1. Introduction — NVRTC 12.9 documentation (original) (raw)

nvrtc

The User guide for the NVRTC library.

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:

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:

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

enumerator NVRTC_ERROR_NO_PCH_CREATE_ATTEMPTED

enumerator NVRTC_ERROR_PCH_CREATE_HEAP_EXHAUSTED

enumerator NVRTC_ERROR_PCH_CREATE

enumerator NVRTC_ERROR_CANCELLED

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

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 ).

nvrtcResult nvrtcSetFlowCallback(nvrtcProgram prog, int(*callback)(void *, void *), void *payload)

nvrtcSetFlowCallback registers a callback function that the compiler will invoke at different points during a call to nvrtcCompileProgram, and the callback function can decide whether to cancel compilation by returning specific values.

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.

Parameters

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

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.

Parameters

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

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.

Parameters

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

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.

Parameters

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.

Parameters

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.

Parameters

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.

Parameters

Returns

nvrtcResult nvrtcGetPTX(nvrtcProgram prog, char *ptx)

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

Parameters

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).

Parameters

Returns

nvrtcResult nvrtcGetProgramLog(nvrtcProgram prog, char *log)

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

Parameters

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

Returns

nvrtcResult nvrtcSetFlowCallback(nvrtcProgram prog, int (*callback)(void*, void*), void *payload)

nvrtcSetFlowCallback registers a callback function that the compiler will invoke at different points during a call to nvrtcCompileProgram, and the callback function can decide whether to cancel compilation by returning specific values.

The callback function must satisfy the following constraints:

(1) Its signature should be:

int callback(void* param1, void* param2);

When invoking the callback, the compiler will always pass payload to param1 so that the callback may make decisions based on payload . It’ll always pass NULL to param2 for now which is reserved for future extensions.

(2) It must return 1 to cancel compilation or 0 to continue. Other return values are reserved for future use.

(3) It must return consistent values. Once it returns 1 at one point, it must return 1 in all following invocations during the current nvrtcCompileProgram call in progress.

(4) It must be thread-safe.

(5) It must not invoke any nvrtc/libnvvm/ptx APIs.

Parameters

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:

3.6. 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.

3.6.1. 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:

Parameters

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:

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:

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

4.7. Predefined Types

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.hand 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);

7. 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 nvrtcGetLoweredNamework 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).

7.1. 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 \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");
..
kernel_name_vec.push_back("f3");
// 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));

7.2. Notes

8. 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_demangleand 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).

8.1. 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 \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
    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(&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());
    ..
    name_vec.push_back(getKernelNameForType());
    ..
    name_vec.push_back(getKernelNameForTypeN1::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));

9. Versioning Scheme

9.1. 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”.

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 theBest Practices Guide.

Some approaches to resolving this issue:

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.

9.2. 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.

10. Caching (CUDA 12.9+)

By default, NVRTC will query a cache for translating the NVVM IR generated by the frontend parser into PTX or SASS (cubin). The cache is implemented in the CUDA driver library (libcuda.so.1/nvcuda.dll). On the first nvrtcCompileProgram() call, NVRTC will load the CUDA driver library and invoke cuInit(). If either of these steps fails, the cache will not be used for the current and subsequent nvrtcCompileProgram() calls.

The use of the cache can be turned off by setting the environment variable CUDA_CACHE_DISABLE, or by adding -no-cache to the NVRTC command line.

Note: In some situations, the use of the cache may slow down compilation:

  1. CUDA persistence mode is turned off (usually affects only non-Windows platforms)
  1. CUDA cache is on a shared or network file system
  1. CUDA cache is too small

See CUDA Enviroment Variables for details on the environment variables CUDA_CACHE_DISABLE, CUDA_CACHE_PATH and CUDA_CACHE_MAXSIZE.

11. Miscellaneous Notes

11.1. 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 nvrtcCompileProgramto 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.

11.2. 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.

11.3. 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.

12. Example: SAXPY

12.1. Code (saxpy.cpp)

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

#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(i); hY[i] = static_cast(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; }

12.2. Saxpy Build Instructions

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

13. Example: Using Lowered Name

13.1. Code (lowered-name.cpp)

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

#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 \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::vectorstd::string kernel_name_vec; std::vectorstd::string variable_name_vec; std::vector variable_initial_value;

std::vector 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"); expected_result.push_back(sizeof(int));

kernel_name_vec.push_back("f3"); 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; }

13.2. Lowered Name Build Instructions

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

14. Example: Using nvrtcGetTypeName

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

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

#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 \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 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(&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::vectorstd::string name_vec; std::vector expected_result;

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

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

name_vec.push_back(getKernelNameForTypeN1::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; }

14.2. nvrtcGetTypeName Build Instructions

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

15. Example: Dynamic Parallelism

Code (dynamic-parallelism.cpp)

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

#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 \n\n" << " 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; }

15.1. Dynamic Parallelism Build Instructions

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

 "%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

This section demonstrates device link time optimization (LTO). There are two units of LTO IR. The first unit is generated offline using nvcc, by specifying the architecture as -arch lto_XX (refer to Code (offline.cu)). The generated LTO IR is packaged in a fatbinary.

The second unit is generated online using NVRTC, by specifying the flag -dlto (refer to Code (online.cpp)).

These two units are then passed to libnvJitLink* API functions, which link together the LTO IR, run the optimizer on the linked IR and generate a cubin (refer to Code (online.cpp)). The cubin is then loaded on the GPU and executed.

16.1. Code (offline.cu)

device float compute(float a, float x, float y) { return a * x + y; }

16.2. Code (online.cpp)

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

#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)

#define NVJITLINK_SAFE_CALL(h,x)
do {
nvJitLinkResult result = x;
if (result != NVJITLINK_SUCCESS) {
std::cerr << "\nerror: " #x " failed with error "
<< result << '\n';
size_t lsize;
result = nvJitLinkGetErrorLogSize(h, &lsize);
if (result == NVJITLINK_SUCCESS && lsize > 0) {
char log = (char)malloc(lsize);
result = nvJitLinkGetErrorLog(h, log);
if (result == NVJITLINK_SUCCESS) {
std::cerr << "error: " << log << '\n';
free(log);
}
}
exit(1);
}
} while(0)

const char *lto_saxpy = " \n
extern device float compute(float a, float x, float y); \n
\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] = compute(a, x[tid], y[tid]); \n
} \n
} \n";

int main(int argc, char *argv[]) { 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 lto_saxpy, // buffer "lto_saxpy.cu", // name 0, // numHeaders NULL, // headers NULL)); // includeNames

// specify that LTO IR should be generated for LTO operation const char *opts[] = {"-dlto", "--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 generated LTO IR from the program. size_t LTOIRSize; NVRTC_SAFE_CALL(nvrtcGetLTOIRSize(prog, &LTOIRSize)); char *LTOIR = new char[LTOIRSize]; NVRTC_SAFE_CALL(nvrtcGetLTOIR(prog, LTOIR)); // Destroy the program. NVRTC_SAFE_CALL(nvrtcDestroyProgram(&prog));

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));

// Load the generated LTO IR and the LTO IR generated offline // and link them together. nvJitLinkHandle handle; // Dynamically determine the arch to link for int major = 0; int minor = 0; CUDA_SAFE_CALL(cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cuDevice)); CUDA_SAFE_CALL(cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cuDevice)); int arch = major*10 + minor; char smbuf[16]; sprintf(smbuf, "-arch=sm_%d", arch); const char *lopts[] = {"-dlto", smbuf}; NVJITLINK_SAFE_CALL(handle, nvJitLinkCreate(&handle, 2, lopts));

// NOTE: assumes "offline.fatbin" is in the current directory // The fatbinary contains LTO IR generated offline using nvcc NVJITLINK_SAFE_CALL(handle, nvJitLinkAddFile(handle, NVJITLINK_INPUT_FATBIN, "offline.fatbin")); NVJITLINK_SAFE_CALL(handle, nvJitLinkAddData(handle, NVJITLINK_INPUT_LTOIR, (void *)LTOIR, LTOIRSize, "lto_online"));

// The call to nvJitLinkComplete causes linker to link together the two // LTO IR modules (offline and online), do optimization on the linked LTO IR, // and generate cubin from it. NVJITLINK_SAFE_CALL(handle, nvJitLinkComplete(handle)); size_t cubinSize; NVJITLINK_SAFE_CALL(handle, nvJitLinkGetLinkedCubinSize(handle, &cubinSize)); void *cubin = malloc(cubinSize); NVJITLINK_SAFE_CALL(handle, nvJitLinkGetLinkedCubin(handle, cubin)); NVJITLINK_SAFE_CALL(handle, nvJitLinkDestroy(&handle));

CUDA_SAFE_CALL(cuModuleLoadData(&module, cubin)); 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(i); hY[i] = static_cast(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)); free(cubin); delete[] hX; delete[] hY; delete[] hOut; delete[] LTOIR; return 0; }

16.3. Device LTO Build Instructions

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

17. Example: Automatic PCH (CUDA 12.8+)

This example demonstrates automatic PCH mode, which is enabled by passing-pch to the nvrtcCompileProgram invocation. There are 2 different programsfirst and second that include the same header:

const char *first = "#include "auto_pch_common.h" \n" "global void foo(double *ptr) {\n" "*ptr = doit();\n}\n";

const char *second = "#include "auto_pch_common.h" \n" "global void other(double *a, double *b) {\n" "*a = *b + doit();\n}\n";

When first is compiled with NVRTC with -pch, the compiler will create a PCH file for the inclusion of the auto_pch_common.h header. When second is compiled with -pch, the compiler will transparently use the previously created PCH file.

Here’s the output of the program when run:

compiling first program "default_program": creating precompiled header file "default_program.pch"

nvrtcGetPCHCreateStatus returned : NVRTC_SUCCESS

compiling second program (expect to use PCH) "default_program": using precompiled header file "default_program.pch"

nvrtcGetPCHCreateStatus returned : NVRTC_ERROR_NO_PCH_CREATE_ATTEMPTED

When compiling the first program, the PCH file is successfully created. When compiling the second program, the PCH file is successfully used, but the compiler chose not to create another PCH file (hencenvrtcGetPCHCreateStatus() returned NVRC_ERROR_NO_PCH_CREATE_ATTEMPTED - this is expected).

17.1. Code (auto_pch_common.h)

device double qqq = 10;

template device double get(T in) { return sin(in) + qqq; }

device double doit() { return get(0.5); }

17.2. Code (auto_pch.cpp)

#include #include <nvrtc.h>

#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)

const char *docompile(const char *progstr) { nvrtcProgram prog;

NVRTC_SAFE_CALL( nvrtcCreateProgram(&prog, progstr, // buffer "", // name 0, // numHeaders NULL, // headers NULL)); // includeNames

const char opts[] = { "-pch" / automatic PCH */ };

nvrtcResult compileResult = nvrtcCompileProgram(prog, // prog sizeof(opts) / sizeof(opts[0]), // 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; delete[] log; if (compileResult != NVRTC_SUCCESS) { exit(1); } nvrtcResult err = nvrtcGetPCHCreateStatus(prog); std::cout << "\n nvrtcGetPCHCreateStatus returned : " << nvrtcGetErrorString(err) << std::endl;

// 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));

return ptx; }

int main() { const char *first = "#include "auto_pch_common.h" \n" "global void foo(double *ptr) {\n" "*ptr = doit();\n}\n";

const char *second = "#include "auto_pch_common.h" \n" "global void other(double *a, double *b) {\n" "*a = *b + doit();\n}\n";

std::cout << "\n compiling first program\n"; const char *ptr1 = docompile(first);

std::cout << "\n compiling second program (expect to use PCH)\n"; const char *ptr2 = docompile(second);

delete [] ptr1; delete [] ptr2;

}

17.3. Automatic PCH Build Instructions

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

18. Example: Explicit PCH Create/Use (CUDA 12.8+)

This example demonstrates explicit PCH mode, where user code can explicitly create and use PCH files with “-create-pch=filename” and “-use-pch=filename” NVRTC flags, respectively. . There are 2 different programs first and second that include the same header:

const char *first = "#include "common.h" \n" "global void foo(double *ptr) {\n" "*ptr = doit();\n}\n";

const char *second = "#include "common.h" \n" "global void other(double *a, double *b) {\n" "*a = *b + doit();\n}\n";

When first is compiled with NVRTC with -create-pch=foo.pch, the compiler will create a PCH file (“foo.pch”) for the inclusion of the common.h header. When second is compiled with -use-pch=foo.pch, the compiler will use the specified PCH file “foo.pch”.

Here’s the output of the program when run:

compiling first program "default_program": creating precompiled header file "foo.pch"

nvrtcGetPCHCreateStatus returned : NVRTC_SUCCESS

compiling second program (expect to use PCH) "default_program": using precompiled header file "foo.pch"

nvrtcGetPCHCreateStatus returned : NVRTC_ERROR_NO_PCH_CREATE_ATTEMPTED

When compiling the first program, the PCH file “foo.pch” is successfully created. When compiling the second program, the PCH file “foo.pch” is successfully used; but no PCH files are created (hencenvrtcGetPCHCreateStatus() returned NVRC_ERROR_NO_PCH_CREATE_ATTEMPTED).

18.1. Code (common.h)

device double qqq = 10;

template device double get(T in) { return sin(in) + qqq; }

device double doit() { return get(0.5); }

18.2. Code (explicit_pch.cpp)

#include #include <nvrtc.h>

#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)

const char *docompile(const char *progstr, const char *pch_mode) { nvrtcProgram prog;

NVRTC_SAFE_CALL( nvrtcCreateProgram(&prog, progstr, // buffer "", // name 0, // numHeaders NULL, // headers NULL)); // includeNames

const char *opts[] = { pch_mode /*create/use PCH */ };

nvrtcResult compileResult = nvrtcCompileProgram(prog, // prog sizeof(opts) / sizeof(opts[0]), // 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; delete[] log; if (compileResult != NVRTC_SUCCESS) { exit(1); } nvrtcResult err = nvrtcGetPCHCreateStatus(prog); std::cout << "\n nvrtcGetPCHCreateStatus returned : " << nvrtcGetErrorString(err) << std::endl;

// 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));

return ptx; }

int main() { const char *first = "#include "common.h" \n" "global void foo(double *ptr) {\n" "*ptr = doit();\n}\n";

const char *second = "#include "common.h" \n" "global void other(double *a, double *b) {\n" "*a = *b + doit();\n}\n";

std::cout << "\n compiling first program\n"; const char *ptr1 = docompile(first, "-create-pch=foo.pch");

std::cout << "\n compiling second program (expect to use PCH)\n"; const char *ptr2 = docompile(second, "-use-pch=foo.pch");

delete [] ptr1; delete [] ptr2;

}

18.3. Explicit PCH Build Instructions

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

19. Example: PCH Heap Resizing (CUDA 12.8+)

The PCH heap is persistent across nvrtcCompileProgram() calls. In memory constrained environments, there may be a need to size the PCH heap to a smaller value than the default. This example shows how to size the PCH heap. First, the heap size is set to a low value (8 KB) 6:

NVRTC_SAFE_CALL(nvrtcSetPCHHeapSize(8*1024));

Then, nvrtcCompileProgram() is invoked with -pch . PCH creation is expected to fail, with nvrtcGetPCHCreateStatus() returningNVRTC_ERROR_PCH_CREATE_HEAP_EXHAUSTED. The required PCH heap size can then be retrieved by calling nvrtcGetPCHHeapSizeRequired(), and the PCH heap can be resized by invoking nvrtcSetPCHHeapSize():

nvrtcResult err = nvrtcGetPCHCreateStatus(prog); std::cout << "\nnvrtcGetPCHCreateStatus returned : " << nvrtcGetErrorString(err) << std::endl;

if (err == NVRTC_ERROR_PCH_CREATE_HEAP_EXHAUSTED) { size_t size; NVRTC_SAFE_CALL(nvrtcGetPCHHeapSize(&size)); ... NVRTC_SAFE_CALL(nvrtcGetPCHHeapSizeRequired(prog, &size)); .. NVRTC_SAFE_CALL(nvrtcSetPCHHeapSize(size)); }

The next NVRTC compilation to request PCH creation with the same file is now expected to succeed. Here is the output of the program:

compiling first program auto_pch_common.h(7): warning #639-D: insufficient preallocated memory for generation of precompiled header file (4481024 bytes required) device double doit() { return get(0.5); } ^

Remark: The warnings can be suppressed with "-diag-suppress "

nvrtcGetPCHCreateStatus returned : NVRTC_ERROR_PCH_CREATE_HEAP_EXHAUSTED nvrtcGetPCHHeapSize() before: 8192 nvrtcGetPCHHeapSizeRequired() reports: 4481024 nvrtcGetPCHHeapSize() after: 4481024

compiling second program (expect to use PCH) "default_program": creating precompiled header file "default_program.pch"

nvrtcGetPCHCreateStatus returned : NVRTC_SUCCESS

6

But not 0, since that disable PCH operation.

19.1. Code (auto_pch_common.h)

device double qqq = 10;

template device double get(T in) { return sin(in) + qqq; }

device double doit() { return get(0.5); }

19.2. Code (pch_resize.cpp)

#include #include <nvrtc.h>

#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)

const char *docompile(const char *progstr) { nvrtcProgram prog;

NVRTC_SAFE_CALL( nvrtcCreateProgram(&prog, progstr, // buffer "", // name 0, // numHeaders NULL, // headers NULL)); // includeNames

const char opts[] = { "-pch" / automatic PCH */ };

nvrtcResult compileResult = nvrtcCompileProgram(prog, // prog sizeof(opts) / sizeof(opts[0]), // 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; delete[] log; if (compileResult != NVRTC_SUCCESS) { exit(1); } nvrtcResult err = nvrtcGetPCHCreateStatus(prog); std::cout << "\nnvrtcGetPCHCreateStatus returned : " << nvrtcGetErrorString(err) << std::endl;

if (err == NVRTC_ERROR_PCH_CREATE_HEAP_EXHAUSTED) { size_t size; NVRTC_SAFE_CALL(nvrtcGetPCHHeapSize(&size)); std::cout << "nvrtcGetPCHHeapSize() before: " << size << std::endl; NVRTC_SAFE_CALL(nvrtcGetPCHHeapSizeRequired(prog, &size)); std::cout << "nvrtcGetPCHHeapSizeRequired() reports: " << size << std::endl; NVRTC_SAFE_CALL(nvrtcSetPCHHeapSize(size)); NVRTC_SAFE_CALL(nvrtcGetPCHHeapSize(&size)); std::cout << "nvrtcGetPCHHeapSize() after: " << size << std::endl; }

// 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));

return ptx; }

int main() { const char *first = "#include "auto_pch_common.h" \n" "global void foo(double *ptr) {\n" "*ptr = doit();\n}\n";

const char *second = "#include "auto_pch_common.h" \n" "global void other(double *a, double *b) {\n" "*a = *b + doit();\n}\n";

//set NVRTC PCH heap to a low initial value (8 KB) (note: don't use 0) NVRTC_SAFE_CALL(nvrtcSetPCHHeapSize(8*1024));

std::cout << "\n compiling first program\n"; const char *ptr1 = docompile(first);

std::cout << "\n compiling second program (expect to use PCH)\n"; const char *ptr2 = docompile(second);

delete [] ptr1; delete [] ptr2;

}

19.3. PCH Heap Resizing Build Instructions

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

19.4. Notices

19.4.1. Notice

This document is provided for information purposes only and shall not be regarded as a warranty of a certain functionality, condition, or quality of a product. NVIDIA Corporation (“NVIDIA”) makes no representations or warranties, expressed or implied, as to the accuracy or completeness of the information contained in this document and assumes no responsibility for any errors contained herein. NVIDIA shall have no liability for the consequences or use of such information or for any infringement of patents or other rights of third parties that may result from its use. This document is not a commitment to develop, release, or deliver any Material (defined below), code, or functionality.

NVIDIA reserves the right to make corrections, modifications, enhancements, improvements, and any other changes to this document, at any time without notice.

Customer should obtain the latest relevant information before placing orders and should verify that such information is current and complete.

NVIDIA products are sold subject to the NVIDIA standard terms and conditions of sale supplied at the time of order acknowledgement, unless otherwise agreed in an individual sales agreement signed by authorized representatives of NVIDIA and customer (“Terms of Sale”). NVIDIA hereby expressly objects to applying any customer general terms and conditions with regards to the purchase of the NVIDIA product referenced in this document. No contractual obligations are formed either directly or indirectly by this document.

NVIDIA products are not designed, authorized, or warranted to be suitable for use in medical, military, aircraft, space, or life support equipment, nor in applications where failure or malfunction of the NVIDIA product can reasonably be expected to result in personal injury, death, or property or environmental damage. NVIDIA accepts no liability for inclusion and/or use of NVIDIA products in such equipment or applications and therefore such inclusion and/or use is at customer’s own risk.

NVIDIA makes no representation or warranty that products based on this document will be suitable for any specified use. Testing of all parameters of each product is not necessarily performed by NVIDIA. It is customer’s sole responsibility to evaluate and determine the applicability of any information contained in this document, ensure the product is suitable and fit for the application planned by customer, and perform the necessary testing for the application in order to avoid a default of the application or the product. Weaknesses in customer’s product designs may affect the quality and reliability of the NVIDIA product and may result in additional or different conditions and/or requirements beyond those contained in this document. NVIDIA accepts no liability related to any default, damage, costs, or problem which may be based on or attributable to: (i) the use of the NVIDIA product in any manner that is contrary to this document or (ii) customer product designs.

No license, either expressed or implied, is granted under any NVIDIA patent right, copyright, or other NVIDIA intellectual property right under this document. Information published by NVIDIA regarding third-party products or services does not constitute a license from NVIDIA to use such products or services or a warranty or endorsement thereof. Use of such information may require a license from a third party under the patents or other intellectual property rights of the third party, or a license from NVIDIA under the patents or other intellectual property rights of NVIDIA.

Reproduction of information in this document is permissible only if approved in advance by NVIDIA in writing, reproduced without alteration and in full compliance with all applicable export laws and regulations, and accompanied by all associated conditions, limitations, and notices.

THIS DOCUMENT AND ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, “MATERIALS”) ARE BEING PROVIDED “AS IS.” NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE. TO THE EXTENT NOT PROHIBITED BY LAW, IN NO EVENT WILL NVIDIA BE LIABLE FOR ANY DAMAGES, INCLUDING WITHOUT LIMITATION ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, PUNITIVE, OR CONSEQUENTIAL DAMAGES, HOWEVER CAUSED AND REGARDLESS OF THE THEORY OF LIABILITY, ARISING OUT OF ANY USE OF THIS DOCUMENT, EVEN IF NVIDIA HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. Notwithstanding any damages that customer might incur for any reason whatsoever, NVIDIA’s aggregate and cumulative liability towards customer for the products described herein shall be limited in accordance with the Terms of Sale for the product.

19.4.2. OpenCL

OpenCL is a trademark of Apple Inc. used under license to the Khronos Group Inc.

19.4.3. Trademarks

NVIDIA and the NVIDIA logo are trademarks or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.

1

Changes to compiler optimizer heuristics in the newer NVRTC shared library may also potentially cause performance perturbations for generated code.

2

These API functions modify the state of the associated nvrtcProgram.