cuModuleGetFunction returns not found

Question

I want to compile CUDA kernels with the nvrtc JIT compiler to improve the performance of my application (so I have an increased amount of instruction fetches but I am saving multiple array accesses).

The functions looks e.g. like this and is generated by my function generator (not that important):

extern "C" __device__ void GetSumOfBranches(double* branches, double* outSum)
{
    double sum = (branches[38])+(-branches[334])+(-branches[398])+(-branches[411]);
    *outSum = sum;
}

I am compiling the code above with the following function:

CUfunction* FunctionGenerator::CreateFunction(const char* programText)
{
        // When I comment this statement out the output of the PTX file is changing
        // what is the reson?!
        // Bug?
        std::string savedString = std::string(programText);


        nvrtcProgram prog;
        nvrtcCreateProgram(&prog, programText, "GetSumOfBranches.cu", 0, NULL, NULL);

        const char *opts[] = {"--gpu-architecture=compute_52", "--fmad=false"};
        nvrtcCompileProgram(prog, 2, opts);

        // 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);

        printf("%s", ptx);

        CUdevice cuDevice;
        CUcontext context;
        CUmodule module;
        CUfunction* kernel;
        kernel = (CUfunction*)malloc(sizeof(CUfunction));
        cuInit(0);
        cuDeviceGet(&cuDevice, 0);
        cuCtxCreate(&context, 0, cuDevice);
        auto resultLoad = cuModuleLoadDataEx(&module, ptx, 0, 0, 0);
        auto resultGetF = cuModuleGetFunction(kernel, module, "GetSumOfBranches");
        return kernel;
}

Everything is working except that cuModuleGetFunction is returning CUDA_ERROR_NOT_FOUND. That error occurs because GetSumOfBranches cannot be found in the PTX file.

However the output of printf("%s", ptx); is this:

// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-19856038
// Cuda compilation tools, release 7.5, V7.5.17
// Based on LLVM 3.4svn
//

.version 4.3
.target sm_52
.address_size 64

    // .globl   GetSumOfBranches

.visible .func GetSumOfBranches(
    .param .b64 GetSumOfBranches_param_0,
    .param .b64 GetSumOfBranches_param_1
)
{
    .reg .f64   %fd<8>;
    .reg .b64   %rd<3>;


    ld.param.u64    %rd1, [GetSumOfBranches_param_0];
    ld.param.u64    %rd2, [GetSumOfBranches_param_1];
    ld.f64  %fd1, [%rd1+304];
    ld.f64  %fd2, [%rd1+2672];
    sub.rn.f64  %fd3, %fd1, %fd2;
    ld.f64  %fd4, [%rd1+3184];
    sub.rn.f64  %fd5, %fd3, %fd4;
    ld.f64  %fd6, [%rd1+3288];
    sub.rn.f64  %fd7, %fd5, %fd6;
    st.f64  [%rd2], %fd7;
    ret;
}

In my optinion everything is fine and GetSumOfBranches sould be found by cuModuleGetFunction. Can you explain me why?

Second Question

when i outcomment std::string savedString = std::string(programText); then the output of the PTX is just:

// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-19856038
// Cuda compilation tools, release 7.5, V7.5.17
// Based on LLVM 3.4svn
//

.version 4.3
.target sm_52
.address_size 64

and this is weird because savedString is not used at all...

Answer 1

What you are trying to do isn't supported. The host side modules management APIs and device ELF format do not expose __device__ functions, only __global__ functions which are callable via the kernel launch APIs.

You can compile device functions a priori or at runtime and link them with kernels in a JIT fashion, and you can retrieve those kernels and call them. But that is all you can do.