)                                        \
do {                                                            \
   nvrtcResult result = x;                                       \
   if (result != NVRTC_SUCCESS) {                                \
      std::cerr << "\nerror: " #x " failed with error "           \
               << nvrtcGetErrorString(result) << '\n';           \
      exit(1);                                                    \
   }                                                             \
} while(0)
#define CUDA_SAFE_CALL(x)                                         \
do {                                                            \
   CUresult result = x;                                          \
   if (result != CUDA_SUCCESS) {                                 \
      const char *msg;                                            \
      cuGetErrorName(result, &msg);                               \
      std::cerr << "\nerror: " #x " failed with error "           \
               << msg << '\n';                                   \
      exit(1);                                                    \
   }                                                             \
} while(0)

const char *gpu_program = "                                     \n\
namespace N1 { struct S1_t { int i; double d; }; }              \n\
template<typename T>                                            \n\
__global__ void f3(int *result) { *result = sizeof(T); }        \n\
                                                               \n";


// note: this structure is also defined in GPU code string. Should ideally
// be in a header file included by both GPU code string and by CPU code.
namespace N1 { struct S1_t { int i; double d; }; };
template <typename T>
std::string getKernelNameForType(void)
{
   // Look up the source level name string for the type "T" using
   // nvrtcGetTypeName() and use it to create the kernel name
   std::string type_name;
   NVRTC_SAFE_CALL(nvrtcGetTypeName<T>(&type_name));
   return std::string("f3<") + type_name + ">";
}

int main()
{
// Create an instance of nvrtcProgram
nvrtcProgram prog;
NVRTC_SAFE_CALL(
   nvrtcCreateProgram(&prog,         // prog
                     gpu_program,   // buffer
                     "gpu_program.cu",    // name
                     0,             // numHeaders
                     NULL,          // headers
                     NULL));        // includeNames

// add all name expressions for kernels
std::vector<std::string> name_vec;
std::vector<int> expected_result;

// note the name expressions are parsed as constant expressions
name_vec.push_back(getKernelNameForType<int>());
expected_result.push_back(sizeof(int));

name_vec.push_back(getKernelNameForType<double>());
expected_result.push_back(sizeof(double));

name_vec.push_back(getKernelNameForType<N1::S1_t>());
expected_result.push_back(sizeof(N1::S1_t));


// add name expressions to NVRTC. Note this must be done before
// the program is compiled.
for (size_t i = 0; i < name_vec.size(); ++i)
   NVRTC_SAFE_CALL(nvrtcAddNameExpression(prog, name_vec[i].c_str()));

nvrtcResult compileResult = nvrtcCompileProgram(prog,  // prog
                                                0,     // numOptions
                                                NULL); // options
// Obtain compilation log from the program.
size_t logSize;
NVRTC_SAFE_CALL(nvrtcGetProgramLogSize(prog, &logSize));
char *log = new char[logSize];
NVRTC_SAFE_CALL(nvrtcGetProgramLog(prog, log));
std::cout << log << '\n';
delete[] log;
if (compileResult != NVRTC_SUCCESS) {
   exit(1);
}
// Obtain PTX from the program.
size_t ptxSize;
NVRTC_SAFE_CALL(nvrtcGetPTXSize(prog, &ptxSize));
char *ptx = new char[ptxSize];
NVRTC_SAFE_CALL(nvrtcGetPTX(prog, ptx));


// Load the generated PTX
CUdevice cuDevice;
CUcontext context;
CUmodule module;

CUDA_SAFE_CALL(cuInit(0));
CUDA_SAFE_CALL(cuDeviceGet(&cuDevice, 0));
CUDA_SAFE_CALL(cuCtxCreate(&context, 0, cuDevice));
CUDA_SAFE_CALL(cuModuleLoadDataEx(&module, ptx, 0, 0, 0));

CUdeviceptr dResult;
int hResult = 0;
CUDA_SAFE_CALL(cuMemAlloc(&dResult, sizeof(hResult)));
CUDA_SAFE_CALL(cuMemcpyHtoD(dResult, &hResult, sizeof(hResult)));

// for each of the name expressions previously provided to NVRTC,
// extract the lowered name for corresponding __global__ function,
// and launch it.

for (size_t i = 0; i < name_vec.size(); ++i) {
   const char *name;

   // note: this call must be made after NVRTC program has been
   // compiled and before it has been destroyed.
   NVRTC_SAFE_CALL(nvrtcGetLoweredName(
                        prog,
         name_vec[i].c_str(), // name expression
         &name                // lowered name
                                       ));

   // get pointer to kernel from loaded PTX
   CUfunction kernel;
   CUDA_SAFE_CALL(cuModuleGetFunction(&kernel, module, name));

   // launch the kernel
   std::cout << "\nlaunching " << name << " ("
         << name_vec[i] << ")" << std::endl;

   void *args[] = { &dResult };
   CUDA_SAFE_CALL(
      cuLaunchKernel(kernel,
         1, 1, 1,             // grid dim
         1, 1, 1,             // block dim
         0, NULL,             // shared mem and stream
         args, 0));           // arguments
   CUDA_SAFE_CALL(cuCtxSynchronize());

   // Retrieve the result
   CUDA_SAFE_CALL(cuMemcpyDtoH(&hResult, dResult, sizeof(hResult)));

   // check against expected value
   if (expected_result[i] != hResult) {
      std::cout << "\n Error: expected result = " << expected_result[i]
      << " , actual result = " << hResult << std::endl;
      exit(1);
   }
}  // for

// Release resources.
CUDA_SAFE_CALL(cuMemFree(dResult));
CUDA_SAFE_CALL(cuModuleUnload(module));
CUDA_SAFE_CALL(cuCtxDestroy(context));
delete[] ptx;

// Destroy the program.
NVRTC_SAFE_CALL(nvrtcDestroyProgram(&prog));

return 0;
}

12.2. nvrtcGetTypeName Build Instructions

Assuming the environment variable CUDA_PATH points to CUDA Toolkit installation directory, build this example as:

  • With NVRTC shared library:

    • Windows:

      cl.exe -DNVRTC_GET_TYPE_NAME=1 host-type-name.cpp /Fehost-type-name ^
   /I "%CUDA_PATH%"\include ^
   "%CUDA_PATH%"\lib\x64\nvrtc.lib "%CUDA_PATH%"\lib\x64\cuda.lib DbgHelp.lib

    • Linux:

      g++ -DNVRTC_GET_TYPE_NAME=1 host-type-name.cpp -o host-type-name \
   -I $CUDA_PATH/include \
   -L $CUDA_PATH/lib64 \
   -lnvrtc -lcuda \
   -Wl,-rpath,$CUDA_PATH/lib64

  • With NVRTC static library:

    • Windows:

      cl.exe -DNVRTC_GET_TYPE_NAME=1 host-type-name.cpp /Fehost-type-name  ^
   /I "%CUDA_PATH%"\include ^
   "%CUDA_PATH%"\lib\x64\nvrtc_static.lib ^
   "%CUDA_PATH%"\lib\x64\nvrtc-builtins_static.lib ^
   "%CUDA_PATH%"\lib\x64\nvptxcompiler_static.lib ^
   "%CUDA_PATH%"\lib\x64\cuda.lib DbgHelp.lib user32.lib Ws2_32.lib

    • Linux:

      g++ -DNVRTC_GET_TYPE_NAME=1 host-type-name.cpp -o host-type-name \
   -I $CUDA_PATH/include \
   -L $CUDA_PATH/lib64 \
   -lnvrtc_static -lnvrtc-builtins_static -lnvptxcompiler_static \
   -lcuda -lpthread

13. Example: Dynamic Parallelism

Code (dynamic-parallelism.cpp)

#include <nvrtc.h>
#include <cuda.h>
#include <iostream>

#define NVRTC_SAFE_CALL(x)                                        \
do {                                                              \
   nvrtcResult result = x;                                        \
   if (result != NVRTC_SUCCESS) {                                 \
      std::cerr << "\nerror: " #x " failed with error "           \
                << nvrtcGetErrorString(result) << '\n';           \
      exit(1);                                                    \
   }                                                              \
} while(0)
#define CUDA_SAFE_CALL(x)                                         \
do {                                                              \
   CUresult result = x;                                           \
   if (result != CUDA_SUCCESS) {                                  \
      const char *msg;                                            \
      cuGetErrorName(result, &msg);                               \
      std::cerr << "\nerror: " #x " failed with error "           \
                << msg << '\n';                                   \
      exit(1);                                                    \
   }                                                              \
} while(0)

const char *dynamic_parallelism = "                             \n\
extern \"C\" __global__                                         \n\
void child(float *out, size_t n)                                \n\
{                                                               \n\
   size_t tid = blockIdx.x * blockDim.x + threadIdx.x;          \n\
   if (tid < n) {                                               \n\
      out[tid] = tid;                                           \n\
   }                                                            \n\
}                                                               \n\
                                                                \n\
extern \"C\" __global__                                         \n\
void parent(float *out, size_t n,                               \n\
            size_t numBlocks, size_t numThreads)                \n\
{                                                               \n\
   child<<<numBlocks, numThreads>>>(out, n);                    \n\
   cudaDeviceSynchronize();                                     \n\
}                                                               \n";
int main(int argc, char *argv[])
{
if (argc < 2) {
   std::cout << "Usage: dynamic-parallelism <path to cudadevrt library>\n\n"
             << "<path to cudadevrt library> must include the cudadevrt\n"
             << "library name itself, e.g., Z:\\path\\to\\cudadevrt.lib on \n"
             << "Windows and /path/to/libcudadevrt.a on Linux.\n";
   exit(1);
}
size_t numBlocks = 32;
size_t numThreads = 128;
// Create an instance of nvrtcProgram with the code string.
nvrtcProgram prog;
NVRTC_SAFE_CALL(
   nvrtcCreateProgram(&prog,                       // prog
                     dynamic_parallelism,          // buffer
                     "dynamic_parallelism.cu",     // name
                     0,                            // numHeaders
                     NULL,                         // headers
                     NULL));                       // includeNames
// Compile the program for compute_35 with rdc enabled.
const char *opts[] = {"--gpu-architecture=compute_35",
                        "--relocatable-device-code=true"};
nvrtcResult compileResult = nvrtcCompileProgram(prog,  // prog
                                                2,     // numOptions
                                                opts); // options
// Obtain compilation log from the program.
size_t logSize;
NVRTC_SAFE_CALL(nvrtcGetProgramLogSize(prog, &logSize));
char *log = new char[logSize];
NVRTC_SAFE_CALL(nvrtcGetProgramLog(prog, log));
std::cout << log << '\n';
delete[] log;
if (compileResult != NVRTC_SUCCESS) {
   exit(1);
}
// Obtain PTX from the program.
size_t ptxSize;
NVRTC_SAFE_CALL(nvrtcGetPTXSize(prog, &ptxSize));
char *ptx = new char[ptxSize];
NVRTC_SAFE_CALL(nvrtcGetPTX(prog, ptx));
// Destroy the program.
NVRTC_SAFE_CALL(nvrtcDestroyProgram(&prog));
// Load the generated PTX and get a handle to the parent kernel.
CUdevice cuDevice;
CUcontext context;
CUlinkState linkState;
CUmodule module;
CUfunction kernel;
CUDA_SAFE_CALL(cuInit(0));
CUDA_SAFE_CALL(cuDeviceGet(&cuDevice, 0));
CUDA_SAFE_CALL(cuCtxCreate(&context, 0, cuDevice));
CUDA_SAFE_CALL(cuLinkCreate(0, 0, 0, &linkState));
CUDA_SAFE_CALL(cuLinkAddFile(linkState, CU_JIT_INPUT_LIBRARY, argv[1],
                              0, 0, 0));
CUDA_SAFE_CALL(cuLinkAddData(linkState, CU_JIT_INPUT_PTX,
                              (void *)ptx, ptxSize, "dynamic_parallelism.ptx",
                              0, 0, 0));
size_t cubinSize;
void *cubin;
CUDA_SAFE_CALL(cuLinkComplete(linkState, &cubin, &cubinSize));
CUDA_SAFE_CALL(cuModuleLoadData(&module, cubin));
CUDA_SAFE_CALL(cuModuleGetFunction(&kernel, module, "parent"));
// Generate input for execution, and create output buffers.
size_t n = numBlocks * numThreads;
size_t bufferSize = n * sizeof(float);
float *hOut = new float[n];
CUdeviceptr dX, dY, dOut;
CUDA_SAFE_CALL(cuMemAlloc(&dOut, bufferSize));
// Execute parent kernel.
void *args[] = { &dOut, &n, &numBlocks, &numThreads };
CUDA_SAFE_CALL(
   cuLaunchKernel(kernel,
                  1, 1, 1,    // grid dim
                  1, 1, 1,    // block dim
                  0, NULL,    // shared mem and stream
                  args, 0));  // arguments
CUDA_SAFE_CALL(cuCtxSynchronize());
// Retrieve and print output.
CUDA_SAFE_CALL(cuMemcpyDtoH(hOut, dOut, bufferSize));

for (size_t i = 0; i < n; ++i) {
   std::cout << hOut[i] << '\n';
}
// Release resources.
CUDA_SAFE_CALL(cuMemFree(dOut));
CUDA_SAFE_CALL(cuModuleUnload(module));
CUDA_SAFE_CALL(cuLinkDestroy(linkState));
CUDA_SAFE_CALL(cuCtxDestroy(context));
delete[] hOut;
delete[] ptx;
return 0;
}

13.1. Dynamic Parallelism Build Instructions

Assuming the environment variable CUDA_PATH points to CUDA Toolkit installation directory, build this example as:

  • With NVRTC shared library:

    • Windows:

      cl.exe dynamic-parallelism.cpp /Fedynamic-parallelism ^
   /I "%CUDA_PATH%\include" ^
   "%CUDA_PATH%"\lib\x64\nvrtc.lib "%CUDA_PATH%"\lib\x64\cuda.lib

    • Linux:

      g++ dynamic-parallelism.cpp -o dynamic-parallelism \
   -I $CUDA_PATH/include \
   -L $CUDA_PATH/lib64 \
   -lnvrtc -lcuda \
   -Wl,-rpath,$CUDA_PATH/lib64

  • With NVRTC static library:

    • Windows:

      cl.exe dynamic-parallelism.cpp /Fedynamic-parallelism  ^
   /I "%CUDA_PATH%"\include ^
   "%CUDA_PATH%"\lib\x64\nvrtc_static.lib ^
   "%CUDA_PATH%"\lib\x64\nvrtc-builtins_static.lib ^
   "%CUDA_PATH%"\lib\x64\nvptxcompiler_static.lib ^
  "%CUDA_PATH%"\lib\x64\cuda.lib user32.lib Ws2_32.lib

    • Linux:

      g++ dynamic-parallelism.cpp -o dynamic-parallelism \
   -I $CUDA_PATH/include \
   -L $CUDA_PATH/lib64 \
   -lnvrtc_static -lnvrtc-builtins_static -lnvptxcompiler_static -lcuda \
   -lpthread